1502 results about "Feature coding" patented technology

The invention discloses a network threat multi-modal detection method based on a large model, and belongs to the technical field of network security, and the method comprises the steps: collecting three types of heterogeneous data of NetFlow flow of a network layer, a system call chain sequence of a host layer and a protocol load of an application layer, and carrying out the desensitization processing and feature coding to generate a unified tensor format; the method comprises the following steps: through network security threat intelligence and MITRE ATTamp; performing supervision fine tuning on the large model by using a CK attack chain sample, and constructing a network threat identification special model; cross-device behavior characteristics are extracted through a model self-attention mechanism, and a dynamic behavior map is constructed; and finally, comprehensively evaluating an attack mode matching degree, a node vulnerability mean value, historical alarm association and an attack path risk by adopting a weighted fusion algorithm, and triggering a high-confidence alarm when a comprehensive score exceeds 0.8. According to the method, through multi-modal data fusion and dynamic graph analysis, the detection precision and response efficiency of the complex attack chain are improved.

The invention relates to the technical field of remote sensing monitoring, in particular to a remote sensing image ocean green tide monitoring method and system fusing physical prior and spatio-temporal evolution. The method comprises the following steps: acquiring a multi-modal remote sensing monitoring image; performing multi-modal feature extraction on the acquired image, wherein the multi-modal feature extraction comprises spectral reflectivity feature extraction, ocean dynamics feature extraction and feature alignment and unified representation; establishing a physical prior of a green tide characteristic wave band by using an ocean optical radiation transmission model; constructing a dynamic space-time diagram based on the extracted multi-modal features to obtain a node global feature vector and a dynamic adjacency matrix; carrying out adaptive graph convolution feature coding based on physical prior and a dynamic space-time diagram; through fusion of multi-spectral images of multiple platforms such as satellites and unmanned aerial vehicles and ocean dynamic data and combination of atmospheric correction and wave band resampling, consistency processing and high-precision extraction of multi-source features are realized, and comprehensiveness and reliability of green tide feature recognition are remarkably improved.

The invention relates to the technical field of multi-modal information processing, and provides a deep semantic collaborative fusion method for heterogeneous multi-modal data. The invention provides a dynamic adaptive fusion framework aiming at the problems that a modal interaction mechanism is rigid and semantic modeling is shallow in the prior art. The method comprises the following steps: carrying out feature coding and alignment on text, audio and video modal data to generate unified-dimension single-modal representation; dynamic interaction is realized through an enhanced multi-head gating fusion module, and double-path features are generated; and carrying out cross-modal depth modeling on the basis of a stacked Transform encoder, and outputting final fusion semantics. Wherein the multi-head attention path calculates cross-modal mapping by taking a text as a query vector and taking an audio / video as a key value vector; the gating path generates a dynamic weight through cosine similarity and a learnable temperature parameter; and the dual-path adaptive fusion adopts a balance factor alpha weighted combination. According to the method, the multi-modal data fusion precision and the system robustness are improved, and the method is suitable for government affair service, man-machine interaction and other scenes.

The invention relates to a multi-modal large model dynamic compression and reasoning optimization method based on a MoE architecture. The method comprises the following steps: establishing an edge computing system conforming to medical equipment specifications, constructing a medical image analysis network based on an improved hybrid expert MoE architecture, and adopting a three-layer cascade structure of a feature coding layer, a dynamic routing layer and an expert execution layer; executing expert module dynamic loading and video memory optimization; executing knowledge graph compensation and domain knowledge injection; executing hardware instruction level optimization and calculation acceleration; executing multi-expert feature fusion and decision weighting; performing diagnosis result generation and confidence evaluation; performing real-time data return and model iterative optimization; executing multi-device cooperation and load balancing; executing system security monitoring and exception handling; and generating a structured diagnostic report. The problem that the precision loss of a multi-modal large model is difficult to meet actual requirements is solved, and medical feature adaptive dynamic compression, medical hardware collaborative energy efficiency optimization and cross-modal compensation of medical knowledge enhancement are realized.

The invention provides a cross-modal remote sensing image-text retrieval method based on multistage semantic collaborative matching, which comprises the following steps of: extracting a region of interest by using a semantic segmentation algorithm through an image preprocessing module, segmenting a remote sensing image into a plurality of regions and generating image blocks; respectively extracting global features, regional features and pixel-level features of the image, and carrying out fine-grained coding on key regions such as fine-grained ground feature edges and the like; the text multi-level coding module is used for carrying out three-level feature coding of documents, sentences and words on texts based on a pre-training language model to ensure multi-level understanding of the texts; in the multi-level matching and fusion module, the similarity between the remote sensing image and the text description is calculated through a cross attention mechanism, weighted fusion is carried out on features of all levels, and finally a retrieval score is output. The method not only improves the accuracy and robustness of image-text retrieval, but also can be widely applied to the fields of remote sensing monitoring, environment change recognition, geographic information systems and the like.

The invention belongs to the technical field of marine environment prediction, and discloses a ship navigation sea wave dynamic space-time prediction method and system based on deep learning. The method comprises the following steps: carrying out space-time alignment, missing value repair and standardization processing on acquired ship AIS data and an ERA5 reanalysis data set, and generating node feature vectors containing latitudes and longitudes, timestamps, wind speeds and significant wave heights; through node feature coding, centrality coding, space coding and time coding, ship trajectory node importance and time-space interaction relation are quantified. Constructing a SeaGraph model, and outputting an effective wave height prediction value of a target waypoint; and performing verification. According to the method, the space-time constraint of a traditional static modeling framework is broken through, the advantages of a self-attention mechanism and a graph network are integrated, the predictive modeling capability of dynamic evolution of a wave field in front of a ship navigation track is enhanced, and high-precision sea wave forecasting support is provided for intelligent ship navigation under complex sea conditions.

The invention relates to the technical field of recommendation, and provides a recommendation method and system based on multi-modal dynamic gating and hybrid coding. The method comprises the steps of obtaining user interaction behavior information and article attribute information, generating collaborative embedding and semantic embedding of articles through an ID-PCA encoder and an LLM encoder respectively, achieving alignment of collaborative signals and semantic signals through bidirectional multi-granularity cross attention, and achieving self-adaptive fusion of multi-expert features through a dynamic weight distribution mechanism; through heterogeneous feature coding, cross-modal dynamic alignment and multi-scale interest modeling, the semantic utilization bottleneck and long-tail generalization limitation in the prior art are broken through, and a new technical path is provided for personalized recommendation.

The invention belongs to the technical field of intelligent monitoring, and particularly relates to an urban underground pipe network real-time monitoring algorithm and system based on multi-source data fusion, and the method comprises the steps: obtaining multi-source heterogeneous monitoring data; performing multi-source data preprocessing; carrying out multi-source heterogeneous feature coding and fusion; carrying out real-time monitoring and anomaly detection on a pipe network state; fault diagnosis and prediction are carried out; and generating decision support information and early warning. The system comprises a data acquisition module, a data preprocessing module, a multi-source heterogeneous feature coding and fusion module, a pipe network state real-time monitoring and anomaly detection module, a fault diagnosis and prediction module and a decision support and early warning module. According to the scheme, multi-source heterogeneous data are integrated, spatial-temporal feature coding and fusion are carried out through deep learning, accurate sensing, early warning and intelligent fault diagnosis of the operation state of the pipe network are achieved, and the safe operation level and maintenance management efficiency of the urban underground pipe network are improved.

The invention relates to the technical field of electrical equipment fault diagnosis, and discloses an electrical equipment multi-modal fault diagnosis method based on dynamic self-adaption, and the method comprises the steps: obtaining an original multi-modal signal flow containing vibration, temperature and current signals, inputting the original multi-modal signal flow into a dynamic self-adaption diagnosis network, obtaining a fault feature matching result set, and determining each modal result. The network is trained by historical fault data, and the data comprises time domain, frequency domain and fusion feature parameters extracted from continuous multi-mode signals, and corresponding fault type labels and confidence coefficients. The network comprises a cross-modal feature fusion module, a spatial-temporal feature coding module and the like, and when the confidence coefficient of at least two modal results exceeds a dynamic threshold value, three-level early warning is triggered. The method improves the comprehensiveness, accuracy and real-time performance of diagnosis, and is suitable for fault diagnosis of electrical equipment.

The invention provides a cross-department data collaborative modeling method and system based on federated learning in the cross technical field of federated learning and privacy computing, and the method comprises the steps: S1, carrying out desensitization operation on department data by each edge computing node, and coding the desensitization data to obtain a feature coding vector; s2, constructing a layered federated learning framework based on a department private feature extractor, a cross-department parameter collaborative aggregation layer and a global task header; s3, extracting data features from the feature coding vectors through a department private feature extractor so as to train the model and extract model parameters; s4, performing hierarchical quantization coding, encryption and compression on the model parameters to obtain compression parameters, and sharing the compression parameters by a cross-department parameter collaborative aggregation layer; and S5, the global task head reads the shared compression parameters to carry out collaborative modeling on the model. The cross-department data collaborative modeling method has the advantages that the precision and efficiency of cross-department data collaborative modeling are greatly improved on the premise that the safety is guaranteed.

The invention provides a traffic participant behavior analysis-based edge calculation signal lamp control method and system, and the method comprises the steps: obtaining video stream data of a target intersection, carrying out the frame-by-frame behavior state analysis of the video stream data through a spatial-temporal feature coding network, extracting the behavior state features of each traffic participant, and carrying out the analysis of the behavior state features of each traffic participant; and inputting the behavior state characteristics into a preset behavior matching model, generating a behavior triggering identifier of the traffic participant in the current time window, generating a signal lamp control instruction of an edge computing node according to the time sequence relevance between the behavior triggering identifier and the phase state of the current signal lamp, and sending the signal lamp control instruction to the edge computing node. And finally, a signal lamp phase switching time sequence of the target intersection is adjusted based on the signal lamp control instruction, so that the traffic participants meeting the traffic rule conflict condition obtain the traffic priority under the signal lamp phase switching time sequence. According to the method, the traffic accident risk is reduced, and meanwhile, the overall traffic efficiency of the intersection is optimized by dynamically adjusting the minimum response period and the phase duration parameters.

The invention discloses a feature attention and bilinear gating fused speech emotion recognition method and device, and the method comprises the following steps: 1, collecting an audio file, obtaining corresponding label information, generating audio waveform and time frequency representation data through preprocessing, and constructing an audio waveform mask and a time frequency mask to mark an effective information region; 2, constructing a dual-path feature extraction module which comprises a time-frequency feature extraction module and a pre-training acoustic feature coding module; wherein the time-frequency feature extraction module models emotion correlation through local convolution and a multi-dimensional attention mechanism, and performs global time sequence modeling based on a bidirectional gated loop network; the pre-training acoustic feature coding module extracts high-level speech representation with high expression ability for emotion distinguishing by using a pre-training model; and step 3, constructing a feature fusion module and an emotion classification module, and combining with a dual-path feature extraction module to form a speech emotion recognition model.

The invention discloses a robot hierarchical reinforcement learning variable impedance control method based on vision and touch, and belongs to the field of robot control. According to the method, an upper-layer planning system and a lower-layer variable impedance control system are included, the upper-layer planning system fuses information from a visual sensor and a tactile sensor, feature coding is conducted on a visual image through a variational self-coding model, tactile feedback and mechanical arm state information are combined, and an action strategy adapting to the current environment and corresponding impedance parameters are formulated. And the lower-layer variable impedance control system realizes variable impedance control of the robot based on a depth kinematics and dynamics model, and allows the robot to dynamically adjust the motion and contact force according to the sensed information when interacting with the environment, thereby improving the adaptability to the environment change and the task execution efficiency. According to the method, the success rate and execution efficiency of the robot for executing the complex assembly task in the unstructured environment are remarkably improved, and high robustness is shown.

The invention provides a shield tunneling machine cutter cylinder state monitoring method based on a multi-sensor information fusion technology. The shield tunneling machine cutter cylinder state monitoring method comprises the steps that multi-physical field synchronous data collection is carried out; self-adaptive signal preprocessing: processing the multi-physics field time sequence signal to obtain multi-physics field synchronization time sequence data; multi-physics field feature space construction: processing the multi-physics field synchronization time sequence data to respectively obtain time domain feature data, frequency domain feature data and time-frequency domain feature data; deep collaborative feature extraction and state evolution modeling: obtaining a multi-physics field coupling feature sequence of the spatio-temporal context through multi-physics field collaborative perception feature coding; global space-time dependency collaborative modeling is adopted, and an enhanced space-time feature sequence is obtained; and state evolution trajectory interpretation and classification decision. According to the scheme, a traditional linear stacking mode of data acquisition-preprocessing-feature extraction-fusion classification is broken through, and a closed-loop intelligent sensing method with physical mechanism guidance, multi-modal data deep coupling and feature self-coevolution is constructed.

The invention relates to the technical field of wireless communication, in particular to a wireless communication anti-interference enhancement system based on multi-mode signal fusion. Comprising an intelligent signal transceiving module, a signal sensing and preprocessing module, an interference feature coding and classification module, an interference suppression decision module, an interference suppression module, a signal fusion module, an interference suppression performance evaluation module, an interference feature library and an interference suppression strategy library, the interference feature coding and classification module adopts a three-layer progressive architecture of multi-dimensional manifold feature extraction, feature transformation and coding inspired by a group theory, and deep learning classification based on chaotic dynamics, extracts interference features from three dimensions of time domain, frequency domain and nonlinearity, and realizes accurate interference identification; the signal fusion module adopts a multi-level fusion strategy, and dynamically adjusts the fusion weight according to the signal quality; the interference suppression module selects an optimal suppression strategy according to the interference type, and executes frequency domain cancellation, space domain cancellation and time domain cancellation; and performance evaluation feedback forms a closed-loop optimization mechanism.

The invention provides an oil and gas pipeline defect three-dimensional contour determination method and device. Acquiring a three-axis magnetic flux leakage detection signal of a to-be-detected target oil and gas pipeline and corresponding space coordinate information of the three-axis magnetic flux leakage detection signal; constructing multi-channel input data according to the three-axis magnetic flux leakage detection signal and the space coordinate information; determining a defect contour prediction result of the target oil and gas pipeline according to the multi-channel input data by using a pre-trained defect contour inversion model; wherein the pre-trained defect contour inversion model comprises a multi-axis feature extraction and fusion module, a feature coding module and a multi-task decoding module; the defect contour prediction result is used for indicating whether the target oil and gas pipeline has defects or not, and the defect contour prediction result is further used for indicating the three-dimensional contour shape of the defects under the condition that the target oil and gas pipeline has the defects. Therefore, high-precision visual reconstruction of the defect position and the three-dimensional form of the oil and gas pipeline is realized, and the accuracy of defect identification and the reliability of evaluation are improved.

The invention discloses a time-space interaction vehicle trajectory prediction method based on speed perception, and the method comprises the steps: obtaining driving trajectory data of a target vehicle and surrounding vehicles in a perception range of the target vehicle, inputting the driving trajectory data into a pre-trained trajectory prediction model, and obtaining the trajectory data of the target vehicle in a prediction time period; the trajectory prediction model comprises an encoder module, a speed sensing interaction modeling module and a decoder module; in the encoder module, a spatial-temporal feature encoder is used for obtaining spatial-temporal feature codes based on the driving tracks of the target vehicle and other vehicles; the scene perception encoder is used for extracting a global spatial dependency relationship between vehicles to obtain a scene perception spatial code; the speed perception interaction modeling module is used for extracting space-time interaction features and scene interaction features of the target vehicle and surrounding vehicles by using a multi-head attention mechanism based on the space-time features and scene perception space codes, and further obtaining global interaction features; and the decoder module is used for obtaining a driving track of the target vehicle in the prediction time period based on the space-time interaction characteristics and the global interaction characteristics. According to the method, the vehicle-scene spatial dependency can be accurately captured, and the trajectory prediction accuracy is improved.

The invention discloses an ocean multi-mode environment perception and intelligent ship navigation decision-making method based on a double-branch vision-semantic encoder. The method comprises the following steps: S1, acquiring a multi-source data image containing a ship and a surrounding environment thereof from an existing public maritime data set or platform; s2, training a double-branch vision-semantic encoder by using the multi-source data image, and inputting a to-be-processed image extracted in real time into a multi-modal feature matrix in the trained double-branch vision-semantic encoder; s3, based on the multi-modal feature matrix, obtaining positioning information of the ship and surrounding environment elements, and constructing a dynamic security domain model; and S4, in combination with the dynamic security domain model and the multi-ship relative position relationship, carrying out quantitative evaluation on the navigation risk, and generating a self-adaptive navigation strategy based on an evaluation result. According to the invention, high-precision ship positioning and environment element identification under complex weather and illumination conditions are realized by using all-weather characteristics and multi-scale visual feature coding of SAR imaging.

The invention relates to the technical field of action recognition, in particular to a somatosensory action interaction recognition method and system based on skeleton coordinate points. The method comprises the following steps of collecting real-time skeleton coordinate data of a human body and performing multi-modal feature extraction to obtain a real-time skeleton coordinate sequence; obtaining a standard skeleton posture corresponding to the target interaction action, performing pre-recording and feature coding, and generating a target posture skeleton feature template library; performing skeleton time sequence filtering and joint mapping and joint included angle calculation on the real-time skeleton coordinate sequence, performing similarity measurement and dynamic binding tracking at the same time, and starting a binding recovery mechanism when binding loss is detected so as to guide the user to execute a preset binding posture and re-establish a binding relationship; and mapping the joint included angle time sequence data to a corresponding joint of the virtual human shape interaction model in real time, outputting a somatosensory interaction instruction, and driving to repeat a human body action so as to trigger a somatosensory action interaction event. According to the invention, the stability of somatosensory action interaction recognition can be improved.

The invention discloses an equipment fault diagnosis method and system based on a large electric power model, and the method comprises the steps: collecting and fusing voltage and current waveform data and alarm log information in a short time window before and after a fault moment, achieving the context integration of a multi-dimensional feature vector through multi-level feature coding and semantic embedding, and achieving the fault diagnosis of a large electric power model. And relevant knowledge fragment retrieval is carried out in combination with a power field knowledge base. On the basis, relevant knowledge fragments, voltage-current waveform fusion feature coding vectors and fault alarm key information semantic embedding coding feature vectors are jointly embedded into a preset Prompt template and then are input into a diagnosis engine based on large model driving, so that intelligent analysis and report generation of fault types, reasons and key features are realized. Through the mode, the limitation of dependence on single data or rules traditionally is broken through, the accuracy, automation and interpretability of fault diagnosis under complex working conditions are remarkably improved, and a solid support is provided for intelligent and high-reliability equipment operation and maintenance.

The invention discloses a multi-modal sequence recommendation method based on a double-gating hybrid expert model and Fourier denoising. The method comprises the following steps: S1, multi-modal feature coding: extracting text / visual features by using a BERT / ViT pre-training model; s2, frequency domain feature denoising: carrying out frequency domain noise filtering by adopting Fourier transform; s3, establishing a double-gating hybrid expert model: adopting a parallel path to realize self-adaptive multi-modal fusion and time sequence interest modeling; and S4, multi-task joint optimization: performing joint optimization on the model by integrating triple auxiliary contrast learning, extracting text and image features by utilizing a pre-training model BERT / ViT, performing frequency domain noise reduction on multi-modal features by introducing Fourier transform, remarkably improving the robustness of modal representation, realizing fine-grained modal interaction through dynamic routing by an input dependent expert, and performing multi-task joint optimization. A shared expert uses time coding with a gating mechanism to model periodic evolution of user interests, the feature fusion quality is optimized, the FT-MSR integrates triple auxiliary learning tasks, and the problem of data sparsity is relieved.

The invention relates to the field of copyright authentication of non-perpetual culture elements, and discloses a block chain copyright authentication method for non-perpetual culture elements, which comprises the following steps: S1, performing multi-dimensional data acquisition through a distributed storage system, performing structured data acquisition and feature coding on the non-perpetual culture elements, and storing the structured data in a database; comprising the following steps: multi-modal data fusion acquisition, dynamic metadata standardization processing, and generation of unique digital identifiers and cultural feature codes; and S2, constructing a hybrid block chain architecture, constructing a hierarchical network architecture comprising a permission chain layer and an open chain layer, and realizing heterogeneous chain data synchronization through a cross-chain interoperation protocol. Through multi-modal data acquisition and dynamic metadata structured processing, the spatial form, the process time sequence and the oral context of non-residual elements are completely recorded, digital twin bodies with microscopic details and semantic association are formed, the problem of fragmentation of traditional archiving is solved, and a unique block chain identifier and two-way index mapping are combined, so that the method is simple and convenient to implement. And the originality of the cultural heritage in virtual restoration and cross-domain propagation is ensured.

The invention relates to an intelligent interview scoring system capable of explaining decisions based on a large language model, and the system comprises a multi-mode resume analysis and feature coding unit, a resume feature adaptive matching unit, an interactive scoring and knowledge enhancement unit, and an answer quality evaluation unit. Text, image and audio features are extracted through a cross-modal attention mechanism of a multi-modal large language model, resume features are encoded into dynamic word vectors, and entity-level feature vectors are extracted; the post description text is encoded into a demand feature vector by a resume feature adaptive matching unit; calculating semantic similarity between the resume entity feature vector and the demand feature vector; the interactive scoring and knowledge enhancement unit dynamically retrieves knowledge fragments to generate a preliminary evaluation report containing a scoring basis; and the answer quality evaluation unit fuses the information density, the fluency and the integrating degree to generate a final score. And the whole-process intelligence from demand analysis to final decision making is realized.

The invention discloses a software project development method based on a low-code platform, which belongs to the technical field of software development, and specifically comprises the following steps: receiving original demand description through an embedded natural language parser, and establishing a semantic incidence matrix to extract demand elements; encoding the demand elements into three-dimensional structured demand vectors; mapping the demand vector to a three-dimensional orthogonal feature space, calculating the coordinates of a functional module, and generating an initial topological structure based on a Delou inner triangulation algorithm; rendering an interactive three-dimensional network map on a visual interface; monitoring user operation, and recalculating the adjacency relation of the affected module when the displacement of the module exceeds a threshold value or the connection edge is reconstructed; compared with a historical version, when a key service path is interrupted or a data flow ring is broken, a compensatory connection bridge is activated; a deployable architecture with a version traceability identifier is output, and a traceability code is built in the module to reversely map an original demand vector.

The invention discloses a three-dimensional scene reconstruction method based on an intelligent LED street lamp multi-mode sensor. The three-dimensional scene reconstruction method comprises the following steps that RGB images are obtained and preprocessed; the information is input to a visual feature coding module, two-dimensional bounding box information is extracted, and an object segmentation module is guided to output a two-dimensional segmentation mask; obtaining point cloud data, projecting the point cloud data to the standardized RGB image, and screening target points in combination with the two-dimensional segmentation mask; complementing the preliminary segmentation result of the point cloud, mapping the result to a standardized RGB image, and extracting a pixel region; carrying out joint coding, implicit representation and neural decoding processing on the object-level RGB image and the point cloud complete segmentation result; and fusing into an original three-dimensional scene, and completing spatial restoration through point cloud registration, attitude optimization and semantic constraint. The invention provides an efficient three-dimensional scene reconstruction method in combination with a multi-mode sensor of an intelligent LED street lamp, and the method has high precision, real-time performance and dynamic target processing capability.

The invention relates to the technical field of computer systems and human-computer interaction, in particular to a text travel element universe virtual role collaboration system and method based on large-space interaction. The method specifically comprises the following steps: tracking a user spatial position in real time through a multi-source positioning technology, and carrying out identity authentication and equipment legality verification by fusing multi-modal features; constructing a virtual-real fused semantic point cloud structural body, binding role interaction nodes, dynamically generating a task collaboration graph by using a graph neural network, and driving multiple roles to collaboratively execute subtasks; calling a graph attention network and a GPT module to deduce personalized plot branches in combination with user multi-modal input and a historical track, and rendering an immersive interactive scene in real time; spatial preferences and content interests are mined by adopting DBSCAN clustering and depth feature coding, a dynamic portrait is constructed, and a task path and content presentation are optimized through a weighted recommendation algorithm. According to the method, cloud rendering, edge calculation and cross-module cooperative control are integrated, and the interactivity and immersion experience of the text travel element universe are remarkably improved.

The invention belongs to the technical field of computers, and particularly relates to a software fault repairing method and system fused with intelligent analysis, which comprises the steps of collecting a multi-level running log and performing structured preprocessing, constructing a dynamic calling graph through a time sequence encoder and a graph neural network, inferring a fault root cause in combination with a Bayesian causal inference model, and repairing a fault fault according to the fault root cause. And matching the repair strategy to generate an atomization instruction sequence, and deploying the atomization instruction sequence to a production system after sandbox environment verification. The system comprises a log acquisition module, a feature coding module, a graph construction module, a causal reasoning module, a strategy matching module, an instruction generation module, a sandbox verification module, a deployment feedback module and the like. Through end-to-end intelligent analysis and a closed loop verification mechanism, the fault positioning precision and the repair safety are remarkably improved, system self-evolution is supported, and operation and maintenance are promoted to be transformed from passive response to active autonomy.

The invention provides a panoramic segmentation method and device in a complex scene and a storage medium, and the method comprises the steps: inputting an initial image into a panoramic segmentation network, and obtaining a panoramic segmentation result of the initial image; the panoramic segmentation network comprises: a backbone network; the feature coding layer comprises a DLK module and a DFF module which are cascaded, and the DLK module and the DFF module cooperate with each other to generate enhanced features E1 '-E4'; the edge guiding layer comprises an edge guiding fusion module and a feature pyramid, the EGF module fuses the C1 feature of the backbone network with the enhanced feature E4'to generate an edge enhanced feature F, and then performs channel splicing with E1 '-E3' to construct a four-level feature pyramid F1-F4; the dual-path decoding layer comprises a learnable query and a pixel updating path; the dual-path feature uses a cross attention mechanism to realize interaction, and a semantic category and an instance mask required by panoramic segmentation are output; the method can meet the adaptability in a complex scene and improve the segmentation quality.

The invention discloses an intention recognition method based on cross attention and multi-scale uncertainty. The intention recognition method comprises the following steps: preprocessing multi-modal data; parallel multi-modal feature coding oriented to intention recognition; the invention relates to multi-scale uncertainty perception decoding. According to the method, a parallelized multi-modal feature extraction path is constructed, and a hierarchical fusion mechanism based on cross attention is designed, so that deep semantic alignment and complementary enhancement of four types of heterogeneous information including the posture, the motion track, the global scene and the local vision of a rider are realized; the problems of incomplete feature representation and insufficient cross-modal correlation modeling caused by dependence on a single information source or adoption of a shallow fusion strategy in a traditional method are solved, so that the accuracy and robustness of intention recognition in a complex traffic scene are remarkably improved. According to the method, a multi-scale uncertainty perception decoding framework is introduced, risk early warning or context auxiliary verification is carried out on a low-confidence identification result, and the reliability of an automatic driving system in a safety critical scene is improved.

The invention relates to an intelligent energy-saving method for a large building refrigerating system based on a large language model, and belongs to the field of intelligence of large building refrigerating systems. A reinforcement learning framework is constructed, a multi-dimensional state space formed by building thermal loads, equipment operation states and thermal environment parameters is defined, and a dynamic reward mechanism is designed by taking improvement of energy efficiency, guarantee of thermal safety constraints and maintenance of thermal comfort as optimization objectives; developing a word embedding model special for the refrigeration field and a cross-modal decision conversion model, and realizing bidirectional analysis of a natural language instruction and a physical control parameter through feature coding and embedding mapping; and performing low-rank adaptive fine tuning on the pre-trained large language model based on the reinforcement learning experience set, and constructing a closed-loop verification system. Generalized migration and multi-target collaborative decision-making of equipment control strategies are achieved, the self-adaptive regulation and control capacity of a refrigeration system under dynamic loads is improved, and a universal intelligent solution for energy efficiency optimization, thermal safety and comfort is provided for a large building.