37 results about "Fusion frame" patented technology

The invention discloses a video monitoring and AI linkage intelligent alarm verification system, and particularly relates to the technical field of video analysis, which comprises the following steps: carrying out dual anomaly preliminary screening by using a flow field structure entropy and a signal track singular value ratio, eliminating environmental noise interference, and constructing a multi-mode normal state baseline; generating a multi-modal event report containing a spatio-temporal context, uploading the multi-modal event report to a cloud, inputting the multi-modal event report into a physical perception cross attention network, configuring a physical embedding vector into a query vector and configuring a visual embedding vector into a key vector and a value vector through an asymmetric feature fusion architecture, and performing multi-modal event report analysis; actively guiding the attention weight distribution of the model on the video picture by using the change trend of the physical parameters, and outputting a confidence score based on the weighted fusion feature; performing closed-loop parameter correction on the baseline model by utilizing online incremental learning based on a verification result; the problems of high false alarm rate caused by lack of physical logic constraints and poor anti-interference capability in a complex environment in traditional monitoring are effectively solved.

The invention provides a non-technical line loss state evaluation method based on multi-scale spatial-temporal feature fusion in a low-voltage distribution network environment, and relates to the technical field of electric power big data analysis and intelligent operation and maintenance of a distribution network. According to the invention, a fusion architecture based on a multi-scale time convolution network and a long short-term memory network is constructed; extracting multi-scale spatio-temporal characteristics of a user from instantaneous electricity utilization abrupt change to a periodic load rule through an MSTBlock unit; designing a cluster balance constraint mechanism to ensure that rare and key non-technical line loss abnormal early warning signals are not covered by mass normal power utilization data; according to the data scale, adaptively selecting a graph segmentation or spectral clustering integration strategy to output a clustering label, and mapping the clustering label into a user power consumption behavior evolution track; according to the method, the power utilization abnormal level can be identified from the original load signal with random fluctuation interference, and the troubleshooting priority is calculated in combination with the transformer area correlation analysis, so that the accuracy and interpretability of the non-technical line loss unsupervised evaluation decision of the power distribution network are remarkably improved.

The invention discloses a method and a system for realizing intelligent control on the basis of a VLM (Visual Language Model) + Actionality fusion architecture. The method aims at solving the problems that an existing end-to-end vision-language-action (VLA) model is scarce in data, high in hardware coupling degree, poor in controllability, weak in compatibility and the like. The core thought is to decouple high-level semantic understanding and bottom-level action execution, finish scene understanding and task planning through a VLM module, map high-level intentions into standardized API calling through an intention analysis and action calling module, execute specific actions through an Action module library packaging a traditional algorithm, and dispatch and monitor through an action execution engine. A large amount of end-to-end training data is not needed, the research and development cost is reduced, the system modularization, expandability and hardware universality are improved, the high delay problem is avoided, meanwhile, the controllability and safety of the control process are guaranteed, and the method is suitable for various intelligent scenes with bodies such as industrial manufacturing, medical assistance and family service.

A multi-modal signal recognition and dialogue method based on a large language model comprises the steps that a signal coding module based on time sequence modeling is constructed, preprocessing and feature extraction are conducted on input I / Q signal data, and semantic alignment pre-training of signal features and modulation type text description is achieved through a contrast learning mechanism; designing a multi-modal fusion architecture, mapping signal features to a hidden space of a large language model by adopting a signal projector, and realizing deep fusion of the signal features and text features through special markers; constructing a dialogue generation module based on a pre-trained large language model, receiving the fused multi-modal input, and generating a natural language answer about signal analysis; performing feature alignment by training a double-layer MLP projector to realize end-to-end multi-modal signal understanding and dialogue ability; an intelligent question-answering system in a reasoning stage is constructed, natural language interaction between a user and the system is realized through a predefined professional prompt word template and a signal feature fusion mechanism, and multi-dimensional signal analysis query requirements are supported. According to the invention, the organic combination of signal understanding and natural language generation is realized, and the accuracy of signal identification and the user interaction experience are improved.

The invention discloses a gas detection precision improving method based on a multi-algorithm fusion framework. A photonic crystal resonant cavity and a tunable band-pass filtering structure are integrated in a Fourier transform spectrometer light path; collecting a wide-spectrum light intensity signal, and constructing a spectral signal database in combination with the enhancement characteristic and the filtering characteristic of the resonant cavity; zero calibration and dynamic baseline deduction operation are carried out, and self-adaptive variational mode decomposition and wavelet transform are adopted to carry out signal denoising on the spectral signals; an environment compensation model is established, and an Arrhenius type correction factor is adopted to suppress water vapor cross interference; constructing an RLS and fuzzy control combined hybrid adaptive filter; separating aliasing spectral signals by adopting a non-negative matrix factorization algorithm; establishing a quantitative relation model; in the online concentration prediction process, the initial concentration value is subjected to recursive optimization by using a hybrid adaptive filter, the deviation is continuously corrected through an environment compensation model, and an accurate concentration value is output. The method can significantly improve the accuracy and stability of multi-gas detection.

The invention relates to the technical field of artificial intelligence, in particular to a customer service reply method and system of a multi-modal fusion architecture and a storage medium, and the method comprises the following steps: obtaining initial input data of a user, preprocessing the initial input data, and generating a standard data source; performing decision compliance analysis on the standard data source to generate a decision judgment result; and generating a corresponding decision scheme according to the decision judgment result. According to the invention, through a space-semantic double-constraint graph network, the technical defects of fragmentation and low standardization degree of multi-modal data identification by traditional intelligent customer service are effectively solved, and the problem of insufficient adaptation of a general identification technology to a government affair format is overcome; in a semantic constraint dimension, spatial features and a government and enterprise professional semantic dictionary are deeply fused based on a node association mechanism of a graph network, conversion from multi-modal data to structured policy elements is automatically completed, and the recognition accuracy is improved by more than 40% compared with that of a traditional NLP technology.

The invention provides a training teaching video intelligent analysis and knowledge point automatic marking method based on multi-modal fusion, and the method comprises the steps: collecting multi-modal data in a training teaching process, and carrying out the time alignment processing; extracting feature information of the modal data, fusing the feature information through a cross-modal fusion architecture, and generating a unified teaching behavior representation vector; based on the teaching behavior representation vector, identifying operation steps in the practical teaching process through a sequence labeling model, and determining the category and the starting and ending time boundary of each operation step; matching the identified operation steps with a preset skill knowledge base, and generating a standardized knowledge point label containing knowledge point content, starting and ending timestamps and confidence information; and storing the standardized knowledge point labels into a database, and constructing a retrieval index. According to the invention, the unstructured teaching video is converted into a searchable, navigable and analyzable knowledge unit, and automatic identification and structured marking of practical teaching operation are realized.

The invention is applicable to the technical field of natural language processing, and provides a multi-granularity event detection method based on adaptive fusion and type perception, which comprises the following steps of: performing multi-granularity dynamic feature fusion, a type perception graph neural network and adaptive context fusion; key challenges such as data sparsity, type imbalance and complex context understanding in event detection in the professional field are effectively solved. The multi-granularity feature fusion mechanism successfully captures event clues in different language units, the type perception graph neural network effectively relieves the problem of insufficient representation of low-frequency event types, and the adaptive context fusion architecture significantly improves the modeling ability of the model for complex long-distance dependency relationships. A powerful end-to-end event detection solution is formed through the synergistic effect of the innovation points, and a new technical thought is provided for solving typical problems in text processing in the professional field.

The invention discloses a transient evaluation method based on SHAP, and relates to the field of transient stability of a power system, and the method comprises the steps: S1, constructing a prediction-interpretation synchronization framework: embedding a Shapley module supporting parallel calculation into a deep neural network, and achieving the synchronous execution of transient stability evaluation and feature attribution in forward propagation; s2, designing a spatio-temporal feature fusion architecture: adopting a time-space-feature three-dimensional tensor as input, constructing a deep SHAP network by referring to a convolutional neural network thought, and enhancing transient process dynamic feature modeling capability; and S3, establishing a quantitative decision support mechanism: based on the quantitative influence of the Shapley value analysis feature on the stability, realizing transparent explanation of an evaluation result, and providing a decision basis for beforehand prevention and control strategy making and post emergency control deployment. According to the method, the balance between the transparency and the accuracy of transient stability evaluation is realized, and a more accurate, quicker and interpretable evaluation result is provided for power grid dispatching.

The invention relates to the field of three-dimensional scanning, and particularly discloses a fixed and handheld integrated laser scanning method and system, and the method comprises the steps: building a high-precision global coordinate system based on the mechanical and magnetic field reference of a detachable fixed module in a fixed mode, and obtaining an initial point cloud; in a hand-held mode, aligning hand-held data to the global coordinate system by identifying space-time cohesion features common in view with fixed scanning; by taking fixed data as an optimization anchor point, constructing a hierarchical factor graph model fused with an uncertainty weight, and carrying out constraint optimization on the aligned handheld data to correct an accumulative error of the handheld data; and finally, outputting a complete and consistent three-dimensional model through closed-loop verification and parameter self-adaption steps. The system comprises a scanning host, a detachable fixing module and a processing unit. According to the method and the system, the architecture error problem of mode switching is fundamentally solved through a standard unification and anchor point optimization fusion architecture, and high-precision and high-efficiency scanning of single equipment in various complex scenes is realized.

This invention belongs to the field of metal pressure processing and intelligent manufacturing, specifically relating to an intelligent prediction and process optimization method for the forming performance of hot-rolled titanium strip. The invention discloses an intelligent prediction and process optimization method for the forming performance of hot-rolled titanium strip, aiming to solve the problems of low prediction accuracy and lack of closed-loop optimization caused by the complex coupling of hot-rolling process parameters. The method includes: collecting multi-source heterogeneous data from hot rolling and constructing a multi-dimensional feature correlation matrix; establishing a prediction model based on a fusion architecture of deep convolutional neural networks and long short-term memory networks, extracting deep features of microstructure evolution and mechanical response; using the prediction model as the objective evaluation function, employing a non-dominated sorting genetic algorithm to perform collaborative optimization of process parameters throughout the entire process; and achieving real-time feedback correction and closed-loop control through an online learning mechanism. This invention achieves accurate performance prediction and global process optimization, improving prediction robustness and finished product qualification rate, and reducing production costs.

The invention relates to a ghost imaging neural differential analysis method under a selected plaintext attack condition in the technical field of computational imaging and optical security. The method comprises the following steps: firstly, designing a differentiated plaintext, and deducing a projection speckle by adopting differential calculation; then, constructing a dimensionality reduction simulation training set and an experimental test set based on the structural features of the projection speckles; then, constructing a one-dimensional signal noise reduction neural network model, and training the one-dimensional signal noise reduction neural network model by adopting the dimension reduction simulation training set; and finally, inputting the experimental test set into the trained one-dimensional signal noise reduction neural network model to obtain deciphered speckles, and based on the deciphered speckles, adopting ghost imaging correlation operation to obtain a cracked plaintext image. That is to say, based on the basic principle of ghost imaging encryption, the inherent linear property of the ghost imaging light path is utilized, and the fusion architecture of the differential attack and the neural network is used, so that the cracking of various ghost imaging encryption methods can be realized, and the universality and reliability of ghost imaging encryption analysis can be improved.

The invention relates to the technical field of automobile starting, in particular to an AMT starting control method, system and device based on a residual network and a storage medium, and the method comprises the steps that the residual network is constructed, historical semaphores needed for clutch and engine control during automobile starting are collected, and clutch sliding friction work, clutch impact degree and engine torque are selected; constructing a data set as the input of the network by the three parameters through a binary converted data image; a self-attention mechanism is introduced into the network to extract different types of image features, importance of different features is dynamically learned by using a gating mechanism, and importance weight distribution is performed on the extracted features; training a network model by using the collected training sample set to obtain a mapping relationship between input and output; and finally, the output value of the network is the optimal starting control parameter after the network decision. According to the method, by introducing the fusion architecture of the residual network and the attention mechanism, the comprehensive performance of AMT starting control is effectively improved.

The invention relates to the technical field of visual language task processing, and discloses a visual language task processing method and device based on hierarchical optimal transmission and a medium. Firstly, an attack specific data set and a prompt specific data set are constructed, and diversified sub-models are constructed from the dimension of attack type confrontation; and text prompt is creatively introduced as a second dimension, so that the sub-models have complementarity in robust characteristics and a characteristic alignment mode. Secondly, a thought of directly fusing all sub-models is abandoned, and a two-stage fusion framework is designed: the sub-models with the same attack type and different prompts are fused firstly, and then fusion results of different attack types are subjected to secondary fusion, so that the semantic similarity between the models during each alignment is effectively ensured, and the fusion efficiency is improved; therefore, the accuracy of the fusion process based on the optimal transmission alignment method is improved. Therefore, the robustness of the visual language model is improved on the basis of ensuring the accuracy of processing the visual language task by the visual language model.

The invention relates to an active sonar target identification method based on multi-pulse accumulation and two-stage fusion. The method comprises the following steps: firstly, acquiring a multi-dimensional feature value of a target, performing multi-pulse (ping) accumulation on each feature to form a feature subset, and performing feature-level fusion through a fusion operator so as to extract stable feature representation with strong discriminability; on the basis, the probability that each fusion feature belongs to a real target and a non-real target is calculated based on prior distribution, and decision-level fusion is performed on probability evidences of all the features by adopting a D-S evidence theory, so that the credibility of a final recognition result is obtained. Through time sequence accumulation and a two-stage fusion architecture, multi-dimensional information and time sequence information of a target are effectively integrated, the recognition accuracy and decision robustness are remarkably improved, and meanwhile, the false alarm rate is greatly reduced.

The invention belongs to the technical field of redundant network verification, and particularly relates to a redundant network verification method fusing frame type dynamic mapping and adaptive optimization, which comprises the following steps: acquiring input combined frames, and judging whether the frame types of the input combined frames are the same or not; detecting whether the input combined frame has an error frame or not; when the frame types of the input combined frames are the same and error frames do not exist, the input combined frames enter a preset first mapping table for verification logic selection, and table items of the first mapping table are updated; otherwise, the input combined frame enters a preset second mapping table for verification logic selection, and table items of the second mapping table are updated; the first mapping table and the second mapping table are used for mapping the frame type and the function algorithm needing to be verified. By presetting mapping tables for different combination conditions, different frame type combinations and verification algorithms are in one-to-one correspondence, the complexity of search and update operation is low, and the method has high flexibility and expandability.

This invention relates to a method and system for automatic identification and classification of defects in power distribution network towers based on deep learning. The method first uses a convolutional neural network to locate and extract the main area of ​​the tower, eliminating background interference. Then, an adaptive contrast enhancement algorithm based on local statistical characteristics is employed to improve the visual saliency of the defect area. In the feature extraction stage, a pyramid-shaped distraction attention module is introduced to fuse multi-scale spatial information and channel attention, and a two-dimensional selective state space module is used to model long-range dependencies. Furthermore, a hierarchical feature fusion architecture and an adaptive anchor box mechanism are used to aggregate multi-resolution features and match targets of different sizes. Finally, an adaptive edge enhancement module is used to strengthen defect edges, and a multi-branch detection head performs parallel defect category determination, location regression, and confidence assessment. This effectively improves the detection accuracy and robustness of tower defects in complex backgrounds, and is particularly suitable for the automatic identification of small-scale defects.

The application provides an end-side RAG implementation method and system based on a fusion architecture, the method comprising: initializing a vectorization engine through a user browser to obtain an initialized vector model; performing vectorization processing on a local text knowledge base based on the initialized vector model to obtain a vector knowledge base; obtaining query text issued by a user, and performing retrieval and query on the vector knowledge base based on the query text to output a retrieval result; intelligently deciding on a full-cycle workload based on a current network environment to realize dynamic calculation offloading; and encrypting, storing and verifying full-cycle data to complete the implementation of the end-side RAG. The application replaces a traditional cloud-side vector calculation architecture, realizes zero server cost operation of an AI workflow through a pure front-end calculation architecture, and solves three major pain points of a centralized AI platform, i.e., high cost of computing power, high delay and poor privacy.

The application relates to the technical field of artificial intelligence, in particular to a customer service reply method and system of a multi-modal fusion architecture and a storage medium, the method comprises the following steps: obtaining initial input data of a user, preprocessing the initial input data to generate a standard data source; performing decision compliance analysis on the standard data source to generate a decision judgment result; and generating a corresponding decision scheme according to the decision judgment result. Through a space-semantic double-constraint graph network, the application effectively solves the technical defects of fragmentation of multi-modal data recognition and low standardization degree of traditional intelligent customer service, overcomes the problem of insufficient adaptation of general recognition technology to government affairs formats, in the semantic constraint dimension, relying on the node association mechanism of the graph network, the spatial features are deeply fused with the government and enterprise professional semantic dictionary, the conversion of multi-modal data to structured policy elements is automatically completed, and the recognition accuracy is improved by more than 40% compared with traditional NLP technology.

This application belongs to the field of motor system fault diagnosis technology, and relates to a motor system fault diagnosis method and system. It collects core electrical data of the motor's three-phase current and voltage, as well as auxiliary data of ambient temperature, and performs timestamp alignment to eliminate phase deviation. Then, it performs signal purification processing to suppress noise and retain effective fault features. Subsequently, it extracts multi-dimensional features and filters core features through correlation evaluation, eliminating redundant features. A fault feature library is constructed and dynamically iterated based on historical data, expert experience, and new data to achieve dynamic tracking of fault modes. A hierarchical multi-model fusion architecture is used to progressively analyze core features, and the stability of the analysis results is ensured by combining result fusion and parameter adaptive optimization. The results are parsed through operating condition adaptation logic, interference is eliminated, and graded early warnings are executed. Finally, based on fault evolution characteristics, historical degradation data, and error correction logic, the remaining lifespan and confidence interval are estimated, filling the gap in the accuracy of lifespan prediction in existing technologies.

This application discloses a radar echo extrapolation method based on temporal motion decomposition and dominant fusion, relating to the field of spatiotemporal sequence prediction technology. The method proposes an innovative network fusion paradigm, specifically including: acquiring implicit long-term / short-term temporal features and explicit global / transient motion features during processing using stacked hierarchical recurrent networks; then, innovatively constructing query vectors using temporal features and key-value vectors using motion features, and fusing them using a heterogeneous cross-attention mechanism to form a fusion architecture "dominated by an implicit state modeling network and referenced by an explicit motion modeling network." Based on this, through a series of collaborative designs such as directional decoupling, reference confidence estimation, and long-term feature constraints on short-term features, noise in long-term / short-term motion features is effectively suppressed, enhancing system stability. This method can improve the problems of blurred texture details and attenuation in high-value echo regions in predicted images.

The invention discloses a target detection multi-model fusion method and device, equipment and a storage medium, and relates to the technical field of computer vision. The method comprises the following steps: acquiring detection results of a plurality of different models on the same image; after sorting according to categories and confidence coefficients, clustering the detection frames based on the overlapping degree to form a fusion frame and a corresponding source frame list; and for each fusion frame, counting the number N of different models related to the source frame of the fusion frame, and optimally calculating the confidence coefficient of the fusion frame based on the ratio r of the N to the total model number M. According to the method, the model consensus factor is introduced into confidence calculation, so that the influence of accidental errors of a single model on a fusion result is effectively inhibited, and the precision and robustness of multi-model fusion are remarkably improved.

This invention discloses a lightweight, partially convolutional neural network-optimized method for detecting foreign objects in coal flow, belonging to the field of computer vision and target detection technology. This method aims to address the problems of high computational cost, poor real-time performance, and insufficient adaptability to multi-sized targets in existing coal flow foreign object detection methods. The technical solution includes: constructing a cross-group feature interaction module to enhance feature diversity through channel recombination and permutation; employing dynamic partial convolution, performing depthwise separable convolution only on some channels to significantly reduce computational burden; designing cross-level residual connections, combining lightweight MLP and regularization techniques to improve feature fusion capabilities; and finally, achieving multi-scale feature aggregation through an adaptive feature fusion architecture. This method reduces the model computational cost from 8.1G to 6.3G and the weight file from 6.2MB to 4.7MB while maintaining a detection speed of 1.3ms, significantly improving the real-time performance, accuracy, and environmental adaptability of foreign object detection in coal flow scenarios.

The embodiment of the invention provides a programming assistance and model training method and device based on a code big language model. According to the scheme, the method comprises the steps of firstly, obtaining a current editing file of a user and historical editing information of the user; then, inputting the coding context information into a trained code big language model to obtain a structured prediction result output by the code big language model; the code big language model is a multi-task fusion architecture and is configured to synchronously support a code completion task, a code editing task and a cursor position prediction task based on single reasoning, and the structured prediction result comprises suggested data for at least one task in multiple tasks; and further, according to the structured prediction result, providing an editing suggestion for the user.

The application discloses a kind of based on time knowledge graph reasoning's troubleshooting intelligent guidance method, system and storage medium, it is related to troubleshooting technical field, comprising the following steps: time knowledge graph construction: troubleshooting process is converted into time knowledge graph TKG, and through TKG quadruple under the scene of troubleshooting;Algorithm model architecture design: adopt the fusion architecture of time knowledge graph and multi-time sequence algorithm branch, algorithm input fusion TKG time sequence feature and graph feature;Troubleshooting intelligent guidance system design: the double check mechanism of algorithm reasoning and TKG rule engine is constructed, filters illegal operation, verifies and strengthens prediction result.The application adopts above-mentioned one based on time knowledge graph reasoning's troubleshooting intelligent guidance method, system and storage medium, realize existing troubleshooting process and expert experience structured reuse, troubleshooting time sequence dynamic nature and topological correlation joint modeling, accurately adapt to different troubleshooting scene.

The invention discloses a multi-view fusion signal classification method based on evolutionary neural architecture search, and belongs to the technical field of signal classification. The method comprises the following steps of: firstly, automatically extracting multi-view features from an original wireless signal by utilizing a plurality of deep learning models so as to comprehensively capture time domain, frequency domain and time sequence information of the signal; in order to solve the problem that a multi-view fusion strategy is difficult to design, an evolutionary neural architecture search is adopted to automatically optimize a fusion structure: a fusion network is represented through binary tree coding, a performance predictor is constructed by using a directed graph convolutional network, and the fitness of a candidate structure is quickly evaluated. In the search process, prediction guidance and real evaluation are combined, and the optimal multi-view fusion network is finally determined through selection, intersection and mutation operation iterative evolution. According to the method, the automatic and adaptive search of the fusion architecture is realized, the limitation of manually designing the fusion rule is avoided, the search efficiency is remarkably improved while the classification precision is ensured, and the method is particularly suitable for a signal identification task in a complex electromagnetic environment.

The invention discloses a sequence recommendation data weight optimization system and method based on reinforcement learning, and aims to solve the problem of insufficient model performance caused by unreasonable data weight distribution in sequence recommendation. The system comprises a strategy network, a memory buffer, a performance evaluation module and a weight updating module, the strategy network generates continuous weight adjustment actions, memory buffer storage states (sample feature embedding and current weight), actions and other empirical data through a multi-level fusion Actor architecture, the performance evaluation generates reward signals based on Recall (at) K and NDCG (at) K, and the weight updating adopts a PPO algorithm to achieve stable iteration; the method comprises the following steps: acquiring data, extracting features through a pre-training model, and initializing weights; constructing an MDP to convert weight optimization into a sequence decision; and iteratively generating a weight adjustment action, cutting a function to limit the weight to [0.1, 1.0], training the temporary model by Bernoulli sampling, and updating the strategy network based on rewards. According to the method, data weight dynamic optimization is realized, and the performance and generalization of the recommendation model are improved.

The invention belongs to the technical field of large-scale semi-structured and non-structured data access, and particularly relates to a method and a system for querying semi-structured and non-structured data under a lake-warehouse fusion architecture, and the method comprises the following specific processes: randomly sampling a data subset d on queried data D, extracting semantic embedding corresponding to the samples in the data subset d and a prediction result of the target sensing model M; according to data types in the data subset d, internal relevance of the data types is extracted to generate multi-dimensional coordinates, and discrete coordinates are converted into high-dimensional continuous relevance embedding through multi-resolution hash coding; semantic embedding and relevance embedding are spliced into total embedding, the total embedding serves as lightweight neural network input, and multi-resolution hash codes and the lightweight neural network are jointly optimized to regress a prediction result of the target sensing model; and applying the trained multi-resolution hash codes and the lightweight neural network to queried data to generate proxy scores corresponding to all the data.

The invention belongs to the technical field of natural language processing and computer vision crossing. Aiming at the problems of single semantic hierarchy, visual noise interference, insufficient cross-modal alignment, collaborative reasoning deficiency and the like existing in the existing multi-modal entity linking method, the invention provides a method based on multi-channel collaborative fusion. According to the method, a fusion framework comprising four parallel channels of text semantics, visual perception, cross-modal alignment and collaborative entities is constructed, double alignment of deep and surface semantics is realized through a hierarchical attention mechanism, visual noise is filtered through a dynamic gating mechanism, deep alignment of text and vision is realized through common semantic space mapping, and the visual perception and cross-modal alignment of the collaborative entities is realized. The multi-mention reasoning capability is enhanced through collaborative feature aggregation, and equilibrium training is realized by adopting a multi-channel consistency objective function combining overall loss and independent loss of each channel. The method is suitable for scenes such as knowledge graph construction, intelligent question and answer systems and multimedia content understanding, and the accuracy and robustness of multi-modal entity linking are remarkably improved.

This invention provides a non-technical line loss status assessment method based on multi-scale spatiotemporal feature fusion in low-voltage distribution network environments, relating to the fields of power big data analysis and intelligent operation and maintenance of distribution networks. The invention constructs a fusion architecture based on multi-scale temporal convolutional networks and long short-term memory networks; extracts multi-scale spatiotemporal features of users from instantaneous power consumption fluctuations to periodic load patterns using MSTBlock units; designs a cluster balance constraint mechanism to ensure that sparse but critical non-technical line loss anomaly warning signals are not obscured by massive amounts of normal power consumption data; adaptively selects graph segmentation or spectral clustering integration strategies based on data scale to output clustering labels, which are then mapped to the evolution trajectory of user power consumption behavior. This invention can identify the level of power consumption anomalies from raw load signals with random fluctuations and interference, and calculates the investigation priority by combining transformer area correlation analysis, significantly improving the accuracy and interpretability of unsupervised assessment decisions for non-technical line losses in distribution networks.