46 results about "Higher-order statistics" patented technology

The invention relates to the technical field of data processing, and particularly discloses a data analysis method and device fusing a large number rule, equipment and a medium, and the method comprises the steps: determining a minimum convergence amount through hierarchical aggregation; calculating a cumulative sample mean trajectory and carrying out convergence diagnosis; performing noise attenuation weighting based on LLN convergence characteristics; performing distribution and component decomposition under steady moment constraint; performing consistency check and re-extraction robustness of LLN guidance; lLN-constrained model fitting and uncertainty calibration output are carried out; according to the method, based on noise attenuation weighting of LLN convergence characteristics, samples which are close to a steady state obtain greater influence in estimation, and unstable or sparse samples are naturally weakened, so that self-adaptive suppression of heterogeneity and small sample noise is realized, and the robustness of overall estimation is improved; steady moment constraint is introduced into distribution / component decomposition, it can be guaranteed that components obtained through decomposition are consistent with observation convergence characteristics in the aspect of high-order statistics, and component mismatching caused by extreme values or local fluctuation is reduced.

The invention provides an AI text recognition method and device based on ensemble learning and advanced semantic statistical feature analysis, and the method comprises the steps: 1, respectively sending a to-be-recognized text into a Bert detector and a high-order natural language statistical feature detector for recognition, the high-order natural language statistical feature detector comprises a word logarithm probability detector, a word ranking logarithm detector, an Entropy detector and a confusion degree detector; and 2, performing election on detection results output by the Bert detector and the high-order natural language statistical feature detector by using an election module to obtain an AI text recognition result. According to the method, an integrated learning strategy is adopted, and a pre-training language model subjected to fine tuning is combined with high-order natural language statistical characteristics, so that when the model detects a large language model to generate a text, the strong expression ability of the pre-training language model can be fully utilized, and a deep rule of the text can be captured through the high-order statistical characteristics; and the detection accuracy is improved.

The invention discloses a radar anti-active slice forwarding interference method based on high-order statistics, and the method comprises the steps: S10, carrying out the preprocessing of a radar receiving signal, carrying out the high-order cumulant construction of the radar receiving signal, completing the signal preprocessing, the calculation of a fourth-order cumulant, and the construction of a feature tensor, the high-order cumulant feature tensor of each time window corresponds to the corresponding time domain signal according to the time dimension and serves as a source of subsequent training input; s20, generating a label mask required by supervision training, and constructing an input feature and a supervision label so as to form a data structure in which the input and the label are paired; s30, constructing a MoE deep neural network based on high-order cumulant characteristics; s40, constructing a loss function and an optimization target, and calculating results of all loss items; s50, constructing an end-to-end blind interference suppression framework; s60, performing model training and parameter optimization; and S70, completing interference suppression and target output generation.

The invention relates to the technical field of laser cutting, and discloses a steel plate laser cutting machine fault detection method and system, and the method comprises the steps: collecting a driving power supply current signal in real time, and filtering a DC component to obtain an AC sequence; separating multi-scale sub-band components containing Gaussian noise by using wavelet transform; performing empirical mode decomposition on the sub-band components, and extracting an intrinsic mode function component set; calculating a kurtosis value and a kurtosis change rate of the component and a phase coupling index based on bispectrum analysis, and constructing a high-order statistical feature vector; a support vector machine is adopted to classify the feature vectors so as to output operation state category labels; generating a fault early warning grade value by combining the tag risk value and the kurtosis evolution trend slope; and the historical state sliding window sequence and the fault confirmation confidence coefficient are used for verification, and the early fault state is confirmed. According to the method, the problem of low nonlinear fault feature extraction accuracy under a complex noise interference background in the prior art can be solved.

The invention discloses an intelligent monitoring method for slope deformation, particularly relates to the technical field of optical vision measurement, and is used for solving the problem that a deformation signal and optical noise are difficult to effectively separate due to the influence of atmospheric disturbance in an existing vision monitoring method. Collecting image data and point cloud data of the surface of the side slope through vision measurement equipment; analyzing high-order statistical characteristics of the data to identify abnormal components; analyzing cluster behaviors of abnormal components under different spatial scales to distinguish a random cluster from a structural cluster; marking according to cluster types or performing correlation analysis with atmospheric environment parameters; evaluating the comprehensive influence of the abnormal component on the deformation signal based on the analysis result; and finally, according to an evaluation result, adopting adaptive filtering to separate abnormal components and deformation signals, and outputting a monitoring result. Noise and real deformation signals caused by atmospheric disturbance can be effectively recognized and separated, and the accuracy and reliability of slope deformation monitoring are remarkably improved.

The invention discloses a video behavior identification method based on a multi-mode high-order statistical adapter, which comprises the following steps: acquiring video data and text description of a tag corresponding to the video data, and carrying out fragment sampling and frame extraction processing on the video data and storing; a comparison language image pre-training model is used as a basic network, a network architecture is expanded on the premise of freezing original visual branch and text branch parameters of the basic network, a high-order statistical adapter comprising a space-time enhancement module and a high-order statistical modeling module is inserted into a visual branch, a parallel text adapter is inserted into a text branch, and the visual branch and the text branch are integrated. Constructing a video multi-mode high-order adapter network, wherein the network realizes video behavior classification by calculating cosine similarity of video features and text features; and performing iterative training on the network by using the text and video data in the training set so as to update parameters, storing the optimal weight of the network according to the accuracy of the optimal verification set, and evaluating the performance of the network on the test set by adopting a multi-view test strategy.

The invention discloses a fault diagnosis method based on integrated empirical mode decomposition and a manifold structure, and aims to research an algorithm model capable of realizing effective fault diagnosis for an early fault with weak characteristics. The main core of the method is to integrate eigenmode function components obtained by empirical mode decomposition, judge the sensitivity of the eigenmode function components to early faults so as to provide a variable reconstruction strategy more sensitive to the early faults, and meanwhile, extract local features and manifold structures by using a neighborhood preserving embedding algorithm so as to improve the robustness of the early faults. And high-order statistical features more sensitive to early faults are constructed in combination with a statistical local algorithm, so that the high-order statistical features are input into a Bayesian classifier, and finally early fault diagnosis is realized. Compared with a traditional method, the method can more effectively distinguish different types of early faults, obtains higher accuracy, and is a more excellent early fault diagnosis method.

The invention relates to the technical field of power system fault detection and positioning, and provides a power grid anomaly detection method based on a free probability theory. Comprising the steps of collecting power grid node power and environmental parameters to construct a dynamic high-order tensor model, and introducing non-exchangeable random variables to describe space-time correlation characteristics; mapping data through R transformation and S transformation, constructing a high-order statistic characteristic matrix, and introducing free entropy to detect anomaly; constructing a spectrum distribution model by using a ring law, and identifying an abnormal signal by combining a free Poisson operator; analyzing abnormal components by adopting a free product decomposition technology and wavelet packet transformation; and constructing a dynamic confidence domain based on the free Fisher information amount, and realizing power grid topology elastic reconstruction in combination with a graph attention mechanism. The power grid anomaly detection method based on the free probability theory can accurately detect anomalies, deeply analyze signals, intelligently make decisions and optimize power grid topology, comprehensively improve power grid anomaly detection and processing capability, and guarantee reliable operation of a power system.

The invention provides an intelligent electric meter calibration method and system, and relates to the technical field of electric meter calibration, and the method comprises the steps: collecting electric energy data; establishing a power sequence of the electric energy data; extracting a high-order statistical moment of the power sequence; based on the high-order statistical moment, an index correction term of excess kurtosis is introduced, and a dynamic confidence boundary used for triggering a calibration process is constructed through a Chebyshev inequality; under the condition that the continuous exceeding frequency of the dynamic confidence boundary is greater than a preset continuous exceeding frequency, iteratively solving the optimal calibration offset based on the dynamic confidence boundary; and carrying out electric energy value calibration by using the optimal calibration offset, otherwise, returning to collect the electric energy data again. The dynamic evolution rule of the calibration offset is accurately captured based on the iterative optimization algorithm of the high-order statistical characteristics, error accumulation and systematic deviation are avoided, the calibration precision and stability are remarkably improved, and high precision and real-time performance can still be kept especially under the complex working conditions of harmonic interference, nonlinear loads and the like.

The invention discloses an intelligent equipment remote maintenance autonomous monitoring method, and relates to the technical field of intelligent equipment operation and maintenance, and the method comprises the following steps: building a unified time baseline, deploying a zero delay bypass sampling mechanism, extracting an unsmooth energy transition track from continuously collected multi-source signals, and generating an event sequence of full time domain resolution; a transition confidence window set is constructed based on an event sequence, a pulse signal is analyzed by using a high-order cumulant, and key signal fragments are screened in combination with a sparse peak value extraction method. According to the method, a closed-loop regulation and control system driven by a unified time baseline is constructed, event sequence capture, high-order statistical analysis and time reversal intervention are fused, real-time recognition and active hedging of equipment transition energy are achieved, the fault prediction and self-healing capacity is improved, the service life of the structure is prolonged, and the operation and maintenance cost is reduced.

The invention discloses a composite image data set distillation method and system based on curvature guidance, and belongs to the technical field of image processing. In order to solve the technical problems of large high-order statistical estimation variance, uncontrollable calculation overhead and insufficient synthetic sample robustness under a sub-sampling condition in the prior art, the method comprises the following steps: preprocessing a real sample and extracting features and kernel functions of the real sample and a to-be-optimized synthetic sample; calculating the sampling weight of a real sample by using a sampler; calculating the first-order loss of the distribution pairing based on the weight and the feature, and extracting the Hessian second-order information to obtain a curvature loss item according to the first-order loss; synchronously updating the synthetic sample and sampler parameters by using first-order loss and curvature loss items to generate a mother set; and finally, cutting and correcting the mother set. According to the method, high-order statistic estimation can be stabilized, gradient variance can be reduced, the structural fidelity and generalization ability of a synthetic sample are improved, and one-time distillation is adapted to high-quality data sets of various scales.

The application discloses a power supply equipment state monitoring system and method based on AI power spectrum, and relates to the technical field of electrical intelligent sensing; the method comprises the following steps: obtaining original analog signals of the power supply equipment through real-time synchronous acquisition, and preprocessing the original analog signals to obtain digital waveform data frames; performing time domain feature calculation on the digital waveform data frames to obtain a time domain feature set; performing frequency spectrum feature extraction on the digital waveform data frames to obtain a frequency domain feature set; performing high-order statistical analysis and nonlinear signal processing on the digital waveform data frames to obtain a nonlinear feature vector; based on the frequency domain feature set, the nonlinear feature vector and the time domain feature set, combining a pre-trained artificial intelligence model, outputting a fault type identifier and a corresponding fault probability, and calculating an abnormal score based on the fault type identifier and the corresponding fault probability; the application can effectively identify the camouflage fault of the power supply equipment and reduce the cumulative risk of potential faults.

The application discloses a power transformer winding fault detection method based on vibration spectrum analysis and relates to the technical field of power equipment detection, which comprises the following steps: collecting original vibration acceleration signals of a transformer and preprocessing the original vibration acceleration signals to form a signal frame sequence; applying an improved Hilbert-Huang transform to each signal frame to obtain a set of intrinsic mode functions, calculating a component instantaneous frequency sequence and extracting a frequency band with an energy proportion exceeding a preset threshold as a candidate fault characteristic frequency band; extracting a subband signal of each frequency band and calculating a high-order statistical quantity feature matrix; loading a health state reference feature matrix matched with a transformer model, capacity and voltage grade; comparing the two in an element-by-element manner to identify an abnormal feature submatrix; spatially clustering the abnormal feature submatrix to form a suspected fault mode cluster; and outputting a specific winding fault mode diagnosis result. The method can accurately extract fault characteristics and improve the accuracy and reliability of fault detection.

The present application belongs to the technical field of target detection, and particularly relates to a multi-direction prediction and high-order statistics target detection method for airborne air-to-air imaging. The method comprises the following steps: S1: obtaining an original infrared image dataset, performing linear fitting filtering on the qth original infrared image to obtain a residual image; S2: selecting a candidate point in the residual image corresponding to the qth original infrared image based on a target threshold; S3: calculating the statistical fourth moment and local gradient level of each candidate point under a multi-scale window, and screening a small target based on the calculation result; S4: replacing the qth original infrared image with the q+1th original infrared image, repeating steps S1-S3, and obtaining the small targets in each original infrared image contained in the original infrared image dataset. The present application does not require target window priori and does not depend on a specific distribution model, can stably operate in a complex and changeable background, and has high engineering practicability and universality.

The invention discloses a bistatic multi-channel synthetic aperture radar (SAR) synchronous signal extraction method and device, and belongs to the technical field of synthetic aperture radar signal processing. According to the method, a bistatic SAR system with a plurality of receiving channels is constructed, and a synchronization signal and a scene echo signal are modeled into a linear mixed signal by using time delay and amplitude differences caused by physical position differences among the channels; a blind source separation algorithm based on high-order statistics is adopted to carry out centralization, whitening and joint diagonalization processing on multichannel received signals, and effective separation of synchronous signals and target echoes is realized. The method does not need to interrupt the imaging process, does not depend on priori knowledge or complex waveform design, remarkably improves the phase synchronization precision and the system flexibility, is suitable for various bistatic SAR systems, and has good practicability and universality.

The invention discloses a coherent mode extraction method based on principal skewness analysis, and relates to the technical field of magnetic confinement nuclear fusion plasma diagnosis, and the method specifically comprises the following steps: S1, calculating Doppler translation according to a Doppler reflectometer system signal; s2, performing short-time Fourier transform on the obtained Doppler frequency shift to obtain a power spectrum matrix of the Doppler frequency shift; s3, constructing a co-skewness tensor according to a row vector of the power spectrum matrix, and solving a feature vector of the co-skewness tensor; and S4, performing noise suppression on the power spectrum matrix by using the feature vector of the co-skewness tensor. According to the method, the non-Gaussian features in Doppler frequency shift are extracted through high-order statistics, the limitation of a traditional method in noise suppression and transient mode detection is solved, the signal-to-noise ratio of coherent mode extraction is improved, and therefore the coherent mode can be displayed from background noise more effectively.

This invention discloses a curvature-guided distillation method and system for synthetic image datasets, belonging to the field of image processing technology. To address the technical problems of existing technologies, such as large variance in higher-order statistical estimation under subsampling conditions, uncontrollable computational overhead, and insufficient robustness of synthetic samples, this invention preprocesses real samples and extracts their features and kernel functions along with those of the synthetic samples to be optimized; calculates the sampling weights of the real samples using a sampler; calculates the first-order loss based on the weights and features, and extracts Hessian second-order information to obtain the curvature loss term; synchronously updates the parameters of the synthetic samples and sampler using the first-order loss and the curvature loss term to generate a parent set; finally, the parent set is cropped and corrected. This invention can stabilize higher-order statistical estimation and reduce gradient variance, improve the structural fidelity and generalization ability of synthetic samples, and achieve single distillation adaptability to high-quality datasets of various sizes.

The application provides a thermodynamic parameter prediction method and system of a solar seawater desalination system, and belongs to the technical field of solar energy utilization and seawater desalination. Firstly, the application synchronously acquires complex impedance spectrum data and inlet and outlet temperature difference sequence data of seawater in a photovoltaic back plate cooling flow channel, extracts resistance and reactance components of the complex impedance, and constructs a multi-dimensional signal matrix with the temperature difference sequence. Subsequently, based on a high-order statistical quantity maximization principle, independent component analysis is performed on the matrix, a pure heat effect component is decoupled and extracted. Then, the component is subjected to multi-level wavelet packet decomposition, energy distribution of each frequency band is calculated, and an energy entropy vector is generated. Finally, the energy entropy vector is input into a trained BP neural network model, and prediction values of instantaneous fresh water production and heat utilization efficiency are mapped and output. The application realizes high-precision prediction of water production and heat efficiency of a photovoltaic photo-thermal system under complex electrochemical interference.

The present application relates to the technical field of identity authentication, and proposes a method for non-intrusive identity authentication of unmanned aerial vehicles, which comprises the following steps: passively externally collecting magnetic signals of the working circuit of the unmanned aerial vehicle to obtain original electromagnetic signals; performing multi-domain feature extraction on the original electromagnetic signals to obtain time-domain pulse features, frequency-domain harmonic features and high-order statistical features; performing weighted fusion and redundancy reduction on the time-domain pulse features, frequency-domain harmonic features and high-order statistical features to obtain a deep fusion feature vector; performing chaotic hash compression and fixed-length encoding on the deep fusion feature vector to obtain a unique hardware fingerprint code; binding the unique hardware fingerprint code with a product serial number of the unmanned aerial vehicle to obtain a unique identity binding code; and obtaining a digital identity certificate of the unique identity binding code by applying for a public trust certificate. The present application can improve the efficiency of non-intrusive identity authentication of unmanned aerial vehicles.

This invention relates to the field of industrial energy management technology and discloses an AI-driven large-scale model-based power consumption tracking and energy-saving control system. The method includes: acquiring multi-cycle power consumption sampling data from industrial equipment and performing phase normalization processing; extracting high-order statistical features at the same phase points to generate cross-cycle energy consumption feature vectors; calculating trend slope and phase synchronization to filter abnormal phase points; encoding abnormal information and inputting it into a large-scale model to generate semantic embedding vectors; using an attention mechanism to calculate semantic distance distribution and generate energy consumption anomaly feature labels; and generating cause diagnosis and energy-saving control instructions through large-scale model-driven inference. This invention achieves an automated closed loop from energy consumption tracking to energy-saving control.

The invention relates to the technical field of metal component nondestructive testing, and discloses a metal component magnetic coercive force damage assessment system and method.The metal component magnetic coercive force damage assessment method comprises the steps that repeated standardized magnetization cycle collection and synchronous average processing are carried out on all measuring points, and high-signal-to-noise-ratio average hysteresis loop data are generated; extracting high-order statistical characteristics of a hysteresis loop local interval, and generating a local skewness value sequence, a local kurtosis value sequence and an inter-cycle micro variation value sequence; a multi-channel high-order statistical feature space field is generated through Kriging interpolation; performing time sequence difference and gradient vector field calculation on the multi-moment multi-channel high-order statistical feature space field; performing DBSCAN density clustering on the cross-channel fusion gradient vector, identifying an early damage active evolution leading edge and extracting leading edge features; and outputting a magnetic coercive force damage evaluation report of the metal component.

The invention provides an electric power system source network load integrated high-order uncertainty cooperative control method and system based on distributed robust optimization, and relates to the technical field of electric power system optimization control. The method comprises the following steps: forming a multi-layer collaborative risk allocation structure by constructing a hybrid uncertainty envelope with high-order statistical constraints and combining cross-layer risk budget and sensitive mapping; based on the real-time operation data and the distribution sudden change identification indexes, self-calibration and safety order reduction control are implemented, and a recoverable control strategy is obtained; and dynamic updating of uncertainty envelope and risk budget is realized through rolling feedback and incremental learning, so that a distributed robust scheduling system with self-repairing and continuous robustness is constructed.

The application discloses a kind of portable three-phase power quality field diagnosis method, the application is first with 0.1~1Hz low-speed acquisition voltage, current signal of the three-phase alternating current loop to be measured, synchronous acquisition ambient temperature and complete phase shift compensation;Current waveform high-order statistical feature is extracted again to realize load type identification, corresponding power quality evaluation threshold is matched and core index is calculated, index is judged as power quality disturbance when threshold, generate local diagnosis prompt and trigger 3.2~12.8kHz high-speed sampling.High-speed sampling stage real-time detects communication link state, link is normal then real-time upload data, interrupt then data is cached to local storage, link recovers after breakpoint and completes supplement transmission;Diagnosis whole process carries out FIFO management to local storage capacity, and high-speed sampling data is preferentially retained.The application realizes local intelligent diagnosis and adaptive sampling, solves the problem that traditional equipment is poor in portability, has no local diagnosis capability, data is easy to lose, sampling strategy is rigid.

The invention discloses an intelligent manufacturing-oriented real-time data analysis method and system based on edge calculation. The method comprises the following steps: collecting original time sequence data at a basic sampling frequency on an edge computing node; inputting the data into a dynamic data governance kernel, extracting high-order statistical features and lightweight convolutional neural network features in parallel by the kernel based on a sliding window, and generating a structured feature vector sequence through adaptive weighted fusion of a dynamic gating fusion unit; inputting the sequence into a time sequence prediction and anomaly detection joint model based on a bidirectional long-short-term memory network and an attention mechanism, and generating a real-time analysis result containing state prediction and classification; and the result and the key summary information are compressed and uploaded to a cloud server through an MQTT protocol in an event-driven mode. According to the method, deep coupling of data governance and analysis decision is realized, and while low delay and low bandwidth occupation are ensured, the accuracy of real-time analysis and the side cloud cooperation efficiency of the system are remarkably improved.

This invention discloses a method for enhancing motor fault data based on a spectral normalization-conditional generative adversarial network (GAN) and high-order statistics for screening, belonging to the field of motor fault diagnosis and data augmentation. The method includes: first, constructing a filter incorporating three aspects: energy efficiency, physical stability, and fault saliency, and a generative adversarial network (GAN) incorporating spectral normalization; second, generating initial fault signal samples using the generator and filtering them through the filter; inputting the filtered fault signal samples and collected real fault signal data into a discriminator for discrimination; and updating the GAN parameters based on the discriminator's discrimination results to obtain a trained generator; finally, using the filter and the trained generator, specifying the fault type, yields new fault signal samples corresponding to that fault type. This invention effectively enhances the diversity of data samples, providing a reference and approach for the practical deployment and optimization of deep learning models in industry.

This invention relates to the field of on-orbit telemetry data analysis, and particularly to a method for in-depth analysis of on-orbit data fault diagnosis results based on high-order statistics. First, historical fault diagnosis conclusion data is segmented by a preset time length, and the types, durations, and switching processes of the diagnosis conclusions are statistically analyzed to form high-order statistical information. Second, different conclusion types are labeled with color bands of different lengths according to their duration, and these bands are then stitched together according to the switching process of the diagnosis conclusion types to form a changing color band diagram. Next, based on manually labeled fault tags, a diagnostic model is trained using a deep learning algorithm. Finally, the statistical analysis and color band diagram drawing process are repeated for new data, and the trained diagnostic model outputs abnormal diagnosis conclusions. This invention can effectively filter out false alarms, uncover deep fault patterns, significantly reduce the burden of manual analysis, and achieve keen early warning of early fault symptoms, thus improving the automation level and reliability of on-orbit data fault diagnosis.

The application relates to the technical field of laser cutting and discloses a steel plate laser cutting machine fault detection method and system, the method comprising the following steps: collecting a driving power current signal in real time and filtering out a direct current component to obtain an alternating current sequence; separating a multi-scale subband component containing Gaussian noise by using wavelet transform; performing empirical mode decomposition on the subband component to extract an intrinsic mode function component set; calculating the kurtosis value, kurtosis change rate and phase coupling index based on bispectrum analysis of the component to construct a high-order statistical feature vector; classifying the feature vector by using a support vector machine to output an operation state category label; combining the label risk value and the kurtosis evolution trend slope to generate a fault early warning grade value; and verifying early fault states by using a historical state sliding window sequence and a fault confirmation confidence degree, so that the method can solve the problem of low precision of nonlinear fault feature extraction in a complex noise interference background existing in the prior art.

This invention discloses a native digital human that continuously grows alongside a real individual and its construction method, comprising: establishing a mapping relationship between a target activity task and multiple predetermined human functional indicators; collecting measurement values ​​and actual performance values ​​of the target individual over multiple time periods to form an individual functional dataset; extracting high-order statistical values ​​from a homogeneous database that are in the top P percentile of a reference group as a top-tier benchmark; generating a first difference between the predicted performance value and the actual performance value, and a second difference between the individual measurement value and the top-tier benchmark; generating intervention target values ​​based on the first and second differences and encoding them into a structured training scheme; continuously injecting new periodic measurement values ​​generated after the real individual executes the training scheme into the dataset and triggering iterative updates, so that the native digital human evolves synchronously with the training results. This invention can construct a native digital human that continuously grows alongside a real individual, enabling lossless experiments in the virtual world and supporting team collaborative simulations.

The invention discloses a data security analysis system applied to wireless communication equipment, and relates to the technical field of data security analysis. Physical layer signal distortion characteristics are extracted through high-order statistics, and a hybrid neural network is combined to capture an attack mode across space-time dimensions; the limitation that high-simulation attacks are difficult to identify through traditional CRC verification and RSS detection is broken through; a frequency offset driven weight distribution and online learning mechanism is adopted, channel quality self-adaptive classification boundary adjustment is achieved, and the false alarm problem of a fixed threshold value in transient scenes such as starting and stopping of heavy equipment is solved; a signal analysis result is directly mapped to a protocol layer defense strategy, closed-loop control from feature recognition to parameter reconstruction is formed, and the technical blank of cross-layer linkage of an existing scheme is filled up; model compression based on knowledge distillation and preprocessing module optimization meet the real-time requirement of industrial-grade embedded equipment while keeping the detection precision.