51 results about "AdaBoost" patented technology

The invention provides a diffusion-driven channel adaptive point cloud semantic communication method, which belongs to the technical field of wireless communication, and comprises the following steps: obtaining a training point cloud data set, constructing a point cloud feature extraction network based on a hypergraph convolutional neural network as a semantic encoder, and constructing a channel adaptive enhancement module as a channel decoder, a channel adaptive recovery module is constructed at a receiving end as a channel decoder, a diffusion reconstruction network is constructed as a semantic decoder, end-to-end joint training is performed on the point cloud feature extraction network, the channel adaptive enhancement module, the channel adaptive recovery module and the diffusion reconstruction network, the diffusion reconstruction network predicts original point cloud distribution, and the point cloud feature extraction network and the channel adaptive enhancement module are subjected to end-to-end joint training. And a loss function is constructed based on the chamfering distance between the predicted point cloud distribution and the original point cloud. In the communication process, the reconstructed semantic features are input into the diffusion reconstruction network to reconstruct an original point cloud structure. The high-order semantic features of the point cloud can be effectively extracted so as to improve the adaptive capacity of the method to the incomplete point cloud.

The invention discloses a breeding performance prediction system based on an intelligent sheep breeding platform, and belongs to the technical field of animal husbandry informatization. The intelligent breeding platform is designed, standardized collection, storage and intelligent management of sheep full-life-cycle data are achieved, and a solid data foundation is laid for follow-up research; secondly, innovatively combining ensemble learning (AdaBoost, GBRT and the like), machine learning (SVR, KNN and the like) and a sorting learning algorithm, and constructing a multi-level breeding sheep breeding performance prediction model system; particularly, a subjective and objective combination weighting method based on AHP-PCA is provided, and the prediction precision of the model is remarkably improved through feature reconstruction; in addition, an Achimedes optimization algorithm (AOA) is introduced into the field of breeding prediction for the first time, and adaptive search of hyper-parameters is realized by using a physical simulation mechanism of the AOA, so that the accuracy, efficiency and robustness of the model are remarkably improved.

The application provides an open vocabulary segmentation method and system for multi-modal model representation optimization, and belongs to the technical field of computer vision. Image data to be segmented is acquired; a pre-trained multi-modal model is used to process the acquired image to obtain a segmentation result. The application better optimizes visual-text representation in a multi-modal task, effectively aligns the same visual-text representation space, proposes a mask-sensitive loss to constrain the classification score and mask quality to be consistent in the parameter fine-tuning process, thereby giving the visual encoder local perception ability and improving the effect of the model in the fine-grained downstream task, introduces the original pre-training feature as a representation compensation to ensure the zero-shot ability of the pre-training visual-language model in the optimization process, and interacts the text representation and the visual representation, so that the text representation can be adaptively enhanced for different input images, and the alignment property of the visual-text in the open vocabulary segmentation can be effectively improved.

The invention relates to the technical field of large language models, in particular to a gearbox degradation trend prediction method of multi-head memory LLM under unsteady excitation, which comprises the following steps: performing feature extraction based on a vibration signal to obtain a time-frequency domain feature set; screening out a plurality of degradation sensitive features from the time-frequency domain feature set; performing feature fusion on each degradation sensitive feature to obtain a fused degradation feature; inputting the fused degradation features into a trained multi-head memory large language model, and outputting a degradation trend predicted value; the multi-head memory large language model adopts a multi-head memory attention mechanism to enhance the model feature extraction capability; a dynamic enhancement factor is generated based on the decoding features through an adaptive enhancement normalization layer, and the normalization strength is adjusted in real time to obtain enhanced normalization features; and inputting the enhanced normalized features into a large model detection head to obtain a degradation trend prediction value. According to the method, the gear box degradation trend prediction precision and efficiency can be improved.

This invention discloses an intelligent management method based on real-time speech recognition and transcription, relating to the field of artificial intelligence technology. The method includes the following steps: Step S1, multimodal audio perception and environmental adaptive enhancement; Step S2, acoustic feature extraction and real-time transcription mapping; Step S3, semantic error correction compensation based on dynamic context weighting; Step S4, structured element extraction and logical association reconstruction; Step S5, intelligent management closed-loop decision-making and task distribution; Step S6, multi-source information backtracking and index construction; Step S7, intelligent management efficiency evaluation. This application can solve the problems of limited recognition accuracy and missing semantic logic under complex sound fields. By improving transcription accuracy through multimodal perception and dynamic semantic compensation, it achieves automated connection from speech recording to structured management decisions, significantly improving office management efficiency.

The present application relates to a kind of multi-loop cable group electromagnetic loss intelligent calculation method and device, belong to power cable operation technical field, method includes: with finite element method to establish the cable group electromagnetic field model including multi-loop, current sample is randomly generated and corresponding cable core and metal sleeve loss are calculated, form current-loss sample set;GWO-RBF-Adaboost Intelligent Prediction Model is constructed, the center point of radial basis function neural network RBF, width and output weight are globally optimized by grey wolf optimization algorithm GWO, and the RBF neural network is used as the base learner of Adaboost, and current-loss sample set is used to train to obtain current-loss mapping relationship model;The current data of multi-loop cable group to be evaluated is input into the GWO-RBF-Adaboost model trained, and the cable core loss and metal sleeve loss of each loop are predicted.The present application realizes the convenient calculation of multi-loop cable group electromagnetic loss by current, and significantly improves the calculation precision of electromagnetic loss.

The invention discloses an adaptive enhanced dynamic graph contrast learning multivariate time sequence classification method, which is oriented to a multivariate time sequence classification task under the condition of few labels, and comprises the following steps: 1) an adaptive enhanced strategy; 2) dynamic graph comparison learning; and 3) carrying out two-stage combined training. According to the method, in a preprocessing stage, trend-periodic decomposition and statistical significance test based on random permutation are combined to judge the strength of time sequence dependence, and an enhanced recommendation strategy matched with data characteristics is generated; and dynamic graph comparison characterization learning and a two-stage training mechanism are introduced, so that the classification performance and the model stability are improved while unlabeled data are fully utilized.

The application discloses a knowledge reasoning path selection and evaluation method, system, device and medium based on reinforcement learning, belongs to the technical field of path selection and evaluation, and comprises the following steps: representing a knowledge graph as a graph structure, adopting a deep reinforcement learning framework; generating node embedding vectors by using graph contrast learning and an adaptive enhancement mechanism; constructing a generation network and a discrimination network to obtain a candidate reasoning path; constructing a multi-objective reward function which fuses a topological connectivity reward and a semantic consistency reward, and performing reinforcement learning; calculating dynamic propagation weights of nodes in a reasoning process by using an information propagation model, and integrating the dynamic propagation weights into node feature representation; and adopting a deep reinforcement learning method to perform end-to-end training on an agent, and outputting a reasoning path and a path evaluation score. The application is suitable for multiple actual reasoning scenes, and provides a technical path for deep knowledge discovery and intelligent decision-making of a large-scale knowledge graph.

The application relates to the technical field of artificial intelligence fault diagnosis, and discloses an alternating current machine fault diagnosis method based on ISBOA-Adaboost. The collected data is pretreated through a VIKOR algorithm, the snake heron optimization algorithm SBOA is improved, a Sinusoidal chaotic mapping is introduced in the SBOA population initialization stage to improve the population initialization uniformity, and the snake heron position is updated according to the current iteration number of the algorithm in the SBOA hunting stage; an adaptive weight factor is introduced in the SBOA escape stage to optimize the snake heron position update, the improved snake heron optimization algorithm ISBOA is used to optimize parameters of an iteration algorithm Adaboost, an ISBOA-Adaboost fault diagnosis model is constructed, and the blindness of parameter selection in the training process is compensated.

The invention belongs to the technical field of medical detection and biological medicine, and discloses a serum protein marker combination for glioma diagnosis and prognosis evaluation and a processing system of the serum protein marker combination. Eight algorithms including KNN, SVM, random forest, XGBoost, AdaBoost, LGBM, Gaussian naive Bayes and decision tree cover traditional machine learning and integrated learning, and model robustness is improved; according to the method, interpretable feature screening is adopted, SHAP and LIME tools are combined, the contribution degree of each marker to a diagnosis result is clarified, and the problem of'black box 'of machine learning is solved; the invention provides a machine learning algorithm-based mass spectrometry system for glioma detection, which is good in detection performance, high in speed, convenient to operate and low in cost, so as to meet the requirements of clinical early diagnosis and non-invasive examination of glioma.

The invention discloses a dual-coding adversarial learning and multi-scale expansion fusion attention image cartoonalization method, and relates to the technical field of image processing, and the method comprises the steps: constructing a generator comprising a dual-encoder structure, and enabling a content feature encoder to extract the structural semantic features of an input real image, a style feature encoder extracts texture style features of the target cartoon image; the style features are injected into the content features through adaptive instance normalization; processing fusion features by adopting a multi-scale expansion fusion attention mechanism, and adaptively enhancing a key channel and a space region; the decoder reconstructs and generates a cartoon image; meanwhile, a double-discriminator framework composed of a global discriminator and a local discriminator is adopted, the overall style consistency and the local detail authenticity of the image are evaluated respectively, and the generation process is optimized through adversarial learning. According to the method, the high-quality cartoon graph can be generated, the color texture is natural, the edge is clear, and details are rich.

This invention relates to the field of temperature field prediction for vehicle transmission components, and discloses a method, system, medium, and device for predicting the temperature field of wet friction components based on a cross-domain transfer learning model. The method includes: acquiring temperature field data and interface morphology data of the friction components, constructing a Transformer-LSTM-AdaBoost hybrid neural network model using heterogeneous physical quantity corresponding sample inputs, and pre-training the model as a feature extractor; constructing a DGDAN cross-domain transfer model, using interface morphology data as the source domain and temperature field data as the target domain, and inputting both into the DGDAN cross-domain transfer model for training, initially extracting common features from the two sets of data through the feature extractor; conducting adversarial training between the feature extractor and the domain discriminator through a gradient inversion layer, explicitly aligning the feature distributions of the two domains using the maximum mean difference metric, and obtaining the final common features; and inferring the full surface temperature field distribution online through the input interface morphology data, achieving dynamic state synchronization between the physical entity and the virtual model.

The invention belongs to the technical field of electric heating load prediction, and particularly relates to an electric heating load prediction method and device, equipment and a medium, and the method comprises the steps: collecting the historical load data and influence factors of electric heating of a user in a region; adopting a maximum correlation minimum redundancy algorithm to select an optimal influence factor set; clustering the users based on the historical load data and the optimal influence factor set to obtain a clustering result; extracting a feature set containing a plurality of feature indexes according to the historical load data, taking the feature set, the historical load data and the optimal influence factor set as input of a corresponding Adaboost-BiLSTM prediction model, selecting a corresponding Adaboost-BiLSTM prediction model according to a clustering result of each user, and outputting an electric heating load prediction value by the corresponding Adaboost-BiLSTM prediction model; according to the method, the adaptive enhancement algorithm and the bidirectional long-short-term memory neural network are combined, different weights are given to a plurality of weak learners, a strong learner is constructed, meanwhile, the time sequence characteristics of the electric heating load data are mined through forward and reverse bidirectional calculation, and finally the load prediction precision is improved.

The invention relates to the technical field of machine learning, and particularly discloses a disease monitoring platform intelligent detection system based on machine learning, which comprises a signal acquisition and processing module, a signal enhancement and model prediction module and a model output optimization module, the system collects original physiological signal data in real time through a multi-modal sensor, and performs preprocessing and preliminary noise suppression to obtain a high-quality physiological signal data set, thereby providing reliable basic data support for disease monitoring; aiming at the initial weak signal, the system adopts a self-adaptive enhancement strategy to improve the identifiability of the initial weak signal, and performs classification processing according to a signal-to-noise ratio so as to generate an accurate disease prediction result; and on the basis of credibility of a prediction result output by the machine learning model, signal processing and machine learning model identification capabilities are continuously improved. According to the system, weak physiological features are effectively enhanced, the prediction accuracy is improved, and adaptive closed-loop optimization is realized.

The invention provides an AI contract tamper-proofing method based on an OCR + large language model, belongs to the technical field of large language models, and improves image quality by performing multi-scale adaptive enhancement and super-resolution reconstruction on a contract scanning image. A semantic bridging fusion model is adopted to deeply fuse an OCR recognition result and an enhanced image under a sparse coding framework to realize context correction of low-confidence characters, semantic normalization is performed on a corrected text, and a term semantic vector sequence is established; the steady-state distribution field of the semantic concentration is calculated based on the fluid dynamic diffusion equation, the concentration gradient matrix is extracted to serve as the text fingerprint, contract tampering detection with robustness to the OCR recognition error is achieved through fingerprint similarity comparison, and the technical problem that the false alarm rate is high due to the fact that contract tampering detection is sensitive to the OCR recognition error is solved.

The invention discloses a data visualization method and system based on a ChatGLM model, and relates to the technical field of data visualization. The invention discloses a data visualization system based on a ChatGLM model. The data visualization system comprises a visual data receiving module and a visual data construction module. According to the method, the visualization effect is intelligently optimized through a self-adaptive enhancement mechanism according to the real-time demand and the data characteristics of the user, and the generated visualization data is ensured to most accord with the expectation and the task target of the user, so that the individuation and the accuracy of data display are improved; a multi-modal learning method is combined with multi-dimensional information such as numerical features, text features and image features, data common features are formed, and more comprehensive support is provided for visualization generation. Through comprehensive analysis of different data types, relevance between data can be deeply mined from multiple angles, the depth and accuracy of data visualization are improved, and a user is helped to better understand complex data.

The method for sensing the temperature of urban solid waste incinerators based on a self-attention hybrid ensemble network belongs to the field of industrial process modeling and intelligent sensing. This method addresses the highly nonlinear, multi-condition variations, and complex physical-chemical reaction mechanisms in the MSWI process. Combining an ensemble learning framework and improved neural network structure, it introduces moving block bootstrap (MBB) sampling based on maximum likelihood estimation (MLE), a self-attention mechanism, and a hybrid ensemble strategy fusing AdaBoost and Bagging to achieve accurate modeling and dynamic prediction of the Fourier transform (FT). This results in good modeling accuracy, generalization ability, and engineering adaptability. Experimental verification under multiple benchmark problems and actual MSWI conditions demonstrates that the proposed FT sensing method exhibits superior performance in both prediction accuracy and robustness, possessing significant engineering application potential and widespread value.

A method for differential diagnosis of gallbladder adenoma and cholesterol polyps based on the fusion of deep learning and clinical features in ultrasound images includes: performing deep learning on ultrasound images using a deep learning network to output an adenoma / polyp risk score; the deep learning network is selected from one of VGG, ResNet, DenseNet, GoogLeNet, Vision Transformer, and Inception; the adenoma / polyp risk score and clinical features are input into a trained machine learning model for learning to obtain the final probability of identifying the polyp as a gallbladder adenoma rather than a cholesterol polyp; the clinical features include: echo intensity, maximum polyp diameter, and age; the machine learning model is selected from one of Logistic Regression, Gradient Boosting, K-nearest neighbors, Gaussian naïve Bayes, AdaBoost, Random Forest, ExtraTrees, and SVM.

The application relates to a water quality parameter modeling method, in particular to a neural network water quality parameter prediction modeling method based on an improved AdaBoost, and solves the technical problems that the BPNN neural network used for existing water quality parameter modeling has the phenomena of overfitting and is prone to falling into local extreme values, thereby causing network training failure. The neural network water quality parameter prediction modeling method based on the improved AdaBoost comprises the following steps: 1) pre-processing measured water quality sample data; dividing the pre-processed sample set into a training set and a test set; 2) obtaining a strong water quality parameter concentration predictor after initializing the sample weight of the training set; and 3) evaluating the prediction effect of the strong predictor by using the test set data, obtaining a water quality parameter concentration prediction model, being capable of proportionally combining the prediction results of multiple BPNN models, taking advantages and making up for disadvantages, thereby reducing errors and improving the stability of the model.

The invention relates to a multi-loop cable group electromagnetic loss intelligent calculation method and device, and belongs to the technical field of power cable operation, and the method comprises the steps: building a cable group electromagnetic field model containing multiple loops through a finite element method, randomly generating a current sample, calculating the loss of a corresponding cable core and a metal sleeve, and forming a current-loss sample set; constructing a GWO-RBF-Adaboost intelligent prediction model, performing global optimization on the central point, the width and the output weight of a radial basis function neural network RBF by using a grey wolf optimization algorithm GWO, taking the RBF neural network as a basis learning device of Adaboost, and training by using a current-loss sample set to obtain a current-loss mapping relation model; and inputting current data of a multi-loop cable group to be evaluated into the trained GWO-RBF-Adaboost model, and predicting cable core loss and metal sleeve loss of each loop. According to the method, the electromagnetic loss of the multi-loop cable group is conveniently calculated through the current, and the calculation precision of the electromagnetic loss is remarkably improved.

This invention provides a sintering furnace temperature control method and system based on a BP neural network prediction model. The method includes: constructing an initial model based on an Adaboost BP neural network; training a prediction model using experimental data and the initial model; the experimental data includes the sintering furnace temperature and power of the target sintering; and predicting and adjusting the required heating power of the sintering furnace in real time based on the prediction model and the sintering temperature curve of the target sintering. This invention's sintering furnace temperature control method and system, based on a BP neural network prediction model, uses experimental data to train an Adaboost-based BP neural network to obtain a prediction model, predicts and adjusts the required heating power of the sintering furnace in real time based on the sintering temperature curve of the target sintering, resulting in a more timely and linear heating process with stronger anti-interference capabilities. Furthermore, it improves the stability of sintered product output and reduces the product defect rate.

This invention belongs to the field of elevator safety technology and discloses an intelligent system for elevator fault early warning and handling with artificial intelligence. It constructs a seven-dimensional data system through a data acquisition and preprocessing module, and combines a blockchain consensus and federated transfer learning dual-drive architecture with a chain-connected learning module to extract multi-dimensional fault features and generate early warning models adapted to different operating conditions. The dynamic early warning module adopts a three-drive mechanism of reinforcement learning decision-making, multi-dimensional threshold dynamic evolution, and fault trend prediction, which can predict fault evolution trends, dynamically optimize early warning thresholds, improve early warning accuracy, and reduce false alarm and missed alarm rates. Through a hierarchical response mechanism, major risks can be directly linked to emergency measures to avoid safety accidents such as people entrapment and elevator falls caused by the escalation of faults. The visual positioning module is based on an optimized YOLOv8 architecture, integrating a small target enhanced detection unit and a low-light adaptive enhancement module, which can accurately identify minor faults such as broken wire ropes and slight guide rail misalignments.

This invention discloses a multimodal fusion graph convolutional fault diagnosis method for slurry pumps under multiple operating conditions, belonging to the field of power equipment fault diagnosis technology. The method includes: acquiring vibration and acoustic emission signals from the slurry pump; generating frequency domain graph nodes through time-domain segmentation and fast Fourier transform; constructing graph structure data based on K-nearest neighbors and Gaussian kernels; constructing a dual-branch network to perform three-layer graph convolution, edge pooling, and fusion-based global pooling on the single-modal and fused-modal graph data respectively; introducing a channel-space joint attention mechanism to adaptively enhance key fault features; and fusing the single-modal and fused-modal diagnostic results through D-S evidence theory to output the final fault type. This invention integrates multi-source sensor information, effectively suppressing the influence of operating condition fluctuations, improving the accuracy and robustness of fault diagnosis under complex operating conditions, and is suitable for intelligent operation and maintenance of rotating machinery such as slurry pumps.

This invention discloses a fast adaptive GNSS interference identification method based on deep learning. The method first preprocesses the GNSS signal samples containing interference by removing DC, slicing to a fixed length, and normalizing power. Second, it converts the preprocessed signal samples into time-frequency images using short-time Fourier transform. Then, it constructs and trains a deep learning interference identification model with a structure-reparameterized convolutional network as the feature extraction backbone, while introducing channel attention and time-frequency dual-axis coordinate attention mechanisms to adaptively enhance the response to key channel features and the representation ability of significant time-frequency regions. Finally, through structure reparameterization, the multi-branch convolutional structure in the training state is equivalently converted into a single-path convolutional structure. The inference model of this single-path structure is used to classify the input time-frequency image and output the interference type identification result. This invention utilizes deep learning to achieve automatic identification and classification of GNSS interference, improving the accuracy and robustness of interference identification while maintaining low inference latency and low deployment resource consumption. It is suitable for real-time GNSS interference identification in complex electromagnetic environments.

This invention significantly lowers the technical threshold for procurement data analysis by introducing LLM-based Text2SQL technology, improving the accuracy of natural language queries and the reliability of generated SQL statements. Through data preprocessing and model fine-tuning modules, it enhances the large language model's understanding of procurement domain expertise. Through query optimization modules and a two-layer feedback-driven adaptive FAQ enhancement method, it achieves continuous learning and performance improvement of the model in specific business scenarios. Through multi-agent consistency reflection mechanisms and SQL syntax checks, it significantly improves the consistency and accuracy of generated results. Through dynamic adjustment of SQL query statements and data authentication mechanisms, it ensures data security and compliance. Through data visualization and insight analysis modules, it provides intuitive visualization of data and the ability to generate analysis reports, thereby improving the efficiency of procurement data analysis and decision support capabilities.

The application relates to a phase unwrapping method based on interactive self-distillation and structure reparameterization, comprising the following steps: constructing a gradient segmentation model, adding an interactive self-distillation mechanism in a semantic segmentation network; adopting a Repvgg module as a basic module of the network; adopting an integrated architecture strong segmenter at the end of the network; adopting a weighted cross-entropy loss and a weighted mean square error loss as a loss function; adopting an adaptive enhancement strategy to train the gradient segmentation model on a pre-constructed data set to obtain a wrapped count gradient prediction model; inputting a current wrapped phase image into the model to output corresponding transverse and longitudinal wrapped count gradients; obtaining a wrapped count image through a least square method based on discrete cosine variation; multiplying the wrapped count by 2pi and adding the wrapped count to the wrapped phase image to obtain unwrapped phase. Compared with the prior art, the application can realize end-to-end cross-resolution phase unwrapping, and can realize high-precision and stable phase unwrapping on different resolution images.

This invention discloses a multi-model fusion method and system for permanent magnet motor fault diagnosis, belonging to the field of permanent magnet motor fault diagnosis technology. It aims to improve the flexibility and adaptability of models in multi-task learning and hierarchical classification by introducing multiple Softmax-layer CNNs and fusing them with multiple models, while effectively solving the problems of weak noise resistance and insufficient generalization ability of single models. Specifically, the technical solution of this invention first collects the vibration and current signals of the faulty permanent magnet motor, then uses Markov transfer fields to adaptively enhance the time-domain signals into images. Using a multi-Softmax-layer CNN and XGboost as base learners and an IWOA-SVM model as the meta-learner, a high-precision fault diagnosis result is finally obtained. This invention combines the advantages of multiple Softmax layers and multi-model fusion, effectively solving the multi-task fault diagnosis problem under complex working conditions, and improving the accuracy, reliability, and adaptability of fault diagnosis.

The invention provides an electromagnetic compatibility high-frequency prediction method based on an improved LSTM, and relates to the technical field of high-frequency prediction methods, and the method comprises the following steps: building a metal shielding box example model; a metal shielding box equivalent circuit is constructed based on a Robinson algorithm, and the shielding effectiveness is solved; and obtaining data of the shielding effectiveness of a plurality of electric fields, and predicting the shielding effectiveness of the metal box body by using a CNN-BiLSTM-AdaBoost method. Compared with traditional LSTM, the model is higher in accuracy when being applied to shielding effectiveness prediction of the metal cavity.

The invention discloses a deep learning enabling industrial manufacturing real-time decision-making and execution integrated system, and particularly relates to the technical field of industrial automation and intelligent control, and the system comprises an image collection and self-adaptive enhancement preprocessing module which achieves the precise synchronization of collection of a trigger signal and the rotating speed of a conveyor belt motor, dynamic illumination compensation and adaptive frame rate adjustment are introduced; the target detection and double data set construction module adopts an improved YOLOv8 algorithm, embeds a CBAM attention mechanism, and constructs a visible data set D1 and an invisible data set D2; the multi-target tracking and data optimization module is used for carrying out target matching and trajectory correction based on an improved DeepSORT algorithm, and constructing a real-time quantitative estimation model based on double data sets; and the dynamic adjustment and decision execution module is used for dynamically adjusting acquisition, algorithm and equipment parameters according to an optimization result and real-time state feedback, generating a decision instruction and issuing the decision instruction to the PLC control system through a visual terminal, so that full-process closed-loop management is realized.

This invention relates to a point cloud instance segmentation method for Torreya grandis seedlings that integrates deep learning and adaptive clustering. Addressing the challenge of instance segmentation caused by the long, dense, and heavily occluded leaves of Torreya grandis seedlings, this invention proposes an innovative framework: First, multi-view images are acquired using consumer-grade devices, and high-precision 3D reconstruction is achieved using 3D Gaussian sputtering technology. Then, a dedicated neural network, Torreya grandisSegNet, is designed for semantic segmentation. Its core includes a dual-attention kernel point convolution module and an adaptive enhanced inverse residual MLP module, enhancing the feature extraction capability for complex needle-like structures. In the instance segmentation stage, the Newton-Raphson optimization algorithm is introduced to improve DBSCAN clustering, enabling automatic search for the optimal parameter combination for initial segmentation. Finally, overlapping leaves are processed through topological skeleton analysis to complete refined instance separation. This method significantly improves the accuracy and adaptability of point cloud segmentation for coniferous plants, providing reliable technical support for the selection of superior Torreya grandis varieties.