67 results about "LTM - Long-term memory" patented technology

The embodiment of the invention provides a question and answer management method and device based on a large model, a storage medium and a program product. In the scheme, a'short-term first and long-term 'progressive recall strategy is introduced, that is, in a storage stage, session abstracts are generated in a segmented manner under the condition that whether the collection frequency of a preset round and the topic category are changed or not, the session abstracts are sequentially written into a short-term memory storage area and sink to a long-term memory storage area after the preset storage duration is reached, and a subsequent recall path is pre-buried; in the acquisition stage, the latest abstract closest to the current round is recalled from the short-term region, and if the abstract is missing, the cross-session or cross-round historical session abstract continues to be complemented from the long-term region. According to the mechanism, session context management of'instant light and thin 'and'long-term consistent' is considered under the condition that the computing power and the bandwidth cost are not remarkably increased, accurate contexts are provided for complex session scenes with long-period, multi-topic and multi-file cooperation, and a large question and answer model can make more accurate and more consistent questions and answers based on complete and related context information.

The invention discloses an intelligent data center energy scheduling optimization method based on machine learning. The method comprises the following steps: acquiring data to form a multivariable time sequence; the improved PatchTST model calls a multi-level memory interaction structure to establish short-term memory and long-term memory, and an attention constraint energy consumption memory module is introduced to output prediction features; inputting the prediction features into a decoding layer of the improved PatchTST model; constructing an energy scheduling multi-objective optimization problem, and generating an equipment control instruction set by adopting an NSGA-III algorithm; collecting an actual execution result to form feedback data; performing deviation evaluation and strategy updating on feedback data through an energy self-feedback correction process; and judging whether the energy consumption prediction error and the energy scheduling deviation meet a preset termination condition or not. The method is suitable for intelligent energy scheduling and operation efficiency optimization in multi-energy coupling scenes such as an intelligent data center.

The invention relates to the field of artificial intelligence, and discloses a multi-modal memory system and method for intelligent interaction. The system comprises a multi-modal information acquisition module, a confidence evaluation module, a scenario aggregation module, a knowledge graph construction module, a hierarchical storage module, an intelligent retrieval module, an active verification module and a memory management module. The core problems that in the long-term user interaction process of an existing artificial intelligence system, multi-modal information management is fragmented, information credibility is not quantitatively evaluated, memory organization lacks semantic association, a retrieval mode is single and passive, and memory life cycle is not adaptively managed are solved. Finally, unified collection, quantitative confidence evaluation, scenario semantic organization, associative intelligent retrieval and adaptive memory optimization of multi-modal information are realized, high-quality and high-efficiency long-term memory support is provided for scenes such as intelligent assistants, smart home, medical health and educational training, and the user experience and practical value of an artificial intelligence system are remarkably improved.

The invention discloses a robot training system and method based on plot memory in the technical field of artificial intelligence and robot learning, and solves the problems that an existing robot training method lacks an effective memory scheduling system, an empirical value quantification mechanism is not intelligent, and the multi-modal information fusion capability is insufficient. The system comprises a sensing module, a multi-modal unified memory encoder, a progressive memory scheduling system, a multi-dimensional value quantitative evaluation mechanism, an intelligent experience hierarchical scheduler, a multi-modal unified code retriever, a strategy generation module, an execution module and an anomaly detection module. The multi-modal unified memory encoder adopts a hierarchical dimensionality reduction Transform architecture, and fuses an RGB image, a depth image, force sensor data and joint angle information into a 576-dimensional unified feature vector; the progressive memory scheduling system comprises a working memory structure, a short-term memory structure and a long-term memory structure. The multi-dimensional value quantitative evaluation mechanism carries out quantitative scoring on experience based on reward evaluation, novelty evaluation and uncertainty evaluation.

The invention relates to the technical field of robot perception and data generation technologies, in particular to a first-view-angle drilling method and device based on memory enhancement and a storage medium, and the method comprises the steps: extracting a plurality of target frames from a historical video stream stored in a space intelligent machine, obtaining a plurality of memory elements based on the plurality of target frames, performing three-dimensional reconstruction on each target frame, constructing a target world model, planning a first visual angle track of the target robot in the target world model, performing imaging simulation on the target robot, performing real-time rendering on a first visual angle video of the target robot, and performing real-time rendering on a second visual angle video of the target robot based on the same time axis and action script as the first visual angle video. A third-person video corresponding to the space intelligent machine is generated, and a drilling result is output; according to the method, the long-term memory data of the space intelligent machine is ingeniously used, high-quality drilling data can be quickly generated, the reliability of the first-view video is improved, and a reliable basis is provided for training and testing of a robot algorithm.

The invention discloses a cable production equipment fault diagnosis method based on machine learning. The method comprises the steps that multi-source time sequence data are collected and preprocessed to generate a standardized data set; executing fractional calculus operation to obtain a fractional response sequence and dynamic memory weight distribution; constructing long-term memory features and forming an exogenous input vector sequence; performing joint training on exogenous input and real observation to obtain a convergent improved NARX neural network model; inputting the exogenous input and historical output lagging sequence into the model, outputting a state prediction value and mapping the state prediction value into a fault risk index; generating a fractional order residual energy index and an abnormal score sequence; and judging and outputting the fault state and the degradation trend grade of the cable production equipment. According to the invention, by introducing the fractional calculus algorithm and improving the NARX neural network model, high-precision fault identification and degradation trend intelligent prediction of cable production equipment under complex working conditions are realized.

The invention discloses a data cleaning scheme generation method and device and a readable storage medium, and the method comprises the steps: carrying out the grouping of pre-training task data through employing a clustering mode, and obtaining a plurality of data groups; extracting data characteristics of each data packet based on a large model; generating a data cleaning scheme for the corresponding data groups by using the pre-training model in combination with data characteristics, and executing the data cleaning scheme; calculating a data cleaning index, and storing the data characteristics, the data cleaning scheme and the data cleaning index as short-term memory into a short-term memory module; evaluating the overall cleaning quality evaluation score of the data cleaning scheme by using a reflection module to obtain a reflection result; the data characteristics, the data cleaning scheme, the overall cleaning quality evaluation score, whether data cleaning succeeds or not and the reflection result serve as long-term memory and are stored in a long-term memory module, the data cleaning process can have traceability, and the data cleaning scheme adjusting efficiency is improved.

The invention discloses a multi-agent-based cross-station intelligent inspection and operation and maintenance method and system, and the system comprises an inspection expert agent, a tool use agent, a code generation agent, an overhaul expert agent and a memory module. Calling a tool using agent to obtain corresponding power station information generation suggestions, and issuing corresponding tasks after user confirmation; a code generation agent receives the task, a tool use agent is called to obtain a corresponding power station model version and hardware equipment information generation script, and script parameters are dynamically adjusted; and the maintenance expert agent obtains the script operation detection model, calls the tool use agent to obtain corresponding power station monitoring information, detects the monitoring information by using the detection model, stores the detection result to the short-term memory module, and gives a maintenance suggestion and adjusts a detection strategy by combining the detection result with the maintenance knowledge information in the long-term memory module. According to the invention, efficient, intelligent and sustainable automatic task management and execution are realized.

The invention discloses an endoscope image intelligent decision-making reflection agent system based on memory guidance. The system comprises an input module; the initialization module is used for establishing an initial multi-modal context; the action execution module is used for selecting an optimal tool from the expert tool set for calling; an expert tool set; a short-term memory module; the self-reflection module is used for forming a reflection feedback result; a long-term memory module; the updating module is used for updating the multi-modal context based on the output and reflection feedback result of the special analysis tool; and the evaluation module is used for judging whether the task is completed or not, and outputting a final analysis result if a termination condition is met. According to the method, traceability of an inference chain and avoidance of redundant operation are realized through short-term memory, cross-task experience accumulation and strategy optimization are realized through long-term memory, and a systematic collaborative inference process is formed by combining various endoscope analysis tools, so that analysis consistency, generalization ability and reliability in a complex scene are improved.

Long-term memory data objects may be generated for orchestrations agents. When a session completes or ends, a long-term memory data object may be generated according to a specified long-term memory type based on turn inputs during the session. When a new session is started, the long-term memory data object may be used as part of inputs to a generative machine learning model to perform or respond to turn inputs of the new session.

A multi-target tracking and monitoring method and system with high trajectory consistency, belonging to the field of computer vision, solves the problems of identity switching and trajectory fragmentation caused by long-term target occlusion and dense interaction in multi-target tracking, especially in UAV monitoring scenarios. The method includes the following steps: 1. Inputting a video sequence, performing target detection and feature acquisition for each frame; 2. Performing preliminary association through a basic tracker to obtain an initial trajectory; 3. Detecting the disappearance and appearance of abnormal trajectories using a position awareness submodule; 4. Storing abnormal trajectory features using a long-term memory submodule; 5. Performing secondary association using a multi-step cross-frame matching submodule; 6. Completing trajectory updates. This invention is applicable to UAV monitoring scenarios such as high-altitude inspection, low-altitude operations, and outdoor scene monitoring.

The invention discloses a large-model-based rail transit operation evaluation agent implementation method and system, and relates to the technical field of rail transit operation management and artificial intelligence, and the method comprises the steps: employing a memory storage assembly to solidify the role information of an agent; constructing a tool set adaptive to rail transit operation evaluation; configuring dialogue memory and long-term memory for the intelligent agent; presetting an operation evaluation execution process for the intelligent agent; and integrating the role information, the tool set, the dialogue memory, the long-term memory and the execution process, and constructing an intelligent agent with the rail transit operation evaluation capability based on a large model. According to the invention, automation, intelligentization and integration of rail transit operation evaluation can be realized, the evaluation efficiency and accuracy are improved, and complex and changeable operation scene requirements are met.

Embodiments of the present disclosure provide a long memory transformer-based time action positioning method and device. Applied to the field of time action positioning; the method comprises: clipping a target video to obtain a short video set; randomly selecting one or more short videos in the short video set and inputting them into a pre-trained short-term Transformer encoder to obtain the features of the randomly selected one or more short videos; using a long memory module to sample from long-term memory to obtain the features of the short videos in the short video set that have not been selected; inputting the features of the randomly selected one or more short videos and the features of the short videos that have not been selected into a time fusion module to obtain fused video features; and inputting the fused video features into a time boundary positioning module to obtain four-tuples of each action instance. In this way, the speed of time action positioning is improved, the memory consumption of processing resources and the training time are reduced, and faster prediction speed and better action classification accuracy are achieved.

The application discloses a kind of federated multi-task continuous learning methods and systems for edge computing, and the method is constructed and maintained by central server dynamic evolution orthogonal main subspace as long-term memory library;Each edge node generates safe update that does not interfere with historical knowledge after local training, by projecting model update to the orthogonal direction of corresponding space;Server then adopts super-collaborative aggregation algorithm to fuse the safe update of each edge node to coordinate the knowledge transfer between multi-task;Finally update global model and expand memory library.The application effectively realizes the interference-free accumulation of knowledge in time sequence and spatial dimension, and adapts the resource constraints of edge node, to provide an efficient continuous learning solution for edge intelligence.

The application discloses an intelligent agent memory and question and answer system, an enhanced retrieval method and a memory management method, and belongs to the technical field of memory and question and answer, and the system comprises: a user interaction module, which receives a query text of a user and presents a question and answer result; a hierarchical memory library, which stores memory nodes and a causal relationship graph among the memory nodes; a causal enhanced retrieval module, which performs enhanced retrieval on the query text in the hierarchical memory library; a question and answer generation module, which generates a question and answer result based on a retrieval result; a memory intervention module, which constructs an anti-fact scenario of deleting a current target memory, generates an anti-fact answer, compares the similarity of the anti-fact answer and the question and answer result, and obtains an influence degree score of the target memory; and an adaptive maintenance module, which manages the memory stored in the hierarchical memory library according to the influence degree score. The application can retrieve causal and relevant memories when a user queries, and quantifies the influence of a single memory, thereby realizing scientific management of long-term memory of an intelligent agent.

The invention discloses a long-term memory enhancement method for a supply chain agent, and relates to the field of information retrieval, and the method comprises the steps: firstly, carrying out the preprocessing of multi-source original data, and obtaining a user feature vector set; the feature vectors are divided into a plurality of clusters through similarity calculation, cluster-level retrieval vectors are generated in the clusters based on historical calling frequency and confidence coefficient weighted fusion, and a unified index database is established. And further constructing a multi-layer cache mechanism which comprises a hot storage layer, a warm storage layer and a cold storage layer, carrying out data migration according to the access frequency and a time decay function, and realizing hot and cold circulation by adopting an LRU strategy. During retrieval, the cluster-level retrieval vectors are quickly matched and indexed through a multi-layer cache to form a result set with cluster tags, and efficient calling is supported. Along with continuous addition of new user data, cluster boundaries and representative vectors are dynamically updated, and labels and cache contents are synchronously evolved, so that the long-term memory and service capability of the intelligent agent is enhanced.

The invention provides an SQL (Structured Query Language) statement execution control method, an SQL statement execution control system and a server, and relates to the technical field of databases. Therefore, the problems of information redundancy and historical preference neglect caused by excessive dependence on correlation recall are effectively avoided; in addition, the method can fully combine short-term memory and long-term memory in the cue word construction stage, so that the execution effect under the scenes of complex query, multi-round interaction and long-term use is remarkably improved.

The application discloses an end-to-end multi-target tracking method and system based on layered space-time memory and cross-modal priori. The application cooperatively deals with the problems of instantaneous instability and long-time appearance mutation in the tracking process by constructing a short-term memory module and a long-term memory module. The short-term memory adopts a ring buffer structure to cache recent trajectory representation, and provides continuous association clues through time proximity and geometric accessibility screening; the long-term memory adopts a novelty-driven strategy to maintain a sparse appearance prototype library, and provides diversified appearance anchor points for occlusion reappearance and re-identification. Further, the method integrates multi-modal positioning priori, encodes it into a feature sequence through space-time alignment and coordinate unification, and performs cross-modal attention fusion with the visual dominant trajectory query. The method unifies the short-term continuity, long-term recognition and absolute position constraint in an end-to-end framework, effectively improving the identity consistency and tracking continuity of target trajectories in complex scenes.

The invention discloses an emotional accompanying robot system based on an end-side AI. The system comprises a main processor unit, a coprocessor unit, a multi-mode sensing unit, an interactive execution unit, a power management unit and an internal communication bus. By adopting a dual-core collaborative architecture of a main processor and a special AI coprocessor, collection, emotion recognition, fusion decision and interactive feedback of multi-mode data such as voice and vision are locally completed on an end side, and dependence on a cloud end is thoroughly eliminated. According to the invention, the core problems of user privacy leakage risk, high feedback delay, strong network dependence, lack of long-term memory and the like existing in the existing cloud scheme are fundamentally solved, and emotion accompanying with high security, low delay, high reliability and anthropomorphic memory ability is realized.

Long-term memory data objects may be generated for orchestrations agents. When a session completes or ends, a long-term memory data object may be generated according to a specified long-term memory type based on turn inputs during the session. When a new session is started, the long-term memory data object may be used as part of inputs to a generative machine learning model to perform or respond to turn inputs of the new session.

The invention discloses a station long-time abnormal behavior identification method based on multi-granularity spatio-temporal context fusion, and the method comprises the steps: collecting multi-source sensing data, carrying out the spatio-temporal alignment, and generating a multi-modal data flow of a unified coordinate system; constructing an atomic event code based on the attitude sequence, generating a dynamic scene graph according to a target-environment relationship, and forming a dual-channel feature primitive; injecting the feature elements into a short-term memory layer STM, abstracting a middle-term behavior pattern, storing the abstracted middle-term behavior pattern into a middle-term memory layer MTM, and fusing a cross-camera scene graph to construct a long-term memory layer LTM to form a layered space-time memory library; performing cross-level retrieval on the associated memory in the STM / MTM / LTM through a deformable space-time attention lens, and outputting context features of multi-granularity fusion; and calculating short / medium / long-term abnormal scores based on the fusion features, and dynamically adjusting a threshold value in combination with the crowd density to realize collaborative judgment. According to the method, the problem of fragmentation of long-time behavior understanding is effectively solved, and instantaneous anomaly and long-time mode anomaly can be accurately identified at the same time.

The invention discloses a production index evaluation method and device based on gating cross attention information fusion and a medium. The method comprises the following steps: firstly, acquiring historical industrial process and process index data to construct a long-term memory sequence, and acquiring current industrial process and process index data to construct a current sequence; inputting the long-term memory sequence and the current sequence into a trained production index evaluation model to obtain a current industrial production index value; wherein the production index evaluation model comprises two encoders and one decoder, inputs of the two encoders are respectively a current sequence and a long-term memory sequence, outputs of the two encoders are respectively used as a Q vector, a K vector and a V vector to be input to cross attention of the decoder, and the decoder uses a reset gate and an update gate to improve original cross attention. The model decision is helped to adopt weighted average value V information to a great extent and reserve query Q information to a great extent, so that an industrial production index estimation task is accurately completed.

The invention relates to an intelligent contract vulnerability repairing method and device, computer equipment and a storage medium. The method relates to the field of computer software, and solves the problem that the intelligent contract vulnerability repairing effect and efficiency are low. The method comprises the following steps: receiving a query request, wherein the query request carries a contract file to be detected; searching a long-term memory storage system, and obtaining an optimal repair scheme matched with the query request; and generating a user response to the query request according to the optimal repair scheme. The technical scheme provided by the invention is suitable for the smart contract, efficient bug repair is realized, and the accuracy of bug repair is guaranteed.

The embodiment of the invention discloses an active interaction method and device of a robot, equipment and a medium, and relates to the technical field of robots. The method comprises the following steps: determining a target state vector corresponding to a target user according to characteristics of the target user detected by a sensing module, and determining a historical memory vector of the target user according with a current scene in a long-term memory matrix according to the target state vector of the target user; determining a current cognitive state and a prediction state of the target user based on an internal hidden state, a target state vector and a historical memory vector of a previous time step of a gated memory enhancement loop unit in the cognitive memory network; and determining an active interaction behavior of the robot according to the current cognitive state and the prediction state, and enabling the robot to interact with the target user based on the active interaction behavior. According to the technical scheme, interaction with the target user can be carried out through the determined active interaction behavior of the robot, and the use experience of the robot is greatly improved.

The application discloses a high-ammonia-nitrogen wastewater denitrification method and system based on a neural network, collects input parameters and output parameters of a denitrification system, and constructs a database through pretreatment; an improved long short-term memory neural network denitrification prediction model is constructed, a long-term history hidden layer output item and a long-term memory weight coefficient are introduced into a forgetting gate of the model, a process type identification parameter is additionally arranged in an input layer, and a total nitrogen concentration prediction value of effluent in multiple time scales is output; based on a prediction value and a dynamic weight multi-objective optimization function, an optimal control parameter is output through a particle swarm optimization algorithm with intermediate constraints; an executing mechanism is adjusted and real-time feedback adjustment is carried out, so that a closed-loop control is formed. The application improves prediction accuracy and control timeliness, adapts to multiple processes, and reduces energy consumption and drug consumption.

The invention discloses a body navigation memory enhancement method and system based on spiking neurons. The method comprises the following steps: (1) constructing random parallel spiking neurons; (2) using a gating linear RNN as a basic structure of a cyclic memory model, and meanwhile, using a random parallel spiking neuron as a discrete gating function of the cyclic memory model; (3) adding the circulating memory model into the body-equipped intelligent agent, and carrying out reinforcement learning based on a long-term memory navigation task; in the actual navigation process of the body agent, the cyclic memory model maximizes the time significance signal-to-noise ratio through a discrete gating function, so that the capacity of filtering information irrelevant to tasks of the body agent is enhanced. According to the invention, the success rate of long-term memory navigation tasks can be improved.

The invention discloses a reconfigurable two-dimensional floating gate memory which is of a layered structure, and a reconfigurable memory mechanism mainly depends on the photosensitive characteristic of a floating gate layer material, that is, illumination can improve the off-state current of the floating gate layer material. In a dark-state small gate voltage range, the floating gate layer is in a closed state, and at the moment, a storage mechanism is a capacitance coupling effect and shows short-term memory; under the action of illumination, the floating gate layer is opened, and a storage mechanism is switched to a Fowler-Nordheim tunneling effect, which is represented as long-term memory. The invention aims to solve the defects of single function and limited regulation and control means of the existing memory, realizes short-term and long-term memory dual-mode in the device through the '2T1C' layered structure design and the photosensitive characteristic of the two-dimensional material, can realize flexible switching of the two modes through external voltage or illumination, has the advantages of high-speed response of a DRAM and long-term maintenance of a Flash, and has the advantages of high-speed response of the DRAM and long-term maintenance of the Flash. And an efficient and flexible storage solution is provided for the fields of intelligent storage, brain-like computing and the like.

The invention discloses a self-adaptive long-term memory management system and method for end-side AI hardware, and belongs to the technical field of artificial intelligence and edge computing. The system comprises a memory acquisition module, a memory coding module, a memory storage module, a memory index module, a self-adaptive retrieval module, a memory evaluation module, a self-adaptive elimination module, a resource monitoring module, a strategy scheduling module and a privacy protection module. Through a hierarchical storage architecture, a multi-dimensional index structure and a mixed retrieval strategy, large-scale long-term memory is efficiently managed on end-side equipment with limited resources; dynamically adjusting a management strategy according to an equipment resource state and a user behavior mode through a self-adaptive elimination mechanism and strategy scheduling; and the privacy security of the user is guaranteed through encrypted storage and access control. According to the method, the memory management efficiency and the intelligent level of the end-side AI equipment are remarkably improved, the storage efficiency is improved by 40% or above, the retrieval accuracy is improved by 30%, the retrieval delay is reduced by 50%, the power consumption is reduced by 60%, and the method is suitable for various end-side AI application scenes such as smart phones, IoT (Internet of Things) equipment and vehicle-mounted systems.

The application discloses a spatio-temporal LSTM network radar echo sequence prediction method based on a self-attention mechanism, and specifically comprises the following steps: dividing a CKIM radar echo dataset into a training set and a test set, and performing pretreatment; adopting a self-attention mechanism to replace a forgetting gate mechanism in an ST-LSTM unit to form an SA-ST-LSTM unit; building an encoding-attention-decoding network; feeding the training set into the encoding-attention-decoding network to perform training, and obtaining a training model; feeding the test set into the training model to perform testing, and obtaining an image prediction result and prediction data. The SA-ST-LSTM unit is proposed, and the encoding-attention-decoding network is designed, the regulation of hyperparameters on long-term memory and short-term memory is adopted to process a catastrophic forgetting problem in the forgetting gate, and in addition, the attention mechanism is added to delay a long-term memory gradual forgetting problem in the network, so that the performance of radar echo sequence prediction is improved.