546 results about "Robustification" patented technology

An advanced model management platform for optimizing and securing generative artificial intelligence systems such as large language models (LLMs) and diffusion models. The platform incorporates various techniques to address the limitations of current generative AI systems, such as hallucination, lack of validation, security vulnerabilities, and inadequate model management. The system employs reinforcement learning algorithms for model optimization, retrieval augmented generation (RAG) for hallucination mitigation, domain-specific validation against expert knowledge, model distillation and similarity scoring for security, adversarial training for robustness, and attention mechanism search and model blending for advanced management and neuro symbolic AI routine combinations. By integrating these techniques, the platform significantly improves the performance, reliability, and security of generative AI across a wide range of tasks and domains leveraging the best elements of symbolic and connectionist techniques alongside automated planning and modeling simulation.

The invention discloses a multi-modal fusion AGV dynamic path planning and cluster scheduling system, and relates to the technical field of multi-modal perception and data fusion, and the system comprises a multi-modal perception module which generates a dynamic obstacle confidence map through multi-source data fusion in combination with a hardware-level time synchronization and Transform feature fusion network; the dynamic path planning module adopts an improved rolling window algorithm, integrates an LSTM space-time conflict prediction model and an adaptive weight cost function, and realizes dynamic obstacle trajectory prediction and non-oscillation global path generation; the cluster scheduling control module is used for optimizing multi-AGV task allocation and conflict resolution in combination with a dynamic priority preemption mechanism and digital twinborn simulation rehearsal based on a distributed contract network protocol of edge computing; and the data conflict resolution module is used for triggering a multi-modal re-calibration process through confidence weighting and sliding window time sequence verification. According to the system, in logistics storage and intelligent manufacturing scenes, the dynamic obstacle avoidance success rate and the robustness and operation efficiency of an AGV cluster are improved.

The invention discloses a building digital twinning three-dimensional reconstruction method and system based on a large model, and relates to the technical field of building digital twinning and three-dimensional modeling fusion, and the method comprises the steps: laying multi-source equipment in a building, collecting building multi-angle data, and carrying out type distinguishing and structured preprocessing. An adaptive deep network model is constructed and trained to identify building features, and the robustness of the model is improved by adopting a multi-type enhancement strategy. Point cloud splicing and coordinate unification are realized by using a fusion algorithm, and twin model updating and synchronization are realized based on a multi-dimensional threshold. According to the method, high-quality input is established for a depth model; building structure feature recognition has the advantages of robustness and precision; and the twinborn model has dynamic updatable capability. The three parts form a complete closed loop, the technical bottlenecks of a traditional reconstruction method in precision, stability and renewability are finally broken through, and a new building digital twinning three-dimensional modeling path facing a complex scene is achieved.

The invention discloses an enterprise big data mining method and system based on artificial intelligence, and the method comprises the steps: carrying out the dynamic mode alignment through employing a multi-mode hypergraph neural network according to enterprise multi-source heterogeneous data, and generating a time-space consistent multi-mode joint embedded tensor; inputting the multi-modal joint embedding tensor into an orthogonal adversarial manifold learning module, and generating a low-dimensional compact semantic embedding vector with enhanced category separability; performing space-time causal association mining on the semantic embedding vector, and outputting a space-time causal meta-path map containing the recessive commercial logic; and inputting the space-time causal element path map into a dynamic game adversarial interpretation framework, and finally outputting an enterprise-level intelligent decision map with anti-factual robustness. By utilizing the embodiment of the invention, the multi-modal data can be efficiently integrated and intelligently analyzed, and the accuracy and effectiveness of the mining result are improved.

The invention discloses a multi-model collaborative operation method based on large-model efficient training, and relates to the technical field of artificial intelligence, and the method comprises the steps: firstly, monitoring model training resource indexes, such as activation tensor memory occupation and gradient calculation intensity, and carrying out the optimization through a three-stage optimization mechanism; constructing a heterogeneous resource space-time diagram to predict resource requirements, and combining multi-modal feature dynamic fusion and prediction error online compensation to improve prediction accuracy; then dynamically allocating resources, determining a migration priority by evaluating a training income gradient, an energy consumption efficiency factor and the like, and implementing video memory block management and pipeline scheduling; carrying out cross-model hierarchical knowledge distillation and self-adaptive cooperative training, and adjusting a cooperative mode according to model similarity; the problems of heterogeneous model resource competition and isolated knowledge precipitation are effectively solved, the resource utilization rate is improved, repeated calculation is reduced, the overall performance and robustness of multi-model collaborative operation are enhanced, the training cost is reduced, and the training efficiency is improved.

The invention discloses a power grid load prediction and scheduling optimization system based on artificial intelligence, particularly relates to the technical field of power system automation, and solves the technical problems of low power grid load prediction precision, poor scheduling strategy robustness and insufficient source grid load storage coordination in the prior art. Multi-source heterogeneous data space-time alignment is realized by constructing a data acquisition layer based on edge calculation, a load prediction result is generated by adopting an AI prediction module fused by a graph convolutional network and an attention mechanism, and a source-network-load-storage collaborative scheduling scheme is generated through a multi-target risk hedging optimization algorithm. And closed-loop optimization is realized by using digital twinborn pre-check and incremental learning. And finally, the load prediction accuracy, the scheduling decision reliability and the system adaptive capability in the new energy access environment are improved.

The invention discloses a mainboard intelligent fault diagnosis method and system based on artificial intelligence, and relates to the technical field of fault diagnosis. Comprises: constructing a task number chain according to a calling sequence relationship; continuously sampling the running state parameters, and constructing a task state feature vector table in combination with the current task number; state feature vector offsets between adjacent task units are calculated pair by pair, and if the offsets exceed a set tolerance threshold value, the offsets are added into an offset anomaly table to serve as doubtful point marks; constructing an abnormal propagation path map through path reasoning, and calculating a path risk score based on the structure association degree and the historical state interaction frequency between the task units; and for the abnormal boundary with the path risk score exceeding a preset value, selecting an optimal alternative number under structural constraint from the task number chain, and deploying the optimal alternative number to a corresponding number position of the current task unit. According to the method, the dynamic substitution of the fault boundary and the local reconstruction of the system state are realized, and the robustness and the continuous operation capability of the mainboard task execution system are enhanced.

The invention discloses a reverse question guide question answering implementation method and system, and belongs to the technical field of artificial intelligence and natural language processing. Context-aware intention dynamic correction is realized through a three-level intention classification system, and cross-modal knowledge matching is realized by adopting a distributed semantic index technology; based on the reinforcement learning strategy, optimizing a cooperative work mechanism of the dialogue strategy and the knowledge base; comprising the steps of intention recognition: analyzing a session of a user by using an intention recognition model, and constructing a three-level intention classification system based on deep semantic understanding, including main class recognition, fine-grained analysis and context perception; question rewriting: constructing a dynamic rewriting engine to rewrite the user question; recalling and cleaning multi-source item knowledge; generating a reverse question; locking items and acquiring item data; generating questions and answers. According to the method, the robustness, the real-time performance and the scene adaptation capability of a professional question answering system can be improved, and the government affair service question answering accuracy, the intention recognition precision and the cross-region recommendation adoption rate are improved.

The invention relates to the technical field of traffic planning, in particular to a transfer connection optimization method in multi-mode traffic integrated planning, which comprises the steps of data acquisition and demand analysis, multi-mode transfer network topology design, multi-mode timetable collaborative optimization, transfer service system development and facility transport capacity dynamic matching. According to the transfer connection optimization method in the multi-mode traffic integrated planning, a dynamic and static combined multi-source data acquisition framework is constructed based on a multi-dimensional data fusion system, the data precision can be improved, the network can be perceived in real time, association modeling of weather events and demand fluctuation is realized based on construction of a multi-agent simulation system, and the optimization of the transfer connection in the multi-mode traffic integrated planning is realized. A dynamic weight mechanism is introduced based on a hypergraph model, the robustness of the algorithm is improved, a layered optimization framework processes global optimization through upper-layer integer programming, dynamic adjustment is achieved through lower-layer deep reinforcement learning, a buffer time intelligent insertion strategy can improve the late absorption capacity, and a facility utility function model is established to improve the transfer efficiency.

The invention discloses an intelligent electric equipment monitoring and optimizing method, and relates to the technical field of electric power, and the method comprises the steps: collecting high-dimensional voltage-current time sequence data and transient event marks, calculating topological invariant features, inputting the topological invariant features to a lightweight neural network model, and recognizing the features of all electric equipment; performing abnormal attribution and anti-fact energy efficiency prediction by using causal reasoning and dynamic regularization regression based on the identified characteristics of each electric device, and constructing a multi-objective optimization function through the abnormal attribution and anti-fact energy efficiency prediction; and based on the multi-objective optimization function, generating an equipment operation scheduling strategy through a deep reinforcement learning agent, based on the operation scheduling strategy, sending a control instruction to the electric equipment, and collecting an operation result feedback in real time for optimization and updating. According to the method, through fusion of topological features, causal reasoning and safety reinforcement learning, the precision, robustness and safety of monitoring and optimization of the intelligent electric equipment are improved.

The invention relates to the technical field of digital twinning, in particular to a digital twinning fused power grid fault prediction system and method, and the system comprises a state sensing module, a path deduction module, a twinning mapping module and a prediction execution module. According to the method, multi-source heterogeneous sensing data are fused, a space-time alignment algorithm is executed to construct a structural state tensor, Laplacian operation is combined with a stress wave propagation model to analyze non-rigid deformation characteristics, a load transfer path atlas is generated to reveal recessive mechanical association, Young modulus is dynamically corrected through a finite element grid to form a structural degradation coefficient matrix, and the structural degradation coefficient matrix is obtained. The time sequence difference and the long short-term memory network are combined to capture the equipment degradation trend, the Kolmogov test is used to quantify the fault space-time distribution probability, the power grid equipment jumps from single threshold early warning to multi-physics coupling prediction, the fault prediction space-time resolution and dynamic adaptability are improved, the artificial experience dependence is reduced, and the fault prediction efficiency is improved. And the robustness of the complex working condition system is enhanced.

The invention provides a text code generation method based on multi-modal semantic embedding and a dynamic knowledge graph, and belongs to the technical field of artificial intelligence. According to the automatic code generation method based on multi-modal semantic embedding, the dynamic knowledge graph, constraint-driven code generation and context-aware repair, the semantic understanding precision and the code generation quality are remarkably improved by integrating the technologies of multi-head Transform, the graph neural network, reinforcement learning optimization, genetic algorithm sequence adjustment, the generative adversarial network and the like. According to the method, the knowledge graph can be dynamically constructed to enhance structured semantic modeling, high-quality codes conforming to specific industry standards are generated, functionality, efficiency and conciseness are considered, meanwhile, the iteration cost is reduced through an efficient error positioning and repairing mechanism, and the method is suitable for large-scale popularization and application. And the robustness and the flexibility of the system in diversified scenes are improved by utilizing adaptive weight optimization and multi-target balance.

The invention discloses a resume analysis method and system based on multiple large language models, and belongs to the technical field of large language models.The resume analysis method and system based on the multiple large language models.The resume analysis method and system based on the multiple large language models comprise the following specific steps that firstly, model parallel analysis is conducted, simultaneously inputting the data into at least two heterogeneous large language models through an application program interface; and each large language model independently performs information extraction and analysis according to the model structure and the training data of the large language model, and outputs a structured data result containing a plurality of preset fields. Through the multi-model parallel analysis and conflict re-judgment mechanism, the misjudgment risk of a single model is effectively reduced, the robustness of the whole system is improved, the resume analysis accuracy is remarkably improved, the model pool is automatically optimized and updated through the dynamic scoring mechanism, and the problem that the model is difficult to select and update is solved.

The invention relates to the field of computer systems, in particular to an operation stability testing method and system oriented to a heterogeneous computing system, and the method comprises the steps: constructing a system overall operation logic model comprising a resource topological structure, a task scheduling rule and a communication link matrix; identifying a key communication path according to the communication link matrix, analyzing a resource competition hotspot in combination with a task scheduling rule, and establishing an interference model library; calling a specific model instance in the interference model library according to a test target, configuring injection parameters, and dynamically applying interference in a system operation process; key performance indexes in the system operation process are collected in real time; based on the real-time monitoring data, a multi-index weighted scoring algorithm is adopted to calculate an overall stability coefficient, a robustness grade and a self-healing capability index of the system, and a stability evaluation result is generated; and generating a visual report according to the topological structure of the operation logic model, and supporting comparison and analysis of historical versions. Unified centralized management is realized, and test efficiency and evaluation accuracy are improved.

The invention provides a distillation method and device based on AI large model knowledge and electronic equipment, and belongs to the technical field of artificial intelligence, in the method, a dynamic adjustment mechanism is adopted for the temperature, and along with the increase of a training period, the preset initial temperature corresponding to the training period is gradually reduced; besides, in each training period, the preset initial temperature is finely adjusted based on the distillation loss, so that the target temperature corresponding to each training sample is obtained, namely the temperature most suitable for the current distillation effect. The target distillation loss and the student loss obtained by recalculating based on the target temperature are more scientific, so that the total loss for updating the weight parameters of the student model is more accurate and robust, the student model is prompted to learn the knowledge of the teacher model more effectively, and a scientific and efficient learning process is learned. The finally obtained student model has good generalization ability, and the robustness of the model is significantly enhanced, so that the overall effect of knowledge distillation is significantly improved.

The invention discloses a federated learning method based on quantum transfer learning, relates to the cross technical field of quantum computing and federated learning, and aims to solve the problem of performance bottleneck of traditional federated learning under non-independent identically distributed data. The local model of each client is fused with a classic convolutional layer and a quantum convolutional layer; during local training, firstly, classic features are extracted by a classic convolutional layer, and then the classic features are coded and input into a quantum convolutional layer to generate quantum features; and the two features are spliced and then classified, and model parameters are uploaded. And the server side adopts a FedAvg algorithm to aggregate parameters so as to update the global model. According to the method, a transfer learning mechanism is introduced to optimize the initialization of the quantum layer, and an experimental result shows that the classification performance and robustness of the method are remarkably superior to those of a traditional federated CNN model on a non-IID image data set, and the method is particularly suitable for a privacy protection cooperative computing scene.

The invention discloses a target multi-attribute identification method based on feature decoupling and cross-task collaboration, and belongs to the technical field of computers of specific calculation models, and the method comprises the following steps: firstly, extracting the initial features of each task through a lightweight backbone network, carrying out feature decoupling in a subspace, and according to the cross-task feature similarity, carrying out feature extraction; according to the target multi-attribute identification method based on feature decoupling and cross-task collaboration, an orthogonal constraint weight is dynamically adjusted, then mutual information confrontation minimization is adopted to further suppress statistical dependence between tasks, task residual errors are injected in a cross-task feature aggregation stage, differentiation enhancement is achieved, and finally unified joint feature representation is formed. According to the method, subspace statistical independence is realized, independence and necessary collaborative information are considered, a stable basis is provided for subsequent fusion, statistical dependence between tasks is further suppressed through mutual information confrontation minimization, complementation information is reserved while independence is ensured, and feature discrimination and robustness are improved.

The invention provides a cue word optimization method and system based on an approximate sub-module function and continuous learning, and relates to the technical field of artificial intelligence multi-mode perception.The method comprises the steps that a candidate cue word set is constructed, a combined objective function based on the property of the approximate sub-module function is designed, and a candidate cue word set is constructed; solving the combined objective function by adopting a greedy selection algorithm combining random disturbance, multi-round iteration and a task self-adaptive mechanism, realizing optimal selection of a candidate cue word set, gradually selecting a cue word with the maximum gain from the candidate cue word set, adding the cue word into an optimal subset, and when a new cue word is selected, selecting the cue word with the maximum gain into the optimal subset. And if the target cue words are selected, carrying out local optimization once, after all the target cue words are selected, carrying out joint optimization on all the cue words in a continuous space by adopting an alternate optimization strategy, and carrying out continuous iteration until the optimized cue words are obtained. According to the method and the device, efficient self-adaptive updating of the cue words of the language model is realized, so that the generalization performance and robustness of the model in zero-sample, few-sample and concept drift scenes are improved.

The invention relates to a source-load multi-subject hierarchical collaborative optimization method based on a fused niche. According to the scheme, multi-subject modeling, a master-slave game mechanism, niche evolution optimization and an interlayer feedback coordination strategy are combined. Firstly, a unified mathematical model of various power supplies and loads in a source-load system is constructed, and optimization targets and constraints of the unified mathematical model are defined. Secondly, introducing a master-slave game (Stackelberg) mechanism, and simulating a dynamic game behavior of source first-onset and load response; thirdly, a niche evolution algorithm is adopted to improve the search diversity and the global optimal solution obtaining capability; and finally, through an interlayer feedback mechanism, realizing mutual guidance and correction of a game result and an evolution process, and outputting a global consistent scheduling scheme. According to the method, the optimization precision, the operation coordination and the scheduling robustness of the source-load system in a complex and changeable environment can be effectively improved, and the method is suitable for multi-source and multi-load collaborative optimization scenes such as a micro-grid and an integrated energy system.

The invention discloses a data governance method and system based on artificial intelligence, and relates to the technical field of data governance, and the method comprises the steps: effectively improving the recognition capability of structural abnormal behaviors in a device file through introducing an abnormal track extraction mechanism combining image enhancement and cuckoo search, and improving the recognition efficiency. The method combines an isolated forest model optimized by a genetic algorithm, achieves the high-confidence screening of candidate abnormal data, remarkably improves the accuracy and robustness of anomaly detection, and achieves the recognition of a redundant entity through the construction of a semantic conflict graph and the introduction of a graph matching network. The problem that in a traditional method, the field similarity is insufficient to support complex conflict judgment is effectively solved, finally, a standard equipment file is generated through a field-level fusion strategy, a governance evolution graph is synchronously updated, a complete closed loop from discovery, evaluation and fusion to optimization is formed, automation, intelligentization and self-adaptive evolution of equipment master data governance are achieved, and the efficiency is improved. And the consistency, availability and governance efficiency of the equipment data are remarkably improved.

The invention relates to the technical field of industrial data processing, in particular to a model-driven FlinkSQL (Structured Query Language) dynamic optimization method, which comprises the following steps of: generating candidate SQL (Structured Query Language); carrying out parallel acquisition and noise reduction processing; constructing a causal dependency graph and embedding the graph; constructing a spatio-temporal feature vector and performing steady-state judgment; performing disturbance sandbox verification and stability judgment; a fault-tolerant strategy is injected, and DML is compiled and produced; and deploying tasks and monitoring and updating in real time. Physical signals of industrial equipment are stripped through a wavelet noise reduction technology, then operation indexes of the Flink platform are subjected to structured expression in combination with causal dependency graph embedding, quantitative classification is performed on current operation and ideal steady state, robustness is verified through multiple cycles, adaptive parallelism and a fusing strategy are injected into SQL, and the method has the advantages of being high in robustness and high in reliability. And the reference base point is automatically updated according to the monitoring result, so that the problems of low SQL execution efficiency, insufficient stability and difficulty in adapting to complex industrial scenes caused by neglecting environment change and system load fluctuation during operation are effectively solved.

The invention relates to the technical field of artificial intelligence, and provides a multi-agent-based task processing method, system and device and a storage medium, and the method comprises the steps: receiving a user request, carrying out the intention analysis of the user request, and generating an initial task sequence according to an analysis result; according to at least one sub-task to be executed in the initial task sequence, scheduling at least one sub-agent to execute the sub-task, and monitoring an execution result of each sub-task; when the execution result of any subtask is monitored to be abnormal, performing reflection analysis on an abnormal reason, and performing task planning again based on a reflection conclusion and the current context information to generate a new task sequence; and rescheduling the corresponding sub-agent to execute the sub-task based on at least one sub-task to be executed in the new task sequence. According to the method, the replanning process is guided by introducing the reflection analysis when the task execution is abnormal, so that the task execution path is dynamically corrected and adjusted, and the robustness and success rate of task execution are improved.

The invention discloses a multi-mode enterprise intelligent brain system based on artificial intelligence, and belongs to the technical field of big data processing. The method and the device are used for solving the technical problems of poor multi-modal data processing precision and poor system robustness in the existing scheme. Semantic alignment of multi-modal features can be realized through a closed-loop process of pre-extraction, comparative learning, dimension alignment, dynamic optimization and verification and fine adjustment, cross-modal semantic association is established through comparative learning, space mismatching is solved through dimension alignment, dynamic matrix optimization fits task requirements, and actual effects are ensured through verification and fine adjustment. Through a progressive process of confidence evaluation, task coding, weight calculation, feature fusion and end-to-end optimization, intelligent weighting of multi-modal features is realized, and through confidence evaluation, reliable data is screened, a decision target is determined through task coding, a gating network dynamically allocates weights, and end-to-end optimization ensures that a strategy meets service requirements.

The invention discloses a task response method based on a large model, and a large model fine tuning method and device, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining first preliminary output data of the large model for a task request; based on the first preliminary output data, determining a knowledge boundary evaluation result of the large model for the task request; the knowledge boundary evaluation result is used for indicating the probability that the task request exceeds or does not exceed the knowledge boundary of the large model; and according to a knowledge boundary evaluation result, deciding whether to call a matched target tool to respond to the task request. Therefore, a decision-making mechanism based on knowledge boundary evaluation is introduced, so that the large model can perform judgment according to the actual demand of the task and does not depend on a tool blindly or excessively confident any more, and the robustness of the large model in different environments and task scenes is enhanced.

The invention relates to the technical field of energy scheduling, and discloses a self-adaptive multi-objective optimization method and device for a distributed energy scheduling system, and the method comprises the steps: S1, taking the scheduling cost, the energy utilization efficiency and the system stability as optimization objectives, and constructing a comprehensive objective function of multi-objective optimization of the distributed energy scheduling system; s2, acquiring operation data of the distributed energy scheduling system; s3, according to the operation data, using a deep neural network to adjust the parameter weight of the comprehensive objective function, and using a genetic algorithm and a particle swarm optimization algorithm to update the comprehensive objective function to generate a candidate scheduling scheme; and S4, iteratively executing the steps S2 to S3 until the maximum number of iterations is reached or the comprehensive objective function is converged, and generating a final scheduling scheme. According to the method, the optimization efficiency and precision of the distributed energy scheduling system can be effectively improved, and the robustness of the algorithm is enhanced, so that efficient and stable scheduling optimization in the complex and changeable distributed energy system is ensured.

The invention relates to the technical field of optimal tracking control, and provides an optimal tracking control method, device and equipment based on adaptive dynamic programming. Constructing an augmented system according to a tracking error and a reference trajectory of the system (the system is constrained by part of unknown disturbance), and designing a corresponding value function; based on the designed value function, the H infinity tracking problem is converted into a zero sum game, and a corresponding Hamilton-Jacobi-Isaacs (Hamilton-Jacobi-Isaacs, HJI) equation is constructed; and the evaluation-execution-disturbance network approximates the HJI equation, the optimal control strategy and the worst disturbance strategy are solved, and the optimal control strategy is applied to the controller. According to the method, the control input and the disturbance input during neural network training are updated by using a double-event triggering mechanism, and the resource utilization efficiency is improved. According to the optimal control input data, external disturbance factors suffered by the controlled system are considered, so that the robust performance of the obtained optimal control input data is improved.

The invention relates to an agent workflow-oriented dynamic fault sensing and risk propagation prediction method, and belongs to the field of artificial intelligence. Comprising the steps of constructing a dynamic dependency graph according to a target workflow; for each tool node, the input and output credible auditor verifies the workflow through a format compliance layer, an execution state layer and a semantic consistency layer, and execution is continued if all verification is passed; when the verification of any layer is not passed, blocking the return of a fault information structured report, and judging an error severity weight and a data sensitivity weight; and downstream nodes of the tool nodes which do not pass the verification are managed and controlled through a risk propagation prediction model based on the dynamic dependency graph. According to the invention, the initial fault can be intercepted in time and prevented from causing cascade propagation and amplification effects in a complex workflow; the normal form of error processing in the intelligent agent tool flow is changed, so that the intelligent agent tool flow is subjected to passive emergency response conversion in-process accurate control after an event, and the robustness and reliability of the complex workflow are enhanced.

The invention discloses an artificial intelligence content generation task multi-server collaborative optimization method based on deep reinforcement learning, and relates to the crossing field of edge computing and artificial intelligence. According to the method, a user equipment-base station-edge server-core network four-layer architecture is constructed, distributed decision is realized through a deep Q network model, and the method comprises the following steps: defining a task and time delay model; a deep reinforcement learning framework is designed, a decision model is constructed through a state space, an action space and a reward function, and training stability is improved by adopting experience playback and a target network slow update mechanism; and providing an adaptive multi-server selection and load distribution strategy, and solving an optimal distribution proportion based on a time delay equality principle. According to the method, the average unloading time delay of the AIGC tasks is remarkably reduced, the task failure rate is reduced, the dynamic environment adaptability and the extreme scene robustness are improved, and the method is suitable for scenes with strict requirements for time delay and reliability.

The embodiment of the invention provides a CSI (Channel State Information) performance monitoring report triggering method based on an AI / ML (Artificial Intelligence / Markup Language) model and a related device, and provides a channel environment adaptive cumulative sum CUSUM algorithm, the algorithm configures different CUSUM parameter sets for different states of UE (User Equipment) respectively, each CUSUM parameter set comprises a relaxation factor k and a decision threshold h, and the relaxation factor k and the decision threshold h correspond to each other. The values of k and h in different parameter sets are different; the UE can select the most matched CUSUM parameter according to the current state (or scene), so that a performance monitoring mechanism of the UE can intelligently balance the detection sensitivity and robustness. Moreover, according to the scheme, performance differences are accumulated through a CUSUM algorithm and then compared with a decision threshold value to judge whether a performance degradation report is triggered or not, instantaneous jitter of a channel is effectively filtered through accumulation of the performance differences, and it is ensured that the report is triggered only when the performance of the AI / ML model continuously declines. Therefore, the uplink signaling overhead caused by mistaken touch and the resource waste caused by subsequent network response (such as mode switching) are reduced.

The invention discloses a large language model jailbreak attack method based on alignment mechanism interference, and relates to the technical field of artificial intelligence. The method comprises the following steps: by constructing a consistency loss function, guiding an intermediate layer representation when a model generates a malicious request to approach an intermediate layer representation of a compliance request, thereby interfering a security alignment mechanism of the model and weakening the rejection capability of the model to malicious input. From the perspective of the internal representation space of the model, a jailbreak strategy which can interfere with the security cognitive path of the model is designed, and the model is disguised to the middle representation of malicious input as benign representation by optimizing prompt input, so that the jailbreak attack with higher concealment and stability is realized, the defense capability of the model to vulnerabilities is conveniently trained, and the security of the model is improved. The method can be widely applied to large model security mechanism research, alignment robustness test, red team evaluation and other scenes.