215 results about "Network embedding" patented technology

The invention discloses an asphalt pavement fatigue damage model calibration method based on AI and digital twinning, and relates to the technical field of damage detection. Compared with the prior art, the problems that traditional asphalt pavement design depends on empirical formulas and static parameters, material aging, environmental coupling and load uncertainty are difficult to dynamically reflect, and cross-scale correlation between microscopic interface behaviors and macroscopic structure performance is lacked are solved; the method comprises the following steps: collecting and preprocessing multi-source heterogeneous data in real time through a multi-modal data fusion and dynamic sensing system, simulating and accurately quantifying asphalt-aggregate interface binding energy and adhesion work in combination with molecular dynamics, and constructing a cross-scale damage evolution model; a physical information neural network is used for embedding an improved Paris formula, actually measured data and a physical rule are deeply fused, and the locality hypothesis of a traditional model is effectively corrected.

The invention relates to the technical field of geotechnical mechanics and engineering, and particularly discloses a soft soil foundation settlement monitoring method and system based on multi-field multi-source information, and the method comprises the steps: collecting multi-source data of a target soft soil area; performing constitutive parameter inversion, data standardization and discrete Fourier transform frequency domain conversion on the data to obtain a standardized frequency domain multi-field multi-source data set; based on the data set, a soft soil constitutive parameter library and a soil mechanics physical equation, constructing an FD-PINN frequency domain physical information neural network, embedding the physical equation as a prior constraint, and training the model by adopting an alternating optimization strategy; inputting the real-time frequency domain data flow into the model, and outputting the current settlement amount, the settlement rate and the multi-physical field frequency domain distribution; and in combination with a pre-established large scale model test result, through IDFT inverse transformation, a time-space domain settlement field is reconstructed, multi-stage early warning is triggered, a targeted reinforcement scheme is recommended, the soft soil foundation settlement monitoring precision and the engineering practicability are remarkably improved, and the method is suitable for construction, operation and maintenance of infrastructures such as high-speed rails and highways.

The invention discloses a perceptually optimized wiring time sequence prediction method for a digital integrated circuit, electronic equipment and a storage medium, and the method comprises the steps: constructing a heterogeneous graph according to a circuit netlist, fusing the heterogeneous graph with feature vectors corresponding to all units and interconnection lines, inputting the fused heterogeneous graph into a trained heterogeneous graph neural network, and obtaining graph network embedding; inputting the feature matrix of the layout into the trained convolutional neural network to obtain convolutional network embedding; splicing graph network embedding and convolutional network embedding to obtain multi-modal local embedding; constructing a feature sequence of the path according to a connection relationship between a logic unit and an interconnection line in the path by multi-modal local embedding, and inputting the feature sequence of the path into the trained Transform network to obtain a path delay residual prediction value after wiring; and adding the path delay value after layout and the path delay residual prediction value after wiring, and finally obtaining a path delay prediction value after wiring. According to the method, the path delay after wiring can be accurately and efficiently predicted before wiring.

The invention discloses an entity digitization and link framework algorithm based on a heterogeneous graph attention network, and the algorithm comprises the following steps: S1, heterogeneous information network construction: carrying out the unified modeling of all multi-source heterogeneous data into a heterogeneous information network containing various types of nodes and various types of edges, s2, meta-path definition and guidance: defining "meta-paths" connecting different types of nodes to capture a complex deep semantic relationship, S3, heterogeneous graph attention network embedding: adopting an attention mechanism to enable a model to automatically learn importance of different neighbor nodes and different meta-paths, generating a final embedding vector of each entity, and establishing a heterogeneous graph attention network model; according to the method, the information fidelity is higher, modeling is directly conducted on different types of nodes and relations on a heterogeneous graph, more abundant and heterogeneous semantic information in data can be reserved compared with a multi-view method, and the end-to-end learning ability is higher; and the complexity of manually designing a fusion strategy is reduced.

The invention discloses an intelligent identification system for an osteoporosis area, belongs to the field of image processing calculation, and aims to solve the problems of limited technical coverage and lack of a dynamic optimization mechanism. In a feature extraction stage, a system combines traditional image processing and deep learning technologies in parallel, extracts bone trabecula multidirectional texture features by using a Gabor filter and a local binary pattern algorithm, analyzes bone contour curvature by combining Sobel edge detection and morphological operation, constructs geometric morphological parameters, and performs feature extraction on the bone trabecula. The local branch focuses on the porosity and arrangement rule of the bone trabecula by adopting a 3D convolutional network, the global branch is embedded into a compression excitation module based on an improved MobileNetV3 network to strengthen the overall morphological expression of the bone, the response intensity of a local microstructure is enhanced by space attention, the weight of global bone topological characteristics is calibrated by channel attention, and a multi-scale splicing strategy is combined, so that the overall morphological expression of the bone trabecula is optimized. And finally, outputting a fusion feature matrix after noise suppression, and remarkably improving the expression ability of pathological features.

The invention discloses an on-orbit spacecraft temperature prediction method and device based on a physical information neural network. The method comprises the following steps: S1, constructing an object model and an orbit heat source model according to a spacecraft to be predicted; s2, solving a space external heat flow of the orbit heat source model, and converting the space external heat flow into a volume heat flow of the object model; s3, constructing a physical information neural network; wherein a three-dimensional transient heat conduction equation is embedded in the physical information neural network; s4, designing a total loss function according to the volume heat flow and the physical information neural network; and S5, training the physical information neural network according to the total loss function to obtain a temperature prediction result of the spacecraft to be predicted. According to the method, the situation that model grids need to be divided in traditional finite element simulation is avoided, the calculation efficiency is improved, the calculation time is shortened, the temperature result can be predicted in real time, and the generalization ability and accuracy of the model are improved.

The invention provides a multi-modal fusion power transmission line inspection method and device, and relates to the technical field of power transmission line inspection application, and the method comprises the steps: collecting a multi-modal image of a to-be-inspected power transmission line at a current moment; inputting a feature pyramid network embedded into a multi-scale context adaptive joint learning network to obtain a multi-modal image fusion feature, and judging whether an abnormal event exists in the image corresponding to the to-be-inspected power transmission line under the action of the multi-modal image fusion feature; if the position information of the unmanned aerial vehicle exists, establishing an incidence relation function between the position information of the unmanned aerial vehicle at the current moment and the multi-modal image fusion features based on the similarity between the multi-modal features representing the position information of the unmanned aerial vehicle; inputting the corresponding image of the to-be-inspected power transmission line with the abnormal event mark into the incidence relation function, and determining longitude and latitude position information of the abnormal event in the to-be-inspected power transmission line; the technical problem that a power transmission line has potential safety hazards due to poor accuracy of an inspection result is solved.

Systems and methods for minimizing dimensionality of a high-dimensionality dataset during feature engineering. The system achieves this by using a tabular neural network to extract non-linear transformations of features without dramatically increasing the dimensionality of the original dataset. The system receives an original dataset for classification and a defined number of final features (e.g., dimensionality) that result from the synthetic feature creation and the neural network embedding process. Once an architecture of a model is determined, a model is fit on a synthetic feature set (e.g., a second dataset comprising synthetic features) with a given classification as a target.

The invention discloses an infrared anti-unmanned aerial vehicle detection method based on an improved YOLO framework. The infrared anti-unmanned aerial vehicle detection method comprises the following steps: S1, preprocessing an infrared image and calculating energy distribution; s2, embedding the backbone network into an improved convolution module and adaptive wavelet basis convolution, and optimizing feature extraction in stages; s3, reinforcing local details and global contours of the feature maps by a multi-granularity pooling mechanism; s4, realizing cross-layer feature fusion through a feature aggregation propagation mechanism, and generating a comprehensive multi-scale feature map; s5, a lightweight detection channel is specially arranged to enhance the small target detection capability, and other standard detection heads are responsible for classified positioning of medium and large targets; and S6, outputting a final prediction result through a context sensing weighted frame fusion algorithm. Through adaptive wavelet feature enhancement, multi-granularity feature optimization, a lightweight detection head and an intelligent frame fusion technology, the problems of small target leak detection, background interference and multi-scale detection real-time performance in an infrared scene are solved, and the all-weather detection capability and the positioning precision are remarkably improved.

A device may receive first information associated with one or more properties, and second information associated with one or more properties, where the first information is of a different information type than the second information. The device may generate a first graph network based on the first information, and a second graph network based on the second information. The device may combine the first graph network and the second graph network to generate a combined network including a portion of the first information and a portion of the second information. The device may generate an embedding based on the combined graph network. The device may apply the feature embedding as input to a housing model to generate an output.

The invention provides a potential predictability embedded deep learning industrial process soft measurement method and device, and relates to the technical field of deep learning industrial process soft measurement. The method comprises the following steps: designing a predictability regularization item according to a probability dynamic model of potential features constructed by a variational cycle network, a time graph constructed by a training data set and a time lag constant, constructing a target-related auto-encoder embedded by a probability potential predictability model, constructing a supervision depth network embedded by the probability potential predictability model, and constructing a target-related auto-encoder embedded by the probability potential predictability model. And obtaining a quality variable prediction value of the industrial production process. According to the method, firstly, a potential feature prediction design regularization item of probability distribution is designed, and network parameters are optimized through a joint training model to ensure overall convergence of the network. Secondly, an improved variational graph recurrent neural network is provided, randomness is brought into an advanced potential space, modeling is carried out on potential predictability of a supervised deep network, and additional graph structure information is added to help to analyze time dependence;

The invention discloses an integrated circuit equipment data optimization monitoring system and method based on big data, and relates to the technical field of integrated circuit manufacturing. The method is used for solving the problems of insufficient multi-physical field monitoring, difficulty in abnormal traceability and lack of closed-loop control in plasma etching. A plasma sheath thickness inversion model and an etching selection ratio model are constructed by collecting radio frequency reflection phase, mass spectrum ion strength and wafer temperature data, and process parameter-plasma state dynamic coupling is established. Interference image distortion features and electron microscope size data are fused, micro-groove geometric parameters are analyzed, morphology instability risk indexes are generated, and nanoscale early warning is achieved. And designing a dual-channel fusion network, embedding a physical constraint attention mechanism, and generating an etching rate optimization instruction. On the basis of incremental learning, model parameters are updated online, a'monitoring-decision-feedback 'closed-loop system is formed, anomaly detection sensitivity and decision reliability are improved, and technical support is provided for intelligence of integrated circuit equipment.

An additive manufacturing real-time process parameter optimization method based on reinforcement learning driving comprises the steps that firstly, a multi-mode online monitoring hardware platform is constructed, and a visible light camera, a thermal imaging camera and an acoustic sensor are integrated to sense the working condition of the additive manufacturing process in real time in an omnibearing mode; secondly, extracting key features, including visible light images, thermal imaging and acoustic signal branches, of modal data on line through a lightweight CNN-Transform hybrid network, and generating a unified low-dimensional state vector through convolution feature extraction and fusion of a Transform encoder; then, a reinforcement learning model is established, a multi-target reward function is designed in combination with the extracted feature data, a process jitter penalty term, an overheating penalty term and the like are covered, and dynamic mapping of parameters and performance is achieved. And finally, an online learning and real-time decision-making system is deployed, the trained strategy network is embedded into manufacturing equipment, self-adaptive adjustment and closed-loop control of technological parameters are achieved, and the stability and product performance of the additive manufacturing process are effectively improved.

The invention discloses a micro-grid group low-carbon optimization operation method based on an improved MAPPO algorithm, and the method comprises the steps: building a micro-grid group optimization operation model with the minimum total cost in a single micro-grid operation period as an objective function; constructing a partially observable Markov decision process of micro-grid optimization operation, and establishing an observation space, an action space and a reward function corresponding to each micro-grid agent; non-linear constraints of carbon treatment cost and collaborative rewards among micro-grids are introduced into a reward function, and a negative sampling experience sharing framework is introduced into an MAPPO algorithm; the LSTM network is embedded into the MAPPO algorithm, and a differential learning rate attenuation strategy is designed to further improve the training speed of the LSTM-MAPPO model; and training the intelligent agent based on an improved MAPPO algorithm to obtain an optimal microgrid group low-carbon optimization operation scheme. The optimization method provided by the invention can effectively reduce the operation cost of each micro-grid and reduce the carbon emission of the system at the same time.

The invention discloses a functional gradient piezoelectric material optimization method based on a physical information network and deep regression. The method comprises the following steps of: 1, constructing a physical information neural network embedded with a piezoelectric constitutive relation; 2, developing a deep regression network based on symbolic regression, wherein the deep regression network is used for converting the piezoelectric phase volume fraction distribution predicted by the neural network into an analyzable mathematical expression; step 3, carrying out joint optimization on the two networks by adopting an alternate iteration strategy; under the condition that parameters of the deep regression network are kept unchanged, a physical information neural network is trained preferentially to meet the precision requirements of a displacement field, a stress field, potential distribution and a piezoelectric phase volume fraction; and then fixing the physical information network, optimizing the expression generation capability of the deep regression network, and finally analyzing the piezoelectric phase volume fraction expression. According to the method, the physical information neural network and the depth symbol regression technology are fused, so that multi-physical field collaborative optimization of component distribution of the functionally graded piezoelectric material is realized.

The invention discloses a central air conditioner control method, system and device, a medium and a program product, and relates to the technical field of data processing and control systems.The method comprises the steps that a central air conditioner system control model is built; the method comprises the following steps: constructing an actor network, and embedding an optimization control model consisting of a global critic network and a local critic network of a GAT (Graphics Attention Transform); and training the optimization control model, and controlling the refrigeration temperature set value and the heating temperature set value of each area according to the current state observation variable of each area by using the trained actor network. A critic network is designed into a global critic network and a local critic network, the problem that regional thermal comfort awards and central air conditioning system total energy consumption awards conflict with each other in the training process is relieved, a graph attention network is fused into the global critic network, coupling information in the system is fully extracted, and the optimization control performance is improved.

The invention discloses a multi-unmanned aerial vehicle cooperative inspection trajectory control method for optimizing medical area coverage and service efficiency. The method comprises the following steps: firstly, constructing a medical multi-unmanned aerial vehicle auxiliary inspection mobile edge computing system model, defining an unmanned aerial vehicle and mobile user set, and establishing a communication model containing A2G and A2A links, an unmanned aerial vehicle mobile model and an energy consumption model; then taking a joint function of a coverage score, a system throughput and an emergency task completion rate as an optimization target, under energy and communication connectivity constraints, proposing an LT-MADDPG algorithm, adopting a CTDE framework, processing a time sequence state by an actor network integrated with LSTM, fusing global information by a commentator network embedded with Transform through multi-head attention, and finally obtaining an emergency task. And modeling a cooperative relationship between the unmanned aerial vehicles and a medical task priority. Experiments show that the method is superior to a traditional algorithm in the aspects of coverage, service fairness and system throughput, and the medical inspection efficiency and reliability are effectively improved.

The invention discloses an artemisinin extraction intelligent tracking and identification method, and relates to the technical field of image data processing. According to the method, an integrated closed-loop system is constructed through cooperation of four core technical means: based on classified filtering of foam stability difference and bimodal adaptive segmentation, dynamic foam is accurately removed, and preliminary interface extraction is optimized; the attention enhancement U-Net network is embedded into a channel attention module and an edge enhancement branch to realize accurate positioning and fluctuation tracking of a layered interface; the attention mechanism dynamically adjusts the bimodal weight, and completes multi-index cooperative tracking in combination with a partitioning strategy and an improved network; a crystal growth model is introduced to correct particle weight, a tracking effect is optimized by adopting layered resampling, and a closed-loop feedback mechanism of a crystal state and process parameters is established. All technical means are deeply coordinated, intelligent and accurate management and control of the whole extraction process are achieved, and reliable technical support is provided for artemisinin extraction production.

The invention provides a cultural creative design assisting method and system based on element fusion. Belongs to the technical field of creative design. The method comprises the steps that design elements are acquired and classified; on the basis of a deep learning algorithm, a semantic network between elements is constructed, the semantic network is converted into a low-dimensional vector space through a network embedding technology, and user preference, emotional tendency and future trend are analyzed in combination with social media data; and generating a fusion scheme, and carrying out continuous iterative optimization on the design. The semantic relation between the design elements is analyzed, so that the internal relation between the design elements can be deeply understood; through a complex network theory, a language network model between design elements is constructed, the relationship between different elements can be visualized, and high efficiency and systematicness of design decision are facilitated.

The invention discloses an angle steel connecting piece shear strength prediction method and system based on a physical information neural network, electronic equipment and a storage medium thereof. The method comprises the following steps: establishing a database containing a plurality of groups of test data based on a numerical simulation result of an angle steel connecting piece finite element model verified by a push-out test; embedding an angle steel connecting piece shear bearing capacity physical constraint condition based on an elastic foundation beam theory into a loss function of the data-driven neural network model DNN; training the neural network embedded with the physical constraint condition by using a database, and adjusting and selecting a physical item weight factor and a model learning rate so as to construct a shear bearing capacity prediction model of a physical information neural network (PINN); predicting the shear strength of the angle steel connecting piece by using the trained physical information neural network (PINN) shear capacity prediction model; according to the prediction method and system, the electronic equipment and the storage medium thereof provided by the invention, the accuracy and reliability of the shear resistance prediction of the angle steel connecting piece can be improved.

The invention discloses a blasting funnel volume solving method based on an improved physical information neural network. The method comprises the following steps: step 1, establishing a mechanical model of blast hole wall blasting load; 2, establishing a blasting physical model in which a columnar cartridge bag is equivalent to a spherical cartridge bag by using a Starfied superposition method; step 3, constructing a blasting funnel volume prediction model based on the wavelet multi-scale synchronous compression transformation enhanced physical information neural network, constructing a solution space of a physical field by using the wavelet multi-scale synchronous compression transformation, and training the physical information neural network in which a blasting funnel physical control equation, an initial condition and a boundary condition are loss functions; and 4, intelligently predicting the volume of the blasting funnel by adopting the trained enhanced physical information neural network. The trained physical information neural network enhances the robustness and generalization ability of the blasting funnel volume prediction model, and provides high-precision theoretical support for blasting design optimization in engineering blasting.

The invention relates to the field of data processing, and relates to a digital twinning-based crane work monitoring method, which comprises the following steps: acquiring real-time operation data of a crane; inputting the real-time operation data of the crane into a physical information neural network embedded with a crane multi-body kinetic equation so as to obtain a high-fidelity state vector representing the real-time working state of the crane; inputting the high-fidelity state vector into a time sequence diagram convolutional network so as to obtain a state sequence track in a future time window; a multi-dimensional risk vector is calculated and generated in combination with a preset failure mode knowledge base; if the mahalanobis distance between the predicted state sequence trajectory and an actual trajectory formed by subsequent real-time data exceeds a preset threshold value, multi-body kinetic equation parameters embedded in the physical information neural network are reconstructed; by the adoption of the method, the crane state monitoring accuracy can be remarkably improved, and prospective risk prediction can be achieved.

The invention belongs to the technical field of multi-source fusion positioning, and particularly relates to a multi-source fusion positioning method based on LSTM-KF, a program, equipment and a storage medium. According to the method, a long short-term memory (LSTM) neural network is embedded into a Kalman filtering framework, nonlinear error compensation of system state prediction and observation updating is achieved, and then an improved Kalman filter with the dynamic noise adaptive capacity is constructed. By loosely coupling GPS satellite positioning, IMU inertial measurement and VO visual odometer multi-source heterogeneous sensing data, the millimeter-level precision of robot pose estimation in a complex dynamic environment and the anti-interference capability of the system are remarkably improved.

The invention belongs to the technical field of ocean prediction, and discloses a chlorophyll a prediction method and system based on heterogeneity perception and equation embedding, and the method comprises the steps: 1, obtaining the related input data of chlorophyll a concentration prediction: 2, carrying out the spatial heterogeneity measurement: carrying out the sub-region division of an original input region, and obtaining the spatial heterogeneity level of the sub-region; step 3, heterogeneity perception space-time evolution modeling: constructing a multilayer adaptive heterogeneity perception network, and outputting heterogeneity perception space-time evolution characteristics; 4, designing a physical constraint guide prediction network embedded by the fluid motion equation, and calculating a space-time evolution characteristic sequence of future chlorophyll a; and step 5, carrying out remodeling and aggregation decoding operation on the space-time evolution feature sequence, and merging the sub-region prediction sequences in corresponding time periods to obtain a final prediction result. By means of the method, space-time evolution modeling of space heterogeneity perception and equation physical knowledge embedding is achieved, and the accuracy and reliability of chlorophyll a concentration prediction are improved.

The invention discloses a medicine sales management method and system based on information monitoring, and relates to the technical field of medicine sales management, and the method comprises the steps: collecting medicine sales time sequence data, carrying out the cleaning and structural processing, and generating standardized sales data; based on the standardized sales data, constructing a sales behavior time sequence causal network, embedding virtual sales intervention nodes, and generating a time sequence causal graph; the intervention strategy is coded to the virtual intervention node in the time sequence causal atlas, the graph neural network is utilized to simulate time sequence propagation of the intervention strategy, and a sales prediction result is obtained; and combining the sales prediction result with the real-time inventory, constructing a multi-level distributed sales dependency network, and forming the sales dependency network. According to the method, the preset intervention strategy is embedded into the causal structure in the node form, so that dynamic modeling of the multivariable causal relationship in the sales behavior is realized, and the prediction result not only reflects the historical trend, but also can respond to the possible influence of different intervention strategies.

The invention provides a bidirectional tension control withdrawal and straightening method for loose edge defects of a cold-rolled 2B plate, which comprises the following steps of: S1, acquiring transverse strain distribution data of the plate in real time through a distributed optical fiber strain sensor array at the sampling frequency of more than or equal to 10kHz; s2, inputting the strain data and the process parameters into a physical-data hybrid drive prediction model, and generating an optimized combination matrix of longitudinal tension sigma long and transverse tension sigma trans by the model through a neural network constrained by an embedded Hollomon elastic-plastic equation; s3, the magnetic field intensity of the magnetorheological fluid clamping roller is dynamically adjusted according to the optimization matrix, so that the transverse tension is continuously adjustable within the range of 0.1-50 N / mm < 2 >; s4, a multi-target reinforcement learning algorithm is adopted, model parameters are updated on line based on real-time defect detection data, and the loose edge elimination rate and the energy consumption index are balanced. The method has the advantages that longitudinal and transverse tension coordinated regulation and control are realized, the loose edge defect elimination precision is improved, and the real-time adaptability of the model is enhanced.

The invention discloses a large-scale sparse bipartite graph-oriented adaptive anchor point diagram attention network embedding method and system, and the method comprises the steps: firstly, building a dynamic anchor point number generation mechanism through analyzing the scale, sparseness, user node number, average degree and other graph characteristics of a graph, and achieving the adaptive adjustment of the number of anchor points; secondly, a full-connection attention mechanism is designed, efficient information transmission between a user node and an anchor node is achieved through bidirectional attention operation of FCAsend and FCArecv, and the problem of information loss of a sparse region is solved; and finally, a bipartite graph attention network with entropy regularization is introduced, a uniform weight matrix and a sparse attention mechanism are adopted to process a dense region, and node features are effectively prevented from being excessively smooth. According to the method, through innovative combination of the self-adaptive anchor points and the double attention mechanism, the embedding quality of the large-scale sparse bipartite graph is remarkably improved, and the performance superior to that of an existing method is achieved in actual tasks such as a recommendation system and link prediction.

The invention belongs to the technical field of medical data processing, and discloses a preoperative multi-complication risk prediction method and system based on structured clinical data, and the method comprises the steps: inputting the causal association between risk factors and complication nodes into the edge of a knowledge graph, calculating the statistical correlation between all complications, and supplementing the statistical correlation into the knowledge graph, and performing network embedding training on the knowledge graph to form a first-stage model, performing preliminary risk assessment on complications, modeling the knowledge graph in a graph neural network mode, and performing joint training with the first-stage model to form a second-stage model to output a final complication probability. According to the method, the interpretability and cross-domain consistency of the model can be improved through deep fusion of the medical knowledge graph and the multi-relational graph convolutional network, stability and calibration performance are still kept in a specialist with scarce sample size, and the problem that a traditional black box model cannot be interpreted is avoided; and the practical application value can be evaluated conveniently.

The invention relates to the technical field of artificial intelligence, in particular to a reinforcement learning green large model training method based on low-carbon contribution feedback, and the method comprises the steps: constructing a knowledge vector library coded by a neural network embedded model; constructing a low-carbon contribution evaluation function by using a first neural network large language model, retrieval enhancement and expert rules, and generating a reward signal; and adopting a reinforcement learning algorithm to optimize a second neural network large language model serving as a learning agent according to the reward signal. According to the method, the dynamically quantified evaluation function is constructed as a reward signal of reinforcement learning, so that the problems that green guidance of a general model is unknown and contribution is difficult to measure are solved, and accurate molding of the specific green behavior tendency of the large model of the neural network is realized.

The method and system for data pruning use the novel heuristic of weighting the selection of images by an internal diversity metric, such as the radius of the cluster, allowing more images to be sampled from clusters that are more internally diverse. This heuristic is added to improve the overall diversity of the selected images and to prevent the over-representation of similar images. By sampling more images from clusters that are more internally diverse, the approach is able to better represent the overall distribution of the data, improving the quality of the resulting pruned dataset.