134 results about "Adaptive learning rate" patented technology

The invention belongs to the technical field of asynchronous federated learning, and discloses a data isomerism-oriented knowledge alignment asynchronous federated learning method. According to the method, a data quality perception aggregation strategy is introduced, and a knowledge distillation mechanism based on the old degree is combined, so that a global model is subjected to balanced training on heterogeneous data of different devices, and the generalization ability of the model is improved. Meanwhile, a self-adaptive learning rate adjustment mechanism based on aggregation frequency and weight is designed, and it is ensured that contribution of different devices to the global model is more fair. According to the method, the training deviation in asynchronous federated learning is effectively relieved, the accuracy and stability of a global model are improved, and the method has a considerable application value for a real federated environment.

The invention provides an electromechanical equipment fault prediction method and system based on multi-source information fusion, and the method comprises the steps: collecting operation data, including vibration data, temperature data and current data, during the operation of electromechanical equipment; performing feature extraction on the operation data based on a principal component analysis algorithm to obtain a fusion feature vector; inputting the fusion feature vector into a fault prediction model based on a deep belief network, and outputting a prediction result; wherein the deep belief network adopts a small-batch stochastic gradient descent algorithm combined with an adaptive learning rate adjustment strategy during training; and judging whether the electromechanical equipment has a fault hidden danger or not according to the prediction result. According to the method, the relevance between different types of data is mined, and the defect of low prediction precision is overcome.

The invention discloses a generative adversarial network architecture search method and system and an image generation method, and belongs to the technical field of network architecture search. The searching method comprises the following steps: performing single-path sampling on a pre-constructed generator super network according to a parameter quantity constraint range to obtain an effective subnetwork; training the generator super-net by adopting a complexity adaptive learning rate optimization strategy to obtain a pre-trained generator super-net; adversarial training is carried out on the generator hypernet and the discriminator to obtain a pre-trained discriminator; generating a generator super-network candidate architecture through a genetic algorithm in the early stage of the evolution stage and through a covariance matrix self-adaptive evolution strategy in the later stage of the evolution stage; and performing multi-target non-dominated sorting on the generator super-network candidate architecture, updating an effective sub-network and keeping a Pareto optimal solution to obtain a searched optimal generator architecture and further obtain a searched optimal generative adversarial network architecture. The method not only ensures the search quality, but also improves the calculation efficiency.

The invention discloses an automatic history fitting method based on multiple data assimilation of an improved set smoother, and relates to the technical field of petroleum engineering. The method comprises the following steps: firstly, constructing a plurality of oil reservoir models, setting a prior model, a real model and an initial expansion factor, obtaining the yield of each oil reservoir model by utilizing oil reservoir model numerical simulation, and calculating a residual error; based on corrected data covariance matrix singular value decomposition, production observation data disturbance enhancement, adaptive learning rate matrix scaling and geological boundary constraint, improving a set smoother, updating the permeability of each oil reservoir model, performing numerical simulation again, updating an expansion factor, judging whether the expansion factor meets a preset condition or not, continuing iteration if the expansion factor meets the preset condition, and if the expansion factor does not meet the preset condition, continuing iteration until the expansion factor meets the preset condition; and if not, updating the expansion factor of the iteration, ending the iteration, and outputting the permeability field of each updated oil reservoir model, thereby solving the problems of parameter overshoot, covariance statistical deviation and low calculation efficiency in oil reservoir history fitting, and facilitating history fitting of complex oil reservoir production parameters.

The invention relates to the technical field of power equipment state operation early warning, and discloses a power equipment state operation time sequence early warning system and method, and the system comprises a data collection module which collects the multi-source operation data of power equipment through infrared, ultrasonic TEV, high-frequency, ultrahigh-frequency and vibration sensors; the data preprocessing module is used for carrying out normalization processing on the collected multi-source operation data by adopting an extreme value method or an averaging method and mapping the data to a [-1, 1] interval; and the time sequence analysis modeling module is used for constructing a Prophet time sequence model for the normalized data, and decomposing a trend item, a season item and a holiday item. A dynamic factor is added in weight updating through a dynamic parameter adjustment method, network training is effectively prevented from falling into a local minimum point, fluctuation in the learning period is reduced, the learning rate is flexibly adjusted according to error changes in combination with an adaptive learning rate algorithm, and slow convergence caused by the too low learning rate is avoided.

The invention discloses a point cloud single tree segmentation method and system. The method comprises the steps of obtaining point cloud data of a target tree and an environment in a power transmission corridor area; constructing a segmentation network model, and performing model training through the optimized loss function and adjustment of the adaptive learning rate to obtain an optimized segmentation network model; inputting the preprocessed point cloud data into the optimized segmentation network model, calculating the category of each point through an activation function, and outputting the category prediction of each point; and point cloud segmentation is carried out based on the category, and a segmentation result is evaluated. According to the invention, point cloud data segmentation is carried out through the segmentation network architecture, the learning rate and the loss function in the architecture are optimized, and the recognition and segmentation precision of the tree monomers is improved in combination with the probability distribution of deep learning; the robustness in a complex environment is enhanced, the problems of noise, sparse point clouds and data imbalance can be effectively solved, and the method is suitable for tree monitoring, accurate positioning and safety evaluation of the transmission line corridor.

The invention relates to the technical field of artificial intelligence and natural language processing, in particular to an online incremental learning AI chat robot response generation system which comprises a cache reweighting module, a semantic offset calculation module, an adjusting module, an updating module, a response generation module and a feedback module. According to the method, cache reweighting based on semantic similarity and time decay, a self-adaptive learning rate based on semantic offset and rollback protection are introduced, low-rank increment updating of LoRA and a weight adjustment closed loop of multi-target weighted decoding and feedback driving are adopted, so that the system can effectively improve the real-time performance of the system while preferentially utilizing aging and related historical contexts, and the real-time performance of the system is improved. And the updating strength is dynamically controlled, and stable parameters are recovered during abnormal offset, so that the problem that the response reliability is reduced due to model updating delay or parameter offset during high-frequency interaction or semantic drift caused by conflict between speed and stability in online incremental updating is effectively solved.

The invention discloses an anti-multipath interference UWB radar signal denoising and positioning enhancement method and system, and the method comprises the steps: deploying a plurality of groups of UWB radar equipment, collecting original signals in a multipath interference scene, marking the original signals, and constructing a UWB radar signal data set containing multipath interference features, target features and marking information; preprocessing the original signal and extracting multi-domain features; a deep learning model containing a denoising model and a positioning enhancement model is constructed, and joint training is carried out through a joint loss function; training the model by adopting a training strategy of dynamic regularization and a self-adaptive learning rate; and finally deploying equipment acquisition signals, inputting the signals into the trained system for processing, and outputting a positioning result. The method can effectively suppress multipath interference, improves the UWB radar positioning precision and generalization capability, and is suitable for indoor navigation, automatic driving and other scenes.

The invention relates to the technical field of unmanned aerial vehicle aerial photography and computer vision, in particular to an unmanned aerial vehicle aerial photography long-term target tracking method based on matching loss confidence. Comprising the steps of performing multi-feature fusion extraction on a tracking target object, calculating the similarity between a target template and a candidate region, performing target search matching by using an improved pelican search algorithm, realizing target scale adaptive updating through a scale pyramid, and dynamically updating the target template based on an adaptive learning rate; and designing a long-term tracker combined with the matching loss confidence, wherein the long-term tracker comprises the steps of judging a target tracking state based on the matching loss confidence and executing rapid positioning redetection when the target is lost. Through multi-feature fusion, an improved pelican search algorithm and a matching loss confidence judgment mechanism, the long-term stable tracking of the target in the aerial photographing scene of the unmanned aerial vehicle is realized, the problems of shielding, scale change, target loss and the like are effectively solved, and the tracking precision, the success rate and the real-time performance are considered.

According to the power transmission line geological disaster time sequence InSAR prediction method based on the improved LSTM model, an improved long short-term memory network LSTM model is constructed, and a basic data set is established for LSTM model training by using obtained power transmission line monitoring area time sequence InSAR deformation result data. And a regularization technology, an attention mechanism and an adaptive learning rate adjustment strategy are introduced in a model training process, so that an overfitting risk is reduced, the generalization ability of the model is improved, and a high-precision deformation prediction model for time sequence InSAR monitoring of the power transmission line is obtained. The method is of great significance to early warning and deformation trend analysis of geological disasters around the power transmission line, and can provide scientific basis and decision support for safe operation of power transmission facilities, thereby effectively reducing potential hazards of the geological disasters to a power system.

The invention discloses a YOLOv11 road disease detection method and system based on multi-scale feature enhancement. The method comprises the steps of road image acquisition and marking, adaptive learning rate adjustment and data enhancement, construction of a YOLOv11 model embedded with a sub-pixel level edge enhancement module and a wavelet transform feature extraction module, staged training, model evaluation, disease detection and the like. Wherein the sub-pixel-level edge enhancement module is used for enhancing fine crack features through multi-operator fusion and a sub-pixel convolution technology; the wavelet transformation module improves the multi-scale feature perception capability through adaptive wavelet basis selection and a multi-stage decomposition and reconstruction mechanism. The system correspondingly comprises a data preprocessing module, a model building module, a training optimization module and an evaluation deployment module. On the basis of keeping the real-time performance of the YOLOv11, the detection precision and robustness of multi-scale diseases in low-contrast, sub-pixel-level cracks and complex environments are remarkably improved, and the method is suitable for road maintenance and safety monitoring scenes.

The invention provides a species detection and identification method based on a hybrid scaling strategy and a cascade architecture, which belongs to the technical field of image identification, integrates multiple technologies such as image preprocessing, neural network model construction and expansion, model transfer learning and fine tuning, optimizes the network structure by means of the hybrid scaling strategy, and improves the identification accuracy. A detection network and a recognition network are organically combined, species recognition tasks are concentrated in a region where detection is meaningful, fine-grained detection and accurate recognition of species are achieved, and the accuracy and reliability of species recognition are greatly improved. The hybrid scaling strategy realizes feature extraction optimization of targets with different scales by dynamically adjusting the depth, width and input resolution of the network; the cascade architecture reduces the influence of background interference on a classification task through progressive processing of positioning and then classification. In addition, an adaptive learning rate adjustment mechanism is introduced, and a verification set early stop strategy is combined, so that the convergence efficiency and generalization performance of the model are further improved.

The invention belongs to the technical field of intelligent oil field and oil and gas yield prediction, and particularly discloses an oil well yield prediction method and system based on a frequency domain enhanced BiLSTM-Transformer hybrid model, and the method comprises the steps: collecting oil well historical production time sequence data, and carrying out the abnormal value elimination and normalization preprocessing; constructing a multi-scale feature system, and integrating a time domain lag term, a rolling statistic, and frequency domain dominant frequency and amplitude features extracted through Fourier transform; introducing an oil well embedding vector to explicitly compensate the inter-well static difference; the features are input into a BiLSTM-Transform hybrid model, a local dynamic state is captured by the BiLSTM, and a global trend is modeled by the Transform; an oil well level adaptive optimization strategy is adopted, sample imbalance is relieved through sample dynamic weighting and adaptive learning rate scheduling, and the generalization ability is improved in combination with gradient cutting and an early stop mechanism; and finally outputting an oil well yield prediction value. According to the method, the problem of prediction deviation caused by yield sequence non-stationarity, inter-well difference and sample imbalance is effectively solved, and the prediction precision is remarkably improved.

The invention discloses a landslide susceptibility evaluation method fusing Bayesian optimization and a sparse residual network, and relates to the technical field of mountain disaster space prediction. According to the method, a landslide multi-source factor set is constructed, space standardization is completed, feature compression is performed by using Bayesian optimization XGBoost, and primary probability features are output; inputting the primary features into a sparse gating residual network, and extracting high-order nonlinear coupling features through a sparse attention and residual contraction dual mechanism; training a network by adopting cross entropy loss and an adaptive learning rate, and outputting a landslide occurrence probability; and dividing susceptibility grades based on a GIS natural breakpoint method and drawing. The method is different from an existing landslide model in the aspects of primary feature extraction and deep network architecture, landslide susceptibility prediction precision and interpretability are remarkably improved, and a scientific basis is provided for traffic line selection and disaster prevention planning in a cold region.

This invention discloses a tracking method based on multi-scale spatial constraints to combat occlusion. It extracts HOG, CN, and grayscale features from candidate target regions and uses PCA to reduce the dimensionality of HOG and CN features, accelerating computation. A channel weight fusion method is employed, training filters separately for each feature layer. Adaptive fusion weights enhance the response of effective feature layers, addressing the problem of insufficient multi-feature response fusion. A stepped spatial constraint method is proposed to optimize the color space constraint model, preventing model errors from obscuring target information and limiting the effectiveness of the spatial domain constraint model. Adaptive learning rate and diffusion search methods reduce irrelevant information learned by the filters and improve tracking accuracy when the target is occluded. A tritree scale acceleration method is proposed, introducing scale filters and transforming the parallel structure of the scale filters into a tritree classification structure.

The invention discloses a virtual power plant load prediction method and system based on big data analysis, and belongs to the field of virtual power plant load prediction, and the method comprises the steps: building a parameter optimization objective function for a preliminarily updated model parameter set, and carrying out the optimization of a parameter optimization objective function according to the prediction error feedback of latest load data; carrying out fine adjustment on weight parameters by adopting a gradient descent method of an adaptive learning rate, and determining optimal parameter configuration after convergence by minimizing a prediction error; reconstructing a load prediction model by adopting optimal parameter configuration, carrying out weighted fusion on a historical load mode feature vector and a current change trend coefficient, and carrying out rolling prediction calculation on a load demand in the next 24 hours to obtain an hourly load prediction value and a corresponding confidence interval; and integrating the new load mode change characteristics into a model knowledge base through a continuous learning mechanism, performing weight decreasing processing on historical mode data of more than 30 days by adopting an exponential decay factor, and constructing a dynamic knowledge base containing multi-mode load characteristics.

The invention discloses an extreme ultraviolet laminated diffraction imaging method based on an iterative frequency mask. Comprising the steps that an extreme ultraviolet coherent light source is used for scanning and irradiating a to-be-reconstructed target, and diffraction intensity patterns of all scanning positions are collected; obtaining an initial probe function and an initial reconstruction target function according to the position information of all the scanning positions and the diffraction intensity pattern; meanwhile, an auxiliary variable tensor, a Lagrange multiplier tensor and an adaptive learning rate are initialized; constructing a multi-level spatial frequency mask; according to diffraction intensity patterns of all scanning positions, a multi-level spatial frequency mask is used for updating the probe function step by step from low frequency to high frequency, a target function, an auxiliary variable tensor and a Lagrange multiplier tensor are reconstructed, and the obtained optimal probe function and the optimal reconstruction target function serve as reconstruction results. The method has remarkable advantages in the aspects of reconstruction quality, convergence speed and robustness to low lamination rate and high noise environment, and has important application value for lamination diffraction imaging.

The invention relates to an electric equipment fault prediction method based on adaptive deep reinforcement learning, and the method comprises the steps: collecting multi-source operation data of electric equipment, carrying out the preprocessing of the multi-source operation data, and obtaining a synchronous data flow; inputting the synchronous data stream into an adaptive feature extraction network based on an attention mechanism, and outputting high-dimensional fault sensitive features; the high-dimensional fault sensitive features are input into a trained fault prediction model, action probability distribution output by an intelligent agent is obtained and converted into an electric equipment fault risk index, the fault prediction model is a deep reinforcement learning model, and the electric equipment fault risk index is an electric equipment fault risk index. The deep reinforcement learning model is obtained by establishing an equipment state space, an action space and an instant reward function and training by adopting an adaptive learning rate algorithm; and according to the electric equipment fault risk index, displaying an equipment health state curve and a risk thermodynamic diagram in real time through a visual interface. According to the invention, high-precision fault prediction can be realized.

The application provides a three-dimensional implant crown intelligent selection and generation method based on U-Net, comprising the following steps: converting a three-dimensional tooth model in STL format into a high-resolution two-dimensional projection image; taking a U-Net network structure with VGG16 as a backbone network, combining a bar pooling module, and designing a neural network model for accurately capturing tooth sequential arrangement features; training the neural network model by using an adaptive learning rate adjustment strategy and a cosine annealing learning rate scheduling strategy, and performing semantic segmentation on the two-dimensional projection image to obtain a 2D tooth classification result image; identifying missing tooth labels based on the 2D tooth classification result image, calculating the center points of left and right adjacent teeth of the missing tooth, calculating the center position coordinates of the missing tooth in the 2D image through the center points of the left and right adjacent teeth, and mapping the center position coordinates to a three-dimensional space to obtain three-dimensional center coordinates of the missing tooth; and moving the template tooth center point to the three-dimensional center coordinates of the missing tooth to generate a three-dimensional crown model that matches the missing tooth position.

The application provides a few-shot bearing fault migration diagnosis method based on multi-working condition supervised contrast learning, which comprises the following steps: taking a plurality of vibration signals of bearings under different working conditions and having a large number of labels as source domain data and taking vibration signals under a working condition different from that of the source domain data and having only a small number of labels as target domain data; obtaining feature representation and prediction probability distribution of each fault category according to a pre-constructed contrast learning training classification model and the source domain data and the target domain data; then, respectively calculating a supervised contrast loss and a cross-entropy loss, and combining them into a total loss function by weighting; and finally, obtaining a trained contrast learning training classification model by updating model parameters through an adaptive learning rate optimization algorithm and back propagation according to the combined total loss function. The application extracts features by using a plurality of different source domain data through an improved supervised contrast loss function, and optimizes model parameters in combination with a cross-entropy loss function, so as to diagnose the target domain with scarce data.

The invention belongs to the technical field of vibration data identification, and discloses a real-time adaptive learning bearing fault classification method and system based on prototype alignment and a parameter-by-parameter adaptive learning rate, and the method comprises the steps: inputting a source domain sample into a source domain pre-training module, obtaining a basic classification model through supervised training, and carrying out the training of a source domain pre-training module; extracting a feature mean value of each category to construct a source domain feature prototype set; in a target domain test stage, generating a relatively stable pseudo tag for a target domain sample by using an EMA Teamer module; then, an online self-adaptive updating process is executed on a target domain sample by using a pseudo tag, and feature alignment is performed in the process by combining symmetric cross entropy loss and prototype-based comparison loss; meanwhile, model parameters are stably updated by adopting a parameter-by-parameter adaptive learning rate strategy; and finally, outputting an accurate prediction label for a target domain sample by the model after adaptive training.

This invention relates to the field of data processing technology, and in particular to a method for dynamic monitoring of reservoir water levels based on machine learning. This method acquires multidimensional monitoring data of the reservoir at each monitoring moment within a preset historical period prior to the current monitoring moment, constructs a three-dimensional spatial coordinate system, fits the data points in the three-dimensional spatial coordinate system to obtain a fitted straight line, and filters at least one suspicious environmental time interval within the preset historical period based on the distance difference between each data point in the three-dimensional spatial coordinate system and the fitted straight line. Based on the fluctuation characteristics of the multidimensional monitoring data within each suspicious environmental time interval, the multidimensional monitoring data within each suspicious environmental time interval is classified to obtain the data classification result corresponding to each suspicious environmental time interval. This result is used to adjust the preset baseline learning rate in the GBDT model to obtain an adaptive learning rate. Based on the adaptive learning rate, an optimized model is obtained, improving the accuracy of dynamic monitoring of reservoir water levels.

The application discloses a photovoltaic array fault diagnosis method based on photovoltaic power station current signal output, collects current signals, and decomposes the denoised current signals into multiple intrinsic mode functions (IMFs) by using an improved AVMD algorithm after denoising; the improved AVMD algorithm introduces an adaptive learning rate adjustment, and combines an improved Adam algorithm to optimize and update bandwidth parameters and frequency parameters; a Lasso regression model is constructed, L1 regularization is introduced, and IMFs with higher importance are screened out as information carriers for diagnosing photovoltaic system faults; wavelet transform is performed on the selected IMFs, low-frequency components and high-frequency components on different frequency bands are obtained, coarse-grained processing is performed on the low-frequency components, multi-scale discrete entropy (MDE) calculation is performed, time scale factors and curves of normalized discrete entropy values are obtained; a threshold is set to judge the health condition of the photovoltaic system, and fault diagnosis is realized. The improved AVMD algorithm is used for decomposing and denoising the current signals, and the method has good robustness and is easy to implement.

The application discloses a kind of bevel gear defect detection artificial intelligence identification method and system, it is related to mechanical part AI detection field.This method is collected multimodal data by industrial camera and vibration sensor, adopts dynamic receptive field attention CNN to extract image feature, extracts vibration feature using adaptive noise auxiliary EMD algorithm, and obtains comprehensive feature by bimodal attention fusion algorithm;DBN model with Dropout-L2 regularization is constructed, combined with cosine annealing-momentum adaptive learning rate training, to realize defect detection.The application solves the problems of single mode dependence, low efficiency and weak small sample generalization of existing algorithms, and is mainly used for efficient and accurate detection of bevel gear surface and internal defects, suitable for quality control of bevel gears in the fields of automobiles, aerospace and other fields.

The invention discloses a defense method for perceptual poisoning, which is used for coping with perceptual poisoning attacks in federal learning target detection tasks by introducing suspicious client detection based on cosine similarity, self-adaptive learning rate adjustment and a dynamic defense interval mechanism, and realizing robustness training and stable defense of a system. The method can solve the technical problems that an existing defense method based on statistical anomaly detection depends on single-round gradient statistical characteristics, it is difficult to keep stable performance under the non-IID data condition, and the method is prone to being affected by noise fluctuation, so that the misjudgment rate is high. And the technical problems that the existing defense method based on trust scoring and weighted aggregation generally adopts a fixed learning rate or a linear attenuation strategy in a trust degree updating process, cannot flexibly adapt to a dynamically changing attack mode, easily causes defense response delay and causes insufficient system robustness are solved.

The application relates to a supporting structure stress state monitoring method based on artificial intelligence and belongs to the technical field of artificial intelligence and data processing. The method comprises the following steps: acquiring and labeling strain data of a supporting structure; after removing abnormal values, normalizing multi-sensor data to generate normalized strain sequences; constructing a state monitoring model, adopting a deep time sequence neural network architecture, including an input layer, an adaptive wavelet attention feature mapping layer, a time domain gated convolution module, a global maximum pooling layer, a dynamic feature importance reweighting layer and a full connection classification layer; inputting normalized data to train the model; optimizing a loss function through a quantile interval adaptive learning rate and a momentum update strategy; after processing real-time monitoring data, inputting the data into the trained model according to a time window slice, outputting four types of probabilities and taking the maximum value as a predicted state; if a plurality of windows are in early warning and danger in succession, terminal alarm is triggered. The application can improve the accuracy of supporting structure stress state monitoring.

The application discloses a kind of energy storage system health state evaluation method and system, mainly related to health state evaluation technical field, to solve the problems that prior art cannot identify key degradation signal on frequency domain, cannot reflect the entropy change law of non-stationary change, the insufficient perception of degradation area of deep neural network.For solving the problems that prior art cannot identify key degradation signal on frequency domain, cannot reflect the entropy change law of non-stationary change, the insufficient perception of degradation area of deep neural network, the application includes: real-time acquisition of the original monitoring data of energy storage system, calculate the normalized data after time domain and frequency domain processing, calculate the characteristics of enhanced monitoring data;Residual attention deep neural network is constructed, and the current adaptive learning rate is adjusted autonomously;Dynamic focal point loss is calculated, and total loss function is calculated;The trained residual attention deep neural network is obtained;Current monitoring data is obtained, the characteristics of enhanced current monitoring data are calculated, and then the trained residual attention deep neural network is used to evaluate the health state of energy storage system.

A multi-parameter BP neural network temperature compensation method for an inertial microsystem comprises the steps of performing full-temperature-range precise calibration and establishing a data set, establishing a multi-physical-effect fused multi-parameter coupling error model, and establishing a BP neural network compensation model with temperature and error coefficients as input and compensation parameters as output. Dynamically adjusting a learning rate optimization convergence process by adopting a self-adaptive learning rate momentum method to obtain an optimal network weight and bias parameter, inputting the optimal network weight and bias parameter into the BP neural network compensation model, burning the updated BP neural network compensation model into the inertial microsystem, and finally obtaining the optimal network weight and bias parameter of the inertial microsystem. According to the method, multi-parameter coupling modeling and online compensation are combined, so that full-temperature-range, high-precision and autonomous temperature compensation of the inertial microsystem is realized, the performance stability and reliability of the inertial microsystem in a complex temperature environment are ensured, and the method is suitable for large-scale popularization and application. And the output precision and stability of the inertial microsystem in a complex temperature environment are obviously improved.

The invention discloses a PCB defect detection method based on deep learning, and relates to the technical field of image processing, and the method comprises the steps: obtaining a standardized PCB image data set; an improved MixUp data enhancement algorithm is adopted to expand the standardized data set; constructing a detection model adaptive to PCB small target defects and complex backgrounds; inputting the training set into an improved detection model, and performing iterative training by adopting an adaptive learning rate adjustment algorithm and a mixed loss function; inputting a to-be-detected PCB image into the trained model; and adding new defect samples in actual production into a training set regularly, updating model parameters by adopting an incremental training mode, and dynamically monitoring and optimizing a model state. According to the method, the problems of poor data quality, low small target precision, weak model adaptability and high industrial landing cost in traditional PCB defect detection are solved through multi-technology cooperation, and high-precision, real-time and sustainable optimization PCB defect detection is realized.

The invention relates to the technical field of voice recognition, in particular to an immersive man-machine interaction method and system based on voice recognition, and the method comprises the steps: voice data collection and fundamental frequency extraction; performing dialect interference quantitative evaluation; analyzing the disharmony trend of the lingual potential; and adaptive learning rate adjustment and model optimization are carried out. According to the method and the device, the fixed learning rate in the online learning model is dynamically adjusted, so that the voice recognition model of the region where the user is located is updated in real time according to dialect characteristics, and immersive man-machine interaction of the user is facilitated.