17results about How to "Reduce computing resource consumption" patented technology

The invention provides a tumor medical image classification method of zero sample evolution NAS based on double-index collaborative evaluation, which comprises the following steps: collecting tumor medical image data, and constructing a data set containing a tumor focus image and a normal tissue image; constructing an extended search space based on a cell structure, wherein the search space comprises a basic operation set, the cell structure and a network overall structure; steady-state evolution algorithm parameters are set, and steady-state evolution initialization is carried out to obtain an initial candidate architecture; taking a double-index comprehensive score calculated by the information rate and the FIM stability as a comprehensive index for evaluating the current candidate architecture, updating the candidate architecture through a steady-state evolutionary algorithm based on a dynamic optimization mechanism until the maximum evolutionary algebra is reached, and obtaining an optimal architecture; and training the optimal architecture by using the data set, and carrying out tumor medical image classification by using the trained optimal architecture. According to the method, the dependence on the annotated data is reduced, the consumption of computing resources is reduced, and meanwhile, the diagnosis efficiency and accuracy are improved.

The invention discloses a navigation mark visual identification and collision early warning method and system based on deep learning, and relates to the technical field of intelligent shipping, and the method comprises the steps: collecting a navigation channel monitoring video stream in real time, and carrying out the preprocessing of a key image frame; constructing a navigation mark detection model based on an improved Officient Det network, and outputting the position and category of the navigation mark in the key image frame; the method comprises the following steps: constructing a ship track prediction model through ship historical track data, predicting a short-time track of a ship, calculating the relative position, speed and course angle of the ship and a navigation mark, and constructing a dynamic collision risk field; and fusing the visual detection result, the radar ranging data and the AIS information to obtain a ship collision risk, and triggering graded early warning based on a preset risk threshold. According to the method, the improved deep learning model and the real-time calculation framework are combined, the robustness of navigation mark identification is improved, an efficient collision early warning mechanism is established, and technical support is provided for intelligent shipping.

The present application belongs to the field of big data processing, and specifically relates to a multi-supplier cooperative store digitized management system and method. Multi-supplier historical delivery time sequence data and store warehouse space time-varying occupation data are collected; a dynamic space-time graph is constructed with supplier nodes and store physical storage location nodes as vertices and logistics time delay and supply-demand dependency as edges; a time sequence attention mechanism is embedded into a graph convolution network to extract space-time coupling features in a concurrent performance state, representing the dynamic mapping correlation between supplier delivery time sequence and store storage location release time sequence; a storage location allocation and delivery time sequence joint optimization model is constructed based on the space-time coupling features, with physical storage location capacity constraints and logistics distribution time sequence constraints as boundary conditions. Dynamic adaptation of cross-supplier logistics time sequence and store physical space is achieved, physical warehouse space occupation and blockage defects caused by concurrent warehousing are overcome, and inventory accumulation and shortage phenomena caused by supply-demand space-time mismatch are eliminated.

The invention provides a power grid dispatching method, a control device and a storage medium, and belongs to the technical field of power grids. The method comprises the following steps: inputting acquired operation data into a multi-objective optimization model to output an optimization scheduling variable; performing constraint condition verification on the optimal scheduling variable, when the optimal scheduling variable meets a preset constraint condition, generating a scheduling scheme of the target power grid, and issuing a scheduling scheme instruction; in the process that the target power grid executes the scheduling scheme instruction, monitoring an operation index of the target power grid; and calculating the deviation between the monitored operation index and the optimization target, and when the deviation exceeds a preset threshold value, adjusting the scheduling scheme of the target power grid. Through a technical path of offline training and online reasoning of the neural network, the generation time of the scheduling scheme is greatly shortened, and real-time response to dynamic change of the power grid is realized.

The application discloses a kind of based on cross-modal data class increment learning's unmanned plane inspection method and system, comprising: by cross-modal feature mixed fusion module, the features of each mode are spliced and reorganized;Each group of processed input feature map is assigned a set of independent convolution kernel, and grouping convolution is carried out, and finally complete feature map is generated;For each pooling window in complete feature map, the average value of all elements in the window is calculated according to adaptive average pooling, the average value is used as the element of the corresponding position of the output feature map, the process is repeated for all regions, and the final output feature map is obtained;The output feature map obtained after being processed by the adaptive average pooling module is flattened and input into the fully connected layer to extract and combine features, generate the final classification output, then update the model parameters according to the gradient of the loss function;The trained model is deployed on the unmanned plane to collect visible light and thermal imaging images in real time, extract features and identify and predict through the model.

This invention provides a task execution method, apparatus, electronic device, and storage medium, relating to the field of embodied intelligence technology. The method includes: determining redundant feature components from the first task reference information based on the feature sources of feature components and the importance values ​​of each feature source, wherein the importance value of each feature source represents the magnitude of the influence of the feature component of that feature source on the successful output of the robot's expected action by the first VLA model; performing redundancy removal processing on the redundant feature components in the first task reference information to obtain processed first task reference information; inputting the processed first task reference information into the first VLA model to obtain the predicted action of the robot output by the first VLA model; and controlling the robot to execute the task based on the predicted action. Applying the solution provided by this invention can improve the robot's motion control frequency, thereby improving task execution performance.

The invention relates to the technical field of damage detection of oil and gas transportation pipelines, in particular to a marine pipeline damage degree intelligent identification method based on multi-source feature deep fusion. The method comprises the following steps: S1, acquiring damage data of a damaged pipeline; s2, performing noise reduction on the damage data by adopting an improved wavelet threshold noise reduction method; s3, performing normalization processing on the data after noise reduction; s4, inputting the normalized data into a preset multi-feature deep fusion white box model for training; and S5, calculating the accuracy to reflect the accuracy of model identification. The method overcomes the defects of a hard threshold function and a soft threshold function, flexibly switches between the hard threshold function and the soft threshold function by adjusting the factor alpha, reduces discreteness, avoids the problem of constancy, is high in adaptability, can be automatically adjusted according to different signal characteristics, effectively removes the noise of experimental acquisition data, improves the signal quality, and greatly improves the identification accuracy of the damage degree.

The invention discloses a steel material temperature field time sequence prediction method and system based on multi-objective evolution, and relates to the field of steel material laser cladding temperature field analysis, and the scheme comprises the steps: obtaining historical temperature field data determined through a finite element simulation strategy; processing the historical temperature field data according to a data set division strategy to obtain a training set and a verification set; carrying out training and parameter adjustment on the deep learning model so as to determine a basic value interval corresponding to each parameter of the model; and determining a target parameter value corresponding to each parameter in combination with a preset multi-target evolutionary optimization strategy so as to retrain the deep learning model to obtain a target deep learning model, and determining predicted temperature field data corresponding to the current process parameter in a future time step based on the target deep learning model. According to the scheme, the deep learning model is introduced into base material laser cladding temperature field data prediction, the prediction precision is kept, meanwhile, the calculation time is greatly shortened, and the overall calculation efficiency is remarkably improved.

The application discloses a kind of force partition weighted railway bed dirt visual detection method and system, belong to railway engineering intelligent detection technical field.The application constructs the dynamic mapping model of interframe displacement and field of view overlap rate by real-time acquisition train motion parameter, and effectively frame is adaptively filtered;Adopt coarse cutting-fine identification hierarchical parallel vision architecture to identify dirty area;Based on the stress characteristics of bed, high stress area and low stress area are divided, and a mechanical weighting model is introduced to calculate the comprehensive risk index;The risk index is associated with the geographic location to generate a risk space distribution map, triggering hierarchical early warning and maintenance decision.The application solves the problems of data redundancy, heavy computing burden, and the disconnection between evaluation indicators and stress state in the prior art, achieving efficient and accurate detection of bed dirt and risk quantification, providing a scientific basis for railway maintenance decisions, and being suitable for heavy haul railways, coal lines and other scenarios.

The present disclosure provides a game processing method, a game processing device, a program product and an electronic device, and relates to the technical field of games. The method comprises the following steps: in the action link of a first virtual character, in response to a first attack event, if it is determined that the attack is successful according to attack resources associated with the first attack event and defense resources associated with an attack object, a target game area is triggered for the first virtual character to obtain; in the action link of a second virtual character, in response to a second attack event of the second virtual character triggering a first defense event of the first virtual character, if it is determined that the attack is not successful according to attack resources associated with the second attack event and defense resources associated with the first defense event, the game area already possessed by the first virtual character is maintained; and in response to the end of a game match, the game area possessed by each virtual character is settled. The present disclosure reduces the consumption of computing resources of a terminal or a server, and improves the game strategy.

The application relates to the technical field of elevator intelligent monitoring, and discloses an elevator abnormal behavior recognition and early warning method and system based on AI vision analysis, which comprises the following steps: receiving a door machine state signal, waking up a static target analysis process when the door is opened, calculating a target physical size based on perspective transformation and outputting a static early warning, reading a high-frequency load sampling sequence to generate a physical dynamics sequence when the door is closed, triggering a hardware interrupt when a first-order difference of the physical dynamics sequence exceeds a threshold value, waking up a dynamic behavior analysis process in response to the interrupt, intercepting a video stream to extract moving foreground pixels and projecting the moving foreground pixels to a two-dimensional grid of a car bottom plane, calculating a visual kinematics acceleration sequence and a spatial distribution variance, using a dynamic time warping algorithm to obtain a time domain matching distance parameter of the physical dynamics sequence and the visual sequence, and constructing a joint confidence based on the spatial distribution variance and the time domain matching distance parameter to output an early warning. Through multi-modal data fusion and dynamic time alignment, the application reduces the consumption of computing resources and improves the recognition accuracy.

The invention discloses an electric power operator behavior edge identification and AR correction method and system based on a double-flow heterogeneous lightweight network. The method comprises the following steps: acquiring a first visual angle video stream and sensor data of a six-axis inertial measurement unit through intelligent equipment; performing normalization and de-noising processing on the video frame, and segmenting the video frame into short-time video clips according to a time window; running the pre-trained double-flow heterogeneous lightweight network, performing real-time reasoning on the video clip, and outputting a current behavior category label; the behavior confidence degree output by the double-flow heterogeneous lightweight network model is compared with a preset threshold value to judge whether the event is a violation event or not; if the event is identified as a violation event, a deviation correction mechanism is triggered immediately; and encrypting and uploading the violation event log to an enterprise security management platform or other third-party platforms for users to check, count and trace. The method can effectively solve the problems of high behavior identification delay, dependence on cloud, low edge equipment operation efficiency, insufficient multi-behavior identification precision and the like in the prior art.

The invention discloses a code abstract generation method and device supporting a low-resource programming language. And translating a source code of a low-resource programming language into candidate codes of a target high-resource programming language by utilizing a large language model, and selecting a code which is closest to the input source code semantically from the candidate codes passing analysis verification as an optimal translation code. And retrieving a plurality of codes similar to the optimal translation code and abstracts thereof from a preset knowledge base as reference examples. And segmenting the optimal translation code and the reference example code into code statements, and screening out core statements. And taking the code, the core statement and the abstract of the reference example and the optimal translation code and the core statement as prompt contexts, inputting into a large language model, and outputting a code abstract of a source code of a low-resource programming language. Through cross-language translation and multi-source information fusion, under the condition that low-resource language training data and zero fine tuning are not needed, the problem that low-resource language corpora are deficient is effectively solved, and high-quality code abstract generation is achieved.

The invention relates to the technical field of aircrafts, and provides an environment sensing method and device and an aircraft. The method comprises the following steps: acquiring a current environment image of an aircraft, and performing feature extraction on the current environment image to obtain image features of the current environment image; obtaining the prior distribution of the obstacle, wherein the prior distribution of the obstacle is generated in advance according to the three-dimensional occupied grid sample set; performing aggregation processing on the image features according to prior distribution of obstacles to obtain aggregation features; based on the aggregated feature, a three-dimensional occupancy grid corresponding to the aircraft is predicted, the three-dimensional occupancy grid being used to indicate an obstacle present in a flight environment of the aircraft. The sensing precision and real-time performance of the aircraft in a complex flight environment can be remarkably improved, and the flight stability and safety of the aircraft are effectively guaranteed.

The present application relates to the technical field of artificial intelligence, and particularly relates to a model fine-tuning method for privacy protection, comprising the following steps: S1, obtaining a basic model to be fine-tuned and injecting an initial low-rank adapter containing a dimension reduction matrix and a dimension increase matrix, and taking parameters of the initial low-rank adapter as current low-rank adapter parameters; S2, obtaining local original training data containing real labels, identifying privacy identifiers in the local original training data and replacing the privacy identifiers with placeholders to obtain training data containing local target training samples and the real labels. In the present application, through clipping and Gaussian noise processing of single-sample gradient in the local training stage and in combination with data desensitization preprocessing, the local original training data does not leave the local client in plaintext form in the whole process of participating in model fine-tuning, so that effective privacy protection is provided for sensitive data of users in a federated distributed training scene.

The application discloses an edge-computing-based lighting control panel energy consumption real-time optimization and early warning system and relates to the technical field of energy consumption optimization.The system solves the technical problem that independent optimization of control panel data uploaded to the cloud causes waste of calculation and bandwidth and poor real-time performance because a large number of control panels have similar states, and that a group aging benchmark is lacking, so that life balance cannot be achieved, resulting in premature failure of some equipment.Through clusters divided based on historical similarity maps, combined with average aging tracks in the clusters and real-time group feature centers, global collaborative optimization parameters are dynamically calculated by a multi-panel collaborative optimization module.The parameters are sent to the edge end, so that control panels in the same cluster with similar behavior characteristics can be collaboratively adjusted, redundant lighting and energy consumption waste in the region can be avoided, and energy consumption optimization at the group level can be achieved.Through construction of the average aging track in the cluster, the normal aging trend of each cluster can be mastered.

The embodiment of the invention provides a vehicle control method and device, electronic equipment and a vehicle. A first planned path corresponding to a rear axle of the vehicle is obtained according to an initial planned path of the vehicle, and obstacle information around the vehicle is obtained; determining a second planned path corresponding to the front end of the vehicle according to the first planned path; according to the obstacle information, determining the minimum projection distance between the obstacle and the first planning path and the minimum projection distance between the obstacle and the second planning path; determining a comprehensive distance between the obstacle and the initial planning path according to the displacement of the obstacle relative to the first planning path and the minimum projection distance; and controlling the vehicle to run according to the corresponding acceleration or the corresponding braking parameters according to the comprehensive distance. According to the embodiment of the invention, the comprehensive influence of the vehicle contour and the obstacle on the planned path can be fully considered while the demand for computing resources is reduced, and the output and response efficiency of the vehicle control instruction is improved.