658 results about "Data dependence" patented technology

The invention discloses an unmanned aerial vehicle image small target detection method based on dynamic filtering and an adaptive sparse Transform. According to the method, an end-to-end target detection framework is adopted, a dynamic filtering module is introduced into a backbone network, global feature interaction is achieved through data-dependent frequency domain operation, and linear calculation complexity is maintained. For feature interaction in a scale, an adaptive sparse Transform module is introduced to enhance the capability of focusing key information on high semantic hierarchy features of a model, and noise interference and feature redundancy are effectively suppressed at the same time. Through the combination of dynamic filtering and adaptive sparse Transform, the model can extract image foreground information more effectively on the premise of not significantly increasing the calculation burden, and the problem that a traditional target detection model is susceptible to complex background interference is significantly relieved.

The invention provides a resource scheduling control method and system for a big data server, and the method comprises the steps: constructing a multi-dimensional resource portrait module, collecting the CPU, memory, network, storage I / O load and task queue length of each node in real time, and predicting a resource demand trend through a time sequence algorithm; extracting characteristics such as calculation intensity, data dependence, memory requirements, network transmission quantity and the like; adjusting the weight coefficients of the resource utilization rate, the task completion time and the energy consumption efficiency according to the system load and the historical effect; establishing a bipartite graph model by taking a resource trend as a node feature and a task vector as an edge feature, and calculating a matching score through graph convolution and a multi-objective optimization function; the scheduling scheme is synchronized by adopting a consistency algorithm; automatic rollback and reallocation are carried out when resources are detected to be insufficient; and optimizing a weight coefficient and a network parameter through reinforcement learning. Through the method, the system resource utilization rate can be improved, the task execution efficiency is improved, the overall scheduling effect stability is improved, and the system fault recovery time is shortened.

The invention discloses a road-oriented full-process digital collaborative management system and method, particularly relates to the technical field of road data management, and is used for solving the problem of abnormal multi-role collaborative conflict recognition. According to the method, dynamic linkage management of task states and role behaviors in the whole highway design process is achieved by building the task-driven atlas and the multi-role scheduling model, and the dynamic linkage management of the task states and the role behaviors in the whole highway design process is achieved by extracting access behaviors, task records and data version states, calculating role collaborative offset and recognizing and controlling the write-in permission of task conflict nodes. A priority factor is constructed based on a stage index, data dependence and a time urgency degree, a nonlinear model is adopted to generate scores, automatic reconstruction of a scheduling sequence and data ownership is driven, role permission and a collaborative view are synchronously updated, a data access control closed loop based on task state evolution is formed, and a three-dimensional responsibility chain is constructed by whole-process behavior traces. The traceable management of collaborative operation is realized, and the intelligence of collaborative scheduling and the accuracy of data management are improved.

The invention discloses a lithium battery thermal runaway control and safety response strategy system, which realizes the real-time monitoring and prediction of the internal temperature and voltage state of a battery through thermal-electric coupling modeling and multi-sensing data fusion, and realizes the real-time monitoring and prediction of the internal temperature and voltage state of the battery based on a model prediction control and threshold enhancement strategy. And triggering graded safety response in the early stage of thermal runaway. The system adopts a multi-stage response mechanism including measures of primary early warning, active cooling intervention, emergency power-off fire extinguishing and the like, a gas suppression path and cooling mode switching function is designed, and heat diffusion and spreading and toxic gas harm are effectively suppressed. Compared with a method depending on complex digital twinning or deep learning, the method is simple in structure, rapid in response and low in data dependence, has good system stability and engineering applicability, and can be widely applied to the field of lithium battery safety management of electric vehicle battery packs, energy storage power stations and the like.

The invention relates to a wind power equipment blade fault detection technology, discloses an unsupervised wind power equipment blade fault detection method based on a phase perception parallel attention mechanism, and solves the problems that an existing wind power equipment blade fault detection method is high in dependence on labeled data, insufficient in generalization ability under strong noise and variable working conditions and high in fault detection efficiency. And a weak transient fault signal and a dynamic change characteristic are difficult to capture robustly. According to the scheme of the invention, the method comprises the steps: collecting a blade operation audio signal, and extracting a dual-channel time-frequency feature containing an amplitude spectrum and a phase spectrum through improved short-time Fourier transform; a deep adversarial auto-encoder is constructed by using an encoder containing a phase perception parallel attention module, a decoder and an auxiliary encoder, and normal working condition feature distribution is learned by reconstructing an error loss, potential representation consistency loss, adversarial loss and phase consistency loss optimization model during off-line training; in the reasoning stage, the fault is judged based on the feature distance score and the reconstruction error score.

The invention discloses an unsupervised semi-pairing cross-modal retrieval method and system based on deep learning, relates to the field of artificial intelligence, and is used for solving the problems of annotation data dependence, asymmetric semantic association and high-dimensional storage efficiency. According to the method, a double-branch visual encoder and a dynamic prompt text encoder are combined, dynamic weighting of visual-text features is achieved through gating cross attention, and modal redundancy interference is restrained. An enhancement strategy is generated through low-frequency semantic guidance, and the long-tail word coverage rate is increased; a dual-stage quantitative hierarchical index is constructed, coarse-grained clustering and fine-grained product quantitative compression feature storage is adopted, and million-level data real-time retrieval is supported. A degradation aware increment maintenance mechanism monitors data distribution offset through a KL divergence threshold, and triggers index reconstruction to maintain long-term update precision. According to the method, limitation of a traditional strong pairing model is broken through, cross-modal sensitive content second-level positioning is achieved, asymmetric semantic alignment is effectively solved, and retrieval efficiency is improved.

The invention relates to the technical field of databases, and discloses a multi-type database performance optimization and operation and maintenance management method and system.The multi-type database performance optimization and operation and maintenance management method comprises the steps that real-time performance indexes of a heterogeneous database cluster are collected, and a noise reduction data set is obtained; generating a cross-library performance coupling degree matrix through the resource competition coupling degree and the data dependence coupling degree; constructing a coupling relation graph; identifying a bottleneck node set based on the coupling relation graph and the abnormal level; generating an optimization strategy; and performing strategy conflict identification according to the optimization strategy, generating a processing strategy, and executing the processing strategy optimization strategy. Through the improved weighting centrality algorithm, the core bottleneck node which has the greatest influence on the whole cluster is accurately identified, a differential optimization strategy is adopted according to the database type, a perfect strategy conflict detection and avoidance mechanism is established, mutual interference in the optimization process is effectively prevented, the optimization success rate is improved, and the optimization efficiency is improved. And the optimization time is shortened.

The invention relates to the technical field of chip design, and discloses a chip dynamic power consumption scheduling method and system based on an intelligent algorithm. The method comprises the following steps of: firstly, acquiring instruction stream data operated by a chip in real time, extracting a feature vector comprising an instruction dynamic change vector and context associated data, and determining a power consumption prediction mapping parameter according to the feature vector; and when the parameter exceeds a preset threshold value, an accurate power consumption prediction result is generated by adjusting the weight of the convolutional neural network. Subsequently, a synchronous timing demand is calculated based on the instruction switching frequency and the data dependency, and an initial power supply configuration is determined. By monitoring task load classification signals, the power consumption distribution proportion is adjusted when the signals are lower than a threshold value, the optimized power supply configuration is obtained, and the improvement index of the resource distribution efficiency is calculated according to the optimized power supply configuration. And finally, according to the index, dynamically adjusting a limiting condition of a scheduling period, and forming a self-adaptive optimization framework, thereby realizing accurate prediction and dynamic optimization scheduling of the chip power consumption.

The invention belongs to the technical field of track detection, and more specifically relates to a track scene thunder-vision fusion anti-intrusion monitoring method and system based on sparse feature fusion. The method comprises the steps that a laser radar obtains three-dimensional point cloud data, a high-definition camera obtains visible light image data, and data preprocessing is carried out; performing external parameter automatic calibration optimization and time alignment compensation on the laser radar and the high-definition camera to obtain time-space aligned multi-modal data; fusing the camera semantic features and the laser radar geometric features based on multi-modal data of space-time alignment to obtain fused features; and after multi-modal feature fusion is completed, a target heat map is generated through a sparse detection head, and graded early warning is realized in combination with a dynamic safety distance model. The method solves the problems that an existing recognition algorithm is high in false alarm rate, and the false alarm rate is higher under low-light conditions such as rainy days or dusk; the data dependence is strong, and the ability to detect unlabeled novel foreign matters is almost zero.

The invention relates to the technical field of cloud edge collaborative computing power scheduling, and discloses a heterogeneous computing power scheduling optimization method based on a cloud edge collaborative architecture. The method comprises the following steps: acquiring real-time computing power state data of all available computing nodes in the cloud edge collaborative architecture; performing heterogeneous type division on the computing nodes according to the real-time computing power state data to generate a three-layer computing power resource pool containing cloud computing nodes, edge computing nodes and terminal computing nodes; extracting task calculation features for the current to-be-scheduled task set, wherein the features comprise calculation intensity, data dependence and real-time requirements; constructing an initial task allocation scheme based on the matching relationship between the task calculation features and the three-layer computing power resource pool; iteratively optimizing the initial scheme by adopting a dynamic load balancing strategy to generate a final task scheduling instruction; the instructions are distributed to the corresponding computing nodes to be executed, and computing power state changes in the task execution process are continuously monitored.

The invention discloses a consumer feedback analysis method based on a multi-agent model, and the method comprises the steps: firstly sampling and generating consumer multi-dimensional feature data, and expanding the data into a semantic portrait through a large model; therefore, an intelligent agent containing a memory module is constructed, and part of the intelligent agent is set as an opinion leader. Commodity attributes, propaganda and competitive products are input into each agent, and initial evaluation is independently output; and then multiple rounds of viewpoint interaction are carried out in the same social evaluation area, the opinion leader has a higher sampled weight, and memory and comments are updated. And after simulation is finished, updating evaluation is generated again and compared with the initial evaluation, attitude changes are analyzed, all data are summarized, and a visual comprehensive report is output. According to the method, accurate simulation and feedback analysis of consumer behaviors are realized by combining a large language model and a multi-agent simulation test technology, and the defects of a traditional method in the aspects of consumer portrait dynamics, agent adaptability, interaction complexity, data dependence and the like are overcome.

The invention discloses a heterogeneous computing power cooperative scheduling system and method for mixed precision training, and belongs to the technical field of artificial intelligence computing. The system comprises a computational graph analysis and operator portrait module which is used for analyzing and dividing a model computational graph and extracting operator features; the heterogeneous hardware capability sensing and matching module is used for managing performance files and real-time states of heterogeneous hardware in the cluster and matching optimal execution hardware for each calculation partition; and the data flow coordination and pipeline parallel controller is used for generating a global execution plan, managing cross-device data dependence and communication and calculating overlapping optimization execution efficiency through communication. According to the method, the problem of low scheduling efficiency of mixed precision training in a heterogeneous environment is solved, automatic and accurate mapping from a calculation task to heterogeneous hardware is realized, the training speed is remarkably improved, the training cost is reduced, and the overall resource utilization rate of a cluster is improved.

The invention relates to the technical field of artificial intelligence and data analysis, in particular to an unsupervised anomaly detection method and system based on comparative potential fusion. The method aims at solving the problems that in the prior art, an unsupervised anomaly detection method is limited in feature expression ability, sensitive in noise, insufficient in potential feature discrimination and lack of statistical interpretability in detection results. According to the method, the global potential features generated by comparison learning and the self-encoder reconstruction residual error are fused, the statistical model is combined for self-adaptive threshold judgment, the problems of insufficient feature expression and high noise sensitivity in multi-source heterogeneous time series data anomaly detection are effectively solved, and the method has the advantages that the detection precision and robustness are improved, and the dependence on labeled data is reduced.

The invention relates to the technical field of multi-peak spectrum demodulation, and particularly provides a fiber bragg grating multi-peak spectrum demodulation method and system. The method comprises the following steps: extracting local spectral features based on an experimental reference spectrum to form initial atoms, and performing translation offset and normalization processing to obtain an over-complete spectral atom dictionary; performing global offset preliminary estimation based on the dictionary, and obtaining preliminary estimation values of peak sites of the measurement spectrum and the reference spectrum through cross-correlation calculation; executing constraint orthogonal matching pursuit sparse recovery based on the estimated value, and recovering atomic displacement from the measurement spectrum by using block sparsity, translation consistency and non-negative constraint; and performing wind speed inversion and calibration based on the atomic displacement, and converting the wind speed into a wind speed estimated value through a nonlinear calibration model to obtain a final result. According to the method, a dictionary based on experimental data is constructed, dependence on large-scale labeled data is reduced, a physical mechanism and sparsity prior are fused, and the problems that a traditional method is insufficient in precision and weak in generalization ability in a complex environment are solved.

The invention relates to a distributed data unified management and intelligent scheduling method based on data braiding, and the method comprises the steps: collecting the real-time state data of a node, and generating a node capability portrait; constructing a dynamic incidence matrix of the metadata and the node capability portraits; establishing a multi-dimensional scheduling evaluation model, inputting a business data dependency relationship, performance data in the node capability portrait and a node future load predicted by the LSTM model, and outputting an initial scheduling scheme; a static threshold value and a dynamic prediction threshold value are preset to serve as scheduling optimization triggering conditions, the initial scheme is optimized through a reinforcement learning algorithm, and an optimized scheduling decision is obtained; and sending a scheduling instruction containing a priority identifier to a corresponding node, collecting feedback data such as response delay and an error rate in real time, updating the association strength of the dynamic association matrix according to the feedback data, and optimizing the parameters of the multi-dimensional scheduling evaluation model. Unified management of distributed data is achieved, scheduling intelligence and accuracy are improved, node state changes can be dynamically adapted, and data processing efficiency and reliability are effectively guaranteed.

The invention discloses a dynamic leakage fault diagnosis method for a flexible hand for a deep-sea submersible vehicle, and belongs to the technical field of flexible hands for the deep-sea submersible vehicle, and the method comprises the steps: obtaining source domain pressure signals from a plurality of different bending angles, and carrying out the preprocessing of the signals, and obtaining an original multi-source domain data set; a fault diagnosis model is constructed and trained, and in the training stage, the fault diagnosis model is a student network and comprises an adaptive frequency spectrum intervener, a multi-scale feature extractor, an adversarial mask generator, a main classifier and an auxiliary classifier; and inputting a to-be-diagnosed pressure signal into the multi-channel feature extractor and the main classifier in the trained student network, and outputting to obtain a fault diagnosis result. The method has remarkable advantages in the aspects of improving diagnosis precision, enhancing generalization ability, improving weak fault detection, reducing data dependence and the like, is suitable for intelligent diagnosis of hydraulic leakage of the deep-sea submersible flexible hand under dynamic and multi-angle working conditions, and has a good engineering application prospect.

The invention discloses a weak supervision pathological image tissue segmentation method based on text prompt learning. The method comprises the steps of feature extraction and initial class activation graph generation; using an MCRM module to optimize the initial class activation graph to obtain a refined class activation graph; and aggregating the plurality of refined class activation graphs to form a fused pseudo mask, taking the fused pseudo mask as a supervision signal, training a segmentation model, and after the training is completed, segmenting the new pathological image tissue by using the segmentation model. According to the method, a text prompt learning mechanism is utilized to focus the model on learning high-discrimination features, so that the influence of tissue co-occurrence is reduced. An initial class activation graph is optimized through a multi-mode class activation graph refining module, and the integrity of boundary segmentation is enhanced. Meanwhile, pseudo masks from different network layers are fused to train a segmentation model, and semantic segmentation of the pathological image is realized. According to the method, high-annotation data dependence is effectively relieved, and the generalization ability of the model is improved, so that application in the field of artificial intelligence-assisted medical treatment is promoted.

The invention relates to a real-time process monitoring method based on data analysis, and the method comprises the following steps: S1, building a three-dimensional component priority evaluation model based on a component operation scene type, a real-time resource occupancy rate and a data dependence degree, and dynamically matching an acquisition strategy; s2, a cleaning rule is adapted according to a data source, a feature extraction dimension is adjusted in combination with process dynamic features, and an improved time sequence decomposition algorithm is used for separating data trends, fluctuations and abnormal residual errors; s3, constructing an exclusive baseline sub-model by using an online learning algorithm according to a scene label, and establishing a scene switching mechanism; s4, in combination with component interaction anomaly features, an anomaly level is judged through a mixed detection model; s5, on the basis of exception processing and user feedback, constructing an error correction model optimization parameter; and S6, generating a report containing an abnormal propagation path, and triggering hierarchical collaborative response of the associated component. The invention aims to solve the problems of single acquisition dimension, no scene adaptability in preprocessing, incomplete abnormal detection and the like in the existing monitoring technology.

The present application discloses a semi-supervised medical image segmentation method and related devices based on contrast manifold regularization. The method includes: importing and preprocessing medical image datasets; initializing a teacher model and a student model, using labeled data to train the teacher model and generate pseudo labels; calculating the similarity between the labeled data and the pseudo labels to obtain a manifold regularization term; constructing positive and negative sample pairs and calculating the contrast loss term; weighted summation to obtain the contrast manifold regularization term, and combining it with the supervised loss as the loss function of the student model; iteratively training the student model to obtain the segmentation result. By combining contrast learning with manifold regularization, the present invention effectively alleviates the data dependency problem in semi-supervised medical image segmentation, improves the model generalization ability and segmentation accuracy, and performs particularly well in small target lesion segmentation scenarios.

The invention provides a simulation deduction method and system based on distributed parallel scheduling and a storage medium, and is applied to the technical field of data processing.The method comprises the steps that proxy services are deployed at a plurality of preset computing nodes, resource state information of all the nodes is dynamically collected, and a computing resource pool is obtained; arranging a pre-registered plug-in based on the data dependency relationship to obtain a simulation task process; in response to simulation scene configuration submitted by a user, generating a distributed scheduling task containing a fragmentation strategy; determining a plurality of sub-tasks corresponding to the distributed scheduling task based on the simulation task process; and based on the real-time load of the computing resource pool and the task priorities of the plurality of sub-tasks, respectively distributing the plurality of sub-tasks to a plurality of target computing nodes for task execution, and obtaining task execution result information. According to the method and the device, the time delay performance of distributed scheduling and the high availability performance of scheduling simulation tasks can be improved.

This application provides a task processing method and chip, relating to the field of computer technology. The task processing method includes: placing all instructions in an instruction sequence used to implement a task into multiple instruction queues corresponding one-to-one with multiple execution units; determining the processing state of a second instruction queue that the first instruction queue depends on in the data dependency relationship indicated by the relation identifier based on a relation identifier corresponding to a restricted instruction at the head of a first instruction queue; and, if the processing state is complete, retrieving the restricted instruction corresponding to the relation identifier from the first instruction queue to execute the instructions following the restricted instruction; wherein the processing state is updated according to the relation identifier; and completing the processing of the multiple instruction queues to obtain the task processing result. This application can reduce the compiler burden when different tasks are executed asynchronously, improving task processing efficiency.

The invention provides a three-axis machine-oriented non-training workpiece edge detection method, relates to the technical field of image segmentation, and solves the technical problems of dependence on high-cost pixel labeling, insufficient model generalization ability and reduction of segmentation precision in the prior art. The method comprises the following steps: acquiring standard and target workpiece images and features thereof; performing screening and position clustering based on the standard and target block feature sets to obtain a positive and negative prompt point set; carrying out range framing to obtain a prompt box; inputting the standard workpiece image, the target workpiece image, the positive and negative prompt point set and the prompt box into an image segmentation model to obtain a preliminary segmentation mask; performing false positive rejection to obtain a foreground mask; and the workpiece edge is obtained through edge processing. The method is used in the workpiece edge detection process, and a non-training workpiece edge detection scheme only needing a single standard image is provided for a three-axis machine. According to the method, the data dependence and the labor cost are remarkably reduced by automatically generating the prompt points, and the segmentation precision is improved by utilizing the feature re-matching module.

The invention discloses a legal affair information analysis method and system, a computer device, a medium and a product. The method comprises the following steps: analyzing legal affair information to be analyzed to generate an initial structured text; analyzing the initial reference law article based on the standard law article database, and updating when a law article analysis result does not meet requirements; logical reasoning verification is carried out on the initial logic information, and when a logic chain analysis result does not meet requirements, updating is carried out; comparing the initial analysis conclusion with a preset legal knowledge graph, determining text segments with potential errors, and performing confidence analysis to obtain a corresponding confidence analysis result; and according to the optimized initial reference law article, the optimized initial logic information and all confidence analysis results, generating a comprehensive, accurate and reliable analysis scheme. According to the method, data dependence can be reduced, the problems existing when a general large model is applied in the legal field are effectively solved, model output illusion is greatly relieved, and legal information analysis quality is improved.

The invention discloses an AI large model intelligent routing and dynamic arrangement method, system and device and a medium. The method comprises the following steps: in response to an input request, performing multi-dimensional feature extraction on the input request to obtain a request feature vector; obtaining the real-time state and metadata of each AI large model, and inputting the request feature vector and the real-time state and metadata of each AI large model into a multi-target optimization model to obtain a candidate model combination corresponding to the input request; according to the dynamic arrangement template, determining a calling sequence and a data dependency relationship of each AI large model in the candidate model combination, and generating an execution plan of the candidate model combination; and the input request is adapted and sent to the candidate model combination, each AI large model in the candidate model combination is called based on the execution plan, and an execution result is obtained and returned to the request terminal. Through the above mode, the candidate model combination can be intelligently selected, the accuracy and efficiency of model selection are improved, and a plurality of candidate models are flexibly arranged to complete a complex task.

The invention provides an inference model processing method and device, electronic equipment and a storage medium. The processing method comprises the steps that in response to an ith layer calculation result of an ith layer calculation operator, the ith layer communication operator starts transmission of the ith layer calculation result, and i is an integer and is larger than or equal to 0; during the period when the ith layer communication operator transmits the ith layer calculation result, the (i + 1) th layer calculation operator is started to process the first data block, the (i + 1) th layer calculation result output by the (i + 1) th layer calculation operator and the ith layer calculation result have a data dependency relationship, and the (i + 1) th layer calculation result output by the (i + 1) th layer calculation operator and the ith layer calculation result have a data dependency relationship; the first data block is a data part which has no data dependence relationship with the ith layer of calculation result in the to-be-processed data of the (i + 1) th layer of calculation operator. According to the processing method, the calculation operator and the communication operator in the reasoning model driven by the processor are executed in an overlapping manner, so that the utilization rate of a hardware unit in the processor for executing the reasoning model is effectively improved.

The invention discloses a multi-data center task scheduling and data routing method and system under network topology, and belongs to the technical field of cloud computing. According to the scheme, multi-geographic position data center scheduling under a complex topology network is modeled as a Markov decision process, and a state, an action, a reward function and a transfer function are defined; the method comprises the following steps of: establishing a model to reflect the change of task and resource states, modeling a data transmission process as a Boolean linear programming problem, optimizing a transmission starting point and a transmission path of data in a data center network topology, and combining the minimum energy consumption cost of data transmission with an MDP reward function to obtain a data center network topology. The intelligent agent can learn to balance the energy consumption cost of task scheduling and data transmission under multiple conditions of complex network topology, resource heterogeneity, task heterogeneity, complex task-data dependency relationship, network bandwidth capacity constraint and the like, and overall optimization of energy consumption of the data center is realized.

The invention belongs to the technical field of fault diagnosis, and discloses a thermal defect identification method and system for a high-voltage switchgear, and computer equipment, and the method comprises the steps: firstly segmenting an infrared image through employing a transfer learning optimized Mask R-CNN model, reducing the dependence of annotated data through sharing pre-training parameters, and achieving the region extraction of pixel-level equipment; secondly, multi-dimensional temperature information is extracted in combination with a gray histogram and a gray co-occurrence matrix, and key features are screened through PCA to enhance noise immunity; and finally, the LSSVM is adopted for classification, so that the training efficiency is remarkably improved. According to the method, the equipment area is automatically segmented through deep learning, temperature distribution is quantified in combination with multi-dimensional features, man-made misjudgment is reduced, pre-training model parameter sharing is utilized, new tasks are rapidly adapted in a small sample scene, the generalization ability and efficiency are improved, feature dimensions are compressed through principal component analysis, the real-time monitoring requirement is met, and the monitoring efficiency is improved.

The invention provides a cloud edge-end collaborative deployment method and system for an image reasoning multi-mode neural network, and is applied to the field of artificial intelligence, and the method comprises the steps: abstracting a preset multi-mode neural network, and obtaining an original calculation graph; based on the calculation node set and the data dependence edge set, splitting the original calculation graph to obtain a plurality of connected sub-graphs, the categories of the plurality of connected sub-graphs including a single-modal processing sub-graph, a cross-modal interaction sub-graph and a task specific sub-graph; based on the topological sequence of the plurality of connected sub-graphs, determining segmentation points of each connected sub-graph; dividing the plurality of connected sub-graphs according to the segmentation points to obtain a plurality of segmentation parts; and integrally dividing the multi-modal neural network based on the plurality of segmentation parts according to a genetic algorithm, and respectively deploying the integrally divided multi-modal neural network to a terminal device, an edge server and a cloud platform. According to the invention, a flexible and efficient model segmentation and deployment mechanism can be realized.

The invention provides an MCM-GPU-oriented resource storage optimization method, device and equipment, and the method comprises the steps: carrying out the static analysis of an operator-level data flow diagram of a calculation task in a compiling stage, and extracting the data dependence relation and life cycle characteristics of each operator; generating a storage optimization scheme based on the data dependency relationship and life cycle characteristics of each operator; in the operation stage, task and data collaborative allocation is carried out according to a data prefetching strategy and a data replacement strategy, task and data collaborative allocation and the calculation task are executed in parallel, and a data flow diagram and an operator dependency relationship of the task are analyzed by utilizing a compiling period; in combination with the data locality characteristics of the MCM-GPU multi-level storage architecture and the concurrency requirements of different modules for data access, an appropriate GPU module is selected to carry out data optimization management, so that frequent data exchange between a GPU and a CPU is reduced, and the performance potential of the MCM-GPU architecture in a super-large-scale task is brought into full play.

The invention discloses a remote sensing image semantic segmentation method based on a multi-stage subtitle-driven diffusion model, and the method comprises the steps: firstly obtaining an original remote sensing image semantic segmentation data set, designing an instance segmentation strategy, and obtaining a remote sensing single-target instance sub-image data set; secondly, providing a two-stage remote sensing semantic subtitle generation algorithm, and migrating a pre-training diffusion model based on Stable Diffusion to a remote sensing scene by combining a cutting instance and a conditional fine tuning strategy to obtain a diffusion model adaptive to the remote sensing scene; thirdly, constructing a multi-layer weighted attention image-semantic mask joint generation framework based on the adaptive model, and generating a high-quality remote sensing target image and a semantic mask; then, providing a cross-scale semantic constraint data synthesis method based on a ground sampling distance to obtain enhanced remote sensing image data; and finally, training a divider by using the enhanced data to realize accurate semantic segmentation of the remote sensing image. The method can effectively alleviate the dependence of annotation data, and improves the segmentation precision and scene adaptability.