272results about How to "Enhance expressive ability" patented technology

The invention discloses a small target detection method based on a Fourier mixed attention mechanism. An improved small target image detection network model based on RT-DETR is researched and designed, and a Basic Block module in a backbone network is replaced by a self-developed Fourier mixed attention enhancement module (FTABlock). The module is composed of a Fourier transform attention module (FTAModule) and a hybrid dynamic convolution feedforward network (DKMixFFN). The FFTModule calculates the attention weight in the frequency domain through Fourier transform, and strengthens small target feature expression in combination with position coding and multi-head attention; the DKMixFFN adopts a dynamic convolution kernel and a multi-scale dynamic convolution layer to realize adaptive modeling of multi-scale features. According to the method, under the synergistic effect of frequency domain attention and dynamic convolution, the small target feature extraction and perception capability of the RT-DETR model is effectively enhanced, and the small target detection precision is improved.

The invention discloses an electroencephalogram feature recognition method for motion intention prediction in a lower limb alternate stepping process. The electroencephalogram feature recognition method comprises the following steps: acquiring an electroencephalogram signal of a subject before the subject executes an action under the requirement of a lower limb alternate stepping motion execution normal form; preprocessing the electroencephalogram signal to obtain an initial signal, performing label labeling on the initial signal according to an action type, and forming a data set by a label and the initial signal; constructing a neural network model, wherein the neural network model comprises a multi-scale grouping space-time convolution module, a self-attention mechanism module for updating external memory based on Top-k gating and a grouping causal time convolution module; and training the neural network model in a supervised manner by using the data set to obtain a prediction model for predicting the action type. According to the method provided by the invention, the movement intention of the patient can be recognized by the movement brain-computer interface before the lower limb alternate movement, the control of the rehabilitation robot on the specific lower limb movement by the brain intention is realized, and the method has an important value on the clinical application of a brain-controlled rehabilitation robot system.

The invention relates to a lightweight multi-scene disease and pest detection method and system based on RT-DETR. According to the scheme, firstly, a standardized image processing link is constructed, multi-band feature extraction is carried out on an input image by using a convolutional backbone network introduced with wavelet transform, an effective receptive field is expanded in a frequency domain through wavelet decomposition and an inverse reconstruction mechanism, and feature capture of a tiny insect pest target is enhanced while calculation redundancy is reduced. Furthermore, a bidirectional feature pyramid network including global and local double-branch collaborative modeling is adopted, cross-level dynamic interaction and gating fusion are performed on multi-scale features, and environmental noise interference such as veins and illumination under a complex farmland background is effectively inhibited. And finally, establishing a homography mapping model from a pixel plane to a geographic space according to camera calibration parameters, converting a visual detection result into a spatial distribution diagram layer with latitude and longitude information, and realizing dimension crossing of pest and disease damage monitoring from single-point identification to region-level risk assessment.

The invention discloses a short-term traffic flow prediction method based on a dynamic multi-dimensional global perception graph convolutional network, and belongs to the field of traffic flow prediction. The method comprises the following steps: acquiring an input dynamic graph feature sequence, and generating space-time element parameters; carrying out feature enhancement on the dynamic graph feature sequence, fusing the dynamic graph feature sequence with the space-time element parameters, and then respectively carrying out space-time feature coding; fusing the spatio-temporal characteristics obtained by coding the spatio-temporal characteristics; and carrying out space-time trend alignment decoding on the fused space-time features. According to the method, the defects of limited modeling precision, insufficient dynamic adaptability, incomplete spatial-temporal feature capture and the like in short-term traffic flow prediction are overcome, and the accuracy and the applicability of short-term traffic flow prediction are improved.

The invention provides a plant disease feature extraction and classification method and system based on deep learning, and relates to the technical field of plant disease identification, and the method comprises the steps: mining the interdependence rule of different disease features through constructing a disease feature symbiosis wake-up network, and setting wake-up conditions; inputting a to-be-analyzed plant disease image into the network, capturing initial disease features, triggering associated feature wakeup, and generating a disease feature symbiotic set; conducting the characteristic information according to the hierarchical relationship to generate enhanced symbiotic disease characteristics; recording morphological change details to construct a disease characteristic evolution sequence; and inputting the disease characteristic evolution sequence into a pre-trained deep learning classification model, analyzing disease essential characteristics, and generating a classification result containing a disease type and a matching basis. The method can comprehensively and accurately extract features, and improves the accuracy and reliability of plant disease classification.

The application discloses an infrared image generation method and system based on a physical prior constraint generative adversarial network, relates to the field of infrared image generation, and aims to solve the problems of insufficient physical reality and lack of representation ability of infrared radiation characteristics of the existing infrared image generation method. The application first designs a physical prior knowledge learning network, which is used to learn the prior representation related to the infrared radiation characteristics from the infrared image, and realizes the prediction of the infrared radiation related parameters through a deep learning network. Secondly, a generative adversarial network based on the physical prior constraint is designed based on the proposed physical prior knowledge learning network, the physical prior features are introduced into the image generation process, and the learning ability of the generator to the overall distribution of the infrared radiation and the representation ability of the generator to the infrared radiation characteristics are enhanced. The method breaks through the limitation that the traditional generator only relies on shallow texture mapping. The application is suitable for the fields of visible light to infrared image generation, infrared vision algorithm research and development.

The invention discloses a charging station dynamic recommendation method and system based on multi-dimensional real-time data, and relates to the technical field of big data analysis, and the charging station dynamic recommendation method based on the multi-dimensional real-time data comprises the steps: obtaining charging data of all users according to charging requests of the users, the charging data comprises historical order data, real-time charging gun data, charging pile data and charging station data; based on the charging data of all users, constructing a multi-source heterogeneous relational graph of an associated order, a charging gun, a charging pile and a charging station, and constructing a corresponding initial feature vector for each node in the multi-source heterogeneous relational graph; and updating all charging station node features based on adjacent order nodes, order node information of each user, charging gun node features and charging pile node features so as to obtain final all charging station node features, calculating recommendation indexes of all charging stations, and recommending the charging stations to the users. And the recommendation result can better meet the real intention of the customer and the station operation collaboration.

The application provides a multi-modal target detection method based on super-resolution assistance, and the implementation steps are as follows: obtaining a training sample set and a test sample set; building a multi-modal target detection network model based on super-resolution assistance; iteratively training the multi-modal target detection network model based on super-resolution assistance; and obtaining a target detection result. The multi-modal fusion module is provided with a symmetrical double-branch structure, realizes efficient fusion of RGB and IR images at the input end, retains color integrity and complementary information of thermal radiation, significantly improves feature discrimination, and effectively improves detection accuracy; the super-resolution assistance branch serves as an internal branch, directly optimizes feature representation of the backbone network through a reconstruction loss, thereby significantly enhancing the expression capability of the network for small targets and fine features, and meanwhile, the super-resolution assistance branch does not participate in the detection process of the test sample, realizes zero additional calculation overhead in the reasoning stage, and improves the detection efficiency while improving the target detection accuracy.

The application relates to a point cloud three-dimensional target detection method and system based on a local attention Mamba model, which comprises the following steps: obtaining a key point cloud subset from a point cloud; inputting the key point cloud subset into a local attention Mamba model composed of N hierarchical stacks, and sequentially performing the following steps on each level: for each point, obtaining a neighborhood point set through a hash table index, generating local attention weights based on element-by-element multiplication interaction, and weighting and aggregating neighborhood features to obtain local geometric features; then modeling global long-range dependencies of the local geometric features by using a bidirectional Mamba module, and outputting global features as the input of the next layer; and inputting the global features output by the last level into a bird's eye view backbone network to obtain a three-dimensional target detection result. The application forms a progressive feature learning mechanism of local perception and global enhancement, and significantly improves the expression ability and reasoning efficiency of a point cloud backbone network.

This invention relates to a graph node classification method based on a graph feature self-attention mechanism using ensemble learning. First, the original graph network data is input, including the node feature matrix H and the graph adjacency matrix A. The LightGBM method is used to preprocess the dataset to obtain model parameters. Finally, the expanded node feature matrix H is obtained by summarizing the tree model parameters. new ; Combine the graph adjacency matrix A and the expanded node feature matrix H new The input graph features are trained on an attention network, and the network's preference information for nodes and features is obtained after training. Finally, the trained model is used to predict the classification of the graph node dataset on the validation and test sets. This invention utilizes an ensemble learning method and incorporates node attention preferences for features, which not only efficiently leverages the high interpretability of tree models but also enriches the model's expression, allowing node features to contain more information, thus contributing to higher performance on node classification tasks.

The invention provides a lightweight human shape detection method and device based on improved RTDETR, a medium and a product, and the method comprises the steps: obtaining a data set for human shape detection, and dividing the data set into a training set, a verification set and a test set; an improved RTDETR model is constructed, an ES Block is used in the improved RTDETR model to replace a Basic Block in the RTDETR feature extraction network before improvement, and the ES Block comprises an OfficientVIT Block and a SimAM attention module which are connected in sequence; the NWD Loss is used as a loss function of the improved RTDETR model; training the improved RTDETR model by using a training set, verifying the improved RTDETR model by using a verification set and testing the improved RTDETR model by using a test set to obtain a trained improved RTDETR model; and inputting the to-be-detected image into the trained improved RTDETR model for reasoning to obtain a human shape detection result of the to-be-detected image. By means of the technical scheme, the human shape detection precision and small target human shape detection in a security and protection monitoring complex scene can be improved.

This invention relates to a physical information neural network-based intelligent performance prediction method and system for automated multi-material fiber placement on complex curved surfaces in aircraft. The invention includes: collecting fiber placement process parameters; performing standardized preprocessing and physical constraint-enhanced sampling to obtain an enhanced training dataset; constructing a high-order nonlinear sparse regression candidate library based on the enhanced training dataset; jointly identifying piecewise smooth physical differential equation systems and their mode switching logic using sparse Bayesian regression and Hidden Markov Models to obtain an embeddable inverse mechanism model; constructing a Physical Information Embedded Generative Adversarial Network (PI-GAN); collaboratively optimizing the generator and discriminator through adversarial training to output surface defect indices and mechanical performance indices; and initiating incremental self-learning when introducing new process scenarios to achieve model self-evolution. This invention achieves end-to-end accurate prediction of multi-dimensional quality indices such as surface defects and mechanical performance.

This invention relates to the field of natural language processing technology, specifically to a method and system for matching the intent of intelligent NPC dialogues in multimodal interaction. The method involves real-time acquisition of multimodal data, generating multimodal semantic features through submodal preprocessing; integrating the semantic features of each modality using a cross-modal fusion module based on semantic association, generating a candidate intent set based on semantic context representation and combined with a semantic parsing module and predefined intent templates; tracking changes in user intent in real time through a continuous semantic learning mechanism and an interactive memory module, combined with a Bayesian update method, dynamically adjusting the confidence level of each intent in the candidate intent set, and filtering the final intent; constructing an NPC semantic cognition model, and performing semantic consistency analysis to perform semantic checks on user input and NPC dialogue state; combining the final intent and a decision engine to generate a dialogue strategy and output synchronized response content; this invention improves the accuracy of dynamic matching of intelligent dialogue intents.

This invention discloses a knowledge- and data-driven time series forecasting method and system, belonging to the field of industrial time series forecasting technology. To address the technical problems of existing technologies when processing strongly non-stationary data, such as key signal attention dilution, systematic underestimation of peak amplitude, incomplete capture of multi-scale dynamic patterns, and difficulty in embedding physical domain knowledge, this invention constructs standardized multi-dimensional time series data, performs heterogeneous feature grouping and parallel encoding to obtain multiple sets of feature representations, then extracts and integrates features from multiple sources to obtain intermediate state features, and finally performs multi-branch prediction and adaptive weighted fusion based on the intermediate state features to obtain the final time series forecast result. This invention can achieve active focusing on abrupt events in industrial sequences and accurate characterization of extreme value amplitudes, significantly improving prediction accuracy, physical consistency, and robustness to extreme conditions.

The invention discloses a traditional Chinese medicine face image color quantitative characterization method fusing color space and texture features, and relates to the technical field of image processing and medical image analysis. The method comprises the following steps: performing illumination normalization and noise suppression processing on a facial image of a patient to obtain an enhanced and denoised image, and constructing a structure response graph and a color fusion feature tensor; introducing a structure-guided color enhancement mechanism, and obtaining a structure-enhanced multi-channel color tensor based on the structure response diagram and the color fusion feature tensor; constructing a yellow saliency guide map and a multi-scale texture response tensor, and generating a joint feature map through a color-texture joint decoupling mechanism under saliency weighting; normalizing the combined feature map, and extracting texture features and colors of the normalized combined feature map; and combining the color and texture features into a joint quantitative feature vector, and quantitatively representing the color of the traditional Chinese medicine facial image. The problems that color distinguishing is unstable and color and texture decoupling is difficult under complex illumination are solved.

The invention discloses an Alzheimer's disease classification prediction method and system. According to the method, firstly, a bimodal brain network diagram is constructed, brain region features serve as nodes, and edge features are constructed through single nucleotide polymorphism data and structural magnetic resonance imaging data; the method is characterized in that closed-loop information interaction between nodes and edges is realized through a mutual feedback backflow graph neural network: firstly, edge features are dynamically updated based on node similarity, and information transmission from the nodes to the edges is realized; carrying out aggregation propagation and key screening among edge features by utilizing a topological neighborhood relationship; and finally, returning the edge feature information subjected to cross-modal fusion to the nodes, and updating node features. According to the method, the limitation that an existing graph neural network only pays attention to edge-to-node one-way information flow is overcome, and by describing dynamic coupling of nodes and edges and complementation of multi-modal information on a connection layer, the capturing capacity of early and tiny pathological modes of the Alzheimer's disease is enhanced, so that the accuracy and interpretability of classification prediction are improved.

This invention relates to the field of intelligent auditing, and more particularly to a graph neural network method and system for mining hidden relationships among multiple sets of enterprise accounts. The method includes: acquiring multiple sets of accounts of a target enterprise, extracting entity and transaction information, and constructing a heterogeneous graph network containing multiple types of nodes such as enterprises, individuals, bank accounts, and transaction vouchers, as well as multiple types of edges such as fund transfers, control, and voucher associations; inputting the heterogeneous graph into a pre-trained GNN, aggregating neighbor features through graph convolutional layers, and updating node embedding vectors; based on the embedding vectors, using an attention mechanism to calculate the confidence of hidden relationships between node pairs, and filtering suspicious related node pairs; performing path reasoning and backtracking on suspicious node pairs, mining multiple paths, and generating a chain of evidence for related paths. This application, by constructing a heterogeneous graph containing multiple types of nodes and edges, comprehensively captures entities such as enterprises, individuals, accounts, and vouchers and their multidimensional relationships, solving the problem that existing methods are unable to comprehensively characterize complex relationship scenarios and capture hidden relationships.

The invention relates to the technical field of social network information spreading prediction, and discloses an information spreading prediction method fusing public opinion field effect and a heterogeneous hypergraph. In order to solve the problems that in the prior art, only pairwise user relations are relied on, multi-user group influences cannot be described, different information is subjected to cascade independent processing, and multi-topic competition is not considered, the invention provides a prediction scheme fusing public opinion field effects and heterogeneous hypergraph learning. A heterogeneous hypergraph is constructed to obtain user multivariate relation representation, then a public opinion field effect is utilized to quantify attraction energy of different information topics to a user, attention competition among multiple topics is modeled, and a more real user propagation tendency is obtained; and finally, realizing joint prediction of user interest features and social influence features through an interactive fusion mechanism. The method can be used for scenes of information propagation trend analysis, public opinion monitoring, marketing recommendation, false information early warning and the like.

This invention discloses a method and system for predicting solar radio bursts based on a complete solar magnetograph, belonging to the field of solar radio technology. It includes: acquiring a complete solar magnetograph; inputting the complete solar magnetograph into a pre-trained improved ResNet model to obtain image feature vectors; dividing the image feature vectors into pseudo-time series and inputting them into an improved LSTM network to obtain solar radio burst prediction results; the improved ResNet model introduces a custom residual layer to extract small-scale features from the complete solar magnetograph, and the improved LSTM network introduces a gating mechanism to process the image feature vectors at different time scales. This method can improve the accuracy of solar radio burst prediction without utilizing solar radio spectrum data; it solves the problem that existing predictions using solar radio spectra are easily interfered with by the accuracy of the input data observations.

The present application relates to a kind of facial expression recognition methods based on attention mechanism and self-distillation, comprising the following steps: step 1, constructing facial expression recognition dataset;Step 2, constructing facial expression recognition model, the model is composed of feature extraction module, adaptive channel attention module, remodeling module and self-distillation network;Step 3, the input facial expression image is preprocessed;Step 4, using the training set image in the above facial expression recognition dataset trains facial expression recognition model;Step 5, using the facial expression recognition model trained to the test set and validation set in dataset are inferred classification.The method designs a kind of reasonable and efficient facial expression recognition scheme, it is well guided model to pay attention to those important feature information by the mode of attention mechanism;By a new self-distillation form, refine compressed network knowledge, improve the robustness of shallow network output feature and reduce the complexity of model in inference stage.

The application provides a mine-used speed reducer working condition mode recognition method and system combined with deep learning, obtains a set of original vibration signals and a set of running load signals of the speed reducer in a preset monitoring period, performs joint deep feature mining processing to obtain deep coupling features of the vibration signals and synchronism correlation features of the load signals, and inputs a working condition characteristic vector mapping network to generate a multi-dimensional working condition characteristic vector through a self-attention weighted fusion module. A pre-trained multi-layer working condition decision network is called to perform hierarchical mode decoding processing on the multi-dimensional working condition characteristic vector to generate a candidate working condition mode set and a confidence distribution. In combination with prior knowledge of fault modes in historical maintenance records of the speed reducer, a final working condition mode recognition result is determined and working condition early warning prompt information is generated. The application can accurately recognize the working condition of the mine-used speed reducer and timely give early warning, thereby improving the equipment operation reliability and safety.

Provided are a simultaneous localization and mapping method and device. The simultaneous localization and mapping method comprises: acquiring a current frame image input by a camera; performing scene recognition on the current frame image to obtain a key frame image having the greatest similarity to the current frame image in a global map; and determining a camera pose of the current frame image according to the key frame image. The simultaneous localization and mapping method comprises: acquiring a current frame image input by a camera; determining a key frame image having a similarity to the current frame image; acquiring a feature point matching relationship between the current frame image and the key frame image; and calculating a camera pose of the current frame image based on the feature point matching relationship between the current frame image and the key frame image. According to the simultaneous localization and mapping method and device of the example embodiments of the present disclosure, the camera pose can be calculated by a traditional method, and the camera pose can also be determined by loop detection and repositioning through a deep learning model.

The application relates to the technical field of electronic reconnaissance and signal processing, and discloses a radiation source individual fine identification method and system based on multi-modal feature fusion, which comprises the following steps: performing preprocessing on obtained radiation source original signals to obtain signal pulse units; extracting time domain sequences of the signal pulse units and performing multi-domain transformation to generate a first time-frequency image, a second time-frequency image and a bispectrum image; inputting the first time-frequency image, the second time-frequency image, the bispectrum image and the time domain sequence into a multi-modal fusion neural network; extracting corresponding image depth features and comprehensive sequence features through three image processing branches and a sequence processing branch; performing cross-modal fusion on the three image depth features and the comprehensive sequence features through a fusion module to output a radiation source individual identification result. The application improves the identification accuracy of radiation source individuals in a complex electromagnetic environment.

The invention discloses an edge terminal-oriented memory efficient model fine tuning method and system, which realize decoupling of a backbone network and a fine tuning process by constructing a parallel side network module, and avoid video memory occupation caused by backbone gradient return. A double-adapter module is arranged, modeling is carried out on same-layer features and cross-layer context information, and the feature expression precision is improved. And constructing a feature fusion module, and performing adaptive weighting on different path outputs through learnable gating. A backbone grouping module is integrated, dynamic grouping is carried out on a backbone network based on interlayer similarity, and redundant modules and calculation overhead are reduced. And through the fine tuning execution module, only back propagation and parameter updating are carried out on the side network, and a trunk freezing state is kept, so that the memory and training cost is reduced. According to the method, the memory efficiency and the training speed of the edge fine adjustment process are improved, high-precision model self-adaption is achieved with extremely low calculation burden, and application in the fields of cloud edge collaborative intelligence and the like of unmanned aerial vehicles, robots, vehicle-mounted terminals and the like is effectively supported.

This invention discloses a specular removal method based on grouped enhanced convolutional attention feature fusion, belonging to the field of image processing technology. The method constructs a specular removal model based on a CycleGAN network, which consists of a generator, a discriminator, and a loss function optimization module. The generator includes detail enhancement depth downsampling, upsampling, and grouped enhanced convolutional attention fusion modules, effectively capturing image details, reducing checkerboard artifacts, and enhancing feature fusion and capture capabilities. The discriminator uses a PatchGAN structure to improve local detail capture capabilities. The optimized loss function integrates multiple losses to improve model performance. Experimental results show that compared with various traditional and deep learning methods, the method of this invention outperforms in PSNR and SSIM metrics, exhibits strong adaptability, and can provide high-quality image data for subsequent detection of elevator door components.

The invention provides an after-myocardial infarction brain injury risk prediction method and system based on multi-modal data fusion, and relates to the technical field of medical information processing, and the method comprises the steps: firstly obtaining a heart image, a brain image and a brain tissue microwave echo signal of a patient; generating waveform data based on the phase difference and the amplitude difference of the echo signals; taking the heart feature data, the brain feature data and the waveform data as nodes respectively, and establishing connection edges among the nodes according to a preset clinical knowledge base to form a heterogeneous graph; processing the graph by using a graph neural network to obtain associated features; and finally, inputting the associated features into a multi-task learning model, and outputting a future brain injury risk probability and a self-regulation function damage degree result. The accuracy and comprehensiveness of myocardial infarction after-brain injury risk prediction are improved.

Embodiments of the present specification disclose a multi-granularity graph contrast learning method, device, equipment, medium and program product. The method comprises: using a granule enhancement method to adaptively refine an original graph to generate a plurality of target granules; the original graph is composed of a plurality of nodes and edges representing the interaction relationship between the nodes; the target granule represents a multi-granularity homogeneous region composed of at least Z nodes in the original graph; Z is a positive integer greater than 1; a positive sample pair and a negative sample pair are constructed based on a plurality of target granules; the positive sample pair is composed of nodes in the same target granule in the original graph; the negative sample pair is composed of nodes in different target granules in the original graph; a contrast loss is determined based on the positive sample pair and the negative sample pair; and an initial vector of each node in the original graph is optimized based on the contrast loss to obtain a target vector corresponding to each of the plurality of nodes.

The invention discloses a complex-material-oriented multi-path full-frequency global illumination neural rendering method and system, and belongs to the technical field of computer graphics, and the method comprises the steps: carrying out the scene query of light emitted by pixels, and obtaining a primary geometric buffer and a multi-path mixed geometric buffer of the multi-time ejection of the light; respectively extracting main frequency features and multi-path sub-frequency features in a neural feature field based on the two features; performing directed graph structured modeling based on the multi-path sub-frequency features to obtain propagation features representing light path propagation information and global features representing overall light path information; performing multi-path frequency joint feature fusion based on the dominant frequency feature, the propagation feature and the global feature to generate a fusion feature with a full-frequency feature; and estimating a final radiation brightness value of the pixel based on the fusion feature, and completing neural rendering of global illumination. According to the method, efficient modeling and high-quality drawing of multi-path full-frequency global illumination under a complex material can be realized, and the sense of reality and dynamic consistency of a rendered image are remarkably improved.

The invention discloses a magnetic levitation track irregularity random evolution modeling method, a medium and equipment, and belongs to the field of track engineering and stochastic dynamics modeling. The method comprises the following steps: acquiring irregularity data of a maglev track collected along a line direction, and acquiring track structure parameters and vehicle load statistical characteristics of a corresponding line; the track irregularity state increment between adjacent sampling points is calculated, a state variable set is constructed based on the track irregularity quantity, the track structure parameters and the vehicle load statistical characteristics, state intervals are divided according to the value range of the state variable set, and the condition statistical magnitude of the state increment is counted in each interval; and identifying drift term and diffusion term functions in a stochastic differential equation by using conditional statistical information through statistical inversion, machine learning or a fusion mode thereof, and constructing an orbit irregularity stochastic evolution model. The method can reflect the non-stationary and state related characteristics of the track irregularity under the action of the track structure and the vehicle load, and can be used for track state evaluation, dynamic simulation and maintenance decision.

This invention discloses a method and system for predicting atmospheric pollutant concentrations for early warning of sudden changes. The method includes: acquiring historical monitoring data and performing hierarchical cleaning; constructing an input feature set containing external influences, periodic time, historical lag, and statistical sliding window features; detecting abnormal events and fusing event features; dividing the dataset chronologically and constructing supervised learning samples under strict temporal causality constraints; constructing and training a CNN-LSTM composite prediction model; constructing a sudden change discriminant based on the predicted values ​​and outputting an early warning when the discriminant exceeds an adaptive threshold. This invention also provides a system for implementing the above method, which can be deployed on an FPGA. This invention improves the accuracy, stability, and sudden change response capability of prediction through end-to-end temporal causality constraints, robust data cleaning, multi-dimensional feature fusion, and the CNN-LSTM model, while ensuring consistency between offline evaluation and online deployment.