180 results about "Co-training" patented technology

The invention provides a private weight self-adaptive heterogeneous data federated cooperative training method and system in the technical field of federated learning and privacy computing, and the method comprises the steps: S1, enabling each client to carry out the differential privacy operation on a local data set based on a private weight, and obtaining a desensitized data set, encoding the desensitized data set through a heterogeneous data encoding model; s2, performing semantic alignment on each coding vector through a contrast learning model to obtain an aligned vector set; s3, training a local model through the alignment vector set, generating a local gradient, extracting local model parameters, and uploading the privacy weight, the local gradient and local difference parameters to a server; and S4, the server trains the global model based on the local difference parameter and the global gradient, extracts the global model parameter and issues the global model parameter to each client for training. The method has the advantages that the compatibility, the flexibility and the efficiency of heterogeneous data federation cooperative training are greatly improved.

The invention discloses a privacy protection data sharing and cooperative training method and system based on federated learning. The method comprises the steps of receiving software development log data, adaptively judging the sensitivity degree according to a data type, dynamically adjusting noise disturbance intensity according to the sensitivity degree to perform data desensitization, and generating a sensitivity index; selecting a feature extraction strategy, extracting time sequence correlation features from the desensitization data, constructing a dynamic graph structure with a weight, and obtaining a time sequence feature vector through iterative fusion; calculating the time sequence correlation of the time sequence feature vector to obtain a data quality score, and setting a contribution weight based on the quality score to perform parameter aggregation; combining sensitivity indexes with data quality scores to construct a security sharing domain, decoupling global training parameters into knowledge fragments in the domain, formulating a recombination rule, and selectively acquiring the required knowledge fragments by all parties for local training. According to the method, deep collaboration is realized on the premise of protecting data privacy, and the collaboration training effect is improved.

The invention discloses a multi-mode driven cross-industry digital twinning universal platform architecture and an implementation method, and relates to the technical field of digital twinning and artificial intelligence. The method comprises the following steps: establishing a multi-modal driven cross-industry digital twinning universal platform, deploying a multi-source heterogeneous data acquisition component in a data access layer to access text, image and time series data, and converting unstructured data into a unified feature space by adopting a Transform-GNN cross-modal encoder in a multi-modal fusion layer; in the large model scheduling layer, feature vectors are analyzed through a multi-modal large model center based on an industry knowledge graph, an algorithm is dynamically matched, and an initial decision strategy is generated; the method comprises the following steps of: establishing a parameterized template library, and supporting security cooperative training of a third-party algorithm scheduling engine on a third-party algorithm, and deploying a lightweight digital twin engine in a twin engine layer: establishing the parameterized template library: pre-defining three templates of geographic space, equipment assets and business processes; deploying the twin model to an edge node by adopting a knowledge distillation method; in the interactive application layer, loading the BIM / GIS model in a lightweight manner through a low-code tool, and completing scene construction through a dragging component; and rendering a twin state in real time through a three-dimensional cockpit, analyzing a natural language instruction and performing corresponding operation.

A system for real-time detection and mitigation of morphing malware in high-density edge networks, consisting of: a data acquisition unit configured to receive, normalize, and encode multimodal telemetry data streams originating from at least one of the following domains: network traffic, process behavior, system call sequences, binary instruction traces, and control flow graphs; the data acquisition unit is further configured to compute feature embeddings over sliding time windows and apply privacy-preserving redactions prior to storage; a generative neural processor that is operationally coupled to the data acquisition unit and configured to generate synthetic morphing malware variants by learning probabilistic transformations of previously observed malicious data representations, maintaining semantic functionality while varying structural and behavioral features; a discriminative neural processor trained adversarially with the generative neural processor, wherein the discriminative neural processor is configured to detect morphing malware by evaluating a probability distribution over multimodal telemetry embeddings and classifying anomalous process and flow behaviors in real time; a coordination processor that is communicatively connected to both the generative neural processor and the discriminative neural processor and is configured to orchestrate adversarial co-training, regulate detection thresholds, calculate reinforcement-based penalties for false negative results, and trigger countermeasures as soon as a detection confidence level exceeds a predefined adaptive threshold; a secure, system-integrated inference and enforcement unit configured to perform low-latency countermeasures at the network edge, including selective packet filtering, flow isolation, process interruption, or system microsegmentation, based on instructions from the coordinating processor; and a hardware-embedded security enclave that is embedded in the system and configured to store cryptographic keys, neural model parameters, and integrity affirmation data to ensure the confidentiality, authenticity, and immutability of model artifacts and policy configurations.

The invention relates to the technical field of intelligent dialogues, and discloses a dialogue marketing strategy optimization method based on artificial intelligence, and the method comprises the following steps: collecting the voice, text, facial expression and physiological signals of a user in real time through a multi-modal perception assembly of a terminal device, constructing a dynamic emotion map, and extracting a multi-modal feature vector; a cross-modal information processing module is utilized to align the multi-modal data through a comparative learning algorithm, and causal relationship description of user behaviors and strategies and strategy risk scores are generated; cooperatively training a global strategy model through differential privacy and homomorphic encryption technologies; combining the dynamic emotion map and a causal model to generate an emotion adaptive dialogue script, and optimizing strategy selection through a reinforcement learning algorithm; and dynamically updating the emotion map, the risk score and the global model through a closed-loop feedback mechanism to form a real-time optimized strategy generation system. According to the invention, the practicability of artificial intelligence to real-time services can be improved.

The invention relates to the field of finance, and discloses an AI-based investment project feasibility intelligent analysis and decision support system, which comprises a multi-modal data fusion engine used for extracting a feature vector of an unstructured text through a BERT variant model, carrying out joint coding on the feature vector and a structured data feature, and outputting fused multi-modal data; the dynamic knowledge graph construction module is connected with the multi-modal data fusion engine; the deep reinforcement learning decision module comprises a feasibility prediction network and a risk assessment network which are cooperatively trained; an interpretable AI module; and a closed loop feedback module. By means of a multi-modal data fusion engine and a deep reinforcement learning decision module, automatic processing and intelligent analysis of data are achieved, and the analysis time of a single project is greatly shortened to 8 minutes. The high-efficiency processing speed enables the investment institution to quickly respond to the market change, timely grasp the investment opportunity, and occupy the decision-making precedence in the financial market which changes instantaneously.

The invention discloses a cross-modal large model construction method and system based on track spatio-temporal characteristics, and belongs to the crossing field of artificial intelligence and dynamic spatio-temporal data processing, and the method comprises the steps: carrying out the sliding sampling and spatial distribution difference judgment through the semantic dynamic segmentation of multi-scale track spatio-temporal data, and generating spatio-temporal data blocks with consistent semantics; designing a space-time encoder of a hybrid architecture, extracting track time sequence association and spatial features, and unifying dimensions; constructing a text space-time fusion mechanism, and dynamically adapting cross-modal features by means of an anchor interface and gating fusion; a staged instruction fine tuning strategy is adopted, semantic alignment of space-time and text features is optimized firstly, then model top-layer parameters are trained cooperatively, complex scene adaptation is enhanced in combination with instruction difficulty progression and hard sample mining, and space-time constraint regular terms are introduced to guarantee output rationality. According to the method, high-precision cross-modal reasoning capability is provided for scenes such as track analysis and track prediction.

The invention relates to a hyper-dimensional calculation model cooperative training method and device based on federated learning, and belongs to the field of distributed machine learning and data privacy protection. The device comprises a central server and M distributed client devices which communicate through network connection, the method comprises the following steps: S1, initializing parameters; s2, broadcasting parameters; s3, local super-dimensional calculation model parameters are updated; s4, the similarity between the two is calculated, and the local training data volume is counted; s5, uploading to a server; s6, calculating parameters of the global super-dimensional calculation model based on a dynamic weighted aggregation method; s7, judging whether convergence is carried out or the maximum training round number is reached; if yes, training is ended, and global super-dimensional calculation model parameters are output; otherwise, returning to the step S3; and S8, the user executes the classification task and outputs a prediction classification result. According to the method, the training efficiency and the global model performance of federal learning under a super-dimensional calculation model can be improved on the premise of ensuring data privacy and security.

The invention provides a cross-modal feature fusion method based on joint attention, which combines multi-modal data from different sources to predict the relative position of a target to be tracked, and comprises the following steps of: firstly, designing a cross-modal data representation network model for the data of different modals; generating a time sequence high-dimensional implicit code retaining key dependency information from the original data through a code conversion algorithm; secondly, a heterogeneous attention network unit is designed, and features of time sequence high-dimensional implicit codes corresponding to different modal data are extracted through two different forms of attention; and finally, through a cooperative training mechanism, information complementation among different modal data is utilized to generate the most effective feature representation for a time sequence prediction model, and accurate prediction of the moving position of the tracked target is realized.

The invention belongs to the field of crossing of remote sensing visual questioning and answering and artificial intelligence, and particularly relates to a remote sensing visual questioning and answering method based on a large language model and a multi-level attention mechanism. Comprising the following steps: 1, constructing a MogaNet-based semantic segmentation network; 2, text coding is carried out through a large language model; 3, constructing a mixed attention guiding module; 4, constructing a bidirectional gating cross attention module; and 5, constructing and cooperatively training a remote sensing visual question-answer network. According to the MLaVQA framework provided by the invention, efficient and accurate visual question-answering processing of remote sensing images can be realized, when large-scale multi-modal remote sensing data is processed, complex semantic questions can be quickly answered, fine monitoring and intelligent management of earth resources are realized, the data utilization efficiency and decision support capability are remarkably improved, and the method is suitable for large-scale multi-modal remote sensing data processing. And technical support is provided for natural resource management, ecological protection and sustainable development.

The invention discloses a robust federated learning method for processing heterogeneous noise and non-independent identically distributed data, and belongs to the technical field of federated learning. The method provides a robust learning framework of two-stage client quality perception. The method comprises the following steps of: 1, constructing a category-level loss vector and clustering by using a Gaussian mixture model, and accurately dividing a clean and noise client set; stage 2, performing differential training: performing standard training on the clean client; dual-network cooperative training, dynamic sample screening and exchange, and a self-distillation and entropy regularization mechanism are introduced to a noise client, so that robust learning is realized; in the global aggregation stage, a distance sensing weighting strategy is further adopted to dynamically suppress the influence of a noise client; according to the method, original data does not need to be shared, the robustness and generalization performance of the federated learning model in the coexistence environment of heterogeneous noise and non-independent identically distributed data can be effectively improved, and the method has wide application value in the fields of medical images, financial risk control and the like.

The invention relates to an access control strategy automatic generation system and method based on multi-task learning, and the method comprises the steps: carrying out the word segmentation, cleaning and embedded vector conversion of an original access control text through a data preprocessing module, and constructing a normative input format; the feature sharing layer module is used for extracting deep semantic features of a text through multi-layer bidirectional coding and an attention mechanism and providing unified representation for downstream tasks; the access control statement identification module is used for judging whether each sentence in the text is an access control statement or not and realizing automatic identification of strategy related contents; and the attribute extraction and annotation module is used for annotating words in the access control statements and extracting subject, object and operation access control attributes. A word coding layer and a sentence coding layer are shared, local and global attention mechanisms are combined, key information of a text is extracted, the semantic understanding ability is enhanced, and cooperative training of statement recognition and attribute extraction is achieved; a conditional random field CRF structure is used for sequence labeling, and the structural rationality of attribute labels is ensured.

The invention relates to the technical field of intelligent coal mine safety production, and discloses an intelligent coal mine safety early warning method and system based on deep learning, and the method comprises the steps: constructing a difficulty evaluation function, and achieving the progressive learning from simple to complex; based on a difficulty assessment result, a model-independent meta-learning algorithm is realized, so that the model quickly adapts to new mining area characteristics; constructing a privacy protection federated learning framework by using the meta-learning model, and realizing multi-mining-area cooperative training; a continuous learning module is constructed, and original experience is reserved when new knowledge is learned; constructing a meta-knowledge evaluation module to realize cross-mining-area safety knowledge sharing; according to the invention, the security risk identification accuracy is improved; multi-mining-area cooperative training is realized on the premise of protecting data privacy; the method has continuous optimization capability and effectively solves the problem of model drift; and efficient sharing and migration of cross-mining-area safety knowledge are realized.

The embodiment of the invention provides an end-cloud cooperative training method, system and device in the field of artificial intelligence, and can be used for stripping an embedding layer to terminal training in an end-cloud cooperative training process so as to improve the privacy security of end-side data. The method comprises the steps that a client uses training data as input of a representation layer to obtain a representation vector, the representation layer is used for obtaining a vector corresponding to the input data from a representation word list stored in the client, and the representation word list is stored in the client; the client sends the representation vector to the cloud platform, so that the cloud platform takes the representation vector as the input of a language model deployed at the cloud platform side to obtain an output feature; the client receives an output feature sent by the cloud platform, wherein the output feature is obtained by inputting a representation vector into a language model by the cloud platform; and the client calculates a loss value by using the output feature, updates the representation layer according to the loss value, and sends the loss value to the cloud platform, and the loss value is used for the cloud platform to update the language model.

The invention relates to a multi-agent reinforcement learning training method, system and device driven by an unreal engine and a medium. The method comprises the following steps: acquiring core attribute data of multiple agents, including real-time communication bandwidth, initial communication topology configuration, training stage progress, single agent computing power and cluster overall computing load state; analyzing the real-time communication bandwidth and the initial communication topology configuration to form an optimized topology configuration scheme; in combination with an optimized topology configuration scheme and a training stage progress, generating a cooperative training progress state by quantifying progress differences among intelligent agents; extracting load balance data according to the single agent computing capability and the cluster load state; integrating load balance data, a collaborative training progress state and an optimized topology configuration scheme, and constructing a multi-layer collaborative training framework; and integrated optimization is carried out to generate a cooperative training scheme for guiding multi-agent communication, training and resource allocation, so that the training efficiency is effectively improved, the resource load is balanced, and the cooperative synchronism is enhanced.

The invention discloses an artificial intelligence network optimization training system and method based on deep learning, and the system comprises a request analysis module which is used for determining a training task type and an initial model hyper-parameter, and obtaining an available edge device list; the parameter monitoring module is used for determining a model index monitoring threshold value, monitoring model indexes in the training process in real time and adjusting corresponding model hyper-parameters; the multi-target optimization module is used for initializing target weights and calculating a multi-target optimal solution set; the self-adaptive training module is used for generating a training strategy and optimizing and adjusting the training strategy; and the cooperative training module is used for allocating training tasks to the available edge devices and feeding back model indexes in the training process to the parameter detection module. According to the invention, the training mode is dynamically adjusted according to the specific parameters of the artificial intelligence network, multi-objective optimization is realized, distributed training of edge devices is supported, and the robustness of the model is improved, so that the training precision and efficiency are improved, and the application range is expanded.

The invention provides a marine image data identification method and system. The method comprises the following steps: step 1, inputting an original marine image; 2, a visual feature extraction module extracts hierarchical features through a multi-scale convolutional neural network; 3, pre-training an LLM semantic reasoning hierarchical attention mechanism semantic mapping multi-task cooperative training framework by the language model agent module; and 4, a dynamic strategy optimization module reinforces learning optimization strategies to adjust identification parameters in real time, and generates a preliminary identification result. And 5, when the user needs to optimize, combining user feedback, a data enhancement interactive learning mechanism and a semantic consistency data enhancement optimization model to realize model updating and parameter adjustment, and returning to the step 2 again. And when the user does not need optimization, outputting a final identification result of the ocean image. According to the invention, a visual coding technology is combined with the semantic understanding capability of LLM, and a recognition solution capable of adaptively processing a multi-modal ocean image is constructed.

The invention discloses an edge caching method for federal deep reinforcement learning based on semantic enhancement, and the method comprises the steps: obtaining request original data and a candidate content set, and carrying out the multi-modal feature extraction; calculating a cache utilization rate, request diversity and a dynamic semantic matching weight; calculating the semantic similarity between the request content and the candidate content, and performing adjustment in combination with a dynamic semantic matching weight to obtain a weighted semantic similarity; generating action probability distribution according to the state of the node, and fusing the action probability distribution with the semantic attention weight to obtain action preference distribution; action execution is selected according to the action preference distribution, and rewards are evaluated; and periodically executing federal cooperative training optimization, locally training a semantic perception strategy network by adopting an SAC algorithm, executing multi-factor dynamic weighted federal aggregation to update a global model, and distributing the global model to each node. According to the method, the edge cache hit rate and the semantic hit rate can be improved, the service response delay is reduced, and the robustness of heterogeneous nodes and a complex network environment is enhanced.

The invention discloses a method for reducing large model illusion of a question answering system based on knowledge graph representation learning, which comprises the following steps of: firstly, based on Ollivier-Ricci curvature calculation, embedding different structures in a knowledge graph into a plurality of approximate geometric spaces so as to analyze geometric patterns in data; secondly, generating a new entity representation by aggregating neighbor information, and transmitting and fusing information in different geometric spaces by using indexes; and finally, a dynamic curvature adaptive adjustment strategy is adopted to promote cooperative training and efficient fusion of multi-geometric space representation. According to the method, a complex semantic relationship is accurately described by combining geometric space and knowledge graph representation learning, geometric distortion in an embedding process is effectively reduced, and a structured knowledge system is constructed. According to the method, the accuracy and credibility of a large model in knowledge reasoning are improved, the illusion problem caused by incomplete or contradictory knowledge representation is solved, and powerful support is provided for a question and answer system.

The invention discloses a voltage sag state estimation method and system based on complex task decomposition reasoning, and the method comprises the steps: firstly solving actual problems in a physical world through guiding LLM by employing a mathematical tool, and giving text output fitting the reality through combining the mathematical tool and the feedback of a thinking tree architecture simulating human thinking behaviors; in order to solve the practical problem of the physical world, the invention constructs a physically-driven DGCN model, and the output of a mathematical tool better fits the practical situation by combining a multi-layer graph convolution layer and taking a power grid topology connection relationship as a physical constraint condition. And finally, introducing an improved DGCN model based on cooperative training and cross-power-grid regional knowledge migration, and improving the overall generalization and adaptability of LLM when voltage sag state estimation is carried out on unknown power grid topology.

The invention relates to the technical field of intelligent operation of a power distribution network, and particularly discloses a power distribution network load prediction and scheduling system based on multi-modal data and semi-modal learning, and the system comprises the steps: dividing monitoring nodes into label nodes and label-free nodes through collecting the multi-modal data of the power distribution network in real time; constructing a power distribution network causal knowledge base containing load type and meteorological condition causal association by using the labeled node data; causal association is used as priori knowledge to configure a semi-supervised meta-learning model comprising a main prediction network and a knowledge query network; carrying out cooperative training by taking the labeled data as a supervision signal and taking the unlabeled data as a supplement; in the prediction stage, the initial load prediction of the main prediction network and the target causal association retrieved by the knowledge query network are combined, and physical rule conformity correction is achieved through calculation; and performing multi-level security check based on the corrected load prediction result, and automatically executing a hierarchical scheduling strategy according to the early warning level.

The invention discloses an urban traffic cooperative scheduling method and system based on a large model and multiple agents. The method comprises the steps that natural language task description is converted into a task semantic graph and a structured prompt; generating a plurality of strategy candidates by using a large language model, and performing language scoring and reasoning arbitration; based on the scoring result, selecting an optimal strategy for multi-agent execution; behavior execution data are collected and evaluated and fed back, and strategy closed-loop updating is achieved through parameter optimization and Prompt fine tuning; and a distributed task embedding mechanism and a lightweight migration module are combined, so that the multi-task adaptability and the training efficiency of the system are improved. The cooperative training system constructed by the invention has the capabilities of natural language interaction, strategy interpretability, behavior controllability and task migration, and is suitable for various complex urban tasks such as traffic jam dispersion, emergency response, signal lamp linkage and the like.

The invention relates to the field of artificial intelligence, can be applied to business system platforms of finance, medical health and the like, and discloses a classification method, device, equipment and medium based on cross-modal and dynamic distillation, and the method comprises the steps: collecting annotated image samples and unannotated text samples, and carrying out cross-modal bidirectional generation type enhancement processing and cross-modal sharing coding; performing teacher-student model cooperative training of dynamic distillation weight through annotated image samples and an extended data set; modal difference analysis is carried out according to the feature vectors of the unified semantic space, and distribution difference analysis is carried out on generated data and real data; and calculating a discriminant loss result according to each item of data so as to update model parameters, and obtaining a trained multi-modal classification model for classification processing. Small sample data distribution is effectively expanded through cross-modal generation of supplementary data and sharing of codes, the model convergence speed is increased through a dynamic knowledge distillation mechanism, the training efficiency is improved while the training effect is ensured, and the prediction classification accuracy is improved.

Provided in the embodiments of the present application are a client-cloud collaborative training method, a system and an apparatus in the field of artificial intelligence, which can be used for offloading an embedding layer to a terminal for training during client-cloud collaborative training, thereby improving the privacy security of client side data. The method comprises: a client uses training data as an input of a representation layer, so as to obtain a representation vector, the representation layer being used for obtaining from a representation word list stored in the client a vector corresponding to the input data, and the representation word list being stored in the client; then, the client sends the representation vector to a cloud platform, such that the cloud platform uses the representation vector as an input of a language model deployed at the cloud platform side, so as to obtain an output feature; the client receives the output feature sent by the cloud platform, the output feature being obtained by inputting the representation vector into the language model by the cloud platform; and the client uses the output feature to calculate a loss value, updates the representation layer on the basis of the loss value, and sends the loss value to the cloud platform, the loss value being used for the cloud platform to update the language model.

The invention discloses a multi-layer semantic perception, distillation and semi-supervised cooperative training target detection method, and the method specifically comprises the following steps: constructing a data set: extracting a first part of images from image data, marking the first part of images to construct a supervised target detection data set, and taking the remaining images as an unmarked image data set, the data volume of the unlabeled image data set is greater than that of the supervised target detection data set; teacher model optimization: performing supervision training on the teacher model on the supervised target detection data set; constructing a teacher model, and executing self-distillation training on the teacher model; pseudo labels are generated for the unlabeled images through a teacher model, and adaptive screening is carried out based on confidence distribution; mapping a pseudo label to a strong enhanced sample through enhanced geometric transformation, carrying out semi-supervised training by using the strong enhanced sample and the pseudo label, and introducing a feature layer distillation constraint at the same time; optimizing a student model; and outputting the target detection model obtained through training.

The invention belongs to the technical field of artificial intelligence and natural language processing, and particularly relates to a rule and regulation question and answer illusion suppression fine tuning method based on an anti-fact negative sample. According to the method, three types of anti-fact negative samples of error units, error thresholds and error versions are constructed, and cooperative training is carried out in combination with boundary maximization loss of evidence perception, numerical value consistency regularization, head rejection and temperature calibration and a controlled generation mechanism, so that the discrimination capability of the model on high-similarity error information is effectively improved; the accuracy of numerical answers is ensured, and the answers can be reliably rejected when evidences are insufficient or unreliable, so that the illusion phenomenon of a large language model in professional questions and answers is remarkably inhibited.

The invention discloses a cooperative training data generation method for an adjustment and optimization problem and electronic equipment, and the method comprises the following steps: carrying out the filtering processing of an input code description text, and obtaining input data; injecting the problem data into the input data to generate first adversarial sample data containing the problem data; gradient information and disturbance amplitude are calculated, and second adversarial sample data are obtained through adjustment and optimization; designing a shared encoder based on the second adversarial sample data to obtain problem data and normal code data features; carrying out separation processing on the two in a feature space; according to the separated features, the shared encoder is promoted to learn through feature comparison, and finally a model generated by code training data is obtained; and designing an evaluation index for judging whether a generation result of the model in the problem data injection scene is correct and stable, thereby completing model training of code training data generation. According to the method, generalization and robustness of the large language model in the process of processing actual problems are improved.

The invention discloses an English listening and speaking interaction intelligent training system based on a large language model, and relates to the technical field of English learning, and the training system comprises a user capability graph module, an interaction feedback module, a personalized learning path generation module, a scene simulation module, an interaction rhythm control module and a community cooperative training module. According to the method, the pragmatic ability, the cross-language communication ability and the linguistic ability are jointly incorporated into the training content, and the adaptive training content is generated in combination with a real cross-language communication scene, so that a user synchronously masters expression specifications and culture adaptation key points in different scenes in the training process; the effect that training content and real communication requirements are deeply matched is achieved, the user is helped to accurately adapt to a native English scene and real phrase habits of people, the requirements of an actual application scene are met, and substantial improvement from basic language knowledge mastering to real communication ability is achieved.

The invention discloses a short text matching method based on enhanced contrast learning and multi-task optimization, and belongs to the technical field of natural language processing. According to the method, a multi-source data set is constructed by fusing general field and medical field texts, a generative language model is adopted to generate a high-quality difficult case to form a training triple, and a character-level noise enhancement mechanism based on a medical similar character dictionary is introduced to simulate an OCR error; aLBERT is used as an encoder, supervised comparative learning and mask language modeling tasks are jointly optimized, multi-task cooperative training is achieved through dynamic gradient balance, and finally a short text matching model with high semantic discrimination ability and noise robustness is obtained. On a test set containing 50% OCR noise in the cardiovascular field, the accuracy rate of the method reaches 94.3% and is improved by 12.2% compared with that of a traditional BERT-base model.

This application discloses a hierarchical obstacle avoidance method and system based on collaborative training of the PPO-TD3 algorithm. This application utilizes a dynamically adaptive hierarchical decision-making system, namely a collaborative architecture of the PPO and TD3 algorithms, to achieve intelligent obstacle avoidance in dynamic environments through a hierarchical reinforcement learning framework. In simple environments, only the high-frequency control network of the underlying TD3 algorithm is run to improve response speed, while in complex scenarios, the complete hierarchical architecture is activated to ensure decision quality. Simultaneously, a semantic instruction interface is innovatively designed, enabling the abstract instructions output by the high-level PPO decision-making layer policy network to be intelligently parsed into executable actions by the underlying TD3 execution layer control network in conjunction with real-time physical constraints. Combined with a hybrid training framework integrating simulation training, human demonstration, and real-world learning, and a resilient safety mechanism based on dynamic risk assessment, this achieves synergistic optimization of safety and operational efficiency in complex dynamic environments.