67results about How to "Efficient fusion" patented technology

The invention relates to the technical field of flood disaster simulation, in particular to a watershed flood dynamic simulation system based on multi-source data fusion, which comprises a multi-source data preprocessing module, a multi-source data fusion module, a digital twin watershed modeling module, a flood dynamic simulation module, a flood risk assessment module and a visualization module. According to the scheme, a space-time cross-modal interaction fusion method is put forward to fuse multi-source data, a cross-modal interaction mechanism is introduced to capture dependency in a modal, high-dimensional interaction tensors are efficiently fused by means of low-rank decomposition, complementarity of space topology, time dynamic and multi-source heterogeneous data is considered, and multi-source data fusion efficiency is improved. The precision and robustness of hydrological process modeling are effectively improved; by constructing a distributed hydrological-hydrodynamic coupling model, whole-process dynamic simulation from runoff production to river channel-flood area flood routing is realized.

The invention relates to the technical field of sewage treatment, and discloses a sewage biochemical treatment process abnormity intelligent early warning method and system based on image recognition, and the method comprises the steps: periodically collecting a sludge image and water quality sensing data; extracting floc form, bubble distribution and color texture features of the image, and performing space-time alignment fusion with the sensing data to construct a multi-source fusion feature matrix; inputting the matrix into a pre-trained digital twin evaluation model, calculating a dynamic deviation degree between a current state and a normal working condition, and identifying an abnormal mode; and generating a graded early warning signal according to the deviation degree and the abnormal mode and pushing the graded early warning signal. The system comprises an online image acquisition unit and a water quality sensor array which are used for data acquisition, and a feature processing module used for feature processing and fusion. According to the scheme, early-stage, accurate and intelligent early warning of abnormality in the sewage biochemical treatment process is realized, the false report and missing report rate is reduced, and the accuracy and the reliability of the system are improved. And the operation stability and the regulation and control efficiency of the system are improved.

The invention relates to the technical field of artificial intelligence, and discloses a credible RAG-based hybrid expert routing method and system, and the method comprises the steps: obtaining metadata of a document associated with an input text to generate a document-level credible RAG vector; obtaining a hidden representation of each token according to the input text and the text content of the associated document, obtaining a credible vector corresponding to each token based on the document-level credible RAG vector, and splicing the hidden representation of each token and the corresponding credible vector to generate an enhanced input vector of each token; taking the enhanced input vector as an input of a gating network, calculating an expert score and selecting at least one expert network for weighted fusion to generate a final output; through the method, the whole routing decision process of organically fusing the credible hierarchical information of the external knowledge into the hybrid expert model is displayed, the spanning from semantic driving to semantic-credible cooperative driving is realized, and the routing accuracy is remarkably improved.

The application relates to the field of sanitation operation scheduling, in particular to a sanitation operation scheduling method based on the Internet of Things, which comprises the following steps: determining an iterative stop condition of a load CEEMDAN model based on time local weighted variance of load data, so as to generate vehicle load characteristics; generating vehicle oil consumption characteristics through an oil consumption CEEMDAN model based on statistical values of load data residuals and oil consumption data residuals, so as to calculate oil consumption channel residual signals; generating comprehensive vehicle carrying characteristics through a feature extraction model based on a multilayer perception machine, a gating mechanism and a cross-modal attention mechanism; and generating corrected vehicle garbage load through a long short-term memory network and a fully connected layer architecture. The application realizes efficient noise reduction and accurate decomposition of non-stationary time series data of oil consumption and load, realizes accurate processing of Internet of Things data of garbage transfer vehicles, improves the operation efficiency of the garbage transfer vehicles and reduces the operation cost.

The invention belongs to the technical field of new energy vehicles, and particularly relates to a hybrid vehicle approximate optimal energy management strategy integrating prediction and control, and the strategy specifically comprises the following steps: (1) constructing a global prediction domain, predicting a global speed curve through employing a bidirectional long-short-term memory network based on traffic data, and obtaining a global speed curve; obtaining a global approximate optimal reference state-of-charge sequence through a dynamic programming algorithm; (2) constructing a mesoscopic coordination domain, performing mesoscale speed prediction by adopting a method of fusing an encoder and a bidirectional long-short-term memory network, and solving a multi-scale information evaluation factor online and performing dynamic correction in combination with a grey wolf optimization algorithm; and (3) constructing a real-time control domain, introducing the evaluation factor to construct a multi-target optimal real-time control framework, and realizing instantaneous optimal mode decision and power distribution. According to the method, the economy, the smoothness and the NVH performance are comprehensively considered, global reference and real-time control are effectively decoupled, and the adaptability of the vehicle to a complex dynamic driving scene and the fuel economy of the whole vehicle are remarkably improved.

The utility model discloses a roof solar water heating system which comprises a flat-plate solar collector, an air source heat pump unit, a heat pump heating circulating water pump, a pipeline system and a heat preservation water tank. The flat-plate solar heat collectors are fixed on a roof foundation through a support, the flat-plate solar heat collectors are arranged in multiple rows and are connected in series through heat collector water pipes, and the two ends of each flat-plate solar heat collector are connected to the heat preservation water tank. The air source heat pump unit and the heat pump heating circulating water pump are connected to the heat preservation water tank through a circulating water pipe. The cold water supply pipe supplements cold water to the heat preservation water tank, the hot water supply pipe conveys hot water in the heat preservation water tank to all water consumption ports, and the hot water return pipe recovers unused hot water to the heat preservation water tank. The solar water heater can solve the problems that the water temperature of the solar water heater rises slowly and is unstable, residual hot water cannot be effectively utilized, and energy is wasted due to electric heating in cloudy and rainy days.

The invention discloses an automobile demand prediction method and system based on large model time sequence cross-modal alignment, and belongs to the technical field of automobile demand prediction. The automobile demand prediction method based on large model time sequence cross-modal alignment comprises the following steps: acquiring time sequence data and prompt data related to a hydrogen fuel cell automobile; inputting the time sequence data into the time sequence coding branch in a reverse embedding manner to generate a time sequence embedding representation; inputting the prompt data into an LLM enhanced coding branch for coding, and generating a prompt embedding representation; and calculating a channel-level similarity matrix between the time sequence embedding representation and the prompt embedding representation, aggregating to the time sequence embedding representation based on the channel-level similarity matrix, generating an aligned cross-modal feature, and generating a demand prediction result of the hydrogen fuel cell vehicle based on a decoder. According to the method, multi-source heterogeneous information is effectively fused, the precision and robustness of demand prediction of the hydrogen fuel cell vehicle are improved, and the capability of responding to emergencies caused by external factors is enhanced.

The invention discloses a blueberry tree branch segmentation method and system, and the method comprises the steps: carrying out the multi-scale feature extraction of a to-be-segmented blueberry tree image, and obtaining the feature representation containing low-layer details, middle-layer semantics and high-layer context; information complementation and semantic alignment of deep and shallow layer features are realized through a bidirectional cross-level information interaction module, and fusion features are optimized by using a space and channel attention mechanism to obtain enhanced features; and performing up-sampling and dynamic reconstruction on the enhanced features by adopting a frequency domain dynamic convolution module, refining branch edges and texture details, and outputting a high-resolution pixel-level segmentation result. The method effectively solves the problems of fuzzy boundary, fracture adhesion and background interference in the segmentation of the slender branches and trunks of the blueberries, and has good precision and generalization ability.

The invention discloses a geotechnical engineering geological investigation system and method, relates to the technical field of geological investigation, and solves the problems of difficulty in multi-source data fusion, low geological anomaly intelligent identification precision and weak regional comprehensive evaluation and visual expression ability in the existing investigation method. According to the method, the unified comprehensive parameter vector is constructed, scattered drilling data, in-situ test parameters and real-time sensor monitoring data are subjected to standardized integration and collaborative analysis, the limitation that various types of data are independently stored in traditional exploration, and a data island is formed is broken through, efficient fusion of multi-source heterogeneous information is achieved, and the method has the advantages of being simple in structure and convenient to use. The problem of data fragmentation is solved, and deep cognition and judgment of complex geological conditions can be more comprehensively carried out.

The application relates to the technical field of file processing, and discloses a file processing method suitable for structured and unstructured data, which comprises the following steps: S1, identifying the type of an input file and classifying the input file, including structured data, semi-structured data and unstructured data; S2, pre-processing and feature extraction are performed on different types of data; S3, the correlation between different data modalities is quantified based on multi-modal mutual information optimization; S4, the interaction consistency between data modalities is verified, and the fusion weight is dynamically adjusted. Through classification and feature extraction of structured data, semi-structured data and unstructured data, efficient fusion of multi-modal data is realized by using mutual information optimization and a dynamic weighting strategy. Compared with a traditional method, the application can significantly improve the precision and robustness of data processing, and provides an efficient solution for multi-modal data fusion in a complex scene.

The invention discloses an event camera video reconstruction method and system based on active aperture modulation, and the method comprises the steps: introducing an aperture modulation strategy for the first time, and reconstructing an initial frame with good quality by periodically adjusting the opening and closing of an aperture and actively triggering a dense global event signal; the problem of low background reconstruction quality caused by sparse events in a static region in the prior art is solved, and good prior information is provided for subsequent dynamic scene reconstruction. By constructing a forward-reverse bidirectional network, rich intensity information is provided for a static scene, and the problems of background disappearance and error accumulation in long-time operation in a traditional method are effectively solved. And meanwhile, high-time-resolution capture of a dynamic region is kept, and high-fidelity and high-dynamic-range video reconstruction is realized.

The invention discloses a multi-modal heterogeneous data increment embedding fusion method, equipment and a medium, mainly relates to the technical field of multi-modal heterogeneous data, and aims to solve the problems of difficult neighborhood structure maintenance, insufficient cross-modal information fusion, single and fixed index structure and the like in the existing scheme. Comprising the following steps: converting modal data in a multi-modal heterogeneous data set into hidden space representation, and fusing the hidden space representation by utilizing a cross-modal attention mechanism to obtain historical embedding representation of a preset d dimension; when the newly-added multi-mode heterogeneous data set appears, calculating a newly-added embedding representation of a preset d dimension; according to the newly-added multi-modal heterogeneous data set, the newly-added embedding representation, the historical multi-modal heterogeneous data set and the historical embedding representation, constructing an adjacency relation graph of the newly-added multi-modal heterogeneous data set; and inputting the adjacency relation graph into a pre-trained incremental graph neural network to obtain an embedded representation fusing the adjacency relation.

The invention discloses a multi-channel enhancement method based on an all-neural beam former, and the method comprises the following steps: carrying out the short-time Fourier transform of an input multi-channel noisy voice signal, and decomposing a time domain signal into the amplitude spectrum and phase of a frequency domain; taking the amplitude spectrum of the multichannel noisy voice signal as input, and mapping the amplitude spectrum into a pure voice amplitude spectrum through an all-neural beam former model based on position sensing cyclic convolution; combining the amplitude spectrum of the pure voice with an original phase, recovering a frequency domain signal to a time domain through inverse short-time Fourier transform and an overlap-add method, and finally outputting an enhanced voice signal; the design does not need complex explicit spatial feature design, breaks through the limitation of a traditional method, solves the defects of existing deep learning, can adapt to complex noise and strong reverberation scenes, improves the target voice extraction precision, and gives consideration to robustness and calculation efficiency.

This invention discloses an infrared weak target recognition method based on skip connections enhanced by Unet, belonging to the fields of computer vision and artificial intelligence. The method includes: extracting features from infrared images using the Unet encoder; adjusting the channel and spatial dimensions of the feature map using an adaptive convolutional embedding mechanism; inputting the feature map into a two-dimensional Mamba backbone network; fusing the original features with the processed features through residual connections; and performing batch normalization and activation function optimization on the fused feature map; reconnecting the feature map to the Unet decoder using the adaptive convolutional embedding mechanism; employing DiceLoss and Adam optimizers for collaborative training; and extracting the target region based on a locking mechanism with a pixel value of 255 in the mask. This invention enhances the feature representation capability of Unet skip connections by introducing an MCA module, enabling better mining of multi-level feature information. Through multi-level feature fusion, it improves the accuracy and robustness of infrared weak target detection.

This application relates to the field of agricultural mechanization and intelligent control technology, and discloses a method for collaborative operation of multiple agricultural machines based on a vertical domain large model. The method includes: a central controller receiving multi-source sensing data uploaded by onboard controllers, fusing and constructing a vertical domain knowledge graph and generating a load distribution prediction model, and issuing an initial collaborative path accordingly; each onboard controller collecting attitude data to calculate deviations, correcting parameters and uploading them if they exceed a threshold, and generating an optimized collaborative path; the central controller receiving torque values ​​to calculate the resultant torque distribution, determining the lever arm length distribution, analyzing the center of gravity offset in conjunction with the weight of the agricultural machinery, and issuing commands to adjust the relative posture of the agricultural machinery under high loads; each onboard controller monitoring hydraulic pressure, fine-tuning actuator resistance and uploading stability parameters if it exceeds the range; the central controller evaluating durability indicators based on historical data, and updating the collaborative path by fusing real-time data when the indicators are too low, obtaining the final solution. This application improves the accuracy of collaborative operation of multiple agricultural machines in complex terrain.

The application relates to the technical field of image processing, and particularly discloses an image registration and fusion method and device, a storage medium and computer equipment. The method comprises the following steps: acquiring PET scanning data and CT scanning data; determining a PET field of view angle corresponding to the PET scanning data and a CT field of view angle corresponding to the CT scanning data; performing three-dimensional image reconstruction on the CT scanning data according to the CT field of view angle to obtain a target CT image, and performing three-dimensional image reconstruction on the PET scanning data according to the target CT image and the PET field of view angle to obtain a target PET image; and performing registration and fusion processing on the target CT image and the target PET image to obtain a target fusion image. The application can more accurately and efficiently integrate PET images and CT images, and can effectively improve the registration effect and overall quality of the PET / CT fusion image.

The invention provides a hybrid deep learning automobile sales prediction method fused with multi-source features and application, and belongs to the technical field of hybrid deep learning fused with multi-source features. The method comprises the following steps: 1, inputting subjective evaluation data into a semi-supervised feature generation layer, and extracting features in the subjective evaluation data through a method based on semi-supervised learning feature clustering; 2, inputting static non-time-series data and the subjective evaluation type data and the time series data in the step 1 into a parallel feature extraction layer, and performing feature extraction; 3, performing deep fusion on the output in the step 2, and inputting the merged feature vector into a fusion optimization layer; step 4; and using an Adam optimization algorithm to train the models formed in the steps 1-3. According to the method, combined feature extraction and fusion of non-time-sequence feature mixed data of time sequence features, static attributes and subjective evaluation classes are realized, so that the automobile sales volume prediction precision is optimized.

The application provides a human action recognition and quality evaluation method and system based on multi-modal data fusion, the application collects RGB visual data, 3D skeleton data, inertial IMU data and electromyography sEMG data of human action, and carries out corresponding pretreatment; through branch parallel feature extraction, RGB visual feature vectors, 3D skeleton feature vectors, IMU inertial feature vectors and sEMG electromyography feature vectors are obtained respectively; based on the modal reliability values of each branch, multi-modal fusion feature vectors are obtained by dynamically distributing and fusing the weights of the cross-modal attention mechanism to the multi-path features; the action recognition result is obtained by using a lightweight action recognition network; and the action recognition result is quality evaluated through the comprehensive action quality score. The application can not only accurately recognize the action category, but also quantize the action quality from three dimensions of form, dynamics and force, thereby improving the robustness and accuracy of action recognition in a complex scene.

This invention discloses an artificial intelligence-based method and system for controlling the amount of liquid fertilizer drip irrigation in agricultural planting. The method involves collecting multi-source datasets through a multimodal sensing network; fusing the multi-source datasets using a STAFN spatiotemporal attention fusion network to obtain a fused dataset; establishing a dual-channel demand prediction model, inputting the fused dataset into the model, and outputting prediction results; calculating the fertilizer concentrate mixing ratio based on the prediction results; adjusting the fertilizer absorption rate of the Venturi fertilizer applicator according to the fertilizer concentrate mixing ratio using a fuzzy PID control algorithm; dividing the planting area into several intelligent irrigation zones; and implementing differentiated fertilization for each zone using a pulse drip irrigation strategy. This method adapts to the differences in soil and crops in different regions, improving drip irrigation efficiency and accuracy.

This invention provides a method and system for waterway monitoring based on real-time attitude compensation. The method acquires sensor data collected by a phased-array radar, an optical camera, and an inertial measurement unit (IMU) under a unified time reference. The phased-array radar, optical camera, and IMU are pre-fixed on a buoy, and their sensor coordinate systems are unified to the buoy's reference coordinate system based on calibration parameters. Based on the real-time angular velocity and acceleration data collected by the IMU, the real-time attitude change information of the buoy is estimated. Motion compensation is performed on the sensor data collected by the phased-array radar and optical camera based on the real-time attitude change information. The motion-compensated sensor data from the phased-array radar and optical camera is then spatiotemporally registered and fused to generate fused sensing data for maritime target identification and tracking. This method improves the accuracy and reliability of waterway monitoring.

According to the RGB-D saliency target detection method based on mixed feature cross fusion, through a mixed feature cross fusion module, a model can fully utilize complementary information between an RGB image and a depth image, and efficient fusion of features is achieved. The module enhances the interaction between RGB and depth features through a self-attention mechanism, so that the connection between different modal features is better modeled and captured, and the problem of feature distortion under the condition of poor depth image quality is solved. Besides, the invention provides a specific training strategy and an optimization scheme, and the model can rapidly converge on a large-scale data set by adopting the combination of an MAE pre-training model and a ViT model, and obtains excellent performance in a saliency target detection task.

The invention relates to the field of soil remediation, and discloses a soil turning remediation device for saline-alkali soil planting soil improvement, which comprises a head end bearing mechanism positioned on a main frame and used for forming a height-adjustable soil turning structure bearing platform; the staggered pressure-bearing mechanism is located on the main rack and matched with the head end transverse rack and the guiding oppositely-arranged rack to form a driving structure which is sequentially inserted in a staggered mode. And the tunneling and turning mechanism is located on the staggered pressure-bearing mechanism and is matched with the jet air knife pipe, the butt joint guide table and the embedded channel to form an insertion type soil turning tool bit structure. Through the staggered state of eccentric linkage convex rods of the staggered pressure-bearing mechanism, rotating power is converted into asynchronous staggered vertical downward pressure, so that multiple sets of tunneling tool bits cut into hard soil in batches, resistance generated when the tool bits enter the soil at the same time is effectively avoided, the whole machine runs more smoothly, meanwhile, an embedded conical hopper at the bottom is matched with a tunneling hopper, and the tunneling efficiency is improved. The mode of puncturing, splitting and then extracting is more suitable for the hardened layer of the saline-alkali soil.

This invention relates to a method for NAION prediction and acute phase differentiation based on deformable convolution and multi-site OCTA, belonging to the field of ophthalmic disease diagnosis and medical image processing technology. This invention acquires OCTA images of ocular samples and extracts data from key sites. Using the OCTA deformable convolution feature extraction module as the core feature extraction backbone network, its optimal performance is verified through single-site experiments, and a multi-site joint diagnostic framework is constructed. Radiomics features are introduced, and deep fusion of deformable convolution features and radiomics features is achieved through a bidirectional cyclic feature interaction module. Finally, a hybrid expert module is added to the three-site joint model to decouple NAION prediction from the acute phase differentiation task of NAION / ON, improving diagnostic accuracy. This invention effectively solves the problems of difficult accurate differentiation between the acute phases of NAION and ON and insufficient early prediction of NAION, providing reliable technical support for ophthalmic clinical diagnosis.

The application discloses a permafrost region railway track roadbed reinforcing and temperature controlling structure, method and storage medium, and the roadbed reinforcing and temperature controlling structure comprises, from top to bottom, a steel rail, a sleeper, a subgrade surface layer, a subgrade bottom layer and an embankment; a transverse ventilation pipe is arranged below the embankment; a plurality of gravel piles are arranged at the bottom of the transverse ventilation pipe; the pile body of the gravel pile is a porous structure formed by gravel filling and is communicated with the transverse ventilation pipe; the gravel pile at the soft soil foundation layer is wrapped by an economical geogrid; the gravel pile at the future maximum soft layer and the safety reserve layer is wrapped by a high-strength geogrid; and the gravel pile at the permafrost layer is not wrapped. The application adopts the composite wrapped gravel pile reinforcing structure to reinforce the foundation based on the degradation position of the permafrost, simultaneously reduces the deep soil temperature through the temperature controlling structure, maintains the stability of the permafrost layer, and improves the stability of the reinforcement.

The invention discloses a visual perception large model distributed training method based on a general calculation fusion network, and relates to the technical field of information and communication, and the method comprises the steps: an edge server initializes and broadcasts a pre-trained general large model; receiving general large model parameters to form a local general model; the edge device continuously receives visual perception data samples, and online learning is carried out through incremental training and a sliding window mechanism; performing clustering analysis on local sample data, and classifying samples into personalized model processing samples and global model processing samples; respectively inputting the classification samples into the personalized model and the global model, and weighting and aggregating output features through a gating network; the edge device uploads the updated global model parameters to a server for aggregation, and a new global model is generated; and the edge equipment periodically receives the latest global model parameters broadcasted by the server, and iterative training is carried out until the model converges. According to the invention, high-efficiency, high-precision and high-reliability distributed model training and visual perception performance are realized.

This invention belongs to the field of seismic data processing and relates to a method and system for consistent processing and adaptive fusion of seismic attributes in multiple contiguous seismic work areas. The method includes: S1: dividing the spatial coordinates of the original seismic attribute data into a grid and establishing a gridded data index; then, based on the gridded data index, executing a localized search strategy to extract the neighboring data point set of the seismic data points to be processed; S2: eliminating systematic value range differences in the original seismic attribute data based on the neighboring data point set and stitching them together to obtain consistent fused seismic attribute data; S3: using a hierarchical interpolation strategy to fill in blank areas to obtain continuous seismic attribute interpolation data; and S4: performing global optimization based on Gaussian pyramid decomposition and an energy minimization model to obtain the final fused seismic attribute data. This method achieves better work area stitching and interpolation, efficiently avoiding the attribute non-fusion effect of traditional work area stitching methods, thereby realizing the goal of consistent seismic attribute stitching across multiple work areas.

This invention relates to the field of flood disaster simulation technology, and in particular to a watershed flood dynamic simulation system based on multi-source data fusion. The system includes a multi-source data preprocessing module, a multi-source data fusion module, a digital twin watershed modeling module, a flood dynamic simulation module, a flood risk assessment module, and a visualization module. This solution proposes a spatiotemporal cross-modal interactive fusion method to fuse multi-source data. This method introduces a cross-modal interaction mechanism to capture intramodal dependencies and uses low-rank decomposition to efficiently fuse high-dimensional interaction tensors, taking into account the complementarity of spatial topology, temporal dynamics, and multi-source heterogeneous data, effectively improving the accuracy and robustness of hydrological process modeling. By constructing a distributed hydrological-hydrodynamic coupled model, it achieves dynamic simulation of the entire process from runoff generation to flood evolution in the river channel and floodplain.

The invention discloses a multi-laser fusion point cloud three-dimensional target detection method based on two-dimensional vision enhancement, and the method comprises the steps: carrying out the three-dimensional universal obstacle recognition of point clouds scanned by a single-line mechanical laser radar and a solid-state laser radar based on a local grid map and a light tracing method; calculating the terrain clearance of the three-dimensional universal obstacle identified by the solid-state laser radar; splicing the images shot by the plurality of cameras to form a two-dimensional panoramic image; performing target detection and semantic segmentation on the two-dimensional panoramic image by using a two-dimensional deep learning detection model and a segmentation model; and finally, projecting an identification result of the three-dimensional universal obstacle to the two-dimensional panoramic image, performing association matching with the two-dimensional detection frame, and performing semantic enhancement processing on the successfully matched three-dimensional universal obstacle to obtain independent three-dimensional targets corresponding to different semantic entities. The method can realize omnibearing perception and accurate measurement of the height of the obstacle, and is low in calculation complexity and strong in dynamic environment adaptability.

The utility model discloses a centralized vacuum feeding system for PVC board production relates to vacuum feeding technical field, including setting on the vacuum department, stirring part and feed part of bottom plate. The vacuum department contains vacuum pipe no. 1, vacuum blanking shell etc. produce negative pressure conveying material through vacuum driving shaft and driven shaft cooperation, and the bottom of blanking shell has the discharge roller and the discharge pipe, and the top is equipped with the filter screen and the vacuum pipe no. 2 connection vacuum housing. The stirring part is detachably arranged on the vacuum pipe no. 1, contains stirring housing, stirring shaft etc. and has multiple feed ports in the top side wall, and after motor drive stirring shaft mixes material, enters the vacuum pipe no. 1 through the discharge pipe. The feed part contains feed housing, and the bottom is through with the vacuum pipe no. 1, and the top has the reversible top cover with sealing washer, and the bottom is equipped with motor -driven distribution roller. Can promote raw material mixing uniformity, realize centralized feeding, strengthen equipment adaptability and operation flexibility, improve production efficiency and board quality.

The invention discloses a fine-grained multi-modal knowledge enhanced image generation method and device, and the method comprises the steps: obtaining the text features of an optimized and expanded cue word through a pre-trained text encoder, obtaining the visual features of a corresponding subject image through an image encoder, and carrying out the recognition of the subject image according to the obtained text features and visual features. Splicing is carried out in the feature dimension direction; using a trainable MLP to carry out dimension reduction on the splicing features, converting the features after dimension reduction into embedding conditions, and inputting a reference image into a reference branch to optimize attention parameters; performing multi-dimensional quality evaluation on the generated image, comprehensively generating relevancy between the image and a text cue word and semantic consistency between the image and a domain knowledge graph to output a final quality evaluation score, comparing the final quality evaluation score with a preset threshold value, and judging whether to trigger an introspection agent mechanism by referring to the current feedback frequency; and sending the image hidden space representation obtained from the output of the trunk DiT branch into an image decoder, and restoring to obtain a final generated image. The device comprises a processor and a memory.