160 results about "Continual learning" patented technology

A computer system implements a unified framework integrating an adaptive elastic funnel (AEF) with a convergent intelligence fabric (CIF) for flexible and contextualized multi-agent AI and human collaboration at scale. The system provides a universal multi-modal key-value subsystem for sharing partial computations across agents, implements a hybrid greedy / non-greedy placement strategy for dynamic memory management, orchestrates dynamic computational workflows and tensor workflows using hierarchical tensor-fragment scheduling, enables cross-agent orchestration with policy-based privacy preservation, and incorporates quantum-resistant secure memory enclaves. The architecture supports continuous learning without catastrophic forgetting, compositional reasoning across modalities, and secure task execution in distributed environments. This integration enables unprecedented computational efficiency, secure collaboration, and adaptive intelligence in high-dimensional decision-making environments while supporting incremental adoption through modular interfaces.

The invention discloses a medical image treatment method and system for multi-source heterogeneous data, and relates to the field of medical image treatment, and the method comprises the steps: firstly, respectively extracting an image embedding vector and a text embedding vector from original medical image data and a description text thereof through a deep learning model; then, the vectors of the two different modes are fused, and a unified semantic embedding vector is formed; based on the unified vector, through semantic similarity calculation with a standard term library, candidate mapping can be automatically generated, and dependence on rigid artificial rules is eliminated. More importantly, a closed-loop mechanism of manual auditing-feedback learning is introduced in the scheme, high-confidence mapping is adopted automatically, low-confidence mapping is audited by experts, and an auditing result is absorbed into a mapping knowledge base, so that the system has continuous learning and self-evolution capabilities, and a semantic gap of cross-mechanism data can be eliminated more intelligently and more accurately.

The application discloses a work clothes identification method combining continuous learning and effectively resisting forgetting disaster, and comprises the following steps: collecting data in an initial stage, constructing a work clothes classification model, and combining data to perform self-supervised training on the work clothes classification model; the trained work clothes classification model is divided into a deep model and a shallow model, an image self-encoder is constructed between the deep model and the shallow model, and the image self-encoder is trained, the intermediate features output by the shallow model are encoded and compressed through the trained image self-encoder, and compressed features are obtained; in a continuous learning task stage, work clothes category data are collected, the compressed features are decoded, and a new model is constructed; based on the new model, a Tt-stage work clothes classification model is constructed through a gradient boosting method, and after training, compression is performed to obtain a final model for work clothes identification; the application alleviates the forgetting degree of the work clothes classification model to old categories, and improves the plasticity of the model.

A federated learning framework including household agents configured to continuously learn model parameters for managing charging periods and discharging periods of household batteries, and microgrid agents to maximize use of local energy based on a pricing policy, including accessing power from other microgrids when there is insufficient local energy to cover local demand. and selling surplus energy to the other microgrids when power generation by the microgrid surpasses the local demand. Each household machine learning agent is configured to control household energy demand from and supply to a microgrid which they are connected in order to minimize household energy cost while adapting to changes in the energy price that is determined based on the pricing policy of the microgrid agent that encourages reduction of carbon emission. A federated learning engine combines the model parameters from the household machine learning agents to update a global household machine learning agent.

This application discloses a system and method for automatically analyzing smart contract vulnerabilities using AI, belonging to the field of blockchain technology. The key technical solutions include a data acquisition and preprocessing module, an AI vulnerability analysis module, a model continuous learning module, an integration service module, and a database module. This application provides comprehensive and in-depth security assessments. Through machine learning and deep learning technologies, this system can identify and predict unknown vulnerabilities, thereby greatly improving audit efficiency and accuracy.

The application discloses a kind of fusion multi-model intelligent reasoning's cause of death chain automatic generation and radical cause of death determination support method and equipment.The method comprises: obtaining and preprocessing electronic medical record data;Adopt four-stage series screening mechanism including semantic matching, context adaptation, timing verification and statistical optimization, map clinical text into candidate ICD coding;Adopt search enhancement generation framework and dynamic multi-round controllable reasoning engine to generate candidate cause of death chain;Adopt the mixed probability model of timing-cause dual constraint to evaluate candidate cause of death chain, calculate comprehensive probability to determine radical cause of death, and the result is visualized and presented.The application can realize the automation of cause of death chain generation and radical cause of death determination, significantly improve accuracy and efficiency, enhance the explainability of decision-making process, and have continuous learning ability.

This invention discloses a dual-constraint incremental axial piston pump fault diagnosis method based on sample selection and playback, belonging to the field of fluid pressure and electrical digital data processing. The method includes: acquiring fault signals at the current moment and labeling them to obtain fault samples, constructing a fault diagnosis model; merging the fault samples at the current moment with old fault samples to form an incremental dataset, where the selection of old fault samples is based on improved sample profile coefficients and thresholds; inputting the incremental dataset into the fault diagnosis models at the current and previous moments respectively, further calculating the total incremental loss function composed of cross-entropy loss, cosine contrast loss, and surrogate loss, updating the model parameters through backpropagation until convergence, obtaining the fault diagnosis model trained at the current moment. This invention overcomes the catastrophic forgetting problem of learned fault knowledge in existing fault diagnosis models, and improves the continuous learning capability of the axial piston pump fault diagnosis model through incremental learning.

This invention discloses a few-sample class-variable audio classification method based on a prototype adaptive network. The steps are as follows: 1) Extracting log-Mel spectrum from input audio samples; 2) In the basic stage, training an initial model consisting of a feature extractor, a prototype classifier, and a class-adjustable prototype adaptive network using base class samples; 3) Extracting representations from the base class training samples, calculating the mean of each base class representation as a prototype, and updating the prototype classifier; 4) In the incremental stage, if the number of categories increases, inputting the representations of the incremental class training samples into the class-adjustable prototype adaptive network to obtain the prototypes of each incremental class and updating the class-adjustable prototype adaptive network; if the number of categories decreases, deleting prototypes of unnecessary categories; 5) Then updating the prototype classifier using the class-adjustable prototype adaptive network; 6) Obtaining the classification category of the test sample. This invention can adaptively expand or shrink the prototype classifier, thereby continuously learning new categories while remembering or deleting old categories, achieving few-sample class-variable audio classification.

The application provides a self-evolution presentation document generation method and system based on multi-modal and knowledge graph, wherein the self-evolution presentation document generation method comprises the following steps: constructing an enterprise-level picture meta-database containing a vector index; parsing a source document to obtain structured content and generating a presentation document outline; searching for matching candidate pictures in the enterprise-level picture meta-database based on the target page core content in the outline; querying a design decision knowledge graph to generate layout and color matching planning; assembling a presentation document according to the planning and synchronously generating design decision metadata; driving knowledge graph updating by displaying the design decision metadata and obtaining user modification feedback on the design; and the self-evolution presentation document generation system comprises functional modules for realizing the above method. The application solves the technical problems of poor picture quality, inaccurate image-text matching, rigid design, high copyright risk and difficulty in continuous learning and evolution from use in the existing automatic presentation document generation technology.

The application discloses a kind of multi-objective collaborative optimization method and device of coal blending of thermal power plant, belong to energy management and intelligent control technical field.Integrating the coal quality of thermal power plant, operation, emission and scheduling data, construct unified feature library, and based on this, set the initial weight of economic, safety, environmental protection target, establish multi-objective optimization model;Adopt parallel computing architecture to solve model, generate multiple candidate blending scheme;Subsequently, combined with the artificial adjustment result of operator and actual combustion feedback data, carry out secondary prediction and comparative analysis;Finally, through continuous learning adjustment behavior and operation effect, drive model automatically update weight and optimization algorithm combination, form closed loop self-learning mechanism, to realize multi-objective collaborative optimization and adaptive decision of thermal power plant.The application improves the intelligence, economy and environmental protection of system operation.

The application discloses an online active false news real-time detection method based on a concept neural network, which comprises the following steps: collecting news samples and performing real / fake labeling to obtain an initial labeling dataset; based on the initial labeling dataset, taking news feature data as the concept connotation and a news sample object set as the concept extension, an initial news false detection model is trained; a current model is used to predict current news samples, and high-confidence samples are screened based on information entropy; a probe model is obtained by training a current model copy using high-confidence samples and pseudo-labels obtained by the current model; the prediction distribution difference between the probe model and the current model is calculated, a high-difference sample set is screened, and artificial labeling is performed; and the current model is updated online by combining the high-confidence sample set with pseudo-labels and the high-difference sample set with artificial labels, so as to obtain a detection model at a next time step. The application can realize continuous learning, online updating and effective detection of false news in a dynamic network environment with limited label resources.

The application discloses a class-incremental continual learning method based on classification layer unification and improved EWC, a neural network is trained through a training set to obtain a recognition model, and the recognition model is used for recognizing the category of things in a picture, the training comprises old task training and at least one new task training, and in the new task retraining process, the model parameters of the old task are taken as a base point to limit the parameter offset. The class-incremental continual learning method based on classification layer unification and improved EWC solves the new category bias problem and improves the average accuracy of the model.

This invention discloses a self-evolving multi-agent collaborative method and system based on a dynamic cognitive graph. The method includes constructing a dynamic cognitive graph and modeling the reasoning process as a directed acyclic graph; initializing a multi-agent collaborative network; executing sub-tasks and dynamically reconstructing the cognitive graph using execution feedback; evaluating the execution process using a case-based reasoning evolutionary mechanism, generating structured cases, and encoding and storing the structured cases in a case library for experience reuse in subsequent tasks; the multi-agent collaborative network adopts a hybrid collaborative architecture, communicating between agents through a blackboard system; monitoring metacognition by real-time monitoring of the confidence level of the reasoning process, and using an arbitration mechanism to arbitrate conflicts when multiple agents reach contradictory conclusions; outputting the execution and evaluating the system performance. This invention significantly improves the processing capability of complex long-link reasoning tasks and the system's continuous learning capability.

The application provides a large language model continuous learning method based on key-value pair replay, comprising the following steps: S1, constructing a task flow; S2, constructing a continuous learning framework, which is used for obtaining the most recent current task sample from the task flow and extracting the compressed key-value cache as the historical task information, inputting the current task sample and the compressed key-value cache into a large language model to constrain the update direction of the model parameters, and retrieving the historical key-value cache with the highest semantic similarity with the compressed key-value cache, and splicing the current task sample and the historical key-value cache to form an output result; S3, sequentially fine-tuning the large language model to obtain a large language model after continuous learning; and S4, performing inference testing on the to-be-tested task sample in a test set to obtain a performance evaluation result of the continuous learning. The application has the beneficial effect of realizing efficient and privacy-protected continuous learning of a large language model.

This invention discloses a multi-task detection training method and system based on an improved YOLOv8 model, comprising the following steps: Step S1, acquiring RGB image data of a road scene in real time using a monocular camera, and labeling, classifying, and constructing an original image dataset; Step S2, resizing and adjusting the resolution of the classified RGB images in the original image dataset from Step S1; Step S3, constructing two independent detection modules on the model that output corresponding task detection results, and configuring independent data loaders to provide image data to the model according to dynamic task batch size and in an alternating manner; allocating RGB image data for the corresponding image detection task to the two detection modules according to the task batch size, so as to realize the model's continuous learning and balanced training of multi-task detection for the two detection modules, and outputting the detection results of the corresponding detection task.

This invention discloses a method and system for automatic extraction and knowledge construction of data quality rules for power equipment, belonging to the fields of artificial intelligence and power data governance technology. To address the problems of low parsing accuracy of complex tables in power technology standard documents, lack of physical constraint verification of extraction results, and catastrophic forgetting of models due to standard updates, this invention utilizes a multimodal large language model to parse documents. A structured document object model is obtained through hierarchical table structure encoding, cross-page context preservation, and semantic binding of table annotations. The large language model, fine-tuned with domain knowledge, extracts entity relationships and performs consistency verification in conjunction with a physical constraint knowledge base in the power domain. The verified entity relationships are assembled into structured quality rule records and stored in a graph database. A differential-adaptive elastic weight consolidation strategy is employed for continuous learning, combined with a human-machine collaborative quality inspection optimization model, achieving high-precision complex table parsing, physical consistency verification, and anti-forgetting continuous learning.

This invention discloses an automatic generation and evaluation method for machine tool fault solutions based on a large language model, relating to the field of equipment fault diagnosis technology. It performs multimodal parsing on structured fault logs and unstructured maintenance reports; extracts key fault information from the parsed historical fault data; and automatically generates fault cause analysis, step-by-step troubleshooting processes, and solutions based on the extracted key fault information, combined with prompt word engineering and a large language model. The method also automatically evaluates and iterates the solutions, employing an expert feedback enhancement system for feedback. This invention's automatic generation and evaluation method for machine tool fault solutions, by integrating historical data mining and large language model generation capabilities, establishes a mapping knowledge base of fault characteristics and solutions. Based on a dynamic evaluation algorithm-driven iterative optimization mechanism and an expert feedback-driven continuous learning architecture, it significantly improves fault handling efficiency and reduces reliance on human experience.

The application discloses a federated continual learning cross-layer optimization method for a UAV relay network, and aims at the problems of easy occurrence of catastrophic forgetting of a model and high system time delay in a streaming task training scene.The application constructs a system architecture comprising a base station, a UAV and a plurality of ground clients; a task stability index is obtained by calculating the gradient similarity between a current task and a historical task of a client, and a system efficiency index is obtained by combining the client computing time delay and the communication time delay, so as to jointly select a target client participating in aggregation; in order to anchor historical knowledge, local compensation updating is performed at the client side based on the historical task gradient, and post-aggregation compensation correction is performed at the base station side based on the historical global gradient; and a cross-layer optimization model about the client computing frequency, bandwidth allocation, UAV trajectory and relay strategy is further established to minimize the maximum completion time delay of the system. The application can improve the model learning precision, reduce catastrophic forgetting and reduce the training time delay.

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.

This invention discloses a method and system for identifying abnormal driver behavior based on continuous learning, relating to the field of driving behavior detection technology. The method includes dividing a driver monitoring image into image blocks and projecting them into feature vectors, adding positional encoding, and combining extracted prompt features with segmentation into prompt vectors to achieve dual coverage of global and key local features, improving feature extraction accuracy. The two types of vectors are concatenated into a fused feature sequence, input to a pre-trained Vision Transformer model with frozen parameters, and enhanced features are extracted through a self-attention mechanism. Then, a global feature vector is obtained through average pooling. Continuous learning is achieved by outputting a vector of all seen task logic values, using a binary mask for filtering, and classifying using the argmax function. This allows for adaptation to behavior category expansion, balancing real-time performance with stability in complex environments, improving the accuracy of similar behavior recognition, and making it easily deployable in in-vehicle scenarios.

This invention discloses an edge device image classification method based on a vision-language model. It obtains the streaming data to be processed within an application scenario and applies data augmentation techniques. A pre-trained model with image-vision dual branches is then loaded for training, incorporating scalable units during training and saving visual prototypes for each category. A category prototype replay strategy is combined to jointly constrain the features of the old and new categories during model updates, effectively mitigating catastrophic forgetting. Furthermore, a two-stage inference method is employed, introducing discriminative descriptions between categories to refine the initial prediction results. This invention addresses the problems of existing methods struggling to fully utilize linguistic contextual information in incremental class learning and being prone to knowledge degradation during continuous learning. While maintaining low computational and storage overhead, it improves the model's recognition accuracy, stability, and generalization ability in dynamic category environments.

The application discloses an AI agent non-inductive smooth evolution method and system based on incremental distillation and layered core replacement, and belongs to the technical field of artificial intelligence continuous learning. The system core comprises a gentle iteration unit, a living water isolation unit and an abnormal rollback unit, and can be extended with a computing power regulation and safety check unit. By dividing the agent model into a read-only locked core layer and an incrementally updatable capability layer, combining double-buffered data isolation, pointer seamless switching, dynamic computing power regulation and abnormal automatic rollback, the smooth evolution of the AI agent is realized without violence, interruption or catastrophic forgetting. The application completely solves the problems of memory loss, service interruption, iteration risk and the like caused by existing violent updates, is suitable for long-term iteration and upgrading of various commercial AI agents, and guarantees the continuity of the core identity of the agent and user interaction.

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.

This application relates to the field of soil analysis technology and discloses a multispectral image-driven intelligent soil quality assessment method and system. By introducing a working condition parameter recording and explicit modeling mechanism, it effectively optimizes the consistency problem faced by multispectral remote sensing data in long-term application and cross-regional promotion. Traditional soil quality remote sensing assessment methods often use single-temporal data for modeling, neglecting the systematic impact of changes in observation conditions on spectral data, resulting in a significant performance degradation of the model when applied in different times and regions. This method eliminates external interference factors through spectral correction and state decoupling techniques, enhances the stability of feature expression through multi-scale feature construction constrained by mechanisms, balances the model's universality and specificity through a hierarchical modeling architecture, ensures performance for cross-regional application and long-term operation through regional adaptation and drift detection techniques, and maintains the model's continuous effectiveness through a continuous learning mechanism.

We provide the system. [Solution] Means for acquiring audio information, A means of converting this audio information into text information, A means of analyzing the converted text information to understand the content of the inquiry or potential threat, A means of generating appropriate responses or warnings based on inquiries, A means for converting the generated text response into audio information, Means for making these operations compatible with multiple languages, A system that includes means for collecting and continuously learning from evaluation results of responses and warnings.

The application relates to the technical field of network intrusion detection, in particular to an automatic intrusion detection system for a dynamic network environment, which has the technical scheme that in the autonomous decision module of Gaussian probability, a contrast loss function taking normal traffic as the center is designed, so that the model can efficiently distinguish the behavior patterns of normal traffic and abnormal traffic; in the automatic continuous learning framework, a double memory bank is designed to adapt to the concept drift scene in the dynamic network, wherein the stable memory bank is used for storing old knowledge and preventing the catastrophic forgetting of the model, and the high-confidence pseudo label generated in the autonomous decision module of Gaussian probability is used to update the adaptive memory bank, so that the real-time updating and fine-tuning of the autonomous decision module of Gaussian probability are realized; in the continuous learning process, the system does not need to rely on manual labeling, can effectively capture the constantly evolving patterns in the dynamic network scene, significantly enhances the applicability of the intrusion detection system to the concept drift, and realizes automatic intrusion detection.

This invention relates to an adaptive modular system and method for professional evaluation and benchmarking across multiple industries. The system integrates advanced artificial intelligence and machine learning algorithms within a modular framework to provide personalized, objective, and real-time assessments. Key components include a data ingestion layer for collecting data from diverse sources, a data processing engine for cleaning and transforming data, a feature extraction module, a machine learning module comprising various models, evaluation modules tailored to specific assessment criteria, an adaptive algorithm controller employing reinforcement learning, and a user interface layer. The system addresses limitations in existing evaluation methods by offering dynamic customization, scalability, efficient data processing, and enhanced personalization. It leverages techniques such as natural language processing, deep learning, explainable AI, and continuous learning to deliver comprehensive evaluations, benchmark performance against industry standards, and support decision-making processes.

The application provides a micro-service routing mechanism optimization method based on SDN combined with reinforcement learning technology. The combination of RL and SDN can form collaborative optimization at different levels. On the one hand, RL can continuously learn and optimize routing strategies based on multi-dimensional operating states, and realize adaptive traffic distribution and strength selection at the application layer. On the other hand, SDN has a global network view and programmable ability through the separation of the control plane and the data plane, can dynamically adjust the underlying forwarding path and flow table rules, and realize fine-grained traffic scheduling at the network layer. When the two are combined, RL is responsible for micro-service strategy decision and global performance optimization, and SDN is responsible for underlying network execution. This hierarchical collaborative mode not only breaks through the limitations of traditional single-layer optimization, but also realizes real-time perception, autonomous decision and continuous optimization in a dynamic environment, significantly improving the stability, resource utilization and service quality of the cloud-native micro-service system under complex load conditions.

The application discloses an automatic driving performance improvement test system based on reinforcement learning and belongs to the technical field of automatic driving.The application regards a vehicle to be tested as a host vehicle and defines a closed-loop test mechanism comprising interactive sampling, closed-loop feedback calculation, host vehicle performance evolution and traffic vehicle confrontation strategy updating.A host vehicle performance evolution module is based on a closed-loop feedback signal, a CARE robust enhancement module and a SCALE continuous learning module are constructed, so that training explicitly focuses on long-tail high-risk samples and inhibits catastrophic forgetting.A traffic vehicle confrontation strategy module adopts a multi-agent strategy with parameter sharing, designs a CAGE confrontation generation strategy, introduces an AIARM confrontation strength adaptive adjustment mechanism, and dynamically generates a confrontation scene that evolves with the host vehicle capability under the feasibility constraint.The application realizes continuous capability improvement of an automatic driving system in a long-tail risk scene, and effectively enhances the robustness and safety of the system in a complex interactive environment.