1292 results about "Memory bank" patented technology

The invention relates to the technical field of industrial defect detection, and discloses a memory matching industrial defect detection method based on adaptive feature fusion, and the method comprises the steps: generating a plurality of unknown defect samples through an enhanced image-level defect simulation strategy, and helping a model to effectively distinguish a normal mode from an abnormal mode; the method comprises the following steps: extracting multi-scale features of a to-be-detected sample image, a normal sample image and a defect sample image, designing an adaptive hierarchical memory bank architecture, embedding the multi-scale features into a grid memory bank in a layered manner, further improving the reasoning speed through an adaptive core set sampling strategy, and combining multi-scale feature fusion and a defect positioning optimization technology to obtain a defect positioning algorithm. And the difference between the multi-scale features and the normal mode is fully utilized, so that the accuracy of defect detection and positioning is remarkably improved. Compared with the prior art, the defect detection precision and the positioning accuracy can be improved, and meanwhile, the real-time detection requirement can be met.

The invention discloses an industrial image anomaly detection method based on deep learning, and particularly relates to the technical field of industrial visual detection. The problems of high false alarm rate, fuzzy fine defect positioning, insufficient real-time response capability, difficulty in model increment updating and the like caused by data distribution drift in an industrial scene are solved. According to the method, robust features are extracted through a multi-scale feature fusion auto-encoder, and a dynamic memory bank is constructed to update a normal sample prototype online; a dual-path detection mechanism is adopted to cooperate with a pixel-level reconstruction error and attention weighted feature matching deviation; efficient edge reasoning is realized in combination with block parallel processing and model compiling optimization; and designing an elastic incremental learning framework to prevent disastrous forgetting. And finally, false alarms caused by environmental changes are reduced, accurate positioning of pixel-level defects is realized, millisecond-level detection requirements of high-resolution images are met, safe and efficient model online evolution is supported, and adaptability and reliability of an industrial quality inspection system are comprehensively improved.

The invention discloses a computer memory bank fault prediction method and system based on deep learning, and relates to the technical field of computer hardware fault diagnosis, and the system comprises a multi-source time sequence data collection module which is used for obtaining memory bank operation state data in real time; the dynamic feature enhancement module is based on a composite architecture of a generative adversarial network and transfer learning, comprises a fault mode generator, and generates synthetic data consistent with real fault distribution by using an LSTM network; aligning feature spaces of different hardware platforms through a maximum mean difference loss function; the multi-modal fusion deep learning model comprises a space-time convolutional network, a graph attention network and an adaptive weight adjustment mechanism; and the fault early warning analysis module is used for analyzing a fault probability predicted value, an interpretable thermodynamic diagram and a maintenance suggestion. According to the invention, passive maintenance is changed into active prevention and control, and preposition and precision of fault management are realized through dual mechanisms of long-term trend prediction and short-term risk early warning.

The invention relates to the technical field of industrial control, in particular to an intelligent centerless grinding machine cooling system and method for difficult-to-machine materials, and the system comprises a multi-source sensing unit, an edge calculation module, a decision center module, a precise execution module and a digital immune memory bank; compared with the defects that in the prior art, monitoring depends on a single temperature sensor, thermal field sensing is incomplete, thermal damage early warning lags behind and the like, according to the scheme, grinding power frequency spectrum, acoustic emission signals, triaxial vibration and coolant mass spectrum data are collected in real time through a multi-source sensing unit, and a micron-sized temperature field is reconstructed in combination with an edge end space-time diagram convolutional network; full-dimensional monitoring of the grinding thermal-mechanical coupling effect is achieved, and thermal anomaly recognition sensitivity and early warning timeliness are remarkably improved.

The invention discloses a quality management and control system for fabricated decoration construction, and the system comprises a sensing layer which constructs a multi-modal data collection network, carries out the three-dimensional real-time data synchronous collection through combining logistics API docking and an OCR recognition system, and constructs a construction process digital twin bottom plate; in the edge calculation layer, an edge node carries out lightweight processing on the original data; the cognitive layer is used for calling a Prolog rule through a process knowledge graph engine to reasone the feature snapshots, carrying out defect instant diagnosis and dynamic constraint propagation, outputting a root cause path with probability weight through three-stage verification, and quantifying intervention influence; the decision-making layer is used for constructing a dynamic prediction model based on a bidirectional LSTM and an attention mechanism, automatically activating a compensation mode when key interference is detected in combination with an anti-fact memory bank and a case-based reasoning compensator, and generating an alternative scheme of optimal cost / optimal construction period / comprehensive balance through a multi-target optimizer; and in the application layer, a Unity engine is utilized to develop the digital twinborn billboard.

The invention provides a long-term video understanding system based on adaptive sparse memory and a language model. The long-term video understanding system comprises a visual encoder used for extracting visual features from a long video; the memory bank module is used for storing and retrieving visual features of historical video contents; and the sparse adaptive module is used for dynamically managing the memory bank module, and the memory bank module interacts with the multi-modal large language model by querying a converter Q-Former and is used for incrementally processing video data and mapping visual features to a language space. By introducing an adaptive sparse memory mechanism, a long-term video sequence can be effectively processed, redundant features can be dynamically compressed, and key information can be reserved, so that efficient analysis of a long video is realized; compared with the prior art, the method has high accuracy in multiple tasks, and can dynamically manage the memory bank and reduce processing of redundant features through a sparse adaptive mechanism, so that the calculation overhead is reduced, and the overall efficiency of the system is improved.

The invention provides an open vocabulary task guiding grabbing method and device, a medium and a program product. The open vocabulary task guiding grabbing method comprises the steps that a user language instruction and a complex scene image are input into a preset visual language model; identifying a target task corresponding to the user language instruction through a preset visual language model, matching a target object associated with the target task in the complex scene image, and determining an optimal grabbing component of the target object based on a plurality of functional components of the target object; constructing a plurality of capture candidate schemes by adopting the optimal capture component and the complex scene image; and performing optimization solution on each grabbing candidate scheme to determine the optimal grabbing pose, so as to solve the problem that the existing method is poor in effect when dealing with an unseen object and a new task in a scene with open vocabularies due to the limited scale of an operability memory bank in the related technology. The invention aims at solving the problem that the existing method is poor in effect when dealing with the unseen object and the new task in the scene with the open vocabularies.

The invention relates to the technical field of robots, and discloses a task planning method and system for a robot, and the method comprises the steps: obtaining a user instruction; based on the large language model and the semantic coding model, retrieving in a memory bank to obtain environment perception information related to the user instruction; inputting a first cue word determined based on the user instruction and the environment perception information into the large language model, and generating an overall action sequence corresponding to the user instruction; and for each sub-action in the whole action sequence, determining a second cue word, inputting the second cue word into the large language model, and generating a detailed action plan corresponding to the current sub-action until each sub-action in the whole action sequence is traversed. According to the method, the natural language or unstructured instruction can be received, the intention of the user can be understood, and the executable overall action sequence and the detailed action plan of each sub-action in the sequence are generated, so that the task issued by the user is completed, it is ensured that the robot can execute the task, and the accuracy and efficiency of task execution are improved.

The invention discloses a few-sample visual defect detection method based on approximate proximity search, and belongs to the technical field of industrial intellectualization. The system mainly comprises an image acquisition module, a feature extraction module, a memory bank construction module, a feature compression module, an approximate nearest neighbor search module and an anomaly detection and result output module. The feature extraction module is used for performing feature extraction on the collected image; the feature compression module is used for reducing the feature data volume and reducing the internal memory and calculation burden; the memory bank construction module is mainly operated during system initialization or model training and is used for constructing a feature memory bank of normal samples; the approximate nearest neighbor search module is responsible for quickly finding normal sample features closest to the input features in the memory bank; and the anomaly detection and result output module is used for converting distance information obtained by neighbor search into anomaly. According to the few-sample visual defect detection method based on the approximate proximity search, the effect of few-sample anomaly detection is achieved.

The invention discloses a distribution network line facility defect detection method based on line magnetic field variable characteristics, and relates to the technical field of magnetic variable measurement. According to the method, high-precision spatio-temporal data are obtained through multi-point synchronous non-contact magnetic field sensing, codes are efficiently preprocessed through self-adaptive voxel division and bytes, and the detection accuracy of the distribution network line facility defects is improved; and the line topology information is deeply fused to Token for representation. A pre-trained large language model is combined with low-rank adaptation (LoRA), sliding window attention and an external memory bank for efficient reasoning, and understanding of long sequences and topological association is enhanced. Characterization quality is improved through mixed mask self-supervised learning, and multi-task (classification, positioning and scoring) output is optimized by adopting uncertainty weighted loss. And privacy protection federal training is supported. And finally, an interpretable diagnosis report is generated, and an alarm and work order process is automatically performed. The accuracy, efficiency, intelligent level and automation degree of distribution network fault detection are remarkably improved, and safe and reliable operation of a power grid is guaranteed.

The invention provides a cyclic feedback scheduling collaboration method based on an intelligent agent and cloud computing power service. The method comprises the following steps: memory recall: the intelligent agent receives a task demand, accesses a memory library to obtain historical experience entries, performs memory matching and memory fusion, generates memory recall information, and outputs the memory recall information to the cloud computing power service; cloud task planning: a cloud big language model understands task demand intentions based on memory recall information, decomposes tasks and generates executable descriptions, and an agent generates a preliminary task plan and calls a proper tool; performing action observation: performing operation by the intelligent agent, acquiring feedback data in real time, comparing an actual result with an expected target adjustment strategy, selecting a more suitable tool, and performing repeated execution until a task is completed; summarizing and reflecting: generating task execution tracks, integrating into experience entries, and storing into a memory bank. According to the method, historical memory, dynamic feedback and cloud computing power are combined, and the autonomous decision-making ability, execution efficiency and accuracy of the intelligent agent in a complex task are remarkably improved.

The invention provides a time-varying graph neural network traffic flow prediction method based on a dynamic memory bank, and belongs to the technical field of traffic prediction. The method adopts a layered deep neural network architecture, and comprises a data embedding layer, a space-time coding layer, a memory enhancement layer and a prediction output layer. The data embedding layer preprocesses a traffic flow sequence, associates a collaborative coding time sequence mode with a road network, and synchronously constructs a dynamic graph structure; the space-time coding layer is subjected to space-time stream decoupling extraction, a space branch models multi-scale space dependence through a time delay graph convolution module and a space Mama module, and a time branch extracts multi-granularity time features through a hierarchical time sequence sensing module and a time Mama module; the memory enhancement layer performs pattern matching and reconstruction on the space-time fusion features by means of a dynamic memory bank; and the prediction output layer generates a prediction result by taking the GCRN as a decoder. According to the method, the space-time dependence of the traffic situation is accurately captured, the prediction curve is highly fit with the true value, and the high-precision prediction of the traffic flow is realized.

Apparatuses, systems, and methods for dynamically allocated aggressor detection. A memory may include an aggressor address storage structure which tracks access patterns to row addresses and their associated bank addresses. These may be used to determine if a row and bank address received as part of an access operation are an aggressor row and bank address. The aggressor row address may be used to generate a refresh address for a bank identified by the aggressor bank address. Since the aggressor storage structure tracks both row and bank addresses, its storage space may be dynamically allocated between banks based on access patterns to those banks.

The invention discloses a wind power plant cluster power prediction method. The method comprises the following steps: step 1, embedding priori knowledge into a learning module; step 2, constructing a memory enhanced adaptive graph structure; step 3, an endogenous feature learning module; step 4, an exogenous feature learning module; 5, performing memory enhancement gating fusion on the MAGF; and step 6, outputting and predicting. According to the method, special learning modules are respectively constructed by decoupling endogenous power signals and exogenous meteorological factors, and a dynamic graph structure adaptive to a sample is retrieved by utilizing a meta-memory library, so that context sensing modeling of a complex space-time dependency relationship is realized. Meanwhile, a memory enhanced gating fusion mechanism (MAGF) is introduced, adaptive semantic alignment and fusion of double-flow features are realized, the precision and robustness of the model in multiple prediction time domains are remarkably improved, and more reliable technical support is provided for power grid dispatching and wind field operation.

A memory device includes a peripheral circuit structure and a cell array structure, in which the cell array structure overlaps the peripheral circuit structure in a first direction, in which the cell array structure includes a merged sub array in which at least four sub array regions are merged, in which the merged sub array comprises a memory cell region including: a plurality of first bit lines and a plurality of second bit lines; and a plurality of first word lines and a plurality of second word lines, in which the peripheral circuit structure comprises a merged bank in which at least four banks are merged, and in the merged bank comprises: a first bit line sense amplifier (BLSA), a second BLSA, a first sub word line driver (SWD), a second SWD, and in which a spare space is defined between the first bit lines and the second bit lines.

The invention discloses a memory enhancement and task planning method and system for large-model multi-agent. The method comprises the steps of performing structured analysis on historical task data, extracting planning experience and an abstract mode, and constructing and indexing an external memory library; retrieving related experience and modes from a memory library based on the new task description, and generating a memory enhanced task planning book; the planning scheme is distributed to a multi-agent network to be executed, an actual execution track is monitored and recorded, and the deviation between the actual execution track and the planning expectation is calculated; and evaluating the planning scheme based on the execution result and the deviation, extracting a planning correction rule, and updating the experience and the rule of the task to a memory bank. By introducing an external memory and closed-loop learning mechanism which can be dynamically evolved, the problems that existing system planning lacks experience guidance, knowledge cannot be precipitated and reused, and planning and execution are disjointed are solved, and the planning quality, the system reusability and the self-adaptive optimization capability are remarkably improved.

An apparatus includes a floating-point gradient register; an integer register; a memory bank; and an array of processing units. Each of the units includes a plurality of binary shifters having an integer input configured to obtain corresponding bits of a 4-bit integer multiplicand, and a shift-specifying input configured to obtain corresponding bits in an exponent field of a 4-bit floating point multiplier. The multiplier is specified in a mantissaless four-bit floating point format including a sign bit, three exponent bits, and no mantissa bits. An adder tree has a plurality of inputs coupled to outputs of the plurality of shifters, and a rounder has an input coupled to an output of the adder tree. The integer inputs are connected to the integer register; the shift-specifying inputs are connected to the floating-point gradient register; and outputs of the rounders are coupled to the memory bank.

The invention provides a knowledge distillation method, system and equipment for a multi-modal large model and a storage medium, and relates to the technical field of artificial intelligence, and the method comprises the steps: inputting multi-modal data sent by a user side into a visual modal analysis unit, a language modal analysis unit and a voice modal analysis unit, correspondingly generating a visual feature vector, a language feature vector and a voice feature vector; carrying out feature fusion to generate a fused feature matrix; on the basis of the fusion feature matrix, migrating knowledge of the teacher model to the student model; obtaining a prediction result output by the student model, feeding back the prediction result to the user side, and receiving feedback information sent by the user side; and the feedback information is stored in a dynamic memory bank, and incremental learning is performed on the student model based on the gradient direction vector stored in the dynamic memory bank. The method has the technical effects that dynamic optimization is carried out according to the actual use condition of the user, so that the model performance meets the actual application requirement.

The invention discloses a reference video object segmentation method and system based on motion modeling and multi-modal interaction, and the method comprises the steps: taking a video sequence and natural language description as input, and generating a preliminary segmentation mask through a text coding and mask decoder; a Kalman filtering motion modeling module is introduced to predict the motion trail of the target object, and time sequence consistency optimization is carried out on the preliminary segmentation mask; fusing the historical track of the object and the action semantics in the semantic features on the basis of a key action semantic coding module to realize action semantic alignment and mask dynamic correction; the segmentation quality of the current frame is subjected to multi-dimensional scoring based on a representative frame screening mechanism, the representative frame is screened out to update a memory bank, and the long-term tracking stability is improved. According to the method, the problems of target drift, insufficient semantic alignment and memory pollution in a complex dynamic scene in the prior art are effectively solved, and the segmentation precision, robustness and semantic consistency are remarkably improved while the light weight of the model is kept.

The invention discloses a storage system supporting Cache and RAM double-form access and an access method thereof, and the system comprises a data storage array which is divided into a plurality of structured storage sub-blocks; the storage mode register is used for storing a configuration value, and the configuration value is used for dynamically configuring each storage sub-block to work in a Cache mode or an RAM mode, so that the data storage array works in a full Cache mode or a full RAM mode or a Cache and RAM coexistence mode; and the storage access control logic is used for arbitrating the storage access request, mapping the storage access request to the corresponding storage sub-block for read-write access according to the access address and the configuration value, activating the storage sub-block mapped by the storage access request, and inactivating the storage sub-block not mapped by the storage access request. According to the method, dynamic configuration is achieved, access of Cache and RAM forms is supported at the same time, part of or all memory banks can be configured to be Cache or RAM, the advantages of the two memory forms are utilized at the same time, parallel access of a plurality of memory sub-blocks is supported, use is flexible and efficient, and power consumption is reduced.

The invention provides a long video understanding method capable of relieving time sequence illusion in a video language large model. The long video understanding method is based on a static bias adaptive frame selection mechanism and a cross-modal feature fusion strategy. According to the static bias mechanism, inter-frame similarity is evaluated through a discriminator, redundant frames are identified, key frames are selected or a complete sequence is reserved, so that calculation overhead is reduced, and spatio-temporal information integrity is kept; a video frame and a text are mapped to a shared semantic space, the single-frame semantic understanding ability is enhanced, then an embedded sequence serves as a soft prompt to be input into a large language model, and a final answer is generated in an autoregression mode. According to the method, the efficiency and accuracy of long video understanding and video question and answer tasks can be remarkably improved; the problem of low training and reasoning efficiency caused by time sequence dependence redundancy and excessive computing resource consumption is effectively relieved; and through a dynamic multi-modal task processing framework and a space-time memory bank compression mechanism, the modeling capability and generalization performance of the model on a long video sequence are further improved.

Disclosed embodiments include a computational memory system. The computational memory system includes at least one computational memory chip including one or more processor subunits and one or more memory banks formed on a common substrate. The at least one computational memory chip is configured to store one or more portions of an embedding table in the one or more memory banks, the embedding table including one or more feature vectors. The one or more processor subunits are configured to receive a sparse vector indicator from a host external to the at least one computational memory chip and, based on the received sparse vector indicator and the one or more portions of the embedding table, generate one or more vector sums.

The invention relates to an intelligent map making system and method based on a large language model, and the method comprises the steps: enabling a client to input data analysis information and a map making demand to a task planning Agent of a server, and carrying out the reasoning and disassembly of the input information into preliminary planning information through a double-LLM mechanism, the method comprises the following steps that Agent is used as a tool library, a tool library document, a knowledge library and a memory library are further inquired, a tool is selected through RAG and sent to a client side, if an existing tool cannot meet requirements, the Agent will generate a special code to complete a task, the client side calls the tool to carry out specific drawing according to received information, and a user can re-plan and adjust the drawing process through dialogue. The memory bank records session information and charting interaction operation information of this time and history. The problem of system fragmentation is solved through a unified architecture, stability and flexibility are balanced by combining large language model driving, tool agent and code generation, domain knowledge is introduced to enhance decision and improve specialty, and a user preference memory mechanism is constructed to realize personalized services.

A memory system includes a memory interface including a first sub-channel interface associated with a first plurality of memory banks and a first plurality of processor in memory (PIM) blocks and further including a second sub-channel interface associated with a second plurality of memory banks and a second plurality of PIM blocks. During a particular mode of operation associated with the memory system, the first sub-channel interface is configured to communicate, with a host device, one or more memory access commands associated with the first plurality of memory banks. During the particular mode of operation, the second plurality of PIM blocks are configured to perform, concurrently with communication of the one or more memory access commands, one or more PIM operations associated with the second plurality of memory banks. The second sub-channel interface is configured to be disabled during the particular mode of operation.

A statically reconfigurable dataflow architecture processor (SRDAP) performs an N-dimensional affine transform specified by a matrix on an input image to produce an output image includes L address pattern memory units (PMUs) comprising a memory arranged as a vector of L banks, and L corresponding data PMUs. Each data PMU receives a copy of the input image. In parallel: each address PMU writes an L-vector of addresses of input pixels to the vector of L banks and reads a single address of the written L-vector of addresses from a predetermined bank corresponding to a PMU number of the address PMU among the L address PMUs, and each data PMU receives the single address from the corresponding address PMU and uses it to read a single input pixel from the data PMU memory. A tree of pattern compute units coalesces the L single input pixels into an L-vector of input pixels.

Memory controller commands to be sent to a memory device may be prioritized based one or more factors. For instance, a memory controller may employ a first-ready, first-come, first-served (FRFCFS) policy in which certain types of commands (e.g., read) are prioritized over other types of commands (e.g., write). The policy may be modified so when deciding which bank to open, the bank with the oldest read command is opened. While continuing to issue commands per this modified FRFCFS policy, the controller may keep track of banks whose pages are ready to receive their respective read commands. Once a number of ready banks meets or exceeds a threshold, issuance of write commands are paused and read commands for the ready banks are issued.

The invention provides a face forgery detection method and system, and the method comprises the steps: determining an image feature sequence of a spatial domain and an image feature sequence of a frequency domain of an input face image of a training set; adopting a preset feature extraction network to determine a first face counterfeiting feature and a second face counterfeiting feature; performing classification prediction by adopting a preset first classification head and a preset second classification head, and determining classification loss; performing comparative learning on the second face counterfeit feature and a preset dynamic memory bank to determine a comparison loss; according to the classification loss and the comparison loss, optimizing a preset feature extraction network, and determining a face forgery detection model; and inputting the actually input face image into the face forgery detection model, and determining the trueness and forgery type of the actually input face image, through the face forgery detection method and apparatus, the spatial domain features and the frequency domain features are fused, the accuracy of face forgery detection is improved, and the robustness and generalization of the face forgery detection model are improved in combination with a classification task and comparative learning.

The invention is suitable for the technical field of memory burning control and automatic management thereof, and provides an intelligent control method and system for memory bank information burning, and the method comprises the steps: obtaining historical burning data of target burning equipment, extracting a plurality of similar burning records consistent with the packaging type of the target heterogeneous packaging memory particles from the historical burning data, and extracting a reference burning record of a plurality of non-heterogeneous packaging memory particles of a specified type from the historical burning data; according to the method, the dual-parameter delay offset factor based on the delay time deviation amplitude and the stable time deviation amplitude is constructed, and dynamic correction of the heterogeneous packaging memory particle page switching control duration is achieved for the first time. When the performance of the target burning equipment is reduced but the failure threshold value is not triggered, the mechanism can accurately identify the abnormal trend in the response evolution of the heterogeneous packaging particles, and the control rhythm is actively adjusted according to the abnormal trend.

The hybrid addressing method and the hierarchical multi-core cascading device provided by the invention comprise the following steps: if a core in a basic module initiates a memory access request, analyzing the memory access request to obtain a row offset high order of an access address, if the row offset high order is located in a sequence region, rearranging the access address to obtain a rearranged address, and if the row offset high order is located in the sequence region, rearranging the rearranged address to obtain the hierarchical multi-core cascading device. The rearranged address is restrained in the memory bank of the basic module, the sequence area is a preset row in the memory bank of the core, and if a row offset high bit is located in an interleaving area, the access address is divided to obtain a divided address; wherein the divided addresses are stored in memory banks of different basic modules, and the interleaving area is an area except a sequence area. A dual-mode design of a local sequence region and a global interleaving region is adopted, and an exclusive sequence region is divided in global interleaving mapping by recoding an address high bit, so that balance between a global shared memory view and local low-latency access is ensured.

The invention provides an intelligent grading and excess subscription management system and method for a GPU video memory, and relates to the technical field of GPU video memory management, and the method comprises the steps: firstly building a grading storage system comprising a first performance region and a second performance region; then receiving a video memory allocation request containing task priority and service quality requirements, and monitoring a data access mode and access frequency; the initial placement position of the data block is dynamically determined through the intelligent engine according to the task priority, the data access mode and the frequency; predicting target access probability distribution of each data block by adopting an LSTM network; then migrating data between the first performance area and the second performance area through a swap-in and swap-out mechanism based on the distribution and periodicity characteristics, and ensuring the service quality of the first performance area; finally, the global oversale proportion is dynamically adjusted through an overpurchase safety management module, an independent virtual address space is distributed for each task, and safe overpurchase can be achieved through the process.