1246 results about "Model learning" patented technology

The invention provides a multi-modal sentiment analysis method based on knowledge distillation and a dynamic fusion mechanism. The multi-modal sentiment analysis method comprises the following steps: pre-training a single-modal teacher model; a single-mode teacher model is used for guiding the learning of a multi-mode student model, and through an interactive knowledge distillation mechanism, the middle layer probability distribution of the teacher model is used as a target to learn the correlation between modes and the cross-mode characteristics; interactive knowledge distillation comprises three loss functions: firstly, calculating the difference of output distribution of output layers of a teacher model and a student model, and defining the difference as cross-modal knowledge distillation loss; secondly, adding alignment loss based on a real label, and constraining a prediction result of the student model to be close to a real emotion label in a cross entropy form; and finally, introducing label smoothing loss to soften the real label. According to the method, the pre-trained single-mode teacher model with relatively good performance is stored and is used for guiding the learning of the multi-mode student model, meanwhile, the loss function is introduced to optimize the multi-mode student model, so that the multi-mode student model is gradually aligned with the output distribution of the teacher model, and meanwhile, the adaptive capacity of the multi-mode student model to the modal heterogeneity is enhanced. The single-mode teacher model greatly reduces the complexity of the model, reduces the calculation amount, and has better performance in the field of multi-mode sentiment analysis.

Disclosed is an industrial question answering model training method based on reinforcement learning and knowledge base matching, comprising the following steps: S1, collecting professional knowledge questions and answers in an industrial field to construct an industrial knowledge base, training a reward model, carrying out, for industrial knowledge questions and answers, matching comparison on outputs of an industrial question answering model and content of the industrial knowledge base, and obtaining reward values on the basis of similarities; S2, sorting the reward values, and using a sorting loss function to train and update parameters of a reward model network; and S3, carrying out industrial question answering model training, incorporating a penalty term for the reward values, and using a reinforcement learning algorithm to train the industrial question answering model multiple times to obtain an optimal strategy. According to the industrial question answering model training method based on reinforcement learning and knowledge base matching of the present invention, the reinforcement learning algorithm is used, and iterative training is carried out multiple times, thereby helping the industrial question answering model to learn and understand industrial professional knowledge and improving the question answering accuracy of the industrial question answering model.

The invention relates to the technical field of motor control, and discloses a brushless motor adaptive control method and system based on artificial intelligence, and the method comprises the steps: collecting the multi-point temperature data of a motor, building a thermal dynamic model, and reconstructing the complete temperature field distribution; identifying a relation function between motor parameters and temperature, and establishing a temperature sensitive parameter model; in combination with temperature field distribution and a temperature sensitive parameter model, learning long-term influence of control actions on temperature distribution, and predicting temperature evolution trends under different control strategies; constructing a multi-objective optimization control strategy, and generating a motor control instruction and a heat dissipation control instruction; dynamically selecting an optimal heat dissipation strategy combination according to the generated control instruction, the current working condition and the predicted heat influence; according to the invention, high-precision control and thermal management optimization of the brushless motor in a temperature change environment are realized.

The invention discloses a multi-modal ship target individual identification method and system, and relates to the technical field of ship target identification, and the method comprises the steps: obtaining multi-modal data of a target region, and constructing a training set and a test set; each training sample in the training set is input into a target individual prediction model, a multi-granularity adaptive loss function is adopted as a loss function to carry out model training, and the target individual prediction model adopts an image feature extraction method based on cross-modal guide enhancement to carry out feature extraction and splicing fusion on each training sample; and a target individual identification result of each training sample is obtained by adopting a visual angle self-adaptive fusion method of embedded type label information. And after training is completed, a trained target individual prediction model is obtained, each test sample in the test set is input into the trained target individual prediction model, and a final target individual identification result is obtained. According to the invention, the generalization ability of model learning individual features can be improved, and the accuracy of ship target individual identification is improved.

The invention discloses a text retrieval enhancement generation method and device, a medium and equipment, and relates to the technical field of retrieval enhancement generation. Comprising the following steps: establishing a database through vectorized corpora, and screening and labeling key information to generate a golden section vector library; extracting an adversarial sample from an external database by using a training problem, and fusing the adversarial sample with related corpora in a golden section vector library to obtain a training set; and adding a noise classification layer to improve the language large model, and training the improved language large model by using the fusion training set. When a user question is processed, relevant corpora are recalled, noise is recognized through the classification layer, context representation is generated through the attention fusion layer, and the output layer dynamically adjusts a decoding strategy and generates an answer. According to the method, a large language model added with a noise classification task is trained through the fusion training set, so that the model can identify and distinguish different types of noise, the adaptability of the model to different types of noise is enhanced, the model is helped to learn how to distinguish related and unrelated information, and thus the method is more robust when an actual problem is processed.

The invention discloses a log abnormal behavior detection method and system based on periodic pattern mining and incremental learning, and relates to the technical field of system abnormality monitoring. The method comprises the steps of obtaining original log data and executing preprocessing operation; main periodic frequency components are extracted through time frequency analysis to form a periodic set, a periodic stable part sequence and a transient change part sequence are divided, and a periodic modeling mechanism and a transient modeling mechanism are used for modeling; a joint prediction model is constructed, Monte Carlo Dropout is introduced to estimate uncertainty, and Bayesian weighting is adopted to generate a prediction value; abnormity is judged through a periodic residual error, a transient residual error and an overall residual error, and an abnormal causal path is analyzed and identified in combination with transfer entropy; a memory sample driven playback and distillation mechanism is adopted to execute incremental training, and modeling structure parameters are dynamically adjusted. The method has the capabilities of periodic rule modeling, unsteady behavior expression, prediction fusion, abnormal causal identification and continuous model learning.

The invention discloses a wire drawing machine operation state evaluation method and system based on deep learning, and relates to the technical field of intelligent manufacturing, and the method comprises the steps: deploying a multi-source sensor at a preset part of a wire drawing machine, and forming a multi-dimensional operation data flow; the deep learning evaluation model is trained based on the standardized time series data set, model parameters are optimized through a cross entropy loss function, the model learns the characteristic difference between a normal working condition and an abnormal working condition, and the deep learning evaluation model is trained based on the standardized time series data set. And inputting a multi-dimensional operation data flow collected in real time into the trained deep learning evaluation model to generate a dynamic evaluation report. According to the wire drawing machine operation state evaluation method and system based on deep learning provided by the invention, a visual operation state score can be output, detailed fault type probability distribution is given, and intelligent support is provided for equipment maintenance decision.

The invention discloses a complex exponential signal joint spectrum reconstruction and parameter estimation method and device, and relates to the field of signal processing, and the method comprises the steps: S1, constructing noise-containing complex exponential signal training data and label data; s2, constructing a dual-module neural network model comprising a super-resolution denoising reconstruction module and a parameter prediction module; s3, training the dual-module neural network model by using the training data and the annotation data to obtain a trained dual-module neural network model; and S4, performing frequency spectrum reconstruction and parameter prediction by using the trained dual-module neural network model, and performing signal post-processing on the output to obtain estimation parameters of the angular frequency, the attenuation factor, the real part amplitude and the imaginary part amplitude. According to the invention, a cascade neural network architecture of a super-resolution denoising reconstruction module and a parameter prediction module is designed, and angular frequency detection is converted into a Gaussian distribution heat map regression task; meanwhile, a sparse activation labeling mechanism is adopted, the parameter truth value is only reserved at the spectrum peak position, and the model learning complexity is remarkably reduced.

The invention relates to the technical field of rehabilitation training detection, in particular to a rehabilitation training detection method and system based on artificial intelligence, a standard action library is constructed through standard action videos shot at multiple angles, and track and angle features of key joints are extracted for user training comparison; the skeleton key points of the user are extracted in real time through a MoveNet network, and efficient posture recognition in a home scene is achieved; analyzing position difference, angle change and acceleration characteristics by combining a space-time sequence matching algorithm, generating a dynamic matching degree index, and positioning a deviation joint to generate a correction prompt; introducing an attention mechanism model, learning the contribution degree of each joint to cycle recognition, dynamically selecting a dominant joint for action counting, recognizing starting and ending points of an action cycle through an acceleration curve, and finishing effective action statistics in combination with a dynamic threshold value, so that the counting accuracy and the self-adaptive capability are improved; therefore, the training cost is reduced, the evaluation credibility is enhanced, and accurate statistics and analysis of rehabilitation training data are realized.

The invention relates to the technical field of electric bicycles, and discloses a data-driven electric bicycle battery fire hazard detection and protection method, which comprises the following steps: S1, multi-modal data acquisition and fusion, S2, data preprocessing and feature extraction, S3, data-driven intelligent detection: classifying normal and abnormal states through a supervised learning model, S4, hierarchical protection strategy, and S5, data-driven intelligent detection. S5, cloud data optimization and model continuous learning; S6, model data updating; S7, intelligent linkage and long-term optimization; S8, cloud risk data acquisition; and low, middle and high three-level protection mechanisms are adopted. Through dynamic risk assessment, charging and discharging power is limited in a medium risk, a charger is automatically disconnected, a user is prompted to stop operation, active protection is triggered in a high risk, and sound-light alarm and remote notification are triggered at the same time, so that accidents are prevented from further spreading.

A federated object detection learning method based on representation enhancement and weighted aggregation under cloud-edge-terminal environment comprises the steps of: 1) building a centralized federated learning framework under cloud-edge-terminal environment; 2) locally conducting representation enhancement training to strengthen model learning for few-shot category after receiving a model from the server at the client; 3) carrying out the weighted aggregation for client models in accordance with sample distribution to obtain the global model after receiving models from all clients at the server. With regard to the problem of existing federated object detection learning on low global model accuracy and weak generalization ability, the present invention can improve the accuracy and generalization ability of global object detection model.

The invention discloses a model optimization method and system based on knowledge distillation and model pruning, and belongs to the technical field of electric power operation and maintenance. The model optimization method comprises the following steps: segmenting and predicting a large amount of unlabeled power operation and maintenance data by using a pre-trained teacher model, generating a preliminary segmentation mask, and generating a domain-specific pseudo tag for the segmentation mask; training a lightweight student model based on a knowledge distillation method by taking the pseudo tag as a supervision signal, and enabling the lightweight student model to learn soft distribution of the teacher model; and adopting a modularized pruning strategy based on sparse training to prune the lightweight student model, and adjusting model parameters based on an annotation data set in the power field to obtain an optimized lightweight student model. The method is used for operation maintenance of power grid equipment, and can reduce the defect omission ratio and reduce the inspection frequency.

The invention relates to the technical field of artificial intelligence, in particular to a large model training system and method based on multi-modal medical data fusion, and the method comprises the steps: collecting multi-modal medical data, and carrying out the feature extraction and pre-calculation of the multi-modal medical data; performing fusion analysis on the information distinguishing value evaluation of the sample features and the label association strength to obtain a context adaptability regulation factor of the sample data; performing modal consistency analysis and modal significance analysis on the multi-modal medical data sample to obtain an information value deepening regulatory factor of the sample data; performing optimization adjustment on the original attention score through a context adaptability adjustment factor and an information value deepening adjustment factor to obtain a final attention score; and multi-modal data feature fusion and model training are performed through the final attention score, so that the accuracy of learning medical knowledge by the model is improved, and the final performance of the model trained by the training system on a target task is improved.

The invention discloses a time sequence knowledge graph retrieval enhancement generation method and device and a medium. The method comprises a preparation stage and an answering stage. In the preparation stage, firstly, a large language model is finely adjusted, so that the large language model can recognize different types of time-related problems; through the design of prompts and task decomposition examples, the model can learn how to decompose a complex time sequence problem into a plurality of subtasks; meanwhile, a query statement template is designed for each subtask, so that the large model learns how to generate a specific query statement according to each subtask; in the answering stage, the model receives the user question and identifies the question type; corresponding prompts are matched and decomposed into a plurality of sub-tasks; the model generates a query statement for each operation and executes query to extract data from the knowledge graph; and integrating the recalled knowledge with the original question to generate a final answer. According to the method, the accuracy and efficiency of time sequence reasoning are improved, the generation capability is enhanced, and the interpretability of the large language model is improved.

The invention discloses a medical image recognition system based on label noise robust learning, and provides a hard sample label refinement strategy based on a confidence perception weighted prototype and an effective noise sample joint correction method to process divided subsets in correction before training so as to obtain training data with higher quality; in progressive hard sample reinforcement learning, data are input according to the learning difficulty of samples for training, the ability of model learning discrimination feature representation is improved by integrating cross entropy loss, consistency loss and credible contrast loss, and the influence of medical tag noise is reduced. According to the method, by integrating label refinement and a progressive hard sample reinforcement learning technology, the accuracy and robustness of medical image recognition under the condition of noise label data are effectively improved, overfitting of the model to noise samples is relieved, and the method can be used for correctly building a disease auxiliary diagnosis model under the condition that the noise label samples exist in training data.

The invention belongs to the technical field of pathological diagnosis, and discloses an artificial intelligence-based lung cancer pathological section analysis and system, which introduces identification basis visual information generated by an artificial intelligence model, and combines the correction operation of a plurality of pathologists on a model identification result and the acquisition of correction reasons, so that the pathological section analysis of the lung cancer is realized. Further performing consistency analysis on the correction results to generate uniform data, and finally performing incremental training on an artificial intelligence model by using the data, thereby constructing a closed-loop feedback optimization mechanism combining artificial intelligence and expert knowledge. The model performance problem caused by insufficient training data coverage, expert labeling efficiency and result inconsistency of an artificial intelligence model in pathological diagnosis is solved, and model learning and performance improvement are achieved.

The invention discloses a dynamic traffic signal control method based on a large language model, belongs to the technical field of artificial intelligence and intelligent traffic control, and aims to solve the problems of poor phase duration flexibility and weak adaptability caused by the fact that most traditional traffic signal control methods are limited to single-stage traffic phase control. According to the invention, the real-time traffic condition is input to the large language model in the form of natural language, and more efficient and intelligent traffic signal control is realized by using the strong generalization ability and the human-like reasoning mechanism of the large language model. The invention provides an efficient fine tuning architecture comprising two stages, the first stage training model learns an answer normal form and a reasoning track of a large parameter quantity model, and the second stage training model promotes the system to effectively learn excellent strategies and keep away from poorer strategies. The method provided by the invention can provide a new thought and technical support for urban intelligent traffic management in the future.

The invention discloses a large language model intellectual property protection method and device based on double-layer nested fingerprints, and belongs to the technical field of model security, and the method comprises the steps: constructing a fingerprint data set which comprises four mutually orthogonal subsets, the three subsets are respectively a normal response subset, a double-layer nested trigger fingerprint subset, a style trigger fingerprint subset and a semantic trigger fingerprint subset; performing embedded learning on a large language model in a knowledge question-answering process by using a fingerprint data set to obtain a fingerprint model, wherein the learning targets are to maintain the consistency of normal response subsets, activate the response of double-layer nested trigger fingerprint subsets, and reduce learning deviation through adversarial learning of style trigger fingerprint subsets and semantic trigger fingerprint subsets; based on the fingerprint data set and the fingerprint model, the double-layer nested trigger meeting the triggering condition is used for verifying the right of any suspected model, the concealment, reliability and effectiveness of double-layer nested fingerprint embedding are improved, and the intellectual property copyright of the large language model is practically protected.

The invention relates to the technical field of automatic fault processing, and discloses an automatic fault processing method based on an AI large model. Comprising the steps that the running state of IT resources is obtained, important fault problems are recognized through an AI large model, the fault problems are combined, and meanwhile alarm information is pushed; a historical knowledge base is retrieved by using an RAG technology, a fault problem is preliminarily analyzed, and a suggestion and a solution are generated by using a large-scale LLM; a solution capable of being automatically executed is confirmed, and the fault processing process is automatically executed through the AI large model; generating a fault report according to the fault processing process, automatically filing the fault report, and learning the fault report through an AI large model; the method has the advantages that faults can be automatically processed, and fault solutions can be archived in time and accurately called at any time.

The invention provides a passive domain adaptive sample forgetting method based on sample sensitivity, and aims to solve the problems that model training is interfered by sample noise and the cross-domain adaptability is poor in a distributed data environment, the method obtains a pure training set by calculating the sample sensitivity of a source domain and screening out samples influencing the model training according to a threshold value, and the sample sensitivity of the source domain is calculated. The method comprises the steps of processing source domain data, training a teacher model by using the processed source domain data, generating a pseudo tag for unmarked data of a target domain, guiding student model learning, transmitting knowledge of the teacher model through a knowledge distillation technology, and aggregating client model parameters by using a federal learning framework, so that data privacy can be protected, and the performance of the model in the target domain can be enhanced. Experiments verify that the method can effectively realize passive domain adaptation, improve the adaptability and accuracy of the model in the target domain, and provide a better model training solution for related fields.

The invention discloses a rapid heat transfer simulation method and device based on a neural network, and relates to the technical field of physical simulation. The method comprises the steps that a hybrid neural network model is trained, the model learns operator mapping from an input function to a temperature or heat flow field, and meanwhile physical constraints such as a heat conduction partial differential equation are coded into a loss function; for a new simulation task, single forward inference is carried out by using the operator mapping, and an initial prediction result is rapidly generated; then, according to physical constraints of coding, calculating a physical residual error of initial prediction, and when the residual error exceeds a preset threshold value, executing a small amount of optimization iteration by taking the prediction as an initial value to carry out rapid local correction; the problems that a traditional numerical method is long in calculation time and an existing neural network method is insufficient in physical fidelity are solved, and high efficiency and high precision of heat transfer simulation are achieved.

The invention discloses a thermal management control method for a power battery. The method comprises the following steps: establishing a basic model for calculating heating and cooling of the battery; basic control logic is set, and the basic model is substituted into different temperatures to select different control modes; data management and analysis including real-time state monitoring, historical data calling, cloud platform data calling and big data comparison rule searching; a model learning algorithm is introduced, historical working conditions, environmental parameters and user behavior data are fused, the system temperature trend is predicted, and temperature management is started in advance; a temperature control model is dynamically optimized through big data; the optimal energy consumption of the control temperature is used as a target function; and 5) fault early warning: obtaining characteristic laws of temperature, voltage, time and the like during a thermal event and each fault event, and realizing early warning when the vehicle has similar characteristics. According to the invention, temperature uniformity control in the whole life cycle of the power battery can be realized, the service life of the battery is prolonged, and timely early warning can be carried out on thermal safety control.

The invention belongs to the technical field of knowledge graph construction, and discloses a supply chain knowledge graph construction method based on time sequence dynamic perception and a large language model. Constructing a time sequence dynamic sensing model, and respectively generating a time-sensitive embedding matrix for the entity and relationship of each sub-graph in a historical time window; inputting the time-sensitive embedded matrix into an aggregator to further mine structure and semantic information of entities and relationships; establishing a dependency relationship of an autoregression model learning sub-graph in a time sequence, and generating a time sequence evolution representation; and inputting the time sequence evolution representation into a pre-trained large language model, generating candidate entities or candidate relationships to complement the fact tetrad, and updating the sub-graph sequence of the current timestamp. According to the method disclosed by the invention, the supply chain domain knowledge is adapted while the general semantic understanding capability is reserved, and the balance of dynamic evolution modeling, long-period dependency capture and efficient utilization of the domain knowledge is realized, so that the reliability and interpretability of a construction result are ensured.

The invention discloses a reward model training method, a large language model optimization method and correlation, and the reward model training method comprises the steps: obtaining a preference training sample pair and a to-be-trained reward model, the preference training sample pair comprising a preferred response sample and a non-preferred response sample; calculating an award score difference between the preferred response sample and the non-preferred response sample based on a to-be-trained award model; constructing a cost matrix based on the reward score difference and the semantic association degree between the preferred response sample and the non-preferred response sample; calculating a loss margin based on the cost matrix; and based on the loss margins, carrying out calculation to obtain paired preference loss values of the band margins, and updating parameters of the to-be-trained reward model by taking minimization of the loss values based on the band margins as an optimization target to obtain a trained reward model. The learning ability and overall generalization performance of the model for difficult samples are improved, excessive dependence on simple samples is avoided, and then the generation quality of the large language model in complex tasks is improved.

The invention relates to the technical field of artificial intelligence, and discloses a vertical field large model training method, which comprises the steps of S10, collecting an initial training data set; s20, pre-training an initial large model based on the project code until the initial large model learns a style and a data format of the project code; s30, performing supervised fine tuning on the initial large model through an initial training data set to obtain a supervised fine tuning stage large model meeting the requirements of the vertical field; s40, constructing a reward model training data set by screening part of data used for supervising and fine tuning and combining output data of the large model in the supervising and fine tuning stage; s50, according to the reward model training data set, reward model training is carried out on the large model in the supervision fine tuning stage, and a reward model which is subjected to sorting scoring training is obtained; and S60, through the reward model and all the collected data, training the large model in the supervision fine tuning stage to obtain a special large model in the vertical field.

The present technology relates to a data creation device, a data creation method, and a program by which a face image dataset suitable for AI model learning can be obtained.The data creation device creates, by partially or wholly changing a creation parameter which is obtained by conversion of a source image of a freely-selected face to a numerical value, a larger number of input creation parameters than a predetermined number from the predetermined number of the creation parameters of the source images, and creates, by creating a face image data item on the basis of a plurality of the input creation parameters, a face image dataset including a plurality of the face image data items. The present technology is applicable to a data creation device.

The invention relates to the technical field of artificial intelligence, in particular to a large language model networking query method, system and equipment based on active triggering, and the method comprises the steps: S1, enabling model learning to recognize a context position needing to trigger networking search in a process of generating an answer text through pre-training or cue word engineering, and autonomously generating and inserting a triggering token; s2, after the trigger token is detected, pausing generation of a model answer text, executing networking search to obtain a retrieval result, and performing abstract extraction on the retrieval result; s3, fusing the summary information with the original question and the historical generation content through a dynamic gating mechanism to form new input; s4, inputting the new input into the large language model again, and continuing to generate the answer text; and S5, when the large language model generates the trigger token again, repeating the steps S2 to S4 until a termination condition is met. And the flexibility and adaptability of answering quality are improved by adopting an active triggering mode.

The invention relates to the technical field of deep counterfeit image detection, and discloses a deep counterfeit detection method and system, a storage medium and computer equipment. The method comprises the following steps: firstly, constructing a reference data set containing a forged image and an original real image; secondly, through an integrated model, generating antagonistic samples for the reference data set, and integrating the successfully attacked antagonistic samples into an antagonistic sample set; and finally, merging the reference data set and the adversarial sample set, and constructing a robustness enhanced data set containing four types of samples. In the model training stage, multi-classification cross entropy loss and comparative learning loss are combined, and expression of the model in a feature space is optimized through comparative learning constraint, so that the model learns discriminative features with more compact intra-class features and more dispersed inter-class features. The model trained by the method not only can effectively defend against attack and improve robustness, but also surpasses original detection performance on clean samples, and has remarkable technical advantages and application value.

The invention discloses a seabed sediment classification method based on multilevel comparative learning and uncertainty measurement. The method comprises the following specific steps: dividing target backscattering intensity data to obtain a training set containing marked samples and unmarked samples and a test set containing marked samples; constructing a feature extraction model, and obtaining a multi-level fusion feature based on the training set; obtaining optimized feature representation based on multi-level comparative learning; screening based on the unmarked samples in the training set and the uncertainty measurement to obtain a target pseudo-tag sample; and carrying out joint learning based on the labeled samples in the training set and the target pseudo-label samples. According to the method, a high-quality pseudo label, namely a target pseudo label sample, is screened out through a pseudo label screening strategy based on uncertainty measurement, and the quality of the target pseudo label sample and the adaptability of model learning are ensured; and supervised and semi-supervised joint learning is carried out through the marked samples and the target pseudo-tag samples in the training set, unmarked samples are fully utilized, and the substrate classification performance is improved.

Methods for training a series of neural networks to output driving behavior control parameters is disclosed. The training dataset for the neural networks includes sensor-based vehicle driving recordings that may be categorized by geographical area, by qualitative driving behaviors, or by some combination, such that various training data subsets are used to train the series of neural networks. By learning either city-specific driving behavior control parameters, qualitative driving behavior specific driving behavior control parameters, or both, the resulting parameters may then be provided to a motion planning model for use in modeling predictive control for an autonomous vehicle. Rather than relying on XYZ trajectories of agent vehicles when planning future trajectories of the ego vehicle, the motion planning model is adaptive, due to the use of the learned driving behavior control parameters.