2391 results about "Gradient descent" patented technology

The invention relates to an unmanned aerial vehicle intelligent dynamic path guidance method and system based on deep reinforcement learning. The method comprises the steps of obtaining three-dimensional terrain point cloud and obstacle information of a flight area, constructing an environment topological structure diagram, extracting features through a three-layer diagram convolutional network, and generating high-dimensional environment feature representation. And a deep reinforcement learning state space is constructed, and a multi-target reward function is designed. And initializing the deep reinforcement learning network by using [-0.1, 0.1] uniformly distributed random parameters to generate an initial path scheme. Optimizing through a dynamic planning algorithm, adjusting a node sequence according to a path length and a task dynamic threshold value, and calculating a target function value. And judging whether the current path scheme reaches the optimal balance point, if not, adjusting the weight of the reward function by using a gradient descent method, and regenerating the path scheme until the path scheme is optimal. And finally, converting the scheme into a waypoint coordinate and speed instruction, performing real-time monitoring during flight, and triggering re-planning when an obstacle exceeds a safety threshold, so as to achieve the purposes of safe flight of the unmanned aerial vehicle in a complex environment and the like.

The invention relates to the technical field of agricultural information perception and decision making, in particular to a multi-source perception fused agricultural monitoring method and system. The method comprises the following steps: converting a multi-source heterogeneous agricultural sensing signal into a space-time tensor, constructing a semantic resonance field to simulate nonlinear coupling between modals, and driving multi-modal data to adaptively aggregate by using a gravitational evolution mechanism to form a fused semantic field; calculating non-linear response to generate an agricultural state emergence index, and according to the index, identifying a potential risk area and constructing a binary risk map; for the risk area, semantic disturbance is mapped into an agricultural variable disturbance vector through a modal decoupling matrix, a minimum intervention strategy is generated in combination with sparse optimization of an operation response matrix, and feasible operation suggestions are output after verification of an agricultural knowledge graph; a drift potential energy function is constructed based on strategy execution feedback, strategy parameters are dynamically updated through gradient descent, and closed-loop self-evolution optimization is achieved in combination with trend prediction. According to the invention, full-link adaptive optimization from multi-source sensing to regulation and control decision is realized.

The invention relates to a task scheduling optimization method and device based on reinforcement learning, equipment and a medium. The method comprises the steps that firstly, system resource state data are collected in real time, dynamic environment characteristics are determined through preprocessing and time sequence analysis, task characteristic data are analyzed at the same time, and a task priority sequence and a resource demand vector are generated through a priority ranking algorithm and a resource evaluation model; and then a state space and an action space are constructed by adopting a reinforcement learning algorithm, an optimal task allocation scheme is generated through strategy iteration and reward function optimization, and if the scheme meets a resource balance threshold, scheduling is executed, and performance indexes are collected. And finally, fusing real-time indexes with historical data, and updating parameters of the reinforcement learning model through experience playback and gradient descent to form a closed-loop optimized improved scheduling strategy. By adopting the method, the accurate mapping of the resource state and the task requirement can be realized, and the problem of insufficient adaptability of the traditional static scheduling to a complex scene is solved.

The invention discloses a solar radiation space-time prediction method and system based on physical information constraint and a neural network, and the method comprises the steps: collecting multi-dimensional time sequence meteorological data, extracting high-dimensional time sequence dynamic features, converting geographic space data into a fuzzy set, and carrying out the defuzzification of the fuzzy set through an inference rule, thereby obtaining geographic space features; and a gating mechanism is adopted to realize deep fusion of the space-time features to generate high-dimensional space-time fusion features. In a model training stage, an energy conservation equation is introduced into an optimization process, a physical residual error is constructed by calculating a time derivative and a space derivative of a predicted value, a physical constraint total loss function is formed in combination with a data loss item, and model parameters are updated by using a gradient descent method. According to the method, the accuracy and reliability of a prediction result are remarkably improved while the calculation efficiency is ensured, and the method is particularly suitable for solar radiation prediction under complex meteorological conditions; according to the method, abnormal prediction caused by data noise can be effectively corrected, and a solution with physical rationality and data adaptability is provided for the fields of solar resource evaluation, photovoltaic power generation power prediction and the like.

The present invention provides a privacy-enhanced structured data simulation and generation method and system. The method comprises: step 1, a data conversion phase: performing normalization preprocessing on data; step 2, a probabilistic graphical model construction phase: on the basis of a Bayesian form, constructing the posterior distribution for variational inference from the data having undergone normalization preprocessing in step 1, obtaining an association relationship between the features of structured data by using a Stein variational gradient descent method, and when introducing differential privacy noise, using a Monte Carlo estimation algorithm to automatically obtain the amount of noise that needs to be added for each update step; and step 3: a data generation phase: using the association relationship obtained in step 2 as a metric set to generate more accurate simulation data than real data. The beneficial effect of the present invention is that: the method of the present invention avoids gradient clipping when applying DP-SGD, thereby avoiding the selection of clipping parameters and reducing the adverse effects of gradient clipping on an inference process.

The invention discloses a multi-model-based intermittent reaction kettle intelligent control method and system. The method comprises the steps of collecting operation data such as temperature, pressure and reactant concentration of a reaction kettle in real time through a sensor, constructing a soft measurement model by utilizing a machine learning algorithm, and performing training based on historical data so as to predict key operation parameters; when operation conditions change, the system updates model parameters in real time through an online learning model by using a recursive least square method or an online gradient descent method; besides, based on a reinforcement learning algorithm, the system can generate and optimize a control strategy, adjust operation parameters such as the opening degree of a cooling water valve and the power of a heater, and finally execute an optimal control instruction through an actual controller to realize high-precision control of the reaction kettle; according to the method, the control precision, the system response speed and the adaptive capacity are improved, the energy utilization is optimized, and the production efficiency is improved.

The invention relates to the technical field of mobile energy storage vehicle energy management, and discloses a mobile energy storage vehicle energy management method based on an intelligent algorithm, and the method comprises the steps: collecting the charging and discharging rate of a battery pack, the environment temperature, the power grid load fluctuation and other multi-dimensional energy state data; a heterogeneous layered architecture of edge computing nodes, a cloud collaboration platform and a vehicle-mounted control terminal is constructed, energy partitions are divided according to peak valleys of a power grid, and low-delay response, global optimization and local closed-loop control modes are configured; fusing the multi-source heterogeneous energy data streams and removing abnormal values; a dynamic optimization decision model is constructed through a battery degradation model and a power grid supply and demand balance equation, and a multi-parameter collaborative constraint relation is solved through simultaneous solving of a multi-target iteration solver and a parallel gradient descent algorithm; and generating a three-level adaptive response instruction sequence of local battery overload alarm, regional power grid frequency modulation early warning and global energy scheduling imbalance pre-judgment based on constraint boundary triggering conditions. According to the method, the accuracy, the collaboration and the robustness of energy management are improved.

The invention provides a large model scientific and technological problem generation method fusing external knowledge, which comprises the following steps: dynamically solving an optimal knowledge fusion strength parameter in a current environment by adopting a gradient descent algorithm through designing and optimizing an objective function according to association representation and difference representation between knowledge units, the knowledge unit balancing module is used for balancing correlation and complementarity among knowledge units; on the basis of real-time updating of knowledge fusion strength parameters, heuristic search is carried out in the knowledge unit association map by utilizing an ant colony optimization algorithm, and candidate fusion knowledge units which achieve target correlation and target complementarity with the current knowledge unit are selected by combining semantic similarity, association degree and difference among the knowledge units; a knowledge mapping ontology is used for guiding fusion of cross-domain knowledge, an ontology inference engine based on description logic is adopted to infer a semantic consistency relationship between fused knowledge units, and potential semantic conflicts are identified and eliminated through a consistency check algorithm.

The invention belongs to the technical field of motor control, and discloses a control method of a motor for an inverter compressor, which comprises the following steps: embedding a distributed thermocouple array in a motor stator slot to acquire temperature data of the motor, synchronously acquiring operation parameters of the motor, generating a dynamic thermal field characteristic spectrum through a data fusion algorithm, coordinates of the hot spot areas with the temperature abnormally increased are marked, a dynamic correlation model is established, and heat conduction characteristic parameters in the dynamic correlation model are updated through a sliding window algorithm; outputting a predicted temperature rise curve through the dynamic correlation model, adjusting a PWM control strategy of the inverter, inputting the PWM control strategy to a motor driving unit to execute preliminary regulation and control, collecting feedback data to generate a multi-dimensional feedback feature vector, iteratively optimizing a combined weight of a switch interval compensation amount and a harmonic injection strategy through a gradient descent algorithm, and performing harmonic injection control. And a global optimal strategy considering the heat dissipation efficiency and the operation stability is formed, and the operation stability and the energy efficiency level of the inverter compressor under various working conditions are remarkably improved.

The invention provides a multi-modal heterogeneous model retrieval enhancement method and system, and the method comprises the steps: building a knowledge and application example double-corpus based on user multi-modal query, and designing a joint retrieval mechanism to obtain a result set; mapping and scheduling to obtain feature representation through special processing channels for texts, images and audios and a Spiking neural network with a segmented trapezoidal topological structure; constructing a three-stage cascade architecture of a basic model, an advanced model and human experts, and obtaining a decision path and answer candidate set in combination with a recursive and discarding decision mechanism; a Hamiltonian graph network is used for representing a multi-modal relation, and a gradient-free descent method is used for rapidly training and optimizing model parameters; an enhanced retrieval result is obtained through cross-modal semantic alignment and dynamic retrieval window adjustment; and high-quality response is obtained through context-aware sorting and retrieval enhanced reasoning. According to the method, the multi-modal information retrieval processing efficiency and the heterogeneous model reasoning response quality are improved.

The invention discloses a virtual machine scheduling method in a distributed environment based on deep reinforcement learning, and belongs to the technical field of cloud computing resource scheduling. According to the method, the defects of a traditional method in multi-objective optimization and mixed action space collaborative decision-making are overcome by constructing a mixed action space joint decision-making mechanism. The method specifically comprises the following steps: establishing a mixed action space containing discrete node selection and continuous resource allocation, filtering invalid nodes by adopting a dynamic mask mechanism, and ensuring resource ratio constraint through projection gradient descent; designing a hierarchical reward function to realize multi-target dynamic balancing, and dynamically adjusting the priorities of energy consumption, load balancing and SLA guarantee based on an adaptive weight strategy; a multi-agent collaborative framework is provided, cross-node topological dependence is captured by using a graph attention network, and dynamic fusion of spatio-temporal characteristics is realized through cross attention in combination with LSTM coding time sequence load characteristics; a course learning strategy and a priority experience playback mechanism are introduced to improve training efficiency and strategy robustness.

The invention provides a high-density RFID tag group scanning system based on a multi-antenna array and beam forming, and relates to the technical field of the Internet of Things. The system comprises a multi-antenna array, an RFID reader-writer, a beam forming unit, a signal processing unit, a feedback unit and a data management unit. Through a digital beam forming technology, the system dynamically adjusts the direction and width of a beam, and achieves the precise scanning of a target area. And in combination with real-time signal feedback, a density peak clustering algorithm and a gradient descent algorithm are adopted to optimize beam parameters, so that the recognition efficiency is improved. The system further utilizes an anti-collision algorithm to dynamically adjust frame parameters, and performs logic grouping based on spatial positions and type characteristics of tags, so that signal collision and missed reading rates are remarkably reduced. Meanwhile, the data management unit stores and outputs the label data in a classified mode, and intelligent management capacity is provided for multi-scene application.

The invention relates to the technical field of image processing, and discloses an image processing method and device for an image acquisition card, and the method comprises the steps: converting an input optical signal into original image data through a sensor array of the image acquisition card, and carrying out the preprocessing, and obtaining a preprocessed image sequence; constructing an image state space model and a color measurement noise model based on the preprocessed image sequence, and performing preliminary denoising on the preprocessed image sequence to obtain a bidirectional time sequence feature tensor; inputting the bidirectional time sequence feature tensor into an additive fusion bidirectional time convolutional neural network for feature extraction to obtain a comprehensive spatial-temporal feature and a confidence map; according to the method, gradient descent image processing parameter optimization is carried out according to the comprehensive spatial-temporal characteristics and the confidence map, optimal processing parameters of the image acquisition card are obtained, and a target output image is generated, and the method improves the image definition and the detail retention capability in a dynamic scene, and improves the robustness of image processing.

The invention relates to the technical field of medical information, in particular to an intelligent prediction model and method for postoperative complications of anesthetized patients, and the method comprises the steps: collecting preoperative to postoperative complete-cycle clinical data of a patient through a medical data interface; analyzing operation codes to generate risk features, extracting vital sign dynamic features, and establishing a complication probability mapping relation through a multi-modal fusion network; combining the complication probability and pharmacokinetic parameters to construct an optimization model, and solving an individualized anesthetic dosage interval by using a gradient descent algorithm; vital signs are dynamically monitored in the operation, a dose re-optimization mechanism is triggered, the infusion rate is adjusted, and a closed-loop control link is formed; and generating a visual decision report. According to the method, through deep integration of complete-cycle clinical data and multi-modal feature modeling, preoperative physiological parameters, operation coding semantic information and intraoperative vital sign dynamic modes are subjected to fusion analysis, a nonlinear mapping relation between dosage and complication probability is constructed, and the risk prediction precision and individualized adaptability are remarkably improved.

The invention discloses a vehicle owner behavior data driving type vehicle insurance risk assessment method based on a large language model, and relates to the technical field of risk assessment. Comprehensive vehicle owner and vehicle behavior data are collected from multiple channels, and deep analysis is performed on the data through the large language model; driving behavior features, historical accident record features and vehicle use condition features highly related to the vehicle insurance risk are extracted, and the influence of various factors on the risk is accurately reflected in combination with a nonlinear conversion function; key features are converted into numerical vectors and stored in the form of a multi-dimensional feature matrix, and high-quality input data are provided for a deep learning model; a risk assessment model constructed by adopting a PyTorch framework is combined with loss function optimization and a gradient descent algorithm, so that efficient model training and prediction are realized; and the risk assessment result is adjusted in real time through dynamic data updating, so that the accuracy of vehicle insurance risk assessment and the response speed to the behavior change of the vehicle owner are remarkably improved.

The invention discloses a multi-mode reinforcement learning driven PM2.5 chemical component vertical profile inversion system and method. The system comprises a deep reinforcement learning model; the deep reinforcement learning model comprises a state space, an Action action space, a reward function, an Actor network and a Critic network; the deep reinforcement learning model is used for calculating a reward value at the current moment in combination with a reward function according to the difference between a vertical profile predicted value of the concentration of each PM2.5 chemical component output by the Action action space and an actual monitoring value; and according to an evaluation result of the strategy evaluation, continuously performing strategy iteration by using a gradient descent method until the deep reinforcement learning model is converged, and obtaining an optimal target deep reinforcement learning model. An Actor-Critic network dynamic optimization strategy is adopted, minute-level model updating is achieved in combination with edge calculation, and the problems of high-altitude blind areas, insufficient real-time performance and the like of a traditional model are solved.

The invention discloses an intelligent metasurface channel optimization method and system based on a beam forming algorithm, and relates to the technical field of intelligent metasurfaces, and the method comprises the steps: collecting real-time environment data, monitoring channel state data, and evaluating the influence of the environment and the channel state on a signal; evaluating the gain strategy of the intelligent metasurface, and evaluating the performance of the current beam forming strategy by combining the influence of the environment and the channel state on the signal; and calculating an optimal beam forming parameter by using a gradient descent method to realize channel optimization. According to the method, through comprehensive utilization of real-time environment monitoring, channel state evaluation, an advanced data analysis technology and a mathematical optimization method, dynamic adjustment and real-time optimization of intelligent metasurface channel optimization are achieved, efficient and stable operation of a communication system in a complex environment is ensured, the speed and reliability of data transmission are remarkably improved, and the method is suitable for popularization and application. The invention provides a more efficient, flexible and adaptive beam forming strategy optimization method, and the performance of a wireless communication system is remarkably improved.

The invention discloses a multi-intersection signal cooperative control method based on deep reinforcement learning, and the method comprises the steps: firstly constructing a comprehensive reward function which comprises waiting time, queuing length and the influence of adjacent intersections, and constructing a mathematical model of a multi-intersection signal control system based on the reward function, so as to achieve the maximization of a comprehensive reward; and secondly, designing a deep Q network cooperative control algorithm for explicit strategy adjustment, and optimizing and adjusting adjustable strategy function parameters by introducing an explicit strategy adjustment function and utilizing a stochastic gradient descent method. And finally, dynamically adjusting the signal control strategy of each intersection in the road network in a complex and changeable traffic environment by designing a parameter sharing mechanism, thereby realizing collaborative optimization control of multiple intersections. According to the invention, the cooperative control efficiency of a multi-intersection signal system is obviously improved, traffic delay is reduced, and the traffic capacity of a road network is improved.

The invention provides a sewage treatment plant terminal effluent carbon source quality detection and evaluation method, which realizes full-flow dynamic evaluation and regulation of carbon source quality through on-line monitoring and intelligent algorithm fusion. According to the method, an online water quality full-spectrum detector is used for collecting original spectrum data flow, and after preprocessing such as variational mode decomposition denoising and mutual information feature selection, organic matter content quantification, variation trend analysis and anomaly detection are completed in combination with algorithms such as a support vector machine and an autoregressive moving average model. An entropy weight method is introduced to dynamically adjust the weight of the evaluation model, model parameters are optimized based on a gradient descent algorithm, a process adjustment instruction is generated through reinforcement learning and fuzzy logic, and an automatic system is linked to execute regulation and control. According to the method, the problems of hysteresis and singleness of traditional offline analysis are solved, multi-dimensional real-time evaluation, abnormal quick response and process dynamic optimization of the quality of the carbon source are realized, the sewage treatment efficiency and the effluent quality stability are improved, and a technical support is provided for continuous standard reaching of the quality of the carbon source.

The invention relates to a true orthophoto generation method based on a three-dimensional Gaussian model, a storage medium and equipment, and aims to solve the problems of low image quality, low generation speed, detail loss and the like in the prior art. The method comprises the following steps: firstly, acquiring image data through an unmanned aerial vehicle or aerial photography, and extracting a camera attitude and sparse point cloud by using a motion recovery structure (SfM) technology; secondly, constructing a three-dimensional Gaussian model based on the sparse point cloud, and performing iterative training through top view orthographic projection in combination with a gradient descent algorithm; in the training process, a densification strategy, a point deleting strategy, a blocking strategy and an image pyramid strategy are innovatively introduced, so that the detail expressive force and the overall quality of the image are remarkably improved. Specifically, according to the densification strategy, fine detail reconstruction is achieved by dynamically increasing Gaussian ball density, and according to the point deletion strategy, rendering efficiency is improved and computing resource allocation is optimized by eliminating redundant Gaussian balls. The image pyramid strategy generates a multi-level visual effect through multi-scale training, and the blocking strategy improves the reconstruction precision through local optimization. Finally, on the basis of the trained three-dimensional Gaussian model, real-time generation of a high-quality true orthophoto in a large-scale scene can be realized. The method has remarkable advantages in the aspects of efficiency, precision and practicability, provides important technical support for the fields of geographic information systems, urban planning, disaster monitoring and the like, and has wide application prospects.

The invention discloses a multi-source-domain multi-teacher knowledge distillation method and system based on reinforcement learning, and the method comprises the steps: obtaining target domain sample data, inputting the data into N pre-trained teacher models, and generating the output features of all teacher models; inputting the target domain sample and all teacher model outputs into a reinforcement learning strategy network, generating a dynamic weight of each teacher model, and calculating a knowledge distillation loss function based on the dynamic weights; constructing a total loss function according to the knowledge distillation loss function and the cross entropy loss output by the student model; student model parameters are updated through gradient descent; and calculating a reward value according to student model performance change, and updating reinforcement learning strategy network parameters. According to the method, the reward function based on student model performance improvement is constructed, the strategy network is continuously updated in a strategy gradient optimization mode, the distillation efficiency is effectively improved, knowledge conflicts among teachers are relieved, and the robustness and generalization performance of the student model in a multi-source complex environment are remarkably improved.

The invention relates to the technical field of battery charging and discharging optimization, and discloses a charging and discharging optimization method, device and equipment for a robot energy storage battery, and the method comprises the steps: collecting a working condition feature data set of the robot energy storage battery; performing state cooperative calculation on the working condition characteristic data set to obtain battery operation state data; establishing a mapping relation model of load power and charge-discharge efficiency, and determining a battery working interval boundary value to obtain charge-discharge characteristic data; inputting the charge and discharge characteristic data into a deep gating cycle unit network, and obtaining a battery performance prediction model through dynamic modeling; based on a battery performance prediction model, a gain factor self-tuning technology is adopted to construct a closed-loop controller, online optimization calculation is carried out on control parameters through a gradient descent algorithm, a control instruction containing a charging and discharging current limit value and a voltage threshold value is generated, and then accurate prediction of the battery performance is achieved. And stable operation of the charging and discharging control system is ensured.

The invention discloses a joint test system, a silicon carbide metasurface grating detection method and a silicon carbide metasurface grating detection device, and belongs to the technical field of semiconductor detection. A synergistic excitation signal is generated by regulating a light source and a microwave source and is applied to the silicon carbide metasurface grating to be measured to form a light intensity and microwave field interference pattern. The method comprises the following steps: synchronously acquiring data through multiple channels, constructing a multi-dimensional matrix through time alignment, marker feature extraction and coordinate mapping, extracting defect features to establish a reference model, calculating deviation degree to judge quality, and carrying out abnormal point screening, candidate region construction and coupling verification on a defect grating to divide a potential defect region. For a potential defect area, excitation parameters are dynamically adjusted through a gradient descent algorithm, and defect grades are divided by combining the characteristic defect degree and the area. According to the invention, the problems of insufficient single-mode detection, low multi-mode alignment precision and blind adjustment of excitation parameters are solved, cross-band cooperative detection and intelligent grading are realized, and the detection reliability is improved.

The invention discloses a fault model construction method for adaptive testing of abnormal values of a system-on-chip. The method relates to the technical field of fault construction of chip testing, and comprises the following steps: testing chips on a large scale under the conditions of different temperatures, voltages, frequencies, loads and the like, acquiring performance parameters such as signal delay and power consumption, screening abnormal values by applying a Z-score or box plot method, and processing by using a sliding window; extracting features through CNN or RNN, and normalizing the features; fault features are classified through K-Means or hierarchical clustering, and the accuracy rate is larger than or equal to 80%; a model is built through an SVM or a decision tree, faults can be predicted, self-updating can be achieved, self-adaptive adjustment can be achieved according to chip parameter changes, independent data verification optimization, one-out-of-leave evaluation and gradient descent optimization are achieved. According to the method, multi-condition testing can be carried out on the system-level chip, abnormal values are accurately screened, features are extracted, faults are classified, a fault model capable of being self-updated is constructed, self-adaptive adjustment and verification optimization are achieved, the accuracy is improved, and the chip testing and fault diagnosis efficiency and accuracy are improved.

The invention relates to the technical field of battery charging and discharging optimization, and discloses an energy management system data analysis method, device and equipment and a storage medium, and the method comprises the steps: collecting a working condition feature data set of a robot energy storage battery through an energy management system; performing state cooperative calculation on the working condition characteristic data set to obtain battery operation state data; establishing a mapping relation model of load power and charge-discharge efficiency, and determining a battery working interval boundary value to obtain charge-discharge characteristic data; inputting the charge and discharge characteristic data into a deep gating cycle unit network, and obtaining a battery performance prediction model through dynamic modeling; based on a battery performance prediction model, a gain factor self-tuning technology is adopted to construct a closed-loop controller, online optimization calculation is carried out on control parameters through a gradient descent algorithm, a control instruction containing a charging and discharging current limit value and a voltage threshold value is generated, accurate prediction of battery performance is further achieved, and stable operation of an energy management system is ensured.

The invention relates to a data-driven lithium ion battery voltage prediction modeling method fusing physical information, and belongs to the technical field of batteries. According to the method, a multi-working-condition battery operation database is constructed, a terminal voltage prediction model fusing a physical model and a neural network is designed, a mathematical constraint relation is established for equivalent circuit parameters and neural network intermediate variables, and data driving characteristics and physical parameters are jointly optimized by using a gradient descent method. The method specifically comprises the steps of collecting battery current, voltage and temperature data; constructing an input feature and an output tag; outputting an intermediate variable through a parameter layer and associating an equivalent circuit parameter; and introducing physical loss function constraint model training to ensure the physical rationality of parameters. The method has high precision of data driving and interpretability of a physical model, improves multi-working-condition adaptability through temperature compensation and parameter constraint, and can be widely applied to battery state estimation and health management of electric automobiles and energy storage systems.

The invention provides a photocuring material performance prediction method based on deep learning, and relates to the technical field of photocuring 3D printing, and the method comprises the steps: collecting the multi-modal data of a photocuring material in real time; performing data cleaning, de-noising and standardization processing on the acquired multi-modal data, fusing the processed multi-modal data, and extracting key index values and printing process parameters of a material formula; constructing an initial deep learning neural network model by using key index values in the extracted multi-modal data; printing process parameters are used for training the initial neural network model, gradient descent is adopted for optimizing the model weight, performance evaluation and optimization are conducted on the model through cross validation, and a material performance prediction model is obtained after verification; and predicting the performance of a to-be-predicted photocuring material based on the material performance prediction model. By comprehensively analyzing the formula, illumination conditions and printing parameters of the photocuring material, the performance of the material under different conditions is accurately predicted.

The invention discloses an embryo growth stage prediction method and system based on self-supervised learning, and the method comprises the following steps: collecting embryo images at different growth stages as a training set, and segmenting each embryo image in the training set into a plurality of blocks with the same size; randomly selecting half of the blocks from each embryo image for mask processing to obtain a mask training set; a prediction model comprising an encoder and a decoder is constructed, the encoder extracts a feature sequence from the mask training set, and the decoder reconstructs an embryo image according to the feature sequence; the kernel functions of the multiple weighted combinations serve as loss functions of the prediction model, the weight of each kernel function is adjusted through cross validation and gradient descent, parameters of the prediction model are adjusted and optimized through back propagation, and an optimal prediction model is obtained; and inputting a to-be-predicted embryo image into the optimal prediction model to obtain a growth stage prediction result of the embryo.

Systems and methods for optimizing large language models (LLM) with domain-oriented model compression. Importance weights for general knowledge in a trained LLM, pretrained with deep learning, can be determined by computing the error when removing a weight from the trained LLM. The trained LLM can be iteratively optimized to obtain a domain-compressed LLM with domain knowledge while maintaining general knowledge by: fine-tuning the trained LLM iteratively with domain knowledge using the importance weights for general knowledge to obtain a fine-tuned LLM; determining importance weights for domain knowledge in the LLM with a regularization term by using gradient descent to optimize parameters when the fine-tuned LLM is trained with domain knowledge; and pruning learned knowledge based on importance weights for domain knowledge. A corrective action can be performed on a monitored entity using the domain-compressed LLM.

The invention discloses a tensor depth semi-supervised learning method for high-dimensional small sample data classification. The tensor depth semi-supervised learning method comprises the steps of preprocessing original high-dimensional small sample data; constructing a deep neural network comprising a feature extraction module and a classifier module; constructing a second-order similarity matrix and a third-order similarity tensor based on the low-dimensional embedding representation obtained by pre-training; combining the second-order similarity matrix and the third-order similarity tensor to construct an objective function containing multi-order smooth constraints; then, taking the low-dimensional embedded representation obtained by pre-training as input, performing iterative optimization on the target function by adopting a gradient descent algorithm through a label propagation network formed by a full connection layer, and generating a pseudo label; and inputting the original high-dimensional small sample data, the low-dimensional embedded features and the pseudo labels into the deep neural network, iteratively updating the network in a semi-supervised mode until convergence, and outputting a final prediction result. According to the method, more accurate label propagation is realized, and the semi-supervised classification precision is improved.