Abstract:Nanofilm thickness is a critical parameter influencing the performance of micro/nano devices, particularly for films on rough surfaces, where spatially resolved measurements encounter significant challenges such as optical scattering interference and imaging non-uniformity. With the rapid development of micro/nano fabrication technologies, complex nanofilms have become extensively utilized in semiconductor devices, optoelectronics, energy, and sensor applications, increasing the demand for high-precision thickness measurements of large-area, rough-surface films. To address the limitations of conventional methods, including limited scanning areas, long measurement times, susceptibility to scattering interference, and imaging non-uniformity, this paper proposes a novel imaging-based film thickness measurement technique using differential reflection spectroscopy. Firstly, the measurement wavelengths are optimized to enhance measurement efficiency significantly while maintaining accuracy. Secondly, a microscopic optical system with conjugate imaging relationships is designed to ensure precise correspondence between object and image planes, achieving uniform microscopic imaging across centimeter-scale areas. Thirdly, a multi-frame averaging strategy is employed to effectively suppress measurement errors caused by system noise, enhancing the signal-to-noise ratio and stability. Finally, signal compensation using effective medium and phase-change models further enhances the accuracy of nonlinear fitting. Based on the proposed method, an experimental system was established, and thickness measurements were conducted on smooth SiO2/Si, smooth TiO2/Ti, and rough TiO2/Ti samples. The experimental results demonstrate that the proposed system achieves high spatial-resolution and accurate thickness measurement of nanofilms within a rough surface area of approximately 1 cm2, validating its effectiveness and strong practical applicability.

Abstract:The cross-scale piezoelectric positioning platform, with its millimeter-scale travel range and nanometer-level positioning accuracy, is considered a core component for enabling atomic-scale manufacturing systems. However, the inherent hysteresis nonlinear characteristics of piezoelectric ceramic actuators severely restrict their trajectory tracking accuracy under micro-positioning mode, making it difficult to meet the sub-nanometer precision requirements. In this context, the precise identification of hysteresis model parameters becomes a crucial step in improving tracking accuracy. To address the issue that traditional particle swarm optimization struggles to obtain the global optimal solution for identifying multi-dimensional complex hysteresis model parameters, a multi-modal Bayesian optimization PSO parameter identification algorithm is proposed. First, a system dynamics model capable of characterizing hysteresis nonlinearity is formulated based on the asymmetric Bouc-Wen model. Then, based on the analysis of the parameter identification results of the traditional PSO algorithm, a modal division mechanism is introduced, and Bayesian optimization is integrated for local search. In this way, the MB-PSO algorithm is developed. This is intended to overcome the limitations of traditional identification algorithms, such as being easily trapped in local optima. To evaluate the effectiveness of the proposed algorithm, a comparative experiment is implemented against PSO, the genetic algorithm, and the differential evolution algorithm. The experimental results show that the parameter fluctuation under MB-PSO is suppressed to within 12.3%. Furthermore, to evaluate the efficacy of the algorithm in improving both hysteresis modeling accuracy and trajectory tracking performance of the piezoelectric positioning platform, comparative tests on hysteresis characterization and trajectory tracking accuracy are carried out. The results show that the maximum displacement error of the hysteresis model identified by MB-PSO is limited to 6.871 nm, and the trajectory tracking error, when incorporating this model into the control system, is reduced to within 0.976 nm, achieving sub-nanometer tracking accuracy. The foundation is laid by this work for atomic-level manufacturing to be achieved with piezoelectric positioning platform systems.

Abstract:The fused silica micro hemispherical resonator is the core structure of the micro hemispherical gyroscope, and the vibration testing of the uncoated transparent resonator is crucial for optimizing the gyroscope process and ensuring its performance. The traditional vibration measurement technology has limitations. The capacitance detection method is restricted by the requirements of surface coating processes and cannot be applied to transparent resonators. Although commercial benchtop laser vibrometers are capable of non-contact measurements, their large volume, high cost, and limited suitability for integration make them challenging to meet the demands of compact system designs. Therefore, this article proposes a new method of vibration measurement based on the fiber Doppler effect and constructs a miniaturized fiber vibration measurement system. The 1 550 nm single-mode fiber laser is employed as the system light source, with its output beam split into local oscillator and signal light via a fiber coupler. The signal light is directed onto the surface of the vibrational resonator, where it is reflected and then combined with the local oscillator beam, which has been frequency-shifted by an acousto-optic modulator, to generate a heterodyne signal. The signal processing module converts the heterodyne optical signal into an electrical signal using a photodetector and then demodulates it using an arctangent algorithm to extract the resonator′s vibration parameters. The system utilizes a six-degree-of-freedom fixture to ensure spatial alignment between the optical fiber and the resonator. The experiments are conducted in a vacuum environment of 70~80 Pa, and the results show that the resonator′s response amplitude is directly proportional to the excitation voltage, with a relative measurement error of less than 1%. The vibrometry system is capable of accurately identifying the first three modal frequencies within the range of 0~9 000 Hz, yielding results consistent with those obtained from the commercial Polytec MSA-600 vibrometer. Additionally, the system is able to detect frequency splitting at the operational mode of 3.9 Hz and the distribution of the four-wave vibration pattern. This fiber optic vibrometry system achieves sub-nanometer level measurements. It is compatible with uncoated resonators, which offers the advantages of compact size and low cost. It provides a reliable solution for micro-hemisphere gyroscope resonator testing and process optimization.

Abstract:Absolute ground-based gravity measurement holds significant applications in various fields such as fundamental science, geological hazard early warning, and national defense security. At present, common terrestrial absolute gravity measurement methods mainly include indoor static measurement, outdoor mobile measurement, and vehicle-mounted static-base measurement, which struggle to balance mobility and measurement accuracy. To overcome the aforementioned issues, a novel vehicle-mounted dynamic-base atomic interferometric absolute gravimeter is proposed. Through the optimized design of the laser system and control system, it enables one-key startup, significantly enhancing the level of system integration and automation. During measurement, the sensor head is rigidly attached to the moving base and coupled to the ground via a support structure, effectively suppressing the impact of carrier vibrations on measurements. In indoor environments, the system achieves a measurement sensitivity of approximately 147 μGal/Hz, with a measurement resolution better than 1 μGal at an integration time of 40 000 seconds. Continuous flow gravity measurement experiments across typical multi-scenario environments are implemented in urban and mountainous field environments, demonstrating stable system operation. Under field fast measurement conditions, the sensitivity is approximately 173 μGal/Hz, which is relatively close to the performance achieved in indoor environments, indicating good robustness. The vehicle-mounted system can be rapidly mobilized to designated points and deployed, obtaining high-precision gravity data within 10 minutes. The test routes covered various typical field terrains, including urban buildings, urban roads, mountainous roads, and mountainous open areas. Test results show that, in complex field environments with altitude changes exceeding 1.6 km and temperature variations of approximately 7.6℃, the gravity measurement residuals are better than 100 μGal, demonstrating excellent system stability. The proposed vehicle-mounted gravimeter achieves high-precision fast measurement under long-distance mobility conditions across diverse field terrains. It exhibits significant application potential in areas such as the fast establishment of gravity reference networks, rapid acquisition of short-wavelength gravity field information, and natural disaster forecasting under field conditions.

Abstract:The unmanned aerial vehicle (UAV)-towed geomagnetic surveying technique provides significant benefits in minimizing platform-induced magnetic interference, making it a highly effective approach for high-precision aeromagnetic surveys. This study introduces a novel towed geomagnetic surveying system based on a Coherent Population Trapping (CPT) atomic magnetometer, which uniquely utilizes a single-probe architecture for the first time. By harnessing the high sensitivity and omnidirectional self-sensing capabilities of the CPT atomic magnetometer, the system facilitates high-precision geomagnetic measurements across all latitudes. To validate the system′s performance, multiple flight experiments were conducted in offshore regions. A comprehensive data processing and magnetic mapping methodology was developed, incorporating key techniques such as heading error compensation, suppression of platform magnetic interference, correction of geomagnetic diurnal variations, adjustment of the geomagnetic reference field, and high-resolution interpolation-based mapping, significantly improving the accuracy and reliability of the measurements. Experimental results revealed that before data processing, the total accuracy of airborne magnetic measurements was 2.517 nT (1σ). After applying extensive error compensation and optimization, the accuracy improved markedly to 0.849 nT (1σ), effectively reducing systematic errors. Additionally, to evaluate the system′s stability under varying measurement conditions, data from two independent survey flights were processed identically and analyzed for consistency. The results showed a correlation coefficient of 99.8% between the two independently generated geomagnetic anomaly maps, with a root mean square error of 1.149 nT, confirming the system′s excellent data consistency and repeatability. These findings demonstrate that the developed CPT atomic magnetometer-based towed geomagnetic surveying system successfully addresses the challenges of platform-induced magnetic interference in conventional aeromagnetic surveys, while achieving high-precision magnetic field measurements and exceptional stability in real-world flight experiments.

Abstract:Mass imbalance in hemispherical resonators—caused by asymmetric defects introduced during manufacturing and assembly processes such as material anisotropy, machining inaccuracies, and assembly misalignments—leads to non-ideal vibration modes. These imperfections result in quadrant-wave standing wave distortion, intensify Coriolis effect coupling errors, and ultimately degrade the output performance and navigation accuracy of hemispherical resonator gyroscopes (HRGs). To tackle this key challenge limiting the development and large-scale production of high-precision HRGs, this study introduces an identification method that combines virtual rotation state parameter demodulation of standing waves with support-beam vibration detection. Laser vibrometry is employed to precisely capture the resonator’s vibration characteristics, enabling quantitative characterization of mass imbalance. A systematic investigation reveals how precision forming, metallization coating, and high-accuracy assembly each impact the mass distribution, highlighting the correlation between process-induced errors and vibration mode coupling. Based on these insights, a hierarchical balancing strategy is proposed. To implement it, an automated balancing system is developed, integrating standing wave control, vibration measurement, and micro-material removal technologies. Experimental validation shows that after balancing, the 4th harmonic frequency split is reduced to 0.056 mHz, and the coupling vibration amplitudes of the 1st-3rd harmonics are kept below 0.03 nm, given a 4th harmonic amplitude of 0.15 μm. These results verify the effectiveness and precision of the proposed balancing method, offering strong technical support for the advancement and mass production of high-precision HRGs

Abstract:High-precision spot localization algorithms are critical for ensuring the accuracy of laser triangulation displacement sensors. To address the impact of unstable laser spot imaging caused by lighting environment variations on sensor precision, a spot localization algorithm based on arctangent edge modeling is proposed. The method first employs bidirectional differential processing to extract envelope signals from initial spot data, effectively eliminating interference such as background noise, speckle noise, and spike noise. An arctangent edge model is then formulated through symmetrical shift processing of the arctangent function, generating a fitting function that aligns with spot waveform characteristics. To enhance fitting accuracy, physically meaningful initial parameter calculation formulas are derived, establishing relationships between fitting parameters and spot features. A gradient algorithm is used to extract physical characteristics from envelope waveforms, forming a feature set from which initial fitting parameters are calculated. Finally, the Levenberg-Marquardt method is applied for least squares fitting to determine optimal parameters, achieving sub-pixel-level spot centroid localization. Verified through laser triangulation sensor experiments the algorithm shows reduced single-point measurement repeatability errors to 0.04 pixels and full-scale nonlinearity errors to 0.02% across various imaging conditions, including submerged, normal, and overexposed scenarios. By mathematically modeling spot intensity distribution, this algorithm lowers sensor requirements for harsh imaging environments and reduces dependence on laser adjustment algorithms. Its real-time processing capability with single-frame handling, moderate complexity, and suitability for integrated development of high-speed, high-precision laser triangulation sensors make it an ideal solution.

Abstract:The atomic interferometer gyroscope (AIG), as one of the next-generation ultra-high-precision inertial sensor solutions, holds significant application value and potential development in defense and fundamental scientific research. However, its complex noise characteristics severely limit its actual performance. To address this critical issue, this article proposes a noise suppression method combining an ARMA model based on time series analysis with the Kalman filtering, aiming to bypass the complex physical modeling of noise sources and directly perform holistic modeling and filtering on the gyroscope′s output signal. First, the gyroscope output data are preprocessed via first-order differencing to meet the stationarity requirement of the ARMA model. The optimal ARMA (2,1) model parameters are determined through calculation and comparison using the AIC and BIC criteria. On this basis, an adaptive Kalman filtering algorithm for measurement noise is designed, which dynamically adjusts the noise covariance matrix by estimating the measurement noise variance in real time, effectively overcoming the parameter rigidity issue of traditional fixed-parameter filters. Experimental results from processing and analyzing 13 hours of atomic interferometer gyroscope output data demonstrate that the proposed adaptive Kalman filtering significantly enhances gyroscope performance. The bias stability improves from 0.076 6°/h to 0.055 0°/h (a 28.2% enhancement), the short-term sensitivity is optimized by 26.7%, and the long-term stability is improved by 20.1%. These improvements are notably superior to those of fixed-parameter filtering (only an 8% improvement). Furthermore, compared with non-model-based filtering methods (such as low-pass filtering and wavelet denoising), the adaptive Kalman filter exhibits superior noise suppression under model-matching conditions. The proposed method provides a practical and effective technical solution to overcome the challenges of complex noise modeling in atomic interferometer gyroscopes and enhance their real-world application performance.

Abstract:The cesium optically pumped atomic magnetometer is a quantum precision measurement device that operates based on optical pumping and magnetic resonance effects. It detects magnetic fields by exciting cesium atoms with a spectral lamp and probing their spin Larmor precession. This device exhibits advantages such as high measurement sensitivity, rapid response, and compact structure, and is widely applied in geophysical exploration, resource detection, and national defense security. This study systematically investigates the development of the cesium optically pumped atomic magnetometer, focusing on the core operational principles and breakthroughs in three key technologies: 1) The design of a magnetically clean probe, which optimizes the geometric structure and material to suppress external magnetic interference and enhance signal-to-noise ratio; 2) A magnetic-free temperature control system, employing high-precision temperature regulation modules to ensure the stability of atomic polarization; and 3) A low-phase-noise self-oscillating circuit, achieved through optimization of circuit parameters and feedback control strategies to minimize system noise and improve measurement sensitivity. During the integrated system development, the spectral lamp, sensor probe, and signal processing modules were co-optimized. Performance metrics were verified through calibration by 1st Class Weak Magnetic Metering Station of National Defense Metrology Station, demonstrating a measurement range of 10 000~120 000 nT and a sensitivity of 0.6 pT/Hz, which meet international advanced standards. Field tests under complex geomagnetic conditions confirmed the device′s stable performance, effectively addressing the demands of weak magnetic field detection for high precision and reliability. This research provides technical support for the domestic development of high-precision magnetic measurement equipment and contributes to the advancement of quantum precision measurement technologies in resource exploration and national defense applications.

Abstract:The frequency locking of a cold atom fountain clock is achieved by synchronizing the center frequency of an externally injected microwave signal with the atomic transition frequency, thereby obtaining a highly accurate frequency reference. In conventional systems, the detected frequency error signal is processed by a digital PID controller to generate a correction term that adjusts the microwave center frequency accordingly. To facilitate optimization of the fountain clock locking process, a simulation model of the frequency locking loop is developed using Python. Two key parameters of the model are determined from experimental data. The standard deviation of the additional Gaussian white frequency noise during locking is σ/τ=1.35×10-13（for τ=2.4 s）, and the proportional coefficient between the measured transition probability difference and the frequency error is C=2.8. Based on this model, a fuzzy PID control scheme is introduced into the frequency locking loop to enable dynamic tuning of the PID parameters, thereby enhancing system robustness and disturbance rejection capability. The simulation framework is first employed to optimize and select relevant experimental parameters, followed by short-term experimental measurements under both conventional PID and fuzzy PID control conditions. Both simulation and experimental results show that the fuzzy PID controller provides superior short-term frequency stability compared to the traditional PID method. Allan variance analysis indicates an improvement of approximately 14.2% in short-term stability, and the strong agreement between simulation and experimental results confirms the validity of the developed simulation model. Furthermore, independent simulations show that the fuzzy PID controller exhibits effective suppression of sudden frequency jumps (±1×10-11) while maintaining a comparable response speed to the conventional PID under systematic frequency steps (±5×10-12).

Abstract:Typical aircraft fasteners are characterised by their extensive variety and substantial production volumes. However, the recognition of these components—which vary in size, exhibit complex geometries, and often appear in disordered orientations—remains a pressing challenge in practical applications. To address this issue, a machine vision-based aircraft connector recognition and classification algorithm is proposed. Firstly, Gaussian filtering is employed to eliminate image noise, followed by a dual-threshold binarisation method to extract edge transition zones. Subsequently, an image coordinate system is established to locate connector edges, partitioning the image into four quadrants. The geometric centres of connectors within each quadrant are then positioned, and key parameters of edge transition zone points are calculated. Finally, the components are identified and classified through a tolerance-based visual recognition algorithm, delineation of enclosed measurement regions for the fasteners, a support vector machine preset value algorithm, and a corner point recognition algorithm based on secondary regions of interest. On a machine vision image processing experimental platform, four distinct types of aircraft fasteners—gaskets, threads, retaining rings, and nuts—were employed as test subjects to validate the detection and recognition accuracy. Building upon this, the algorithm′s solution process is demonstrated. Experimental results indicate that the average detection time per image is 2.14 seconds, with an average classification accuracy of 95.02%. Individual part recognition takes 0.54 seconds, with an error rate of only 4.98%. The SIFT and Hu algorithms achieved average classification accuracy rates of 90.29% and 72.42%, respectively, with individual part recognition times of 1.16 seconds and 1.34 seconds. The recognition time differences between the two detection methods were 0.62 seconds and 0.80 seconds, while the accuracy differences were 4.8% and 22.65%. These results demonstrate that the proposed method meets the requirements for rapid and accurate recognition of aircraft fasteners.

Abstract:Due to the unique imaging mechanism of synthetic aperture radar (SAR), existing deep learning detection algorithms struggle to achieve an optimal balance between accuracy and speed. To address the requirements for edge applications, this paper proposes a lightweight SAR image aircraft target detection network, SAERFDnet, which integrates pruning techniques for optimization. Based on YOLOV8n, SAERFDnet utilizes re-parametrized large kernel convolutions for feature extraction, while the neck of the network incorporates an adaptive multiscale discrete feature fusion module, providing a larger effective receptive field with a shallower network depth. Additionally, a deformable convolution is introduced in detection head classification branch to enhance the network′s focus on the geometric feature differences of different target classes. A frequency-adaptive dilation convolution is employed in the regression branch to strengthen the model′s ability to locate targets in high-frequency image regions. Finally, model pruning is applied to further reduce the model size and improve computational efficiency. Experiments conducted on three publicly available datasets demonstrate that the proposed method achieves 96.3% mAP50 and 72.5% mAP50-95 on the SAR-AIRcraft-1.0 dataset, with 0.5M parameters and 2G FLOPS, representing a reduction of 83.3% in parameters and 75.3% in FLOPS compared to the YOLOv8n model, while improving detection accuracy by 0.7% mAP50 and 2.2% mAP50-95. Compared to other models, the proposed method effectively improves detection efficiency in SAR image aircraft target detection while maintaining high detection accuracy. Furthermore, transfer experiments on the SADD dataset and GaoFen-3 aircraft target dataset show that the proposed method exhibits excellent generalization and robustness.

Abstract:To address the demand for high-precision machine vision measurement of the relatively large-sized parts, monocular telecentric systems suffer from a limited measurement range, while parallel telecentric systems risk lens interference. This paper therefore proposes an orthogonal telecentric machine vision measurement method and system. Firstly, the orthogonal telecentric system is established, with a detailed explanation of its composition principle and workflow. Subsequently, the calibration method for this system is investigated, introducing a mirror coordinate transformation model into the monocular telecentric imaging model and unifying the coordinate systems of the two non-overlapping-field-of-view telecentric systems using a calibration artifact. Next, key algorithms for the system are studied, including image quality evaluation based on multi adjacent pixel gradients, two-step corner localization based on image gradient information and coordinate linear optimization, sub-pixel edge detection, and automatic measurement point recognition and localization based on skeleton extraction and depth first search strategy. Finally, experimental studies and a comprehensive error model are employed to verify the system′s measurement accuracy and analyze key influencing factors. Experimental results show that the absolute error does not exceed 0.012 mm when measuring Grade 0 gauge blocks. For glass mold bottleneck dimension measurements, the maximum absolute error, minimum absolute error, and root mean square error (RMSE) are 0.035, 0.002, and 0.014 mm, respectively. The proposed orthogonal telecentric system demonstrates high measurement accuracy and consistency, enabling high-precision measurement of the relatively large-sized parts.

Abstract:Lane detection is a core task in autonomous driving perception systems, holding significant application value in complex traffic environments. While existing methods perform well under normal conditions, lane detection still faces challenges such as blurriness, disconnection, and occlusion in adverse scenarios like low light, backlighting, heavy fog, rain, and snow. To improve lane detection performance in these harsh conditions, this paper proposes the α-SimADNet detection network, built upon the ADNet framework. This model performs anchor point extraction and parameter regression using ADNet, while enhancing the backbone network′s feature discrimination and environmental adaptability by introducing negative sample contrastive learning and a mask twin network with an alternating optimization strategy. These enhancements significantly improve the model′s feature representation capabilities in challenging environments, without increasing computational overhead during inference. Additionally, to address the insufficient gradient response from traditional IoU loss in the regression of difficult samples, we introduce the power-adjusted α-GLIoU loss function to improve the model′s ability to fit broken and occluded lane lines. To thoroughly assess the proposed method′s performance, we constructed a high-quality lane detection dataset, HardLane-F100, focused on harsh environments, which includes 106 video segments and 10 600 image frames. This dataset effectively mitigates the current public datasets′ lack of extreme environmental samples. Experimental results show that α-SimADNet achieves an F1@0.5 score of 83.2% on the HardLane-F100 dataset, outperforming mainstream methods ADNet and RVLD by 2.7% and 1.2%, respectively. Under the more stringent F1@0.7 metric, it scores 60.9%, improving by 3.8% and 3.2% compared to ADNet and RVLD, respectively. This method demonstrates superior performance across various challenging scenarios, fully proving its effectiveness in harsh environments.

Abstract:To address the inability to model the long-distance dependence of features due to the limitation of local receptive fields, and the destruction of topological structure caused by the window division strategy for point cloud data in most 3D object detection, this article proposes a global voxel feature interaction-based 3D object detection method. First, a long-range context feature extraction module based on the Hilbert space-filling curves and Mamba is designed. It employs Hilbert curve ordering to serialize the voxel space while preserving spatial locality among voxels, and leverages the capability of Mamba in processing long sequences to capture point cloud context features with long-range dependencies, significantly enhancing the ability to model global contextual relationships. Secondly, an adaptive voxel diffusion module based on feature map intensity is introduced, which facilitates large-scale long-range feature interactions between voxels by dynamically generating diffused voxels to enhance the semantic representation capacity of target center voxels. Furthermore, a spatial feature recovery operator is proposed to compensate for information loss during serialization and aggregation, leveraging the local structure preservation of submanifold convolution and the global modeling capability of Mamba to further synergistically optimize both local and global feature representations. Experiments on the KITTI dataset show that the method achieves state-of-the-art performance, with 82.36%, 61.96%, and 66.05% accuracy on the car, pedestrian, and cyclist classes at moderate difficulty, while maintaining a high inference speed of 19 frames per second (FPS). The proposed method represents a superior balance between accuracy and efficiency. In addition, by comparing our method with others in real road scenes intuitively. It demonstrates that the proposed method has strong generalization ability and practical application potential.

Abstract:At present, wind turbine blade inspection suffers from high-resolution imagery, extremely small defects, and complex morphologies, making accurate identification and localization of surface flaws difficult. To address these challenges, this article proposes a small object perception-enhanced defect detection method for turbine blades. Firstly, a dynamic channel spatial convolution module is constructed to improve the YOLOv8 detection network. By using spatial and channel reconstruction modules, the computational load of the model is reduced and feature extraction redundancy is lowered, thereby enhancing the detection performance of the model. Secondly, a small object perception enhancement network is designed, which consists of a multi-scale Transformer block, a feature fusion module, and a small object detection head. The multi-scale Transformer block assists the network in understanding the semantics of the surrounding areas of small objects, including a multi-scale fusion module, a multi-layer perceptron, and a query selection module, to achieve coarse extraction of small object defect features. Subsequently, bilinear interpolation and context-guided attention fusion mechanisms are employed to align the size and semantics of shallow and deep defect features, enhancing the model′s perception of small object defects. Finally, an adaptive distribution powerful IoU Loss function is introduced to improve defect localization accuracy and reduce the impact of class imbalance on detection accuracy. Experiments implemented on a self-built offshore wind turbine blade dataset demonstrate that the proposed defect detection network achieves an average precision of 0.815. Compared with the YOLOv8 and RT-DETR models, it shows improvements of 0.134 and 0.182, respectively. Moreover, it achieves an inference speed of 14 frames per second on an RTX3090 GPU, meeting the requirements for real-time detection and further proving its potential for application in wind turbine blade defect detection.

Abstract:Multi-modal vehicle trajectory prediction, as a bridge between perception and decision planning, plays an important role in autonomous driving systems. Aiming at the problems of insufficient feature fusion and difficulty in balancing the prediction accuracy and efficiency of existing methods, a vehicle multimodal trajectory prediction model based on Hierarchical Feature Fusion and End-point Induction (HFF-EI) is proposed. Firstly, One-dimensional residual convolution and a feature pyramid network (FPN) are used to encode the vehicle historical trajectory information, thereby fully extracting the relevant features. Then a hierarchical feature fusion structure is constructed, and local feature fusion is carried out for the vehicle and the map, followed by global feature fusion, achieving efficient and comprehensive fusion of scene features across all elements. Secondly, a multi-layer perceptron (MLP) based on the dynamic weight model is introduced for trajectory endpoint prediction, enhancing the adaptive ability of the model under different traffic scenes. Finally, an endpoint refinement module based on endpoint information interaction is proposed, which uses the attention mechanism to interact trajectory information in a longer spatial and temporal ranges. Ablation and comparative experiments were conducted on the public dataset Argoverse1. Results of the ablation experiments show that the three modules of the HFF-EI model effectively improve the performance of trajectory prediction, and reduce the minimum average displacement error, minimum final displacement error, loss rate and minimum final displacement error with penalty by 8.87%, 13.52%, 31.07%, and 8.93%, respectively. On the test set, the minimum final displacement error is 1.134 m, the minimum final displacement error with penalty term is 1.773 m and the inference time is 10.22 ms, which proves the effectiveness of the proposed model by its comprehensive performance advantages compared with the 10 benchmark models.

Abstract:To address the critical limitations of conventional methods for detecting trace moisture content in transformer oil-such as destructive sampling and poor anti-interference capability-this study proposes a non-invasive, high-precision detection technique based on the integration of broadband ultrasonic time-frequency analysis and deep learning. The approach combines broadband multi-frequency ultrasonic scanning with deep neural network modeling to dynamically characterize and quantitatively predict complex acoustic signatures associated with trace moisture in oil. A dataset comprising 240 transformer oil samples was established, with ground-truth moisture content calibrated using a precision trace moisture detector (220 samples for training, 20 for testing). Each sample was subjected to broadband ultrasonic excitation at six distinct center frequencies. Echo signals were processed via Continuous Wavelet Transform (CWT) for time-frequency analysis, extracting a 128×1 000-dimensional joint high/low-frequency feature matrix as model input, with actual moisture content as output. The core innovation lies in constructing a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) deep learning model: The CNN branch efficiently extracts spatial patterns from the CWT time-frequency spectrograms, while the LSTM branch captures temporal dynamics and cross-frequency dependencies within ultrasonic features. This synergy establishes a robust nonlinear mapping between complex acoustic characteristics and moisture content. Comparative experiments with multiple models demonstrate the superior performance of the CNN-LSTM framework, achieving an exceptionally low mean absolute error (MAE) of 1.33 mg/kg, a mean absolute percentage error (MAPE) of 7.167%, and a high coefficient of determination (R2) of 0.958. This research provides a novel, industrially viable solution for online, non-destructive, and high-accuracy monitoring of trace moisture in transformer oil.

Abstract:To address the autonomous decision-making challenges of a spacecraft manipulator performing a rotating target capture task in a complex space environment, this article proposes an improved distributed deep deterministic policy gradient decision-making method to further enhance the autonomous decision-making capabilities of the capture task. The capture spacecraft is equipped with a three-degree-of-freedom manipulator for capture, while the target spacecraft is fixed and rotates at a constant angular velocity. To improve the exploration capability of the space capture system in complex environments, this article designs an internal reward exploration mechanism based on state entropy maximization. This mechanism calculates the Euclidean distance between the current state and each state in a minibatch, selects the minimum distance, and converts it into an internal reward through entropy calculation. This reward is then linearly superimposed with the external reward to form the final total reward, thereby improving the algorithm′s convergence speed. Furthermore, this article constructs a dual-network architecture. Two value networks evaluate candidate actions in parallel, and two policy networks select and execute the action with the best value. A reward reshaping function is introduced to reshape the reward signal to reduce estimation bias and improve sample efficiency. Finally, simulations and comparisons with several mainstream reinforcement learning algorithms, evaluate the effectiveness and superiority of the proposed method. Specific experimental results show that the improved D4PG algorithm has increased the reward value by 32.25% and the convergence speed by 3.08%, significantly improving the autonomous decision-making ability of the spacecraft robotic arm in performing space capture missions.

Abstract:Downhole inclinometers are susceptible to interference from various factors during geomagnetic field measurements, including drill tool magnetization, the borehole environment, and external electromagnetic noise. This can lead to significant errors in the drill tool azimuth angle calculation. Existing multi-point analysis methods can compensate for magnetic interference to a certain extent. However, the compensated azimuth angle still exhibits significant nonlinear characteristics, making it difficult to meet the requirements of high-precision downhole measurements. Therefore, this article proposes a regularized extreme learning machine (RELM) error compensation method based on improved chaotic game optimization (ICGO). First, a multi-station analysis method is used to perform preliminary correction for magnetic interference. Subsequently, a CGO-RELM compensation model is formulated, which uses triaxial accelerometer and magnetometer data as input and outputs correction residuals. The model is compared with RELM models based on particle swarm optimization (PSO) and genetic algorithms (GA). To address the problems of insufficient population diversity and local extrema in the original CGO algorithm, three random seed mutation methods, including chaos, Gaussian, and wavelet, are introduced. A three-stage fusion strategy is proposed to improve global search capability and convergence efficiency. Experiments show that the proposed ICGO-RELM method significantly outperforms competing algorithms in both fitting performance and generalization accuracy. Compared with the traditional RELM model, the mean absolute error (MAE) and root mean square error (RMSE) of azimuth angles are reduced by 94.1% and 93.1%, respectively, and the coefficient of determination is increased to 0.999 4. These results demonstrate that this method significantly suppresses nonlinear errors, improving downhole azimuth angle solution accuracy and providing a reliable technical approach for trajectory control and measurement while drilling in complex downhole environments.

Abstract:To explore the factors influencing the electromagnetic wave resistivity logging response, numerically simulations are performed on conventional electromagnetic wave resistivity logging signals. The effects of coil operating frequency, coil source distance, and coil spacing on the amplitude ratio and phase difference of the resistivity logging signal are analyzed. A three-layer horizontal formation model is established to simulate the azimuth electromagnetic wave resistivity logging signal. The electromagnetic field distribution in any formation under this model is deduced and calculated, and the relationship between the azimuth logging response and formation boundaries is analyzed. Additionally, an experimental model with a single transmitter and dual receivers for resistivity logging electromagnetic response is constructed. Relevant signal data are collected, and the amplitude ratio and phase difference of the resistivity logging signals are computed using the all-phase spectrum analysis method. The influence of different coil parameters on the amplitude ratio and phase difference is analyzed, and the results are compared with the previous numerical simulations for validation. The results show that the numerical simulation aligns well with the experimental response behavior. Specifically, when the formation resistivity remains constant, the amplitude ratio increases with the coil operating frequency and coil spacing, but decreases as the coil source distance increases. Conversely, the phase difference increases with the rise in coil operating frequency, coil source distance, and coil spacing. By applying the all-phase spectrum analysis method, the study provides more realistic and accurate formation resistivity and interface information, which can more effectively support logging data processing and real-time geological steering.

Abstract:In response to the problem that traditional obstacle avoidance algorithms cannot guarantee trajectory similarity and smoothness when processing robot reconstruction trajectories, this paper proposes a robot reconstruction trajectory smooth obstacle avoidance scheme based on dynamic artificial potential field. Firstly, Gaussian mixture model and Gaussian mixture regression are used to model and reconstruct the sampled trajectories. The gravitational potential field of the attractor and the repulsive potential field of the obstacle are constructed separately, and a V-shaped extended potential field is superimposed in space to confine the trajectories within the demonstration area, thereby improving the similarity of the trajectories. On this basis, the dynamic artificial potential field method is adopted to guide the generation of obstacle avoidance trajectories, with trajectory tracking and obstacle avoidance achieved by designing the motion mode of attractors and the repulsive force model of obstacles. Finally, a forward-backward fusion planning strategy based on sigmoid function is proposed, which integrates the latter half of the forward planning trajectory and the latter half of the backward planning trajectory together to further improve the smoothness of the trajectory. To verify the effectiveness of the proposed scheme, simulation experiments were conducted on a human handwritten letter dataset and physical experiments were conducted on the trajectory obstacle avoidance of a six-axis robot. Physical experiments have shown that the average curvature of the obstacle avoidance trajectory generated using this scheme is only 0.035 cm-1, and the average tracking error is only 2.96 cm. Compared with the rapidly-exploring random tree method, these values are reduced by 20.1% and 66.9% respectively, and compared with the dynamic motion element algorithm, they are reduced by 28.5% and 20.8% respectively. This study achieved smooth obstacle avoidance while preserving the shape and features of the reconstructed trajectory. During the obstacle avoidance process, both trajectory similarity and smoothness were taken into account, enabling robots controlled by demonstration learning techniques to be more flexibly applied in complex industrial scenarios.

Abstract:As a vital component of coastal zones, the effective analysis of tidal flat areas holds significant implications for economic development and resource utilization. However, due to the influence of surface heterogeneity and meteorological sensitivity in mudflat areas, it is difficult to characterize the true spectral features of mudflat objects. Therefore, there is an urgent need to develop efficient, flexible, and accurate methods for acquiring and predicting spectral radiation characteristics. Based on this, this paper proposed a multi-parameter BP network-based method for radiation characteristics predicting of tidal flat features. Firstly, aiming at the detection requirements of the tidal flat areas, this paper constructed a set of unmanned aerial vehicle (UAV)-borne multi-spectral data acquisition system, in which the spectral channels include 450, 555, 660, 720, 750, and 840 nm, and the 7-element meteorological instrument that records data such as light intensity, temperature and humidity. Secondly, we proposed a spectral data preprocessing method based on standard gray boards, which obtains standardized spectral data through multi-spectral meteorological data, time stamp alignment, and standard gray board correction. Finally, a multi-parameter BP neural network based characteristic prediction method is designed, enabling the prediction of different tidal flat features under the constraints of meteorological conditions such as light intensity, temperature, and humidity. Based on the UAV multi-spectral remote sensing system, this paper collected three types of spectral radiation data including beach forests, beach water edges and sand gravel. The results of spectral data prediction and clustering show that the proposed algorithm can effectively fit the variation law of different spectral radiation values. The minimum MAE, MSE, and RMSE for the prediction of coastal inter-tidal flats reach 0.214 9, 0.184 3, and 0.429 3, respectively, providing reliable data support for remote sensing monitoring and radiation characteristic prediction of tidal flat features.

Abstract:Point cloud registration is a key technology for light detection and ranging (LiDAR) positioning. To address the challenges of the traditional iterative closest point (ICP) algorithm in point cloud registration in scenarios with sparse geometric features and degenerated environments, this paper proposes an optimization method for LiDAR point cloud registration considering the reliability of geometric features. This method consists of two modules: 1) A curvature-based feature extraction module that constructs a local neighborhood covariance matrix and decomposes its eigenvalues to calculate the maximum and minimum principal curvatures of the point cloud, while introducing the principal curvature difference and normal vector angle to construct a curvature consistency error function. By quantifying geometric feature differences, it effectively filters out ambiguous candidate matching points in repetitive structures, resolves matching ambiguities under noise interference, ensures the geometric consistency of feature registration, and provides high-quality feature pairs for subsequent registration; 2) A feature reliability-based point cloud registration module that uses three factors—feature fitting error, local curvature, and spectral entropy—to quantify feature fitting quality, forms a unified reliability weight through a Bayesian model, and constructs a weighted point cloud registration optimization framework based on this weight. The weight adjustment mechanism suppresses the interference of low-quality features, thereby reducing the errors introduced into the registration results. Experimental validation on both self-built and public KITTI datasets shows that the proposed method reduces 3D positioning errors by 47% compared to conventional point-line (PL)/point-plane (PP) ICP methods in indoor degraded scenes, significantly improving positioning accuracy and robustness in challenging environments. Ablation experiments further confirm the importance of multi-feature collaborative optimization, and the algorithm maintains optimization times within 20 ms for ten thousand-level feature pairs, meeting real-time requirements.

Abstract:In response to the demand for precision linear displacement measurement in fields such as precision manufacturing, aerospace industry, and military applications, an absolute linear time-grating displacement sensor based on secondary coupling is proposed. The fixed scale employs a discretely distributed passive design, which avoids the distributed inductance and capacitance caused by interlayer connection through-holes in traditional designs, thereby improving the continuity and uniformity of the induced magnetic field. Both excitation and sensing signal processing are concentrated on the slider side, requiring only the arrangement of leads on the slider side, with magnetic coupling confined to the excitation coil area. This design effectively suppresses interference from electromagnetic coupling between uncovered coils on the stator and ambient electromagnetic waves, while expanding the sensor′s application scope. First, a planar transient magnetic field coupling model is established, and a dual-array absolute sensor measurement model with its sensing mechanism is constructed. By using a coprime pole-number absolute displacement calculation scheme, the allowable error range in absolute displacement measurement is expanded, and the accuracy of absolute displacement calculation is improved. Additionally, a novel same-frequency modulation signal decoupling method is proposed, which achieves a high signal-to-noise ratio while significantly reducing ADC sampling rate requirements by optimizing the signal processing mechanism, thus resolving the contradiction between ADC sampling rate and resolution. Through electromagnetic finite element simulation, theoretical verification and error analysis of the sensor are conducted, determining the optimal installation gap to be 0.5 mm. Finally, sensor prototypes are fabricated using PCB technology, and experimental studies are performed. The experimental results show that the proposed sensor can achieve absolute displacement measurement within a range of 203 mm, with original measurement errors ranging from -12.62 to +3.23 μm.

Abstract:Magnetic labeling of biomolecules is a crucial technique in biomolecular analysis and manipulation. However, existing detection systems for magnetically labeled molecules often suffer from limited detection targets and restricted application scenarios. Magnetically labeled molecules can be driven by external magnetic fields, and their motions can be monitored using a thickness-shear quartz oscillator sensor. These molecular motions are influenced by both the intrinsic properties of the molecules and the characteristics of the carrier fluid. By analyzing the corresponding motion signals, the properties of the molecules or the carrier medium can be effectively characterized. In this work, a magneto-acoustic integrated multi-parameter detection system for magnetically labeled biomolecules is developed, providing enhanced flexibility for diverse detection applications. The system employs direct digital synthesis (DDS) signal generators as excitation sources for both the magnetic field and the sensor oscillation. Coordinated control between the DDS modules and computer software allows automatic tuning of excitation frequencies to achieve the optimal signal-to-noise ratio (SNR). The sensor output signals are demodulated to extract molecular motion components, which are then processed and analyzed to derive multiple signal features for multi-parameter molecular characterization. Experimental results demonstrate that the system successfully extracts modulated molecular motion signals from the sensor output, and the frequency optimization algorithm effectively enhances the SNR of the sensor signals. Consequently, the system enables quantitative detection of molecular concentration at the ng/mL level and hydrodynamic size at the nanometer scale. The developed instrument serves as an open and extensible platform for multi-parameter detection and biochemical process monitoring of magnetically labeled biomolecules, representing a novel and versatile technology for advanced biomolecular sensing.

Abstract:Accurate calibration of the installation angles of vehicle-borne inertial navigation systems (INS) is essential for ensuring the precision and robustness of GNSS/INS integrated navigation systems. To address the problem of unknown installation angles, this study proposes an online automatic calibration method for arbitrary installation angles in vehicle-borne GNSS/INS systems. The method adaptively executes static and dynamic alignment based on the detection of stationary, moving, and high-speed straight-driving vehicle states, completing the calibration in three stages. During the stationary phase, a coarse static alignment is performed to compute the horizontal installation angles and estimate the accelerometer and gyroscope biases through static filtering. After the horizontal alignment, dynamic alignment is conducted when vehicle motion is detected, calculating the heading angle and separating it through dynamic filtering. Finally, during high-speed straight driving, the precise online calibration of installation angles is achieved using the relationship between the velocity and body coordinate frames. To verify the effectiveness and generality of the proposed method, two sets of vehicle experiments were conducted using tactical-grade and MEMS-grade IMUs, and the installation angle estimation accuracy was compared with the conventional engineering reference method. Experimental results show that the proposed method achieves rapid convergence under different IMU accuracy conditions, with mean installation angle errors of 0.389° and 0.287°, significantly outperforming the reference methods. Moreover, the method is applicable to devices such as vehicle-mounted smartphones, enabling recalibration when the IMU position changes, thereby enhancing system adaptability and robustness. This approach provides critical technical support for high-precision autonomous positioning in intelligent vehicles.

Abstract:Traditional quasi-type-I phase-locked loops (QT1-PLLs) in sensorless control systems for permanent magnet synchronous motors (PMSMs) suffer from decreased rotor position estimation accuracy as operating conditions change. They also struggle to balance dynamic response speed and steady-state accuracy. To address these challenges, we propose a parameter-adaptive phase-locked loop (PLL) method based on an improved QT1-PLL structure. Building on the conventional QT1-PLL topology, we introduce a frequency-adaptive hybrid filtering mechanism by cascading an adaptive notch filter with a moving average filter. This combination effectively suppresses estimation errors and noise disturbances caused by frequency fluctuations while maintaining high filtering efficiency under complex operating conditions. By coupling this hybrid filter with the QT1-PLL, we create the adaptive hybrid-filtering QT1-PLL (AHF-QT1-PLL), which coordinates back-EMF filtering and harmonic suppression. This significantly enhances the system′s robustness and stability, especially under low-speed operation, wide speed variations, and load disturbances. Compared to the traditional QT1-PLL, the proposed method significantly improves dynamic tracking performance during fast speed changes and ensures high rotor position estimation accuracy during steady-state operation, effectively balancing dynamic and steady-state performance. Simulation and experimental results show that the AHF-QT1-PLL outperforms the traditional QT1-PLL across a range of operating conditions, demonstrating higher rotor position estimation accuracy, reduced steady-state errors, faster dynamic adjustment, and stronger disturbance rejection and harmonic suppression capabilities. These results confirm the effectiveness and reliability of the AHF-QT1-PLL under both dynamic and steady-state conditions, offering strong engineering application potential and practical value for optimizing sensorless control strategies in PMSM.

Abstract:This study presents a fault diagnosis and isolation framework for cable wear in cable-driven manipulators, aiming to improve the system reliability and safety during operation. During the fault diagnosis phase, by integrating the mapping relationship between manipulator joint angles and cable tension, a cable tension model is created to accurately predict tension changes under various working conditions. Based on this, a physics-informed fault diagnosis network is developed. By leveraging the dynamic features of tension signals and joint angle data and optimizing with a multi-objective loss function, accurate fault diagnosis of faulty cables is achieved. In the fault isolation phase, this research designs a fiber Bragg grating sensor based on a synchronous compensation method to stably and precisely collect joint movement information and introduces a three-driven cable synchronous compensation strategy. This strategy sequentially performs multiple equidistant synchronous contraction operations on the three driving cables of each joint, thereby reducing the impact of dynamic errors on isolation accuracy. The optical fiber sensor is used to collect the dynamic responses of the manipulator′s pitch and yaw angles, and two-dimensional feature maps are generated using the Gram Angle Difference Field algorithm. An ensemble classification model is then constructed to achieve precise isolation of manipulator cable damage. Experimental results show that the proposed method achieves a fault diagnosis accuracy of 98.48% in the cable fault diagnosis experiments and a fault isolation accuracy of 94.89% in the cable fault isolation experiments. The approach demonstrates significant advantages in cable tension prediction, fault diagnosis, and localization, providing robust support for the safe operation of cable-driven manipulators.

Abstract:As a novel traction and load-bearing component derived from steel wire ropes, the elevator belt has been widely applied in modern elevator systems due to its flexibility, corrosion resistance, and high transmission efficiency. However, long-term cyclic stresses can induce hidden defects such as cord wear, broken wires, and strand fractures, which are difficult to detect in time and pose serious safety risks. Existing methods often suffer from slow response, low efficiency, or dependence on shutdown inspection, making them unsuitable for online detection of broken wires. To overcome these limitations, this study proposes an eddy current-based detection method for broken wires in elevator belt steel cords. A segmented mutual-inductance probe with a spatially wound, openable configuration was designed, and finite element simulations were conducted to optimize coil turns and lift-off parameters. Based on the optimized design, a prototype was fabricated and a test platform established. Sensitivity was first evaluated by introducing different numbers of broken wires into a single steel strand, followed by detection experiments on elevator belts. Results show that the probe can stably detect at least four broken wires at a belt speed of 0.5 m/s. The induced signal amplitude increases with the number of broken wires, and the system exhibits good dynamic detection capability. These findings provide a practical reference for online detection of broken wires in elevator belt steel cords and offer technical support for enhancing elevator safety and enabling preventive maintenance.

Abstract:This paper proposes an efficient and reliable modeling method to address the longstanding reliance on empirical expertise and the lack of systematic theoretical guidance in the design of pipeline permanent magnet in-line inspection tools. This method accurately characterizes the localized saturation magnetic field and the coupling mechanism with pipeline wall defects, enabling a quantitative analysis of the inspection performance. On the basis of the traditional magnetic circuit method, the theory of magnetic field segmentation is introduced, and an improved magnetic network model suitable for internal detection of pipeline leakage is constructed. The model fully considers the edge effect and the nonlinear magnetization characteristics of ferromagnetic materials, combining the key parameters of the detection device with the magnetic leakage field caused by defects. The internal relationship between the two is revealed, and then the quantitative calculation of the dynamic evolution process and spatial distribution of the local saturated magnetic field is realized. The improved magnetic network method proposed in this study shows good versatility and model adaptability while ensuring high computational efficiency. In order to verify its effectiveness, finite element numerical simulation and experiments were carried out respectively, and the calculation results of the model were compared and analyzed. The results show that the improved magnetic network model aligns well with the finite element simulation results in terms of magnetic field distribution characteristics and key signal response, with an error controlled within 5%. Compared to experimental measurement results, the overall relative error is less than 30%, demonstrating acceptable accuracy for engineering applications. The model effectively captures the magnetic field disturbances caused by inner wall defect, accurately reflecting their influence on the output performance parameters of the detection device, thereby compensating for the shortcomings of traditional magnetic circuit design theory. The proposed method exhibits both high computational accuracy and efficiency, making it suitable for the rapid performance evaluation and iterative optimization design of in-line inspection devices in engineering practice. Moreover, it provides a reliable theoretical basis and model support for the parameter design, performance prediction, and engineering applications of such equipment.

Abstract:To address issues in gearbox fault diagnosis of doubly-fed wind turbines, such as the lack of coverage of doubly-fed grid-connected conditions in existing stator current methods and the difficulty in comparing frequency band aliasing under variable operating conditions, this article proposes a gearbox condition monitoring method based on stator current signals from doubly-fed wind turbines. Firstly, a phenomenological model is formulated based on the stator current signal of the generator of the doubly-fed wind turbine physical simulation platform, and the composition and law of the stator current signal were determined. Then, the stator current signal phenomenological model is used to link the mechanical part of the fan with the electrical characteristics and determine the influence of the mechanical side of the fan on the electrical side. The current-energy ratio algorithm was proposed. For the problems of frequency band overlap and inability to compare caused by variable working conditions, the method of segmentation and reorganization is used to avoid it. Rotor speed estimation is achieved by leveraging motor cogging harmonics, enabling the calculation of the current energy ratio using only stator current signals, with a mean relative error of 0.277 2 and a mean squared error of 0.114 6. To validate the feasibility of the method, stator current signals under various operating conditions are collected from a physical doubly-fed wind turbine simulation platform for both normal and multiple fault states of the gearbox. The results show that the current energy ratio under gearbox fault conditions is significantly higher than that in the normal state, and the severity of faults in the same component is positively correlated with the current energy ratio—specifically, more severe faults correspond to higher current energy ratios. Comparative analysis of the current energy ratio between the parallel stage and the secondary planetary stage further shows that the algorithm maintains strong universality even when wind turbine parameters change. It can effectively monitor faults in both the parallel and secondary planetary stages of the gearbox.

Abstract:Wind turbines are complex electromechanical hybrid systems that convert wind energy into electrical power. Their transmission systems include numerous high-precision mechanical components such as gears and bearings. These components are prone to fatigue and wear during prolonged operation, making the transmission system vulnerable to various failures, which significantly impacts both the safety of the units and their power generation efficiency. To ensure the safety of wind turbines, it is necessary to efficiently and accurately detect abnormal states in the transmission system. This paper addresses the limitations of existing research in capturing complex operational states and the low accuracy in component maintenance and white noise rejection. A state monitoring model based on a learnable decomposition and dual-attention network using autoregressive embeddings is proposed. First, the data collected by the supervisory control and data acquisition (SCADA) system is preprocessed, and correlation analysis is performed to select feature parameters such as temperature that are strongly correlated with fault evolution. Next, an autoregressive embedding module is introduced, utilizing dynamic tokens to better capture multidimensional time-series features. This allows for the prediction of relevant temperature variables in the transmission system, achieving dynamic modelling of fault characteristics. Then, a dual-threshold discrimination monitoring network is proposed, combining the residuals and information entropy of relevant variables to determine the dual health thresholds, further eliminates abnormal signals such as white noise and providing accurate warning times. Finally, the effectiveness of the proposed model is verified through two actual wind turbine transmission system failure cases. Compared to traditional SCADA systems, early warning signals can be detected approximately 6 to 10 days in advance when a fault occurs in the wind turbine.