Abstract:Circular ripples are surface defects in the manufacturing process of contact lenses, resulting from uneven distribution of hydrogel materials, causing a concentric contraction along the edge of the lens. These defects are challenging to detect in projection inspections, leading to poor product quality. Detecting circular ripple defects poses a technical challenge in the production of contact lenses. In this study, a circular ring illumination imaging system was constructed based on the characteristics of this defect. An image model database of circular ripple defects was collected, and a lightweight detection algorithm for contact lens circular ripple defects based on an improved RT-DETR was introduced. Initially, the original ResNet18 backbone of RT-DETR was modified by replacing the BasicBlock with a lightweight FasterNetBlock. Subsequently, the SimAM three-channel attention mechanism was integrated into the Neck part of RT-DETR to enhance the model’s accuracy. Finally, the GIoU loss function was replaced with the MPDIoU loss function to accelerate convergence and improve detection accuracy. Experimental results demonstrate that the improved RT-DETR algorithm achieved a mAP@0.5 of 94% on the contact lens circular ripple database, a 3.1% improvement over the original RT-DETR algorithm. Params and FLOPs were reduced by 15.6% and 13%, respectively, compared to the original algorithm. This algorithm effectively reduces computational complexity, enhancing the mean average precision of contact lens circular ripple defect detection. It holds promise for overcoming technical challenges in online detection of circular ripple defects in contact lenses.

Abstract:The aerospace engine test bench is a key equipment for verifying engine reliability, and its health status assessment is of great significance for ensuring the safe operation of the engine. The gas circuit system of the engine test bench has the characteristics of complex and variable fault modes, strong correlation between multi-point and multimodal sensing information, etc. Moreover, there are issues such as uneven distribution of collected health status samples, high signal noise, human resource waste caused by manual monitoring of the operating status of the gas pipeline system, and high false alarm rates. To this end, a health assessment model for test benches based on adaptive reconstruction of phase space and support for high-order tensor machines is proposed. This method first involves designing stability criterion for E1(m) to achieve adaptive phase space reconstruction of the gas path system. Secondly, tensors are used to characterize the multi-point and multimodal data of the pneumatic system. Then, a high-order tensor machine is used to mine the multi-source sensor correlation information and fault modes in tensor samples, achieving a health status assessment of the test bench pneumatic system. Finally, the proposed method is compared with the support vector machine, decision tree and plain Bayesian algorithms based on the actual test data from the engine test bench of a China National Aviation Corporation (CNAC). The results show that the proposed method has a good evaluation capability in a weak data environment, with an overall evaluation accuracy of 89.7%, and the accuracy drop is kept within 8% in an extremely weak data environment.

Abstract:Phase-sensitive optical time domain reflectometer (Φ-OTDR) usually uses coherent detection to achieve long-distance, distributed and high-sensitivity vibration detection. To accurately obtain the position and phase information of the vibration signal, the quadrature demodulation algorithm is an important technology widely used at present, but the algorithm has the limitation of time-consuming. In order to solve this problem, a fast demodulation scheme of Rayleigh scattering signal based on field programmable gate array (FPGA) is proposed. The pipeline structure is used to realize the synchronization of sensor data acquisition and data demodulation. Two orthogonal signals are obtained by digital quadrature mixing technology. The finite impulse response low-pass filter is used to remove the high-frequency component, and the coordinate rotation digital algorithm (CORDIC) vector mode is used to realize the hardware demodulation of the vibration phase, which can improve the overall real-time performance of the coherent detection Φ-OTDR system. The experimental results show that the scheme can successfully realize the positioning and phase reduction of the vibration signal under the condition of a detection distance of 40 kilometer. When the detection distance unchanged and the pieces of data acquisition increased to 4 000, the FPGA demodulation scheme only takes 1.60 seconds, which is 145.61 seconds shorter than the traditional host computer CPU demodulation scheme, thus providing a reference for the real-time demodulation of Φ-OTDR vibration sensing data.

Abstract:The operating conditions of planetary transmission are mostly non-stationary operating conditions. During the operation, the gear meshing vibration signals are coupled with each other, which leads to the aliasing of test signals, and the difficulty of hidden fault diagnosis increases. At the same time, when applying complex neural network models for fault diagnosis and prediction, most of them will be limited by the hardware of industrial field edge computing equipment. Aiming at the related problems, an intelligent recognition model based on smooth and pseudo Wigner-Vile distribution (SPWVD) and knowledge distillation is proposed to reduce the parameters of the network model while ensuring the accuracy of planetary transmission fault diagnosis. First, the ensemble empirical mode decomposition (EEMD) method is used to decompose the multi-component vibration signal, then the single-component signal is selected for SPWVD calculation and linearly superimposed to obtain a two-dimensional time-frequency diagram as input. The ResNet101 is used as the teacher model to guide the student model MobileNet for training. The complex teacher model imparts the knowledge in the data to the student model, which improves the accuracy of the student model.The method is compared with similar methods. The results show that the storage cost of the model is reduced to 24.55% of the teacher model at the expense of 2.43% accuracy, which is 9.61% higher than that of MobileNet without knowledge distillation. This research method provides an effective and feasible solution to improve the practical application of deep learning model in engineering and reduce the deployment cost of edge computing equipment.

Abstract:Existing test water lubrication bearing characteristics of the important characterisation parameter water film pressure of the many methods, because the sensor distance from the real measurement point is far away, or trauma to the shaft system is large and other reasons, the accurate water film pressure empirical data is difficult to obtain, restricting the bearings of further research and development. In response to these challenges, a new monitoring method is proposed in which thinfilm sensors are embedded in bearing shingles, while pressure data are transmitted via wireless sensing. Firstly, the physical model of axial tile grooved bearing is established, and the location, structure and number of grooves are determined by finite element analysis of axial tile deformation near the grooves; A physical model of the bearing fluid domain and solid domain is established to simulate and analyse the water film pressure distribution; Then, a thin film sensor calibration method is proposed to calibrate it accurately; Finally, a multi-operating condition bearing water film pressure test experiment is carried out, it is compared and analysed with the simulation results and existing methods. The results of the study show that it is feasible to embed a thin film sensor in the groove of the bearing shaft tile and transmit the data wirelessly, and the measured data of water film pressure deviates less than 10% from the simulation results, which is more accurate than the measured data of existing methods. The water film pressure decreases along the axial direction, and there exists a part of lubrication film inside the bearing in a mixed lubrication state.

Abstract:Ultrasonic detection is a common method of steel defect detection. The classification model established by machine learning algorithm can realize effective defect identification. Neural network is the most commonly used algorithm at present, but it has the problem of complex model structure and large amount of training data. In this paper, an ultrasonic defect recognition method based on random forest is proposed, which can realize intelligent and accurate identification of defect types to solve the problems of complex model structure and large training data requirements. Firstly, ultrasonic detection experiments were carried out for defects of different shapes, sizes and depths in the specimen. Based on the experimental data, an ultrasonic defect recognition model was established using random forest algorithm. Then, the defect recognition effect of the model is analyzed, and compared with support vector machine, K-nearest neighbor classification algorithm, AdaBoosting algorithm and convolutional neural network. Then the defect identification verification experiment is carried out with the verification specimen to further verify the validity of the established defect identification model. The results show that the proposed method has the highest accuracy compared with other algorithms, and the accuracy of defect classification reaches 94.6% in the verification experiment.

Abstract:Addressing the challenges of low measurement accuracy and restricted applicability range in existing passive sound source localization algorithms, this paper proposes a passive sound source localization estimation method based on an optimal quadruple-array. This method constructs an optimal quadruple-array structure to enable multi-element point sharing, aiming to achieve fusion localization estimation of the target sound source with a reduced total number of elements, thereby enhancing localization accuracy. Spatial target localization equations are derived from the array model, transforming the problem of solving position coordinates into that of determining time delay differences between array elements. Subsequently, a second-order fractional low-order covariance algorithm is employed to resolve the corresponding time delay differences between array elements in an impulse noise environment. After obtaining the self-fractional low-order covariance of array signals and the mutual-fractional low-order covariance between two array elements, the mutual-fractional low-order covariance of both is recalculated to further mitigate the impact of impulse noise and improve time delay estimation accuracy. Finally, the obtained time delay estimation information is incorporated back into the localization equation set to achieve localization estimation of spatial sound sources. The feasibility of the proposed method and the superiority of the array structure are validated through numerical simulations and field experiments. In the field experiments, the estimation error of sound source localization is only 0.085 1 meters, demonstrating the method’s capability to achieve high accuracy in sound source localization under impulse noise environments. This work extends the application scenarios of passive sound source localization algorithms and holds practical application value.

Abstract:To address the decline in drive system efficacy due to uncertainties in the model, variations in parameters, and interruptions from external sources affecting permanent magnet synchronous motors (PMSMs), a novel control strategy is proposed. First, to reduce the reliance on the system’s mathematical model, a new super-twisting algorithm is formulated for the speed loop of the PMSM. Secondly, based on the new super-twisting model of the speed loop, a novel model-free super-twisting fast integral terminal sliding mode controller (MFSTFITSMC) is designed by integrating a new type of integral terminal sliding surface and an improved super-twisting control law, achieving precise control of the motor speed. Furthermore, a non-singular fast terminal sliding surface and a dual-power approaching law are used to devise an improved extended nonsingular terminal sliding mode disturbance observer (IENTSMDO). This observer accurately detects and provides feedforward compensation for unknown disturbances, effectively suppressing parameter perturbations and external disturbances, thus enhancing the system’s robustness and improving both dynamic and steady-state performances. Finally, through simulation and experimental comparison with traditional control methods, the proposed algorithm has been verified to improve speed overshoot resistance by 0.412 5% and enhance the torque response speed by 0.013 s. The results indicate that the proposed method possesses strong robustness and good interference rejection capabilities in the presence of unknown disturbances.

Abstract:Currently, most image denoising models based on convolutional neural networks cannot fully utilize the redundancy of image data, which limits the expressive power of the models. Moreover, edge information is often used as a priori knowledge for effective denoising, while texture information is usually ignored. To address these issues, a new image denoising network is proposed, which firstly uses the attentional similarity module to extract global similarity features of the image, and smooths and suppresses the noise in the attentional similarity module through average pooling to further improve the network performance; secondly, the dilated residual module is used to extract both local and global features of the image; finally, a global residual learning is utilized to enhance the denoising performance from shallow to deep layers. In addition, a texture extraction network is designed to extract local binary patterns from noisy images to obtain texture information, which can be utilized as a priori knowledge to preserve the details in the evolved images during the denoising process. The experimental results show that compared with some advanced denoising networks, the newly proposed denoising network has a great improvement in image vision, higher efficiency and peak signal-to-noise ratio by about 2 dB, and structural similarity by about 3%, which is more conducive to practical applications.

Abstract:In the existing algorithm for fault diagnosis in analog circuits, artificial intelligence-based fault diagnosis algorithms require a large amount of training data and long training time, making it difficult to achieve parameter identification. Traditional circuit analysis methods require multiple test points and involve complex calculations. Based on this, a fault diagnosis algorithm for analog circuits based on optimized matrix perturbation analysis is proposed. Firstly, the Laplace operator is used to convolve the output response matrix of the tested circuit, thereby enhancing the perturbation pattern between matrix elements and circuit component parameters. Secondly, the trace and spectral radius of the matrix are selected as fault characteristics, and a matrix model is established using this perturbation pattern. Then, an improved diagnostic algorithm is used to verify examples in Sallen_Key bandpass filter circuits and leapfrog low-pass filter circuits. The results show that with only one test point, the proposed method can achieve parameter identification of faulty components. The fault diagnosis rate reaches 100%, with parameter identification error controlled within 1%, and computation time controlled at millisecond level. Therefore, this method is easy to implement for online testing and suitable for situations requiring high accuracy in fault localization and precise parameter identification.

Abstract:In order to solve the problems of output saturation and signal amplification difference of the traditional two-dimensional tri-stable stochastic resonance system driven by dualinput signals (DTDTSR), a novel system, coupled piecewise symmetric tri-stable stochastic resonance system (coupled piecewise symmetric tri-stable stochastic resonance system) driven by dual-input signals, is ingeniously proposed. A novel system is proposed: coupled piecewise symmetric tri-stable stochastic resonance system driven by dual-input signals (DCPSTSR). Firstly, the problem of output saturation of the system is studied in depth, which provides a key theoretical foundation for the optimization of the system performance. Secondly, the output spectral amplification (SA) function of the system is derived within the framework of the adiabatic approximation theory. The influence of system parameters on it is analyzed in detail, which provides theoretical support for deeper understanding. Further, a comprehensive comparison of the DCPSTSR, coupled piecewise symmetric tri-stable stochastic resonance system (CPSTSR) and DTDTSR systems is carried out through numerical simulations, and the results clearly indicate that the DCPSTSR system is significantly superior to the other systems in terms of output spectral amplification function. Finally, the system parameters are precisely optimized by genetic algorithm and successfully applied to bearing fault detection. The experimental results verify the excellent performance of the DCPSTSR system and provide strong theoretical support and feasibility verification for future theoretical research and engineering applications. This design and its successful application in bearing fault detection provide a new direction and example for further research and practical application in the field of resonance systems, which has important scientific and engineering value.

Abstract:The current pedestrian detection algorithm is a research hotspot in the field of driverless driving, but the pedestrian occlusion problem has not been well solved due to factors such as relatively small sample size, diverse occlusion situations, and reduced visual features. Aiming at the problem of missed detection caused by pedestrians blocking each other or pedestrians being blocked by other objects, a pedestrian detection method based on inter-frame directional gradient histogram feature correlation is proposed. First, a tracking method is added based on the YOLOv7 baseline network model to discover missed pedestrians and estimate their location information; the nearest local image containing missed pedestrians is used as the new information, using directional gradient histogram features and support vectors, a machine-based method is used to detect pedestrians at the estimated position of the missed target to improve the missed detection phenomenon caused by partial occlusion. Experimental results compared with the baseline network, the precision (P) value of this method increased by 6.25%, and the average precision (AP) of occluded pedestrians increased from 26.67% to 53.42%. Experiments show that the pedestrian detection method based on inter-frame directional gradient histogram feature correlation can improve pedestrian detection accuracy, has low computational complexity, does not significantly increase the computational overhead of the original method, and has certain application value.

Abstract:In response to the complex imaging environment, irregular deformation of batteries, and uneven diffuse reflection of metal surfaces encountered during the automated disassembly process of retired cylindrical power lithium batteries, existing visual recognition methods are unable to accurately extract contour and pose information. We propose an accurate contour extraction method based on the Fr-chet distance similarity function and a pose detection method based on rectangles and edge morphology features. By establishing a Lambert diffuse reflection model for cylindrical lithium batteries and using morphological operation methods to obtain the rough localization contour of lithium batteries, as well as based on the similarity function defined by the Fr-chet distance, the contour is accurately extracted by classifying each pixel band in the rough localization image; Subsequently, utilizing the positive and negative terminal features of cylindrical lithium batteries, feature contours of the positive and negative terminal ROI regions are extracted employing an adaptive threshold segmentation algorithm. Finally, by comparing the rectangular values of the two end regions, the pose information of the lithium battery can be calculated. The experimental results show that in the Self-built retired cylindrical lithium battery image dataset that includes deformation, corrosion rust spots, and uneven lighting conditions, the proposed method has high accuracy in identifying lithium batteries of different models and poses. The diameter length detection error is less than 3%, and the pose detection accuracy is higher than 94%, which can meet the actual needs of automated disassembly and detection.

Abstract:Unsupervised person re-identification aims to identify the same person from non-overlapping cameras under unsupervised settings. Aiming at the problem that the existing unsupervised person re-identification network cannot fully extract pedestrian features and the difference between cameras leads to pedestrian retrieval errors, we propose an unsupervised person re-identification of adversarial disentangling learning guided by refined features. A feature refinement information fusion module is designed and embedded into different layers of ResNet50 network to enhance the ability of the network to extract key information. A disentangled feature learning method is designed to minimize the mutual information between pedestrian features and camera features, and reduce the negative impact of camera differences on the network. At the same time, the adversarial disentangling loss function is designed for unsupervised joint learning. Using the Market-1501 and DukeMTMC-reID public datasets, we tested the proposed method. The mean average precision increased by 4.6% and 3.1% respectively. Compared with the baseline algorithm, it has strong robustness and meets the needs of pedestrian recognition in unsupervised background.

Abstract:To address the problems posed by redundant features in transformer fault recognition and the low accuracy of traditional methods, a transformer fault recognition method leveraging kernel principal component analysis (KPCA) in conjunction with chaotic sparrow search algorithm (CGSSA) is introduced. Initially, KPCA is employed to preprocess the transformer fault data, aiming to mitigate the correlations among features. Subsequently, CGSSA is improved by incorporating the improved Tent map and Gaussian mutation to increase the search accuracy and convergence speed of the algorithm. Comparing the results involving CGSSA, SSA, GWO and WOA. Utilizing the data extracted through the KPCA as the model input, CGSSA is then used to select the kernel function parameters and regularization coefficient of KELM, thereby establishing the KPCA-CGSSA-KELM transformer fault recognition model. The experimental results demonstrate that, with the identical input data, CGSSA has the best results in terms of convergence speed and optimization accuracy. In addition, the proposed method shows the fault recognition accuracy of 95.7%, which is 18.6%, 10%, and 15.7% higher than WOA-KELM, GWO-KELM, and SSA-KELM, respectively. These findings suggest that the proposed method effectively manages the impact of redundant features and enhances the precision of transformer fault recognition, thus verifying the validity and feasibility of the proposed method for transformer fault recognition under the feature redundancy.

Abstract:To address the limitation of standard attention mechanisms that can only generate coarse-grained attention regions, failing to capture the geographical relationships between remote sensing objects and underutilize the semantic content of remote sensing images, a structured image description network named GRSRC (geo-object relational segmentation for remote sensing image captioning) is proposed. Firstly, considering the highly structured nature of remote sensing image features, a feature extraction method based on structured semantic segmentation of remote sensing images is introduced, enhancing the encoder’s feature extraction capability for more accurate representation. Simultaneously, an attention mechanism is incorporated to weight the segmented regions, enabling the model to focus more on crucial semantic information. Secondly, taking advantage of the well-defined spatial relationships among objects in remote sensing images, geographical spatial relations are integrated into the attention mechanism, ensuring more accurate and spatially consistent descriptions. Finally, experimental evaluations are conducted on three publicly available remote sensing datasets, RSICD, UCM, and Sydney. On the UCM dataset, BLEU-1 achieved 84.06, METEOR reached 44.35, and ROUGE_L attained 77.01, demonstrating improvements of 2.32%, 1.15%, and 1.88%, respectively, compared to classical models. The experimental results indicate that the model can better leverage the semantic content of remote sensing images, demonstrating its superior performance in remote sensing image captioning tasks.

Abstract:A mathematical method is proposed to evaluate the delay performance of RF-Mesh-Networks that can realize data classification transmission and ensure real-time data requirements, for deficiency in evaluating and analyzing the delay performance of different types of transmission data when the large-scale wireless advanced measurement instruments are applied in smart distribution grid. Based on the analysis of the WMNs architecture of the smart distribution grid, the functional relationship between the initiation stages of the two consecutive time slots is established using Markov chain modulation techniques. In order to avoid the difficulty of solving high-order differential equations during the process of obtaining steady-state solutions, the solution method based on error iteration to obtain steady-state operating points is proposed, and the detailed solution process is also provided. On the basis of obtaining the steady-state working point, the analytical formula to evaluate the average delay performance of real-time data and non-real-time data for uplink and downlink transmission. To verify the effectiveness of the proposed delay performance evaluation method for WMNs applied in smart distribution grid, the delay performance of real-time and non-real-time data is simulated and tested，where they are set up three different transmission rates. The experimental simulation and test results show that the proposed method can achieve performance evaluation and analysis of transmission delay for different types of communication data in intelligent distribution networks, and can improve the transmission performance of WMNs.

Abstract:When the preferred measurement signal for equipment condition monitoring in industrial field is acoustic signal for various reasons, it is especially necessary to propose an equipment condition monitoring method based on acoustic signal. In this paper, a certain type of centrifugal pump is taken as the basis object, and the Mel-scale frequency cepstral coefficients（MFCC）are extracted from the acoustic signals collected in the field as the initial features of the signals, then the dispersion entropy（DE）values of these MFCC initial features are calculated, and the matrix is downscaled by principal component analysis（PCA）, so as to construct the feature matrix. The penalty coefficients and kernel function parameters of the support vector machine（SVM）are optimized by using the bat algorithm（BA）to carry out diagnosis of various fault conditions of centrifugal pumps and compared with various diagnostic methods. The experimental results show that the model optimized by BA improves the diagnostic accuracy by 21.7%; on the basis of this model, the deep mining of the signals extracted by MFCC using DE improves the diagnostic accuracy of the model by 2.05%.

Abstract:With the increase of aircraft service time and the extreme service environment, fatigue cracks and other defects may occur in the multi-layer metal riveting structure of aircraft. It is of great significance for damage assessment and maintenance to find defects in time and obtain information such as defect depth and direction. However, due to the concealment of defects caused by multi-layer structure, the detection signal characteristics of conventional eddy current probes are indistinct, and conventional eddy current probes are not sensitive to defects in certain directions, making it difficult to determine the direction of fatigue cracks. To address these problems, a cross-runway-type differential eddy current probe is designed, which is mainly composed of a cross-runway-type excitation coil and two sets of differential detection coils. The feasibility of the new eddy current probe is investigated by establishing a three-dimensional finite element model for defect detection of aircraft multi-layer metal riveting structures, including the optimization of the structure of the probe, and simulations are conducted to analyze the different directions of defects, the buried depths and the lift-off heights, respectively. The results indicate that the new probe can effectively detect deep defects with a buried depth of 6 mm and dimensions of 10 mm×1 mm×1 mm, and it can obtain the direction information of the defects. Compared to traditional probes, designed probes have advantages such as no missing defects in all directions, resistance to lift off effects, and high resolution. The research results can provide some reference for the design of eddy current probes for aircraft multi-layer metal riveting structure detection.

Abstract:Large scale plant protection machinery in non-standard orchards in mountainous areas has poor accessibility, and small wheeled plant protection robots have broad application prospects. A path planning algorithm for wheeled plant protection robots based on improved ACO-DWA algorithm is proposed to solve the problems of visual information misjudgment caused by closed orchard branches and leaves, as well as delayed obstacle avoidance caused by complex working terrain. Firstly, the orchard environment information is obtained through LiDAR, and the voxel grid method is applied to simplify the point cloud density. The grid method is used to segment the ground point cloud, and the K-means algorithm is used to extract the robot’s inter row passage area. Combined with the kinematic model and job specification constraints of the plant protection robot, a series of candidate trajectory sets are generated using the model based prediction algorithm (SBMPO). Then, using the improved ACO-DWA algorithm, the robot’s travel cost is integrated into the objective function of the search node, and path planning is carried out online based on the environmental map. Finally, simulation validation and real-world deployment experiments were conducted using MATLAB R2021 simulation platform and robot ROS operating system, respectively. The experimental results show that this method can significantly improve the traffic capacity of robots in complex orchard scenes, and the path planning effect and operational efficiency are significantly improved.

Abstract:Aiming at the problem of irregular and random distribution of defects on the surface of texture images, such as scratches and cracks, which leads to low accuracy of defect detection, a self-supervised defect detection method based on the bi-radial fusion of positive and negative sample difference features is proposed. Firstly, Otsu threshold segmentation is used to extract image foreground information, and Perlin noise is superimposed on the data-enhanced positive samples or the texture images, from the DTD dataset, to simulate defects on the positive sample images and synthesize the negative samples. Then, the mean-square error is calculated for feature matching using the intermediate features output from the encoder, while the coordinate attention (CA) and path aggregation network (PANet) are combined to enhance the information fusion of the matched features. Finally, the fused features are input into the decoder together with the low-level and high-level features output from the encoder, and the weights of Focal, L1, and Dice loss functions are optimized and adjusted to realize the prediction of the defective masks more accurately. Experiments show that the average imagelevel and pixel-level AUROC of the proposed model on the texture category of the MVTec AD dataset reaches 0.995 and 0.968, respectively, which improves the classification and segmentation accuracies compared with the other defect detection models, demonstrating the effectiveness of the proposed method in texture defect detection.

Abstract:The effective monitoring of hydrocarbon gas content is an important aspect of safety assurance in oil and gas exploration and production processes. Infrared spectroscopy, as a safe and efficient detection method, has attracted the attention of on-site engineers. However, it mainly uses offline models for measurement, which cannot cope with the complex working conditions and various nonlinear influencing factors on site, making it difficult for this non updated model to maintain high prediction accuracy. A weighted kernel partial least squares method based on fusion of similarity measurement criteria in just-in-time learning for quantitative analysis of alkane gases is proposed in this paper. Firstly, a similarity criterion based on fusion of multiple similarity measurement criteria is designed to effectively select historical samples for online modeling. Secondly, nonlinear kernel functions are introduced into local PLS models to effectively extract nonlinear features and compensate for the nonlinear processing ability of linear partial least squares models. The experimental results on the multi-component mixed gas infrared spectral data have verified the effectiveness of this method, with a goodness of R2 of 0.994 1. Compared with that of the PLS model, the RMSE and MRE of the proposed model have improved by 43.6% and 85.8%, respectively. It can be effectively used for online updating and high-precision prediction of infrared spectral quantitative analysis models for hydrocarbon gas.

Abstract:The process of proton exchange membrane fuel cells (PEMFC) involves strong coupling of multiple physical fields, components, and factors, inevitably leading to prolonged performance degradation and local performance fluctuations during operation. However, effectively identifying key features from the multitude of parameters under the multiple couplings and capturing the overall performance degradation trend becomes exceptionally challenging. In response to these issues, a PEMFC degradation prediction model based on XGBoost and Self-Atten-LSTM is developed. First, a wavelet threshold denoising method is employed to remove noise interference from the original PEMFC data. Then, the XGBoost algorithm is used to select the main features significantly affecting PEMFC performance from the numerous parameters, achieving precise feature selection. Finally, the introduction of the self-attention mechanism in LSTM addresses its limitations in global modeling and complex interaction among multi-dimensional vectors when dealing with long sequences. Through adaptive weighting, it more effectively utilizes PEMFC degradation information. Compared to traditional LSTM, Bi-LSTM, and GRU models, the developed model can more accurately predict fuel cell degradation under both steady-state and dynamic conditions. The model exhibits a reduction in the average mean absolute error by 56.34% to 77.04%, with a predictive accuracy of up to 99.09%. This approach can find broad applications in developing vehicle operation and maintenance strategies and enhancing system reliability.

Abstract:The non-contact voltage measurement method is not in direct contact with the metal conductor of the line and can adapt to the voltage monitoring in a variety of application scenarios. This paper designs a system which uses the improved non-contact voltage measurement technology to measure the line voltage and applies the measured voltage waveform to the line fault voltage diagnosis. Based on the topology analysis of the traditional non-contact voltage measurement technology and the improvement of the measurement circuit topology, the voltage on the line can be measured accurately without being affected by the coupling capacitance. Because of the limitation of the current single fault feature extraction method, in order to accurately identify and diagnose the line fault voltage by using the voltage waveform measured by the non-contact voltage measurement technology, in this paper, a fault voltage state identification system based on integrated learning is proposed, and a variety of feature extraction methods are used to extract the voltage waveform features obtained from non-contact voltage measurement. The identification results are used for early warning and processing of line faults. In this paper, aiming at the voltage monitoring system, the measurement accuracy and fault identification test are designed, and the steady-state average error is 0.9%, and the highest fault identification accuracy is 93%, which shows that the voltage monitoring system has high accuracy and fault identification accuracy.

Abstract:The improved Harris Hawks optimization algorithm (IHHO) is proposed to solve the problem that analog circuit fault diagnosis is difficult due to multiple fault types, unstable fault states and redundant fault data. IHHO optimized back propagation (BP) neural network to realize fault feature selection and diagnosis of analog circuits. Firstly, the nonlinear adaptive factor, Cauchy variation and stochastic difference perturbation are introduced into the Harris Hawks optimization algorithm to improve the convergence speed and accuracy. Secondly, IHHO is used to select the characteristics of the single fault and the combined fault simulation data of the analog circuit to complete the data preprocessing. Finally, IHHO-BP algorithm is used to train and test the preprocessed fault data to realize the fault diagnosis of analog circuits. The diagnostic results show that the proposed method improves the diagnostic accuracy by 5.5% compared with other algorithms.