Abstract:Variable fractional delay filters are widely used in delay compensation technologies due to their ability to achieve arbitrary fractional delay transformations. However, due to the high complexity of solving its filter coefficients and poor adaptability, its application in engineering is severely limited. To address this challenge, this paper proposes a method to flexibly change the performance of fractional time-delay filters by adjusting the filter parameters under the condition of low computational complexity, and completes the FPGA simulation verification. The method precisely controls the window shape by adjusting the width factor of the window function, thereby optimizing the time-frequency characteristics of the filter at different orders and providing more accurate frequency selectivity compared to traditional methods. In addition, this paper adopts the orthogonal triangular decomposition least squares matrix method to solve the filter coefficients, and the designed filter requires only one matrix inverse under the condition of guaranteeing the accuracy of group delay, which effectively avoids the complex mathematical operations such as partial derivatives and double integration. Simulation results show that the method proposed in this paper reduces the computational complexity by one order of magnitude compared with the existing methods under the condition of maintaining the same delay accuracy, with the maximum magnitude error reaching -104 dB and the maximum group delay error reaching 2.34×10-4. FPGA verification results show that the design method has low hardware computational resource consumption, greatly improving the efficiency of the design of the Farrow filter.

Abstract:The 20th Central Committee′s Third Plenary Session emphasized the comprehensive implementation of water conservancy reform tasks, with a focus on residential drinking water as a key livelihood task. The coagulation process is a critical step in drinking water treatment. Due to the significant time delay characteristic of the coagulation process, conventional PID control cannot achieve satisfactory results for control systems with frequent changes in raw water quality. Therefore, a linear active disturbance rejection controller (LADRC), which does not rely on an accurate system model, is applied to the system. An extended observer is used to estimate and compensate for disturbances in the coagulation control system. Additionally, an adaptive Smith controller, combining a Smith predictor with a fuzzy controller, is designed to mitigate the impact of large time delays on control effectiveness, leading to the proposed Fuzzy-Smith-LADRC controller. To address the difficulties in adjusting the controller parameters, an improved dung beetle optimization algorithm (MSIDBO) is introduced for parameter tuning. This improved algorithm optimizes issues such as uneven initial population distribution and the tendency to fall into local optima found in the DBO algorithm, enabling MSIDBO to converge quickly and better balance global exploration and local exploitation capabilities. When the system model is accurate, this control method reduces the settling time by 279 s and decreases overshoot by 8% compared to PID control, it also reduces settling time by 40 s compared to DMC control. When the system model changes, it demonstrates better anti-disturbance and robustness compared to LADRC.

Abstract:Addressing the trajectory tracking problem of a quadrotor UAV, a command filter backstepping control strategy based on predefined time is proposed to mitigate the impact of model uncertainties and unknown external disturbances on system stability. Firstly, a predefined-time disturbance observer is designed to accurately estimate the system uncertainties and unknown external disturbances in real-time. Secondly, to effectively alleviate the "differential explosion" issue in the backstepping control strategy, a predefined-time command filter is designed. Based on this, position and attitude controllers are further designed using the backstepping method, enhancing the system′s control accuracy and response speed. Finally, the Lyapunov theory is employed to verify the stability of the proposed control strategy. Simulation experiments validate the effectiveness and superiority of the proposed control strategy.

Abstract:Aiming at the problems of insufficient consideration of time delay in the modeling process of servo motor drive system, lack of analysis given in the premise of system stability in the calibration of control gain, and relatively high conservatism of system stability criterion, this paper proposes a time delay related stability criterion with lower conservatism. Firstly, the time-varying delay is considered in the inertia structure model and the state space equations of the time-delayed closed-loop systems are established by designing the corresponding controllers, and then the Lyapunov generalized functional analysis method is used in combination with the techniques of free weight matrix, time delay splitting and integral inequality to reduce the conservatism of the stability criterion. The applicable time delay range of the controller gain for system stability guaranteed by this stability criterion is derived by running the MATLAB program, and it is concluded that this paper′s criterion improves the upper bound of the maximum time delay for system stability by 46.33%, which verifies that this paper′s criterion has a lower conservatism. The stability analysis of this paper provides theoretical reference for the analysis and control of more complex servo motor drive systems.

Abstract:In response to the problems such as the limited applicable scenarios of the automatic white balance algorithm (AWB), an adaptive white balance algorithm based on histogram adjustment is proposed, and the proposed algorithm is implemented in hardware using a field-programmable gate array (FPGA) to meet the real-time processing requirements of embedded systems. The algorithm statistically calculates the histograms of different color channels in a color image, uses the similarity of histogram shapes between channels as the judgment condition, and combines an adaptive histogram adjustment strategy to perform white balance correction for images in different scenes. The experimental results demonstrate that this algorithm exhibits excellent adaptability in both colorful and image scenarios containing large areas of monochromatic blocks. Compared with traditional white balance algorithms, the image color restoration effect is remarkable, the average accuracy of calibration is increased by 6%, and it can achieve real-time processing at a resolution of 1 280×720@30 fps on embedded devices, presenting a promising prospect for engineering applications.

Abstract:FairMOT, a multi-object tracking algorithm, proposes a balanced learning strategy between the detection branch and the re-identification branch, effectively balancing the tasks of object detection and re-identification, thereby improving tracking accuracy. However, due to the limited feature extraction capability of its DLA34 backbone network, the model′s tracking performance often declines in complex real-world scenarios, leading to missed detections and incorrect tracking. To enhance the backbone network′s feature extraction capability, this paper designs a deep aggregated backbone network based on an element-wise multiplication structure and proposes the FairMOT-Star algorithm. This algorithm leverages the principle of hidden dimension enhancement brought by the element-wise multiplication structure to achieve concise and efficient object feature extraction. Additionally, EIoU_Loss is used as the regression loss function for the bounding box regression task, more precisely describing the positional and shape relationships between detection boxes and ground truth boxes, thus improving prediction accuracy. In the matching and association part, the Kalman filter algorithm predicts target motion information, and the Hungarian algorithm associates and matches targets and trajectories across frames in the temporal dimension. Experimental tests on the MOT16 dataset achieved an MOTA accuracy of 86.0%. The model′s weight parameters amount to 19.59 M, reducing parameter count by 9.7% compared to the FairMOT model, while increasing MOTA accuracy by 3.5%, effectively optimizing the computational parameters and tracking accuracy of the FairMOT algorithm.

Abstract:Aiming at the limitations of small targets in remote sensing images, such as complex image background, dense distribution of small targets, and diverse target scales, this paper proposes an improved algorithm based on YOLOv8n. Firstly, a multi-scale null attention module is designed to introduce a multi-scale null attention mechanism in the backbone network in combination with the C2f module to effectively capture multi-scale semantic information and reduce the redundancy of the self-attention mechanism; secondly, a residual fast convolution module is designed to reduce the model computation and improve the feature extraction capability; finally, the PIoU v2-Iou loss function is used instead of the CIOU loss function to improve the detection accuracy of the model. The experimental results on DOTA, RSOD and VisDrone2019 datasets show that the improved YOLOv8n model improves the mAP by 2.7%、3.3% and 3.8%, respectively, and reduces the computation by 0.5 GFLOPs compared with the original model YOLOv8n, which validates the effectiveness of the new algorithm.

Abstract:To tackle complex numerical optimization problems, this paper proposes an improved hiking optimization algorithm based on Cauchy distribution operator and random differential mutation strategy (CDHOA). The algorithm enhances diversity through effective population initialization, balances global search with local exploitation using the inverse cumulative Cauchy distribution operator, and employs a random differential mutation strategy to boost exploitation and reduce local optima risks. Experimental results show the average performance of CDHOA on the CEC2017 test set is better than that of eight comparison algorithms. The statistical test further confirmed that the performance difference was significant. Nine representative test functions are selected from the CEC2017 test set, and the effectiveness of the three enhancement strategies in the algorithm is verified by comparative experiments. Additionally, it is applied to the parameter identification of photovoltaic model, and a small root mean square error of 2.43×10-3 is achieved, which has the best result of all comparison algorithms. In two kinds of engineering design problems, the algorithm achieves the minimum objective function value, which is better than the comparison algorithms. Overall，CDHOA performs well in global search ability, convergence speed and accuracy, which effectively improves the performance of solving complex numerical optimization problems.

Abstract:To address the issues of large parameter quantity, high computational complexity, and high resource demands on the computing platform for the steel surface defect detection model, a lightweight improved algorithm has been proposed. Firstly, using ShuffleNetV2 as the improved backbone layer has achieved remarkable results in reducing model complexity and computational load. Secondly, a sufficiently flexible and lightweight channel attention mechanism (CA) was incorporated after the SPPF module, while the bidirectional feature pyramid network (BiFPN) was utilized to enhance feature fusion, thereby improving the efficiency of feature information flow. Finally, the lightweight dual convolution kernel (DualConv) was employed to replace the convolution layer in C2f, and the parameter quantity was reduced through the grouping convolution strategy. Experimental results indicate that, compared with the original YOLOv8n, the improved model achieves lightweighting while maintaining detection accuracy. The parameter quantity is 56.2% of the original, and the volume and computational load have decreased to 3.6 MB and 4.8 GFLOPs, respectively, representing a reduction of 42.86% and 41.47% compared to the original model. The lightweighting of the model reduces the deployment cost and is suitable for practical deployment and application.

Abstract:To address the issues of inconsistent feature distributions and the influence of noise components on the transfer effect in gearbox vibration data collected under different operating conditions, this paper proposes a deep transfer learning fault diagnosis method that integrates an attention mechanism with domain adversarial transfer networks. First, labeled and unlabeled vibration signals are constructed into datasets using a fixed-length data segmentation method. Second, to reduce the negative transfer impact caused by noisy samples, a convolutional block attention module (CBAM) and a discriminative loss term are used to assist the feature extractor in extracting discriminative features and enhancing the classification decision boundary. Finally, to solve the problem of inconsistent data feature distributions, a multi-kernel maximum mean discrepancy (MK-MMD) is employed to align the global distributions of the source and target domains, and an adversarial mechanism is used to align the subdomain distributions between the two domains. Experimental validation on a publicly available variable-condition gearbox fault dataset demonstrates that the proposed method achieves an average recognition accuracy of over 96.25%. A comparison with other diagnostic methods further validates the effectiveness and superiority of the proposed approach.

Abstract:To address the high false alarm rate and delayed fault detection in traditional principal component analysis (PCA) methods for industrial process fault detection, this paper proposes an iterative modeling-based sliding window PCA fault detection method. To improve detection real-time performance, an iterative approach is used during the modeling process to progressively remove outlier samples from the PCA model data, optimizing the PCA model. To reduce the false alarm rate, a sliding observation window is employed to count the number of outlier samples, and a second confidence limit is constructed for fault detection. To enhance fault detection accuracy, a composite statistic is used as the detection index, which considers anomalies in both the principal component direction and the residual space. To validate the effectiveness of the proposed method, simulation experiments were conducted using numerical examples and the Tennessee-Eastman (TE) process. In the numerical examples, the fault detection accuracy reached 89.20% with a false alarm rate of 1.33%. For the TE process, the fault detection accuracy reached 99.39% with a false alarm rate of 3.12%.

Abstract:When diagnosing rotor bar faults using motor current signature analysis, the fault characteristic frequencies and amplitudes on both sides of the fundamental frequency are crucial parameters for determining whether a fault has occurred and its severity. The diagnostic capability of FFT algorithm heavily depends on the length of the analyzed data. Although the least squares prony analysis algorithm has short-time data analysis capabilities, it is highly sensitive to noise levels and suffer from insufficient fault feature extraction, and failure may occur when the motor operates at low frequencies and low loads. To address these issues, an improved method combining singular value decomposition and LS-PA algorithms for diagnosing rotor bar faults is proposed. Initially, the SVD matrix is reconstructed using truncated data of original current signal, and effective order is determined based on the difference quotient of singular values. Subsequently, pre-processes technique is used to moderately suppress noise in stator current signal. Finally, the LS-PA algorithm is applied to identify and diagnose fault features from the preprocessed signal. Finite element simulation and experimental results demonstrate that the proposed method can effectively suppress signal noise and has the diagnostic performance of short-time data with high resolution. It achieves stable diagnosis of rotor bar faults under full load conditions, from light to full load, both in constant frequency and variable frequency power supply scenarios, outperforming traditional FFT methods.

Abstract:In order to improve the detection distance and section location accuracy of rail fracture with ultrasonic guided wave and ensure the safety of rail during service. In this study, Barker code encoding and Kaiser window function were used to modulate the excitation signal, expand the excitation signal time width, enhance the energy of the transmitted signal, and improve the resolution of the defect signal through pulse compression processing of the detected signal. The finite element simulation was carried out, the rail 3D simulation model was built, and the rail fracture damage was simulated. When the rail fracture degree is 20%, the peak-to-peak ratio of the main wave packet signal and the adjacent wave packet signal of the direct wave signal is 1.065 by using multi-period sine wave signal excitation, while the peak-to-peak ratio of the main wave packet signal and the adjacent wave packet signal is increased to 2.542 by using Barker13 coded signal excitation. The characteristics of defect echo signal under different fracture degrees are analyzed, and the amplitude of the echo signal intensity changes above 40 dB by Barker13 coded excitation signal. The finite element simulation shows that the encoded excitation and pulse compression technology can effectively improve the resolution and signal amplitude of the detected signal. Combined with the offline long-distance rail detection experiment, the detection signal obtained by using multi-period sine wave signal has complex components and difficult calculation of time difference. However, pulse compression technology can effectively improve the signal-to-noise ratio of detection signal and improve the signal resolution of the initial wave signal and the echo signal, and the accuracy rate of section positioning is 99.3%. Coding excitation and pulse compression technology can improve the amplitude of detection signal and the accuracy of detection and positioning effectively, and provide technical means for long-distance rail section positioning.

Abstract:Student classroom behavior recognition is important for improving the quality of teaching and learning. Most of the current mainstream research is based on video or sensor technologies, however, these methods suffer from privacy invasion and high cost, which constrain their wide application. Therefore, this paper proposes a method of student classroom behavior recognition based on WiFi CSI. This method first collected CSI signals for four typical classroom behaviors (hands up、stand up、sit down and turn over books) in a real classroom environment. Then combined with the characteristics of WiFi CSI data, Hampel filter and wavelet transform are used to denoise the CSI signal, and all subcarrier features are fused by designing principal component analysis algorithm. Subsequently, the CSI action intervals are intercepted according to the fusion features designing the local outlier factor detection algorithm, and the intercepted CSI signals are converted into amplitude energy maps by introducing three-dimensional mapping. Finally, the amplitude energy map dataset is feature extracted and classification recognized by designing a transfer learning model based on residual network. The experimental results show that the accuracy of the method is 98.89% and 99.07% in the step classroom and small classroom, and it can reach more than 98% in the test for different people. It is proved that the method has high recognition accuracy and good robustness, which provides a new idea for the research of student classroom behavior recognition.

Abstract:Addressing the issues of urban drainage pipeline defects being susceptible to background interference, the variability of characteristic scales, and the low detection accuracy and high false positive rate of existing detection models, this paper presents an improved defect detection algorithm based on YOLOv8. Initially, the DSK module is designed and embedded within the C2f module of the backbone network to expand the receptive field and improve the ability to extract multi-scale defect features. Subsequently, the Slim-neck network structure is introduced to refine the neck network, effectively utilizing and fusing defect feature information, which also contributes to the lightweightification of the model. Finally, the FocalEIOU loss function is adopted to enhance the detection performance for smaller defect targets and the convergence speed of the model. Experimental results on a pipeline defect dataset indicate that the proposed improved algorithm achieves a mean average precision (mAP) of 67.5% at a detection rate of 70.4 fps. Compared to the original YOLOv8 algorithm, the mAP value and detection speed are respectively increased by 3.8% and 1.7 fps, demonstrating superior detection performance. For the purpose of practical application, this paper has developed a system software capable of real-time detection of pipeline defects based on an improved algorithm. Through actual project detection, the enhanced algorithm proposed in this paper has been validated to meet the requirements of high precision and real-time detection for the task of urban drainage pipeline defect inspection.

Abstract:The use of low-dose CT scans has significantly reduced radiation exposure during examinations. However, this reduction has led to increased noise and artifacts in CT images, compromising image quality and diagnostic accuracy, which can affect physicians′ judgment during the diagnostic process. Recent advancements in generative models have demonstrated excellent performance in addressing these issues. Nonetheless, these models still face challenges, such as generating confusion and structural deficiencies during the generation process. To tackle these problems, a conditional diffusion denoising network model has been developed. This model incorporates a trainable curve matching module to correct different noise levels and includes a joint loss function. Experimental results indicate that the proposed algorithm achieves superior denoising outcomes compared to comparative algorithms, with test results of PSNR 35.70 and SSIM 0.912 8 on the dataset, representing the optimal performance among the selected methods. Additionally, it demonstrates good generalization across low-dose CT images with varying radiation doses, maintaining excellent denoising levels.

Abstract:Regarding the perennial safety production problems that factories constantly encounter, such as the strict prohibition of smoke and fire in the workshop area, the need for constant attention to the behavioral safety of workers, and whether workers wear masks in adverse working condition scenarios, an improved worker behavior and fire detection algorithm FDH-DETR based on RT-DETR was proposed. Firstly, through the fusion of the Deep Faster feature depth fusion module and FasterNet, the number of parameters and the amount of computation of the algorithm were reduced. Secondly, through the DRBC3 module size convolution kernel conversion mechanism, the inference cost of the model was decreased. Finally, through the HiLo-AIFI high-low frequency scale withinfeature interaction module, the extraction ability of high-low frequency features was enhanced. Experimental results indicate that the improved algorithm achieved an average accuracy of 93.8%, a reduction of 31.6% in parameters, a reduction of 61.4% in computation, and an FPS of 150 frames per second. Inference experiments were conducted in real working condition scenarios, verifying the effectiveness of the algorithm.

Abstract:Indoor location technology based on WiFi fingerprint database has attracted much attention because of its high precision and easy deployment, while the quality of offline fingerprint database is a key factor to determine the location accuracy. To solve the problem of high acquisition cost of offline fingerprint database, a denoising fingerprint database enhancement model (FASRGAN-DAE) based on denoising autoencoder super resolution generation adductive network is proposed. The method enhances the location accuracy by enhancing the sparse fingerprint database. Specifically, firstly, the fingerprint data is mapped to the corresponding fingerprint image; then, on the basis of deleting the batch normalization layer (BN layer), the generator network improves the perception loss function to generate high-resolution fingerprint images, and reduces the hidden layer and output layer of the autoencoder to improve the quality of the generated images. Meanwhile, in the discriminator network, the BN layer is deleted and the output of the convolutional layer is used as the authenticity score of the input image. The mean square error loss function is used to optimize the discriminator network to enhance the ability of distinguishing between real and generated images. Finally, the fingerprint image is restored to the fingerprint data through the mapping module to realize the enhancement of the fingerprint database. Through the localization experiment in the real underground parking lot environment, compared with the original fingerprint database, the average localization error was reduced by 5.69% after FASRGAN-DAE enhanced data.

Abstract:In complex road environments, existing algorithms for road multi-target detection suffer from poor recognition performance, large number of parameters, and high computational complexity, making them unsuitable for deployment on resource-limited mobile devices. To address these issues, a lightweight road multi-target detection algorithm combining non-adjacent features is proposed based on YOLOv7-tiny. First, the design of the Tiny-AFPN combines non-adjacent features of different scales, reducing the loss of features caused by scale differences and achieving richer cross-scale information interaction. Secondly, with the introduction of DSConv, the ELAN was redesigned and named ELAN-DS, improving the expression of features while optimizing the efficient layer aggregation network and reducing the complexity of the model. Finally, the use of the MPDIoU loss function improves the accuracy of bounding box regression and enhances the network′s target detection capabilities. In the experiments on SODA10M, compared with the original YOLOv7-tiny model, the improved algorithm increased accuracy, mAP@0.5, and recall by 1.4%, 1.4%, and 5.9%, respectively. It also reduced the number of parameters and computation by 8.2% and 41.5%, respectively. This effectively reduces the number of parameters and the computational complexity, substantially improves the detection speed of the model, and provides the possibility for deployment on edge devices.

Abstract:Affected by factors such as water attenuation and scattering, underwater optical images suffer from severe color distortion, blurriness, and other issues, resulting in a significant decline in quality and poor target resolution,which in turn is not conducive to carrying out underwater target detection tasks.To address the above problems, in order to improve the accuracy of underwater target detection and reduce the incidence of misdetection and missed detection, this paper proposes an underwater target detection algorithm based on improved YOLOv8n: ESA-YOLOv8. Firstly， the algorithm introduces the ESP module in C2f to improve the Bottleneck structure, the ESP module optimizes network efficiency and reduces the number of model parameters and computations in YOLOv8n;secondly, a small target detection layer is added to improve the detection capability of underwater small targets; finally, the lightweight up-sampling operator CARAFE and the attention mechanism ECA are successively introduced into the neck network to improve the target detection accuracy and realize the enhancement of up-sampling feature fusion.The experimental results show that on the underwater biological dataset DUO,the ESA-YOLOv8 algorithm designed in this paper achieves a mAP@0.5 of 84.7% and a mAP@0.5：0.95 of 65.5%, with a reduction in model parameters. These results represent an improvement of 1.7% and 1.8%, respectively, compared to the base model YOLOv8n.The high accuracy detection results and the reduction in the number of model parameters demonstrate the effectiveness of the improved YOLOv8n and its potential application in underwater target detection.

Abstract:Obtaining a panoramic view of the seafloor through image stitching is of great significance for understanding deep-sea topography and geomorphology. Due to the challenges posed by the deep-sea environment, seafloor image features are often blurred, making the continuous stitching of sequential images require a stable and efficient stitching network. To address the issue, this paper proposes a deep-sea sequential image stitching network called AP-LG, which combines an improved ALIKED with LightGlue. Firstly, Deformable ConvNets v2 is used to replace the original deformable convolutional networks in ALIKED, introducing an adjustment mechanism to enhance the network′s feature capture capability. Then, multi-scale feature fusion is achieved through feature pyramid networks, improving robustness of the network to environmental changes. Finally, LightGlue is employed as the core feature matching network, and based on homography estimation strategies, continuous alignment and stitching of multiple sequential images are achieved. The experimental results indicate that on the UIEBD and DISD datasets, the AP-LG network achieved matching rates of 32.91% and 49.41%, respectively, enabling 86.00% and 93.60% of the image pairs to be matched with over 100 valid feature points. The proposed method can stably extract seafloor image features, achieve feature matching, and effectively complete the stitching of sequential seafloor images.

Abstract:The ground current field characterized by field strength is widely applied in fields such as geophysical exploration, seismic monitoring, and through-earth communication. However, due to the variability of regional ground current field strength, it is challenging to establish a regional fingerprint database for ground current field strength. This paper proposes a method for constructing a regional fingerprint database of ground current field strength based on RBFNN. By employing time-division cross-injection to construct a regional ground current field and using orthogonal electrodes to detect ground current field signals at different detection points, fingerprint features of the ground current field strength are extracted. RBFNN is used to fit the field strength variation function model in Kriging interpolation, and Kriging interpolation is then employed to estimate fine-grained fingerprint features of the ground current field strength. Based on the estimation results, a regional fingerprint database of ground current field strength is constructed. An experiment to construct the fingerprint database was conducted in a natural environment of 150 m×50 m. The results show that the constructed fine-grained (0.1 m×0.1 m) regional fingerprint database of ground current field strength achieves an average construction accuracy of 89.84%, with the highest accuracy reaching 95.46%.