Patents
Literature
Patsnap Eureka AI that helps you search prior art, draft patents, and assess FTO risks, powered by patent and scientific literature data.

51 results about "Knowledge retention" patented technology

Knowledge retention is a process for identifying. critical knowledge that is at risk of loss, prioritizing what is at risk based on potential knowledge gaps and their impact upon overall organizational performance, and then developing actionable plans to retain that knowledge.

Fine tuning method and system based on power failure multi-modal model, and medium

The invention relates to the technical field of smart power grids, in particular to a fine tuning method and system based on a power failure multi-modal model and a medium, and the method comprises the steps: generating a knowledge retention mask according to a pre-training weight matrix of a multi-modal large language model; generating a sparse low-rank matrix according to the space-time sparse characteristic of the structured multi-modal data set; generating a conflict mitigation regularization item according to the knowledge retention mask and the sparse low-rank matrix, constructing a power task joint loss function according to the conflict mitigation regularization item, and performing fine tuning training on the multi-modal large language model to obtain a fine tuning multi-modal large language model; and inputting the real-time operation data of the power grid into the fine-tuning multi-mode large language model for reasoning to obtain a power grid fault diagnosis result, and performing isolation control on a power grid fault section according to the power grid fault diagnosis result. According to the method, efficient fine tuning is performed on the multi-modal large language model by using the power grid data space-time sparsity modeling and the conflict mitigation regularization strategy, and rapid and accurate diagnosis of the power grid fault is realized.
Owner:WENZHOU ELECTRIC POWER BUREAU +1

Enterprise knowledge retention and access system

An enterprise knowledge retention and access system is disclosed. In various embodiments, data comprising a plurality of content items associated specifically with a user is stored. Generative artificial intelligence techniques are used to generate, based at least in part on the plurality of content items associated specifically with the user, a generated content reflecting information derived from the plurality of content items with respect to a specific subject.
Owner:JELLED INC

Learning progress monitoring method and system for online education platform

The invention provides a learning progress monitoring method and system for an online education platform, and is applied to the field of online data processing. Through multi-dimensional data acquisition and intelligent analysis, the learning behavior monitoring accuracy is improved, the learning achievement degree of the user is accurately judged, the evaluation objectivity is ensured, the learning behaviors of the user are comprehensively tracked in combination with multiple types of learning data, the limitation of single-dimensional evaluation is avoided, and the user experience is improved. Meanwhile, through monitoring of active learning, passive learning and non-learning behaviors, the learning state of the user is further refined, the authenticity of data is improved, personalized learning habit data of the user is collected, an Ebbinghaus forgetting curve is dynamically constructed, accurate review time recommendation is achieved, the review time period can be adaptively optimized in combination with the change trend of test results, and the review efficiency is improved. The user can review at the best time, and the knowledge retention rate is improved.
Owner:SHENZHEN BOMAOYOU EDUCATION TECHNOLOGY CO LTD

Method for evaluating fatigue life of main beam based on continuous increment of sample

The invention relates to the technical field of engineering structure health monitoring, and discloses a method for evaluating the fatigue life of a main beam based on sample continuous increment, which comprises the following steps: acquiring multi-modal monitoring data including physical parameters of strain, displacement and acceleration and external factor data of environmental temperature and humidity and traffic flow information; performing multi-source data fusion by using an adaptive feature extraction network; constructing a dual-knowledge consolidation incremental learning framework, and keeping the memory of historical knowledge when newly added data is processed; performing uncertainty quantification on a prediction result based on Bayesian deep learning, and establishing a risk early warning mechanism; according to the method, the problem of disastrous forgetting in the model updating process is effectively solved, and a relatively high historical knowledge retention rate is kept; a scientific basis is provided for bridge maintenance and maintenance decision making, and maintenance resource allocation is optimized; the method has excellent adaptability and generalization ability, and realizes continuous monitoring and evaluation in the whole life cycle.
Owner:HENAN MINE CRANE

Field adaptive large model fine tuning method and system based on enterprise private data

The invention relates to the technical field of artificial intelligence, in particular to a field adaptive large model fine tuning method and system based on enterprise private data, and the method comprises the steps: extracting the feature representation of local data by each participating client through a unified pre-training model, generating a prototype set representing local region category features by adopting a mean value, clustering or random sampling method; gaussian random noise is added to the prototype set, differential privacy protection is achieved, and then the prototype set is uploaded to a server; the server summarizes prototypes of all the clients, constructs a global prototype training set, performs global training of a single communication round on the adapter module by using supervised learning classification loss under the condition that similar prototypes of different clients are not averaged, obtains a global adapter model and issues the global adapter model to the clients; and the client initializes a local model based on the global adapter model, performs fine tuning by using a small amount of data with labels, and meanwhile, prevents catastrophic forgetting through knowledge distillation constraint by taking the global model as a teacher, so that balance between global knowledge retention and local performance improvement is realized. According to the method, the problem that the expression of the model in a cross-domain reasoning task is remarkably reduced due to the fact that domain characteristics of different enterprises cannot be effectively aligned in the prior art can be solved.
Owner:YOUWEI TECH (SHENZHEN) CO LTD

Study performance precise teaching management method based on time sequence behavior modeling

The invention provides a learning performance precise teaching management method based on time sequence behavior modeling. The learning performance precise teaching management method comprises the following steps: S1, carrying out multi-modal learning behavior data acquisition; s2, aligning and segmenting the time sequence data; s3, performing knowledge retention rate differential modeling; s4, carrying out time sequence feature extraction; s5, constructing a mixed time sequence model; s6, performing dynamic learning ability evaluation; s7, carrying out adaptive resource recommendation; s8, carrying out cross-correction knowledge diffusion optimization; the method has the following advantages: the method is real-time and accurate; the data is acquired to acquire the intervened closed-loop delay lt; compared with the traditional method, the time is increased by 40 times. And cost optimization: the workload of teachers is reduced by 58% and the resource purchase cost is reduced by 42% through an automation strategy. The scale effect is that federal learning supports ten-thousand-person-level concurrence, and the model updating period is shortened to the hour level from the quarter level. Education fairness: cross-school knowledge diffusion enables the superior rate of weak schools to be improved by 29%.
Owner:XINHUA WINSHARE PUBLISHING & MEDIA CO LTD

Incremental OS-ELM cable defect online identification method and system

The invention relates to the technical field of cable defect identification, and discloses an incremental OS-ELM cable defect online identification method and system. The method comprises the steps of collecting cable joint monitoring signals to extract time-frequency features, calculating feature importance weighting initialization model weight, screening samples to adjust forgetting factor recursion update weight, constructing mixed batch optimization weight from a multi-scale time window, and fusing multi-classifier evaluation values to output a defect identification result. According to the invention, the accuracy and stability of cable defect online identification and the long-term knowledge retention capability are improved.
Owner:STATE GRID ZHEJIANG ELECTRIC POWER CO LTD NINGBO POWER SUPPLY CO +1

Large model privacy knowledge decoupling forgetting method and system based on influence function decomposition

The invention relates to the technical field of privacy protection in large language model application, and discloses a large model privacy knowledge decoupling forgetting method and system based on influence function decomposition. The method comprises the steps of obtaining a privacy sample set and a knowledge retention set; a random orthogonal projection matrix is constructed, Hessian matrix calculation of a large language model parameter space is projected to a low-dimensional subspace, and based on an influence function of a sample in a space set, the influence on the model parameters after the sample is removed is quantified; performing principal component extraction on the influence function of the knowledge retention set, and constructing a knowledge subspace base; projecting the influence function of the privacy sample to a knowledge subspace base, and decomposing to obtain a knowledge component and a privacy component; and designing a parameter correction strategy according to the data, updating model parameters according to the strategy, and compensating knowledge loss along the knowledge component direction while eliminating privacy influence along the privacy component direction, so as to iteratively optimize the model. According to the method, the optimal balance of sufficient forgetting of the privacy data and the model general capability is kept.
Owner:DATA SPACE RES INST

Attack detection method and system based on Bayesian incremental learning and storage medium

The invention provides an attack detection method and system based on Bayesian incremental learning and a storage medium, and the method comprises the steps: 1, collecting a data set, and dividing the data set into a plurality of tasks according to years; 2, using a Bayesian continuous learning framework, taking posterior distribution obtained by learning of a previous task as prior distribution of a current task, and adopting a gradient projection method to project a gradient of the current task to an orthogonal subspace of an old task feature space to obtain projection parameters; step 3, performing label deviation and noise processing on the task; 4, minimizing new task loss to obtain parameters, and adopting a training strategy according to a label deviation and noise processing result; 5, finding an optimal combined solution; and step 6, taking the combined model parameters as initialization parameters of the next task, returning to the step 2, and entering the next round of iteration until training of all tasks is completed. The method has the beneficial effects that the knowledge retention capability can be remarkably improved, and the problem of disastrous forgetting is effectively solved.
Owner:HARBIN INSTITUTE OF TECHNOLOGY (SHENZHEN) (INSTITUTE OF SCIENCE AND TECHNOLOGY INNOVATION HARBIN INSTITUTE OF TECHNOLOGY SHENZHEN)

Method and system for improving memory capacity of large model

The invention relates to the technical field of large models, in particular to a method and system for improving the memory amount of a large model, and the method comprises the steps: deploying a monitoring module in a training or reasoning process, carrying out the statistics of the number of memory errors and the total task amount of each component, and calculating the initial memory performance of the components; based on the performance evaluation result of each component, quantifying the overall initial memory amount of the model as an optimization starting point; utilizing a dynamic sensitivity analysis algorithm to mine the influence degree of each component parameter on the memory performance, and calculating a dynamic sensitivity factor; and in combination with the sensitivity factor, the current parameter state and the memory performance, model parameters are adaptively adjusted through dynamic parameter optimization and an iterative lifting algorithm, an iterative verification mechanism is introduced, and the memory ability of each component and the whole machine is subjected to closed-loop optimization. According to the method, fine quantization and parameter-level dynamic regulation and control on the memory performance of the model are realized, and the memory stability and long-term dependence processing capability of the large model in tasks such as knowledge keeping and context understanding are effectively improved.
Owner:BEIJING INFORMATION TECH BOTE INTELLIGENT TECH CO LTD

Music teaching system based on artificial intelligence

The invention discloses a music teaching system based on artificial intelligence, and particularly relates to the technical field of artificial intelligence, and the music teaching system comprises a neurocognitive adaptation module, a music DNA map construction module, a cognitive load monitoring module, an anti-AI dependence adjustment module and a multi-source data fusion unit. According to the music teaching system based on artificial intelligence, a dynamically evolved personalized teaching model is constructed through multi-modal data fusion and real-time analysis driven by artificial intelligence. The neurocognitive adaptation module is combined with electroencephalogram feature analysis and physiological signal monitoring to accurately capture cognitive preferences and ability bottlenecks of the learner; the music DNA map is based on the quantum enhancement modeling technology, the skill development trajectory is continuously updated and predicted, the AI teaching strategy can realize millisecond-level dynamic adjustment according to the neural feedback and behavior data of the learner, the skill mastering efficiency and the knowledge retention rate are remarkably improved, and the limitation of staticization and simplification of a traditional teaching system is broken through.
Owner:PINGLIANG VOCATIONAL & TECH COLLEGE (PINGLIANG SPORTS SCHOOL)

Influence function decomposition-based large model privacy knowledge decoupling forgetting method and system

The present application relates to the technical field of privacy protection in large language model application, and discloses a large model privacy knowledge decoupling forgetting method and system based on influence function decomposition. The method obtains a privacy sample set and a knowledge reservation set; a random orthogonal projection matrix is constructed, the Hessian matrix of the large language model parameter space is calculated and projected to a low-dimensional subspace, and the influence function of the samples in the space set is used to quantify the influence on the model parameters after removing the samples; the principal components of the influence function of the knowledge reservation set are extracted, and a knowledge subspace base is constructed; the influence function of the privacy sample is projected to the knowledge subspace base, and the knowledge component and the privacy component are obtained by decomposition; a parameter correction strategy is designed accordingly, and the model parameters are updated according to the strategy, so that the privacy influence is eliminated along the privacy component direction while the knowledge loss is compensated along the knowledge component direction, thereby iteratively optimizing the model. The present application realizes the optimal balance between sufficient forgetting of privacy data and preservation of model general ability.
Owner:DATA SPACE RES INST

Remote sensing large model incremental pre-training optimization system and method for small sample scenarios

The application discloses a remote sensing image processing and deep learning technical field small sample scene-oriented remote sensing large model incremental pre-training optimization system and method, which comprises the following steps: obtaining initial parameters of a remote sensing large model and a small sample data set of a specific task; introducing a modulation self-attention mechanism into a remote sensing large model feature extraction network, fine-tuning the remote sensing large model based on the small sample data set, and obtaining fine-tuned model parameters; performing interpolation processing on the initial parameters and the fine-tuned model parameters by using a dynamic parameter interpolation method, and balancing new and old task weights; calculating a weighted total loss of an original task loss and a specific task loss through a knowledge reservation mechanism, and reserving model knowledge on the original task; and performing incremental pre-training on the interpolated remote sensing large model based on the weighted total loss. The application solves the problems that model fine-tuning is prone to overfitting and incremental pre-training is prone to forgetting original knowledge in the prior art under a small sample scene.
Owner:SIWEI SHIJING TECH (BEIJING) CO LTD

Bearing fault diagnosis method and device, medium and equipment

The invention discloses a bearing fault diagnosis method and device, a medium and equipment, and relates to the technical field of mechanical fault diagnosis. Comprising the following steps: constructing a plurality of data sets containing different bearing fault types, and obtaining a fault diagnosis model based on the data sets; and repeatedly carrying out trimming operation, data set merging operation and network training operation on the fault diagnosis model to obtain the fault diagnosis model for learning all bearing fault features. In order to balance the old knowledge retention capability and the new data learning capability of the model, old samples and new samples are selected from historical data to be combined, and partial parameters in the trained fault diagnosis model are frozen, so that the model avoids forgetting old task knowledge while updating the parameters of the model; meanwhile, in order to enhance the learning ability of new fault features, the model is pruned to overcome the negative influence of a fixed network model structure on new feature learning, and the learning ability of a new network model on new fault features is enhanced.
Owner:ZHEJIANG OCEAN UNIV

Model distillation method based on singular value decomposition and sinker macro distance

The invention belongs to the technical field of machine learning and deep learning, and particularly relates to a singular value decomposition and sinker macro distance-based model distillation method, which comprises the following steps of: S1, selecting a sparse hybrid expert model as a teacher model; s2, copying weights of all layers structurally shared by the teacher model and the student model from the teacher model to the student model; s3, overlapping the feed-forward network weights specific to a plurality of experts in the layers to form a single dense feed-forward network in the student model; and S4, refining the student model by minimizing the ocker macro distance between the teacher and the student output distribution. According to the method, the feed-forward network weight matrix of each expert in the layer is processed through the singular value decomposition technology, elements are selected, the knowledge retention amount is controlled, key knowledge in the model is accurately captured and retained, and the knowledge of the teacher model is effectively migrated to the student model.
Owner:SHANGHAI XIAOKABANG INTELLIGENT TECHNOLOGY CO LTD

Explosive ordinance disposal and humanitarian demining training method and system

The explosive ordinance disposal and humanitarian demining training method and system is a product to simulate explosive hazard incidents and instructional material in a repeatable, cost effective, high fidelity data rich environment in order to assess knowledge retention, skills, understanding, and capture lessons learned. The primary focus is perceived realism and ability to export data in a clean and efficient way to simulate an environment and gather the user data to validate learning objectives, identify trends, and diagnose training and environment issues. The intended user of this product will be EOD technicians and humanitarian deminers. The methods and systems of the present invention increase fidelity of a VR / XR training environment by incorporating additional verified data details, elements, features, and relationships between objects and elements into a 3D training environment thereby increasing and improving the fidelity of training scenarios.
Owner:VALENTINE CHARLES ALAN

Data learning server for the creation and use of learning models

A data learning server that includes the following: ◯ a system for distinguishing between content data and background data within a minimal dataset; ◯ a module for generating synthetic training examples from a clean template; ◯ a co-learning model to predict knowledge retention and the propensity to follow instructions; ◯ a system for personalizing learning experiences by optimizing hints based on real-time analytics; ◯ a knowledge-driven dialogue system for dynamic and context-related interactions; ◯ IoT connectivity for remote monitoring and control of learning models.
Owner:ARUMBAKA RAMESH BABU HYDERABAD +7

Generative pre-trained transformer robot for

A generative pre-trained transducer (GPT) robot executes a knowledge retention engine configured to generate and retain a summary of instant dialog history during and between chat sessions for both single-topic discussion and multi-topic discussion. The GPT robot also executes a recommendation engine configured to recommend references to past discussion topics with the most recently accepted summary and solutions based on the retained summary.
Owner:SCHNEIDER ELECTRIC USA INC

A method and system for fine-tuning a large model based on domain adaptation of enterprise private data

The present application relates to the technical field of artificial intelligence, in particular to a domain self-adaptive large model fine-tuning method and system based on enterprise private data, the method comprising: each participating client extracts the feature representation of local data using a unified pre-training model, and generates a prototype set representing the category features of the local domain by using mean, clustering or random sampling methods; uploading the prototype set to the server after adding Gaussian random noise to realize differential privacy protection; the server aggregates all client prototypes to construct a global prototype training set, and performs global training of the adapter module in a single communication round using supervised learning classification loss without averaging the same prototypes of different clients, obtains a global adapter model and distributes it to the client; the client initializes the local model based on the global adapter model, fine-tunes it using a small amount of labeled data, and uses the global model as a teacher to prevent catastrophic forgetting through knowledge distillation, achieving a balance between global knowledge retention and local performance improvement. The present application can solve the problem that the existing technology cannot effectively align the domain features of different enterprises, resulting in a significant decline in the performance of the model in cross-domain reasoning tasks.
Owner:YOUWEI TECH (SHENZHEN) CO LTD

A personalized federated continual learning method and system based on dynamic clustering and multi-scale prototypes

The application discloses a kind of personalized federated continuous learning method and system based on dynamic clustering and multi-scale prototype, to solve the problem of spatiotemporal catastrophic forgetting in personalized federated continuous learning.The server initializes global model containing prompt parameters and cluster-level multi-scale prototype library, and the client generates routing histogram representing local data distribution and incremental prototype through a gated routing network and uploads it.The server calculates JS divergence based on routing histogram to detect drift, triggers K-Means online re-clustering, and dynamically updates the multi-scale prototype library containing short-term, long-term and drift prototypes in combination with the drift signal.The client uses long-term prototype to construct a joint loss optimization model.The application balances new knowledge learning and old knowledge retention while protecting data privacy and reducing overhead, and adapts to scenarios where data distribution continues to evolve.
Owner:TIANJIN UNIVERSITY OF TECHNOLOGY

Autonomous framework for learning and teaching processes

An information technology system for an autonomous teaching and learning automation framework is disclosed. The framework is used for autonomous teaching and learning process by leveraging and mining a wide variety of historical data and real-time data to deliver a dynamic, automated, autonomous, customized, personalized and user-centric learning curve in a way that boost engagement, learning and knowledge retention while generating more historical and statistical data thru various important reports. The framework includes an engine for identifying new autonomous processes, wherein the processes are generated for asynchronous teaching and learning. Real-time feedback is used along with validations for each task. An orchestration module is used for orchestration tasks along with validations and execution of the tasks to ensure learning of each learner / participant. In some embodiments, the framework is used for autonomously deploying software such as an operating system by asynchronously developing the processes for different types of deployment executors.
Owner:WATAT INNOCENT LEGRAND

Fault diagnosis method and system for train transmission motor, gearbox and axle box

The invention discloses a train transmission motor, gearbox and axle box fault diagnosis method and system, and belongs to the technical field of train transmission system fault diagnosis, and the method comprises the steps: designing special expert networks for key components of a train transmission system; a multi-head self-attention gating network is introduced, and features output by the special expert network are dynamically fused; a joint loss function including cross entropy loss, KL divergence regularization loss and load balancing loss is adopted to train a special expert network and a gating network; based on a teacher-student model architecture and knowledge distillation, newly added fault category adaptation and historical knowledge reservation are realized; and inputting a to-be-diagnosed multi-component vibration signal into the trained model, and outputting a fault category with the highest probability as a diagnosis result through feature extraction, gating network fusion and classifier prediction. According to the method, the problems of multi-component independent modeling, incremental learning delay, catastrophic forgetting, poor multi-working-condition adaptability and the like in train transmission system fault diagnosis are solved.
Owner:QINGDAO UNIV OF TECH

A semi-supervised continual learning method and system, and a disease recognition method based on medical images

The application relates to a kind of semi-supervised continuous learning method and system, and disease identification method based on medical image, the system includes two core modules of double memory playback (DMR) and unbiased contrast learning (UBCL). The method is based on NNCSL, and the traditional mechanism is optimized by data classification and feature deviation correction, efficient knowledge retention and unbiased feature learning are realized. By designing double memory playback (DMR) strategy, a double buffer area with labels and without labels is constructed, random sampling and loss-based confidence sampling are used respectively, and the overfitting of labeled data and the catastrophic forgetting of unlabeled data are simultaneously relieved. An unbiased contrast learning (UBCL) method is proposed, which combines metric learning and balanced KL regularization techniques to eliminate feature representation bias and help the model obtain unbiased and discriminative features.
Owner:NORTH CHINA ELECTRIC POWER UNIV

Model processing method and device, electronic equipment, storage medium and computer program product

The invention discloses a model processing method and device, electronic equipment, a storage medium and a computer program product. The method comprises the steps that first feature data is acquired, and the first feature data comprises at least two types of sample data associated with a first task; using the first feature data to determine first information corresponding to at least two regions associated with a first model, the first information representing an importance degree of each region in the at least two regions to each category of sample data in the at least two categories of sample data, the at least two areas are at least used for knowledge learning; using the first information to determine a growth strategy of the first model, the growth strategy being used for indicating a to-be-grown area in the at least two areas; based on a knowledge retention algorithm, a second model is obtained through the growth strategy and the first model, and the second model is used for executing the first task.
Owner:CHINA MOBILE COMM LTD RES INST +1

Human Capital Economics: Automated infrastructure and method for determining employee life cycle costs with return on investment calculation and decision making framework

An automated computer infrastructure for synthesizing computer data regarding employee life cycle costs for human capital economic analysis and management. A computerized framework receives and processes employee life cycle financial data elements in an automated manner by computer and transforms them into actionable synthesized human resource economic analysis. The computerized framework comprises computer data entry and analysis software adapted for receiving, processing, and transforming individual employee and organizational life cycle financial data elements relating to costs of hire, costs of training & onboarding, and costs of offboarding. The automated computer infrastructure is adapted to transform life cycle financial data elements to, for instance, percentage of recruiter time spent on each of plural employee tiers, recruiter cost per hire for plural employee tiers, hiring salary overhead, salary time costs for hiring, contract cost, cost of training, cost of mentoring, knowledge retention expense, offboarding expense, productivity loss, and cost of attrition.
Owner:SCARANGELLA MARIA

Remote sensing large model increment pre-training optimization system and method for small sample scene

The invention discloses a small sample scene-oriented remote sensing large model increment pre-training optimization system and method in the technical field of remote sensing image processing and deep learning. The method comprises the following steps of: obtaining initial parameters of a remote sensing large model and a small sample data set of a specific task; after a modulation self-attention mechanism is introduced into a remote sensing large model feature extraction network, fine tuning is carried out on a remote sensing large model based on the small sample data set, and fine-tuned model parameters are obtained; adopting a dynamic parameter interpolation method to carry out interpolation processing on the initial parameters and the fine-tuned model parameters, and balancing weights of new and old tasks; calculating the weighted total loss of the original task loss and the specific task loss through a knowledge retention mechanism, and retaining the knowledge of the model on the original task; and performing incremental pre-training on the remote sensing large model after interpolation processing based on the weighted total loss. The method solves the problems that in the prior art, model fine adjustment is prone to over-fitting in a small sample scene, and original knowledge is prone to being forgotten in incremental pre-training.
Owner:SIWEI SHIJING TECH (BEIJING) CO LTD

Generative pre-trained transformer bot for conversational context retention

A generative pre-trained transformer (GPT) bot executing a knowledge retention engine configured to generate and retain a summary of immediate conversation history during and in between chat sessions for both single-and multiple-topic discussions. The GPT bot also executes a recommendation engine configured to recommend references of past discussion topics with the latest accepted summary and solutions based on the retained summaries.
Owner:SCHNEIDER ELECTRIC USA INC

Flow time series anomaly detection method and system based on spectral purification and dynamic memory

The application discloses a kind of based on spectral purification and dynamic memory's flow time series anomaly detection method and system, belong to time series anomaly detection technical field.The present application is to solve the problem of memory pollution caused by abnormal noise in flow online detection, catastrophic forgetting caused by normal mode evolution and limited memory capacity, by frequency domain transformation to flow time series window, retain low-frequency component and utilize learnable complex spectral weight to carry out adaptive filtering, obtain spectral purification sequence;Spectral purification sequence is embedded and context time series coding, and obtain query feature;According to the similarity of query feature and normal mode memory item, memory retrieval, expansion and weighted combination are carried out, and projection reconstruction is carried out by fusing query feature and memory enhancement feature, and the abnormal detection result is determined according to reconstruction error.The present application can reduce the interference of anomaly and noise on online memory, consider new normal mode absorption and historical knowledge retention, improve the accuracy and stability of flow time series anomaly detection.
Owner:INSTITUTE OF INFORMATION ENGINEERING CHINESE ACADEMY OF SCIENCES

An efficient fine-tuning and knowledge editing combined optimization method for large models

The application discloses a kind of high-efficiency fine-tuning and knowledge editing combined optimization methods for large model, belong to artificial intelligence technical field, to solve the problem such as low efficiency, task conflict and serious forgetting caused by fine-tuning and knowledge editing separation processing in prior art, the integrated optimization framework is proposed, the method first freezes the basic parameters of pre-training large model, and injects low rank adaptation (LoRA) module as the only trainable parameter carrier;While constructing the joint optimization objective function of fusion fine-tuning loss, knowledge editing loss based on zero space projection and knowledge retention loss based on KL divergence, then in a single training cycle, the joint loss is optimized synchronously by gradient descent, to drive LoRA parameter collaborative update, to simultaneously realize the efficient adaptation of downstream task, accurate editing of target knowledge and effective protection of original knowledge.
Owner:HOHAI UNIV

Method for evaluating fatigue life of main girder based on sample continuous increment

The present application relates to the technical field of engineering structure health monitoring, and discloses a method for main beam fatigue life evaluation based on sample continuous increment, comprising: collecting multi-modal monitoring data, including physical parameters of strain, displacement and acceleration, and external factor data of environmental temperature and humidity and traffic information; performing multi-source data fusion by using an adaptive feature extraction network; constructing a dual knowledge consolidation incremental learning framework to maintain the memory of historical knowledge when processing new data; quantifying the uncertainty of the prediction result based on Bayesian deep learning, and establishing a risk warning mechanism; the present application effectively solves the problem of catastrophic forgetting in the model updating process, maintains a high historical knowledge retention rate; provides a scientific basis for bridge maintenance and repair decision-making, optimizes maintenance resource allocation; has excellent adaptability and generalization ability, and realizes continuous monitoring and evaluation within the whole life cycle.
Owner:HENAN MINE CRANE