888 results about "Model inference" patented technology

The present invention relates to the field of industrial intelligent predictive maintenance, and in particular to an interpretable deep feature fusion network-based industrial intelligent predictive maintenance method, comprising: acquiring gearbox vibration data comprising noise; performing preliminary extraction and noise suppression on features of the acquired data by establishing an interpretable feature extraction module having a physical information constraint; integrating multi-scale features comprising long-distance and local dependencies by means of a dual-branch feature fusion module having global and local feature fusion capabilities; performing dimensionality reduction on a high-dimensional feature and generating an output by means of a classifier to obtain a final fault identification result; and performing interpretability analysis on a diagnosis process of a model. In the present invention, by embedding the signal processing technology having a well-defined physical theory support into a deep neural network, the interpretability and reliability of model inference results are effectively improved while the fault identification accuracy of the model is improved.

The invention discloses a welded pipe conveying abnormity prediction method and system based on large model reasoning, and aims to solve the problems that multi-source data is difficult to align, cross-station false correlation is caused, prediction lacks executable positioning and time sequence, and linkage control reliability is insufficient. Event alignment is carried out by taking a controller edge signal and an encoder zero position as time anchor points, a production line topology semantic graph containing time delay, capacity and interlocking attributes is constructed, and topology reachability and physical time delay constraints are applied in a self-attention long sequence model to carry out multi-step rolling prediction. And outputting a risk probability, refining the risk probability to spatial positioning of a roller way section or a shaft and the minimum executable intervention time, and generating a risk interval in combination with uncertainty estimation and calibration so as to drive an upstream beat self-adaptive speed reduction, shunting or stopping strategy. The technical effects of improving accuracy and interpretability, reducing false alarm and missing alarm, ensuring that linkage can be executed in advance and meeting edge time delay budget are achieved.

The invention discloses a MoE large model reasoning optimization method and device based on expert module dynamic scheduling, and relates to the technical field of large model reasoning optimization. The method comprises the steps of performing pre-reasoning through a MoE large model according to a dialogue text data set to obtain a first evaluation score set; performing value optimization on the GPU candidate expert quantity value set through an expert cache hit rate calculation method to obtain a GPU candidate expert optimized value set; cPU expert module dynamic scheduling is carried out based on the relation curve of the expert score ascending value-the transmission expert number, and a CPU expert scheduling index list is obtained; gPU expert module dynamic supplementation is carried out according to the first evaluation score set and the CPU expert scheduling index list, a GPU expert supplementation index list is obtained, re-reasoning is carried out through a MoE large model, and a reasoning result text data set is obtained. The invention relates to a large model reasoning optimization method for balancing model reasoning speed and output result quality of a MoE large model.

The invention relates to the technical field of behavior recognition, and discloses a lightweight abnormal behavior recognition system and method based on edge computing, and the method comprises the steps: obtaining multi-dimensional heterogeneous behavior observation data covering a target region through an edge side behavior collection node system, and constructing a behavior dynamic feature vector set which can be iteratively updated; performing structure self-adaptive decoupling processing on the behavior dynamic feature vector set, and recording the fluctuation convergence rate of an abnormal category in real time in a model iteration process; judging the behavior recognition stability in the model reasoning stage, and constructing a behavior transfer trajectory map in combination with the environmental transaction interference factors; extracting density disturbance parameters of the abnormal behaviors, and generating a behavior intervention evaluation result set; and performing hierarchical risk judgment on the current identification behavior result, and automatically generating an edge execution regulation and control instruction set. The method has the advantage of improving the efficiency.

The invention discloses a power equipment fault diagnosis system and method based on edge cloud cooperation. The system comprises an edge end module and a cloud end module. The edge end module is deployed in a power equipment site, is embedded with a lightweight diagnosis model, collects power equipment operation state parameters in real time, executes localized preliminary fault identification and alarm judgment, extracts key characteristic quantities, and uploads a processing result to the cloud end module through a network communication protocol; the cloud module is deployed in a data center or a control platform, and performs complex model reasoning, cross-device historical data comparison, fault depth research and judgment and model updating and distribution according to the characteristic quantity data uploaded by the edge module; the edge end module and the cloud end module perform data interaction through an FTP, MQTT or 5G protocol, a model trained by the cloud end can be automatically distributed to an edge end to realize rapid deployment and switching, and equipment fault diagnosis is realized. According to the invention, through a task division cooperation and data interaction mechanism, an efficient and low-delay fault diagnosis process is realized.

The invention provides a lightweight multi-model collaborative edge intelligent target detection system and a lightweight multi-model collaborative edge intelligent target detection method, belongs to the technical field of edge target detection, and aims to solve the problem that a traditional target monitoring scheme depends on a cloud server to carry out deep learning model reasoning; in order to solve the problem that high-precision and low-delay real-time target monitoring cannot be realized on a resource-limited edge device, the detection system comprises an edge calculation module, an image acquisition module and a storage module. The detection method comprises the following steps: acquiring a to-be-detected image, and calling an IPS module through a performance core and an energy efficiency core to pre-process the acquired image; optimizing the YOLO model, and inputting the preprocessed image into a deep learning model; performing target detection on the preprocessed image data through a deep learning model; and a detection result is input into the storage module for storage and is uploaded to equipment connected with the peripheral interface for visual display and a remote monitoring center.

According to the lightweight helicopter fault diagnosis method and device based on cloud edge cooperation and the medium, a dynamic cooperative distributed intelligent diagnosis system is constructed between the cloud end and the edge end, so that efficient lightweight reasoning of the edge end and centralized optimization training of the cloud end are realized; therefore, an aviation health management platform with self-learning, self-adaption and online updating capabilities is constructed. The whole system adopts a cloud-edge-end three-level architecture, wherein the end side is responsible for data acquisition and preprocessing; the edge side undertakes real-time diagnosis and lightweight model reasoning; and the cloud is responsible for global training, model scheduling and strategy optimization. Cyclic interaction of model parameters, task instructions and diagnosis results is achieved between the cloud and the edge through a secure communication link, and a closed-loop intelligent updating mechanism is formed. According to the method, the real-time performance, the computing power efficiency, the model generalization ability and the system adaptability of a fault diagnosis system are remarkably improved.

The embodiment of the invention provides a model reasoning data caching method and device for a caching system and a storage medium, and the method comprises the steps: configuring a first virtual address mapped to a high-performance memory of the caching system for a data caching pool, and generating a first reasoning data according to the data size of the first reasoning data generated by a target reasoning model; determining a target memory space of the high-performance memory mapped by the first virtual address, and enabling the data volume of the first reasoning data to be in direct proportion to the memory space volume of the target memory space mapped by the first virtual address, thereby creating a dynamic data buffer pool of the target reasoning model, and then caching the first reasoning data based on the target memory space corresponding to the dynamic data buffer pool, thereby realizing on-demand allocation of the high-performance memory in the cache system, and improving the utilization efficiency and the use flexibility of the high-performance memory in the cache system.

The embodiment of the invention provides a model reasoning cache management method based on software and hardware collaboration, which comprises the following steps: a user side sends a model access request, executes reasoning calculation, and stores a newly generated key value cache in a video memory cache layer; in the video memory cache layer, updating the access popularity corresponding to each unit-level key value cache, and sorting according to the access popularity from high to low to form a first popularity sequence; and updating the hierarchical access popularity corresponding to each hierarchical key value cache, and sorting according to the hierarchical access popularity from high to low to form a second popularity sequence. And when the space utilization rate of the video memory cache layer reaches a set threshold value, determining and evicting a unit-level key value cache and / or a hierarchical key value cache to be evicted based on the first heat sequence and / or the second heat sequence. The cross-layer data migration efficiency is optimized by utilizing the characteristics of a multi-level memory and through mechanisms such as cache elimination prediction and the like, and efficient utilization of a video memory, reduction of reasoning delay and improvement of the throughput capacity of a system are realized.

A model deployment tuning system (MDTS) receives a trained ML model, specified constraints, and model evaluation data and applies a plurality of model utilization techniques to the trained ML model to produce a plurality of useable model versions of the trained ML model. The MDTS executes each of the plurality of useable model versions of the trained ML models on a plurality of different compute instance types using the model evaluation data to produce model evaluation results for a plurality of different combinations. The MDTS filters the model evaluation results based on the specified constraints to indicate one or more of the different combinations satisfying the specified constraints. The MDTS deploys one of the plurality of useable model versions of the trained ML model to a compute instance types according to a selected combination satisfying the specified constraints.

The invention discloses a model reasoning scheduling method and electronic equipment, and relates to the technical field of artificial intelligence, input information to be processed is acquired, and the actual length of the input information is determined, so that searching is performed in a pre-configured preset length set based on the actual length, and a length decomposition strategy matched with the actual length is obtained; the length decomposition strategy comprises a target combination of one or more preset lengths and is used for scheduling reasoning calculation of input information.

The invention provides a continuous neural state monitoring method based on brain-computer interaction, and relates to the technical field of brain-computer interaction. The method sequentially comprises the steps that electroencephalogram, electro-oculogram and body movement signals are obtained and preprocessed, and a multi-channel time sequence segment and artifact marks are generated; constructing a target state scale sequence in combination with task geometry, physiological prior and eye movement events; performing time delay estimation and forward alignment on the input and the scale; setting space-time consistency constraint and drift penalty in the multi-scale state space model, and training to obtain a neural state mapping parameter set; extracting a session invariant subspace based on historical session data and executing small-step increment updating to form a multi-scale state space model subjected to individualized updating; and outputting a neural state vector and a confidence interval in model inference, generating a continuous monitoring result with a time index and an abnormal prompt, and storing the continuous monitoring result. According to the invention, high-real-time, high-stability and high-reliability continuous monitoring of non-invasive brain-computer interaction is realized.

The invention discloses an interpretable false news detection method and system based on retrieval evidence enhancement, the interpretable false news detection method combines a big language model with an external knowledge base retrieval mechanism, that is, factual supplementary evidence is provided for the big language model to generate final reasoning through a retrieval module. Screening candidate text evidences by utilizing a clustering algorithm, and performing credibility scoring and filtering on the candidate evidences in combination with a large language model; a unified reasoning Prompt is constructed to input a plurality of large language models, true and false judgment is output respectively, and detailed explanatory texts are generated with the assistance of natural language explanation, so that the false news detection method with factual support, language expression and user understandability is realized. And the judgment results of the models are fused through a majority voting mechanism, and explanation is generated from multiple perspectives, so that the risk caused by reasoning deviation of a single model is reduced, and the accuracy and stability of system output are effectively improved.

The invention provides a multi-modal emotion recognition fusion method based on a multi-head attention mechanism, and the method comprises the steps: carrying out the feature extraction and fusion of various data, capturing the internal relation of a modal through the multi-head attention mechanism, achieving the information complementation between modals through cross-modal interaction, and dynamically adjusting the weight according to the quality of the modals. And a self-built database containing a large amount of Chinese data is constructed, and a culture adaptation optimization strategy is combined, so that the generalization performance and culture adaptability of the model in Chinese user groups are improved. The system is deployed on a cloud server, optimized hardware and software configuration is adopted, efficient model reasoning and multi-user concurrent processing are achieved, and the large-scale real-time application requirement is met. The emotion of the user can be monitored in real time and early warning can be provided in scenes such as psychological counseling and group emotion monitoring, professionals are assisted in better understanding the emotion state of the user, and the service effect is improved.

The invention provides a kernel function trajectory-based deep learning inference service performance analysis method, which comprises the following steps of: based on service indexes and hardware theoretical computing power acquired from a production cluster, defining floating point operation times per request (FPR) index to quantify service resource efficiency, and identifying high FPR hotspot services; positioning a reasoning iteration candidate boundary based on a GPU kernel function trajectory, verifying iteration integrity through fingerprint matching and chi-square test, and calculating a second reasoning iteration number IIPS and a model reasoning efficiency MIE; aiming at calculation-intensive operators on the key path, combining a dynamic Roofline model to estimate an operator theoretical performance upper limit, and based on actual execution time, calculating efficiency and a BottleScore index to identify a key bottleneck operator; and outputting targeted optimization suggestions according to analysis results of service efficiency analysis, model efficiency analysis and operator efficiency analysis. According to the method, the inference behavior pattern can be automatically identified from massive kernel trajectories, and the efficiency loss of each level is quantified.

The invention discloses a multi-model inference chain data analysis method and device based on natural language input, a medium and a program product. The method comprises the steps that an analysis problem text in a natural language form is analyzed into a semantic task structure, and the semantic task structure is mapped into a structured analysis model conforming to a preset specification; registering the structured analysis model as an instance to an analysis model library, allocating a unique identifier ID to the structured analysis model during registration, and writing attribution relation information of the structured analysis model; an upstream model ID and a downstream model ID in the attribution relation information of the instance are read, the attribution relation information of the upstream model and the downstream model is analyzed, layer-by-layer expansion is carried out in the mode, and a multi-model reasoning chain of the instance is constructed according to the dependency relation between the models; and according to a sequence determined by the multi-model reasoning chain, executing each model instance in sequence or in parallel, and integrating the obtained final output into a comprehensive analysis report. According to the invention, automatic execution from a natural language to full-process data analysis can be realized.

The invention provides a model reasoning method and device, electronic equipment, a storage medium and a program product. The method comprises the following steps: performing attention calculation on inferred tokens by adopting first numerical precision through a large language model to obtain a first attention score corresponding to each inferred token; screening a target tokens from the inferred tokens based on the first attention score; reasoning the input sequence by adopting second numerical precision through the large language model to obtain a reasoning result output by the large language model; the input sequence comprises a target token and tokens to be input corresponding to the token to be inferred; the tokens to be input are tokens pre-selected from the inferred tokens according to a preset rule; the first numerical precision is lower than the second numerical precision. According to the method, the sparse processing of the token is realized by adopting mixed precision calculation, so that the reasoning efficiency of the large language model is improved.

The invention relates to the technical field of model testing, in particular to a model testing method and device and electronic equipment, and is used for solving the problems of long test data combination time consumption and low coverage rate of a model testing mode in the related technology. The multiple features are classified, a core feature set, a common feature set and an edge feature set are obtained, and the model can be a deep learning model; generating a test set of the model according to a differentiated data generation strategy of feature importance on the basis of the rule set of the service type adapted to the model, the core feature set, the common feature set and the edge feature set, and performing model performance test on the model by using the test set to obtain a test report; therefore, the efficiency of model algorithm testing is improved.

A model inference acceleration method, used for improving the efficiency of model inference. In the method, two models having different parameter quantities are deployed separately on different accelerator sets, a first model having a smaller parameter quantity is used for processing an inputted prompt word, so as to generate tokens one by one for conversion into an output result, and the second model having a larger parameter quantity is used for verifying the accuracy of the tokens generated by the first model and correcting an erroneous token, so as to ensure the accuracy of the finally obtained output result. The time required for the first model to process the prompt word and generate the tokens one by one is short, and the time required for the second model to verify the tokens generated by the first model is far shorter than the time required for the second model to generate the tokens. Therefore, the cooperation mode of the first model and the second model can shorten the model inference time while ensuring the accuracy of an output result.

The invention discloses a multi-modal large model reasoning method and system based on self-driven feedback and symbol collaboration, and the method comprises the steps: carrying out the structural representation of multi-modal information through a knowledge graph, carrying out the entity recognition and relation extraction in combination with a large model, constructing a unified knowledge graph, and generating a knowledge ternary set; defining a symbol logic expression, constructing a diversified symbol logic rule by using the knowledge ternary set, and calculating a symbol consistency award of a reasoning path; constructing a symbol-human feedback collaborative reward mechanism to obtain a mixed reward function; in the process of interacting with the multi-modal environment, sampling a group of outputs for specific tasks, and constructing an intra-group relative reward optimization strategy network objective function in a multi-task scene in combination with a mixed reward function; the system interacts with the environment to realize autonomous evolution cycle to generate a training sample, and iterative cycle realizes self-driven feedback without a large amount of manual annotation data; and the logicality, the interpretability and the autonomous evolution ability of the multi-modal large model in a complex reasoning task are promoted.

The invention provides a large language model quantification method, related equipment and a computer program product, and the method comprises the steps: carrying out the reasoning of a to-be-quantized large language model through calibration data, obtaining the activation of the large language model, and carrying out the statistics of the activation distribution of a target layer according to channels; calculating a smoothing factor of each channel according to the activation distribution; compensating the weight of the target layer channel by channel according to the smoothing factor to obtain a compensated weight; performing 4-bit quantization on the compensated weight; in the model reasoning process, activation of a target layer is smoothed, and 16-bit quantization is carried out on the smoothed activation. According to the method, a W4A16 quantification scheme is adopted, compared with W8A8, the weight storage amount is compressed by half, meanwhile, the precision loss is controlled to be smaller than 3%, and deployment of a large language model on edge equipment is facilitated.

The invention discloses a blind signal source separation method based on a multi-resolution attention separation network, and belongs to the technical field of signal processing. The method comprises the following steps of signal preprocessing, multi-scale feature extraction, attention module training, multi-resolution feature fusion, signal source recognition model training, model performance real-time monitoring and multi-scene adaptation optimization, and microcosmic and macroscopic features of signals are effectively captured through a multi-scale channel feature coding network. The adaptability of the model to non-stationary signals and complex structure signals is enhanced, the separation precision is improved, the attention mechanism can dynamically adjust feature weights, highlight target signal features, suppress background noise and remarkably improve the signal separation quality, meanwhile, calculation resource allocation is optimized, the model reasoning time is shortened, and the method is suitable for large-scale popularization and application. The MRAS-Net adopts a classification-oriented consistency synthesis strategy to ensure that the synthesized virtual signal features and the real signal features are highly consistent in classification tasks, and the generalization ability and the result interpretation of the model are enhanced.

The invention provides a lightweight large model operation method based on end-side deployment. The lightweight large model operation method comprises the following steps: acquiring equipment operation data and storing a model file in a mixed precision format according to quantization precision supported by terminal equipment; determining a mixing precision quantification strategy and carrying out strategy analysis on the equipment operation data; a Key-Value cache file obtained through model reasoning in the historical dialogue is reserved; calculating a semantic embedding vector of the second user request, calculating a cosine similarity between the semantic embedding vector and the first V vector, and judging whether the second user request has a similar intention or not according to a calculation result; and if yes, performing incremental reasoning on the second user request by multiplexing the Key-Value cache file. By monitoring the state of the terminal equipment in real time, dynamically adjusting the quantization level, multiplexing the cache data of the semantic related requests and merging batch request reasoning, the effects of balancing the energy consumption and performance of the terminal equipment, reducing repeated calculation and improving the service time of the equipment and the user experience are achieved.

The invention discloses a thinking data-based large model reasoning ability evaluation and analysis method and device. According to the method, a data flow diagram is constructed through a variable use-definition chain during running of a specific language LMCL in the dynamic monitoring field, logic variable LVAR nodes are extracted, redundant edges are eliminated, and a thinking map with a direct dependency relationship is generated. Based on the thinking map, a five-dimensional evaluation system including reasoning efficiency, key node recognition capability, reasoning generality, multi-path reliability and accumulative hierarchical reasoning is designed, and an internal mechanism of a model reasoning process is quantitatively analyzed. A semantic rule is extracted through a frequent mode of mining successful and failed thinking data, reasoning path probability distribution is integrated in combination with a path aggregation strategy, an interpretable cue word optimization strategy is generated, and a large model reasoning process is dynamically injected. Through the rule guidance and path equalization strategy, the model reasoning accuracy is improved, and the problems of thinking data isomerism, single evaluation and black box enhancement in the traditional technology are solved.

The invention discloses a KV Cache compression and hierarchical management method and device for RAG acceleration and a medium, and belongs to the technical field of large model reasoning acceleration. In order to solve the problems that in an existing RAG KV Cache management technology, storage occupation is too large, and long-term importance and instant accessibility of data cannot be considered at the same time, the invention provides a hierarchical management strategy, and according to the method, a multi-dimensional popularity vector containing global popularity, session popularity and context popularity is defined for a KV Cache block. The method is characterized in that the compression precision of a cache block is independently determined by using global popularity so as to balance fidelity and performance; and meanwhile, the session heat and the context heat are used for independently determining the placement positions in heterogeneous hierarchies such as a GPU (Graphics Processing Unit), a CPU (Central Processing Unit) fixed memory and a CPU paging memory. According to the method, placement is guided through the real-time dimension, compression is guided through the importance dimension, context intelligent prefetching is achieved, and the reasoning performance of an RAG system, especially in a session and in a context pursuit scene is remarkably improved.

The invention relates to a multi-mode large model and light-weight small model collaborative road surface ice condensation state prediction method, and belongs to the field of road traffic safety monitoring and prediction. The method aims at solving the problem that real-time early warning and accurate prevention and control are difficult in the prior art. According to the invention, a multi-modal data coding system is constructed, road surface monitoring images, meteorological time sequence data and historical ice condensation text cases are integrated, and feature fusion is realized by adopting visual-physical feature joint coding, meteorological time sequence feature enhancement and text semantic mining; a pseudo label is generated through large model zero sample reasoning, and a lightweight small model is trained through cross-modal knowledge distillation; and finally, on the basis of a dynamic trigger type double-model reasoning framework, calling a cloud large model for fine judgment when the small model is low in confidence coefficient or high in scene complexity, and outputting the icing starting moment and thickness through confidence coefficient weighted fusion. According to the method, the prediction accuracy and real-time performance are improved, the model generalization ability is enhanced, and reliable support is provided for road traffic control in winter.

The invention discloses a multi-model reasoning method and device based on a graph structure, a medium and a program product, and the method comprises the steps: querying a task target node in a graph database, starting from the task target node, traversing the graph database according to a corresponding driving model, and carrying out reasoning path extension, each directed acyclic graph is obtained by mapping metadata of the registered analysis model; generating a task topology execution graph according to the model nodes involved in the expanded reasoning path and the corresponding driving model; and according to the task topology execution graph, actual execution of each model node is scheduled in sequence, and an interpretation result corresponding to the task target node is obtained. According to the method, the model attribution relation is explicitly analyzed, and the reasoning path is optimized by utilizing graph calculation, so that the reasoning execution efficiency is improved, and the intelligence, performance and interpretability of multi-model collaborative analysis are improved.

The invention provides a sparse Transform model reasoning acceleration method based on a GPU platform, and the method comprises the steps: building a first storage format in a sparse Transform model, and carrying out the coding of a sparse mask of any type through the first storage format; wherein the first storage format comprises an inner layer block and an outer layer block; an outer layer block adopts a block compression sparse row structure, and row pointers and column indexes of a completely effective block and a partially effective block are recorded; the inner layer block adopts an unsigned 64-bit integer value bitmap and is used for representing masks in the block; based on the sparse mask matrix coded by the first storage format, performing sparse calculation on the query, the key and the value, and outputting a corresponding context vector; compared with a traditional sparse mask storage structure in the prior art that calculation sparsity and memory access regularity are difficult to consider when the masks present random, global or mixed distribution characteristics, the method has the advantages that support for any type of masks is achieved, storage overhead is reduced, memory access continuity is improved, good GPU parallelism is kept, and thread divergence and memory access conflicts are avoided.

The invention discloses a deep learning inversion AOD-assisted atmospheric correction-based method and device and a medium, and relates to the field of atmospheric remote sensing and deep learning cross technologies, and the method comprises the steps: constructing a multi-type training sample set fusing actual measurement and physical simulation data, and carrying out mass screening and layering processing to obtain a multi-type training sample set; obtaining a pre-training sample, a fine tuning sample and an extreme scene sample matched with the satellite remote sensing data; establishing a deep learning network model taking data loss and inverse operator constraint loss weighted fusion as a total loss function, and adopting a cross-satellite transfer learning strategy to sequentially complete pre-training, satellite exclusive fine tuning and extreme scene enhancement training to obtain a trained aerosol optical thickness inversion model; radiometric calibration, cloud detection, geometric correction preprocessing and input model reasoning calculation are carried out on satellite remote sensing original data, post-processing and quality grading are carried out on an output result, and an aerosol optical thickness inversion result used for assisting atmospheric correction is obtained.

The invention relates to the technical field of brain tumor image segmentation, in particular to a brain tumor segmentation method and system based on anatomical perception symmetric comparison and cross-modal migration. The method comprises the following steps: carrying out data preprocessing on acquired multi-modal MRI image data; constructing a brain tumor segmentation model based on anatomical perception symmetric comparison and cross-modal migration; performing model training based on a two-stage decoupling training strategy; and performing model reasoning by using the trained model, and outputting a brain tumor segmentation result. Through a self-supervised learning framework, pre-training is carried out by using unmarked MRI data, dependence on a large-scale marked data set is greatly reduced, the problems of time consumption and high cost of medical image marking are solved, and the applicability of a model in a limited data scene is improved.