557 results about "Linear transform" patented technology

The invention provides a cellular network flow space-time prediction method and system based on a hybrid neural network, and belongs to the technical field of intelligent communication. The method adopts a layered deep neural network architecture, and comprises a data embedding layer, a space-time coding layer, a feature fusion layer and an output layer. The data embedding layer maps a historical traffic sequence, cross-domain external data and metadata into high-dimensional features; the space-time coding layer is used for respectively fusing one-dimensional causal convolution and a Mama neural network to extract multi-scale time features and densely connecting convolution and a multi-head attention mechanism to capture multi-scale space features through time and space modeling branches; the feature fusion layer realizes adaptive weighted fusion of spatial-temporal features, cross-domain features and metadata features by using a gating fusion mechanism; and the output layer performs linear transformation on the fusion features to generate a final prediction result. According to the method, the spatial-temporal dynamic capture of the service traffic is accurate, the prediction curve is highly fit with the true value, and the accurate prediction of the multi-service traffic of the cellular network is realized.

The invention discloses a dynamic feature retrieval generation and management method based on a cross attention mechanism, which belongs to the technical field of natural language processing, and comprises the following steps: step 1, converting knowledge base document fragments into atomic knowledge units, each atomic knowledge unit comprising question and answer pairs, codes as key value pairs, and adding dynamic priority weights; self-attention query is replaced with a double-channel query structure, one channel is used for cross attention, a cross attention query vector is generated through linear transformation, and key value pairs of atomic knowledge units are used; and step 3, training a cross attention adapter, freezing the weight of the language model, optimizing parameters of the adapter, and dynamically adjusting a loss function by using pre-judgment parameters based on input sequence context complexity. By means of the method, deep correlation description of user input and knowledge fragments can be achieved, semantic ambiguity is eliminated, knowledge injection and context information are balanced, distortion is avoided, and the strict requirement for semantic precision in the professional field is met.

The invention discloses an acute pancreatitis complication assessment method based on artificial intelligence, and belongs to the technical field of medical information, and the method specifically comprises the steps: firstly, obtaining a clinical data set containing real-time clinical indexes and basic disease historical records; performing time sequence analysis on the historical records of the basic diseases, extracting long-term influence characteristics of the historical records of the basic diseases and generating basic disease influence factors; carrying out standardization processing on the real-time clinical indexes and the basic disease influence factors, and generating a comprehensive feature vector after calculating association weights among features; inputting the vector into a trained artificial intelligence evaluation model, and outputting a complication risk probability through layer-by-layer nonlinear transformation; and finally, mapping the risk probability value into a risk level, and integrating patient information to generate a structured risk assessment report. And an acute pancreatitis complication evaluation scheme integrating the acute stage index and the chronic basic disease influence is established.

Disclosed in the present invention is a deep learning-based power distribution network load prediction method, comprising: acquiring regional load data and renewable energy power generation data to form a data set, and preprocessing the data set to obtain a first data set; using a convolutional neural network to extract a time series feature in the first data set, and converting the time series feature into a data form of a deep learning model by means of an embedding layer to obtain one-dimensional time series data; performing fast Fourier transform on the one-dimensional time series data to obtain a frequency curve, extracting amplitude values on the frequency curve to calculate corresponding periods, and selecting the corresponding periods to slice the one-dimensional time series data and form same into two-dimensional matrixes; using a two-dimensional convolutional network to perform feature extraction, reshaping the two-dimensional matrixes that have undergone feature extraction into one-dimensional arrays, and performing adaptive fusion on the one-dimensional arrays to obtain a first time series feature; and inputting the first time series feature into a fully connected layer for weight calculation and linear transformation to obtain a power distribution network load prediction result, thereby improving the stability and reliability of electric power supply.

The invention discloses a ship route planning method and system in a polar region sea ice environment and a program product. The ship route planning method comprises the steps of preprocessing satellite remote sensing radiation intensity data, performing image reconstruction of inter-ice water channel detection through an improved Vision Transform model, planning and visualizing a ship route and the like. The deep features of the image blocks are extracted through a Transformer encoder which comprises a multi-head self-attention mechanism and is processed by a multi-layer perceptron; and the Transform encoder firstly performs layer normalization on the input, and then performs multi-head attention mechanism and multiple linear transformation and regularization processing. And realizing path planning in an inter-ice water channel environment by fusing a dynamic planning cost function through a fast marching method. Compared with a traditional algorithm, the method has the advantages that the calculation precision is higher in the same scene, the calculation efficiency is greatly improved, and the planning result obtained after multiple costs are comprehensively considered is given instead of pursuing the shortest distance.

The invention discloses a large model dynamic batch processing method based on sequence splicing. The method comprises the following steps: receiving input sequences of a plurality of users in a batch; converting the input sequence into a corresponding token sequence; carrying out heterogeneous splicing on all token sequences along a sequence length dimension to form a unified joint token; the spliced joint tokens pass through a normalization layer, standardization operation is executed on the joint tokens, and data distribution is unified; carrying out linear projection on the joint token through a shared linear transformation layer to generate a joint query vector, a joint key vector and a joint value vector, and splitting the joint query vector, the joint key vector and the joint value vector into sub-vector groups corresponding to each user; executing multi-head attention calculation to obtain attention output of the user; and carrying out linear transformation on the attention output, and inputting a transformed result into a shared MLP to carry out nonlinear feature extraction and enhancement so as to obtain a final output corresponding to each user. According to the method, the problems of efficiency bottleneck and resource consumption when a large model processes mass data are effectively solved.

The invention provides a coupling space-spectrum conversion operator transformer multi-physics field distribution prediction method and system, and the method comprises the steps: mapping the initial condition distribution of multi-physics field distribution to a potential space of a unified dimension through independent linear transformation, generating an initial potential feature vector, and carrying out the prediction of the initial potential feature vector; performing spatial expansion on the cross-field coupling state of the multi-physical field distribution in combination with the GRU, splicing the obtained current global cross-field coupling state and the current potential feature input, and obtaining the potential feature input of the next layer of the GRU through a space-spectrum conversion operator layer; and inputting the potential features of all the current physical field distribution and the global cross-field coupling state into a GRU unit, updating the cross-field coupling state including cross-physical field interaction, minimizing weighted error optimization parameters of a physical field distribution prediction result, inversely mapping the reconstructed final potential features to a target space, and generating the prediction result of the physical field distribution. The method has higher generalization ability, the modeling precision is improved, and the calculation complexity is reduced.

The invention discloses a CT (Computed Tomography) image pulmonary embolism segmentation and classification method combined with quality evaluation, which relates to the technical field of image processing, and comprises the following steps: inputting a 256 * 256 pulmonary embolism CT image and a quality score thereof into a quality score guide encoder, expanding a quality score dimension through linear transformation, and carrying out point product fusion with a feature map extracted by ResNet34 layer by layer to obtain a final product; generating multi-scale coding features; performing wavelet domain decomposition and reconstruction on the coding features through a wavelet transform fusion jump link module, and optimizing feature transmission; a multi-scale cross enhanced decoder is adopted to carry out multi-scale deconvolution fusion on the features, a segmentation result is output in combination with an efficient channel attention mechanism, meanwhile, pulmonary embolism existence judgment is output through a classification head, and the method provides powerful support for early diagnosis of pulmonary embolism, development of an image auxiliary diagnosis system and clinical application, and has good application prospects. Wide application prospects and profound social significance are realized.

Disclosed are apparatuses, systems, and techniques that may use machine learning for implementing generative text-to-speech models. The techniques include identifying a mapping of speech characteristics (SC) on a target distribution of a latent variable using a non-linear transformation for at least a subset of the SC. Parameters of the non-linear transformation are determined using a neural network that approximates a statistics of the SC with a statistics predicted for the SC based on the identified mapping and the target distribution of the latent variable.

The invention relates to a virtual power plant resource flexible aggregation method and device based on linear programming projection, and the method comprises the steps: constructing a linear state equation corresponding to each to-be-aggregated resource according to the resource characteristic data of each to-be-aggregated resource, and carrying out the linear transformation; aggregating the energy use flexible feasible region of each to-be-aggregated resource, and solving a coordinate projection problem to obtain a preliminary aggregated energy use feasible region; constructing a network security constraint and embedding a solution process of the preliminary aggregation energy use feasible region to obtain an aggregation energy use feasible region considering the network security constraint; and constructing a multi-parameter planning model, and fusing an aggregation cost function into a solving process of an aggregation energy use feasible region considering network security constraints to obtain a virtual power plant resource flexibility aggregation result. Therefore, the problems that in the prior art, energy use feasible regions and adjustment cost of massive heterogeneous resources cannot be aggregated at the same time, network security constraints cannot be embedded, and high-dimensional aggregation errors are large are solved, and higher calculation efficiency and compatibility are achieved.

The invention relates to the technical field of slope monitoring, in particular to a mine slope monitoring method based on unmanned aerial vehicle inspection. The method comprises the steps that a slope image is collected and preprocessed, a disturbance excitation expansion mechanism is introduced, time-space domain nonlinear transformation is carried out on the preprocessed slope image, disturbance excitation response is calculated, and a slope image after expansion transformation is obtained; introducing a residual energy diffusion modeling mechanism based on the slope image after expansion transformation, and constructing a residual energy tensor; based on the residual energy tensor, an entropy density splitting mechanism is introduced, a stack product space projection function is constructed, and nonlinear landslide information response intensity is obtained; and based on the nonlinear landslide information response intensity, constructing a stack product type discriminator, and carrying out landslide risk category discrimination to obtain a monitoring result. The problems that a traditional mine slope monitoring method is not accurate in processing of image data obtained by an unmanned aerial vehicle, so that the positioning and judgment accuracy of a landslide risk area is low, and self-adaptability and sensitivity are insufficient are solved.

The invention relates to the technical field of photovoltaic power generation, in particular to an ultra-short-term offshore photovoltaic power prediction method and system considering multi-energy coupling influence, and the method comprises the steps: employing a bidirectional long-short-term memory network to extract the time sequence dynamic characteristics of wave, photovoltaic and fan power; a space-time collaborative attention mechanism is introduced, and feature enhancement is realized through time dependence and spatial relevance in multi-head self-attention capture features and residual connection and layer normalization technologies; generating a modal self-adaptive gating weight through a full connection layer, and carrying out linear transformation on the enhanced features; a cross-modal multi-head attention mechanism is introduced, and coupling rules among wave-photovoltaic-fan three modals are excavated and fused; mapping context features of the fused features to a prediction time domain by using a long short-term memory network, and designing an adaptive output gate structure; and a multi-head self-attention mechanism is applied to capture time step association of a prediction time domain, and an ultra-short-term photovoltaic power prediction value is output through full-connection layer regression.

The invention discloses an image fusion method and system based on multi-semantic guidance and mixed experts. The problems that in the prior art, robustness is insufficient, and visual fidelity and semantic integrity are difficult to balance are effectively solved. According to the method, infrared and visible light images to be fused and task identifiers are obtained, and firstly, a CLIP network is utilized to extract high-level semantic features as guide vectors; and then inputting the image and the guide vector into a pre-trained hybrid expert (MoE) fusion network. According to the network, intermediate features are extracted through an encoder, a gating network dynamically activates part of expert subnets according to task identifiers and semantic vectors and calculates routing weights, adaptive nonlinear transformation and weighted fusion are carried out on the features, and finally a high-quality fusion image is reconstructed through a decoder. The method can adapt to different task requirements, and the calculation efficiency is remarkably improved while the image fusion quality and the semantic consistency are improved.

The invention discloses a method for effectively simulating general quantum computing based on a classical circuit network, and relates to the technical field of quantum computing simulation, and the method comprises the following specific steps: converting data into a voltage signal through input voltage coding, constructing a single-bit gate and a two-bit gate by using an operational amplifier and an analog multiplier, and constructing a single-bit gate and a two-bit gate; according to the invention, the general quantum computation is simulated by a classical circuit network, an n-bit unitary matrix is decomposed and reconstructed by means of input voltage coding quantum state and voucher / two-bit gate modular design, and the high-efficiency simulation of general quantum computation is realized. Quantum gate linear transformation is reproduced, quantum calculation universality is achieved, voltage signals and data are converted through a reflection formula, signal distortion is avoided, multi-level gate circuit cascading is supported, and a flexible and stable platform is provided for quantum algorithm verification and the like.

The invention belongs to the technical field of data processing. The invention discloses a matrix multiplication calculation task processing system, method and device, a storage medium, a program product and a chip, and the system comprises a storage module which is used for storing a preset weight matrix in advance; the linear transformation control module comprises a finite-state machine used for matrix multiplication operation scheduling and is used for generating an input request signal so that the matrix multiplication calculation task processing system can receive an input matrix; the operation module is used for executing matrix multiplication operation on the input matrix and a preset weight matrix to obtain a matrix multiplication operation result and generate an output matrix; and the data flow control module is used for controlling a plurality of groups of data flows formed by the input data in the input matrix according to rows to sequentially enter the operation module according to the input request signal until all the data in the input matrix execute and complete the matrix multiplication operation. The problems that in the prior art, a matrix multiplication accelerator lacks flexibility in a linear layer implementation task of artificial intelligence hardware, the data migration cost is high, and the energy efficiency ratio is low are solved, and the method is suitable for self-defined extension based on an RISC-V instruction set architecture, can be used as a coprocessor of an RISC-V processor, and can be used as a coprocessor of the RISC-V processor. And linear layer calculation in the neural network is accelerated, especially in an artificial intelligence task, the calculation efficiency can be remarkably improved, the power consumption can be reduced, and the real-time processing requirement can be met.

The invention relates to the technical field of logistics automation and intelligent scheduling, and discloses a multi-objective optimization logistics distribution path dynamic planning method, which comprises the following steps: S1, mapping a logistics distribution area and vehicles, clients and constraint entities in the area to a continuous coordinate space; s2, in the continuous coordinate space, optimizing cost, time, priority and constraints based on the mapped clients and constraint entities, and respectively constructing corresponding decision force fields; s3, based on a preset strategy adjustment matrix, carrying out linear transformation and combination on the plurality of constructed decision force fields so as to synthesize total decision force acting on the space; and S4, taking the synthesized total decision-making force as an external driving force. According to the method, through the strategy adjustment matrix and the path dynamic evolution equation, the high-level strategy and the path physical characteristics are deeply coupled, and refined, smooth and efficient regulation and control of multi-target dynamic path planning are realized.

The invention discloses a large language model machine forgetting algorithm based on representation spatial offset. The algorithm comprises the following steps: constructing a forgetting set and a retention set, executing causal tracking on a large model by using a knowledge exploration data set, and identifying a feedforward neural network having a significant contribution to correct prediction of the model to determine a forgetting layer; estimating an input feature space by using the reserved set approximation, and constructing a null-space projection matrix based on the feature space; inputting a forgetting set and a retention set, obtaining representation output of the training model and the original model in the last forgetting layer, and exporting a parameter updating gradient through a loss function; and carrying out null-space projection transformation on the updated gradient and then updating the last linear transformation matrix of the feedforward neural network in all forgetting layers. According to the method, the knowledge storage characteristics of the feedforward neural network in the large model are utilized, knowledge removal is guided in the representation space, interference on non-target knowledge is effectively suppressed while accurate forgetting is achieved, and the effectiveness and controllability of the forgetting process of a large model machine are remarkably improved.

The invention relates to a multi-modal sentiment analysis method based on text enhancement and modal completion perception fusion, and belongs to the field of natural language processing. The method comprises the following steps: 1, extracting text semantic features by utilizing a pre-training language model, and performing linear transformation on non-text features to form unified multi-modal input representation; 2, injecting a cross-modal enhancement module into the pre-training language model, fusing non-text information by taking a text as a core, and performing multi-modal input representation; 3, introducing a modal completion module, generating a missing modal completion representation through reconstruction loss and random modal discarding, and suppressing noise features in combination with a convolution gating structure; and guiding full-connection network learning modal weight to realize joint optimization of sentiment classification and regression. According to the method provided by the invention, on the premise of keeping the dominance of the text, through cross-modal interaction and dynamic weighting and in combination with a modal completion mechanism, the model has relatively strong emotion recognition capability.

The invention relates to an efficient concrete crack detection method, device and system. The method comprises the following steps: acquiring an image of a to-be-detected concrete beam; generating a hierarchical grid backbone network of redundant features based on linear transformation, and performing feature extraction processing on the to-be-detected concrete beam image to obtain multi-scale feature information; performing feature fusion processing on the multi-scale feature information based on a geometric perception dynamic feature pyramid network to obtain fused feature information; based on task alignment of the dynamic detection head and classification and positioning processing of the fusion feature information, the target crack information corresponding to the to-be-detected concrete beam image is obtained, the concrete crack detection accuracy is improved, the weak texture feature extraction and retention capability is significantly enhanced, the leak detection risk of fine cracks is effectively reduced, and the detection efficiency is improved. The anti-interference capability under a complex background is enhanced, the false detection rate is reduced, mismatching of classification and positioning tasks is avoided, and the crack positioning precision is improved.

The invention provides a segmentation threshold self-generation alumen ustum detection method and system based on a convolutional neural network, and the method comprises the following steps: constructing an algorithm model of a specific convolutional structure, dividing the operations of convolution, ReLU activation, pooling and the like into one layer, sequentially passing through three layers, and finally obtaining four thresholds after adaptive average pooling, flattening and linear transformation, an optimal binarization threshold value is screened out; high-speed cameras are deployed at multiple angles of the sewage treatment pool to collect multidirectional alumen ustum images, the algorithm carried by the upper computer is input, binarization processing is conducted on the input images through the optimal threshold value output by the algorithm, and the binarized alumen ustum images are output. According to the method, the alumen ustum image can be efficiently processed, an accurate binary image is provided for alumen ustum state analysis in the sewage treatment process, and the requirements of image recognition and processing in a sewage treatment scene are met.

The invention discloses a multi-robot affine formation path optimization method driven by a virtual navigator, and the method comprises the steps: firstly setting a point set composed of the virtual navigator and followers, enabling the virtual navigator to form a convex hull and meet an affine localization condition, so as to guarantee that the followers are always located in the convex hull; generating a collision-free reference trajectory of the mass center of the virtual navigator based on a fast random tree algorithm; constructing a nonlinear constraint optimization problem NCO which contains linear transformation constraint and meets smoothness, obstacle avoidance, speed and angular speed constraint; and establishing a communication graph of the virtual navigator and the follower, calculating balance stress, designing a smoother follower distributed differential speed control law, and realizing formation path optimization and obstacle avoidance control. Compared with an existing method, the formation path generation method has the advantages that the smooth and stable formation path fitting the reference trajectory can be generated while obstacle avoidance and physical feasibility are guaranteed, and the cooperative motion capability and the control performance of a multi-robot system in a complex environment are improved.

The invention provides an image registration method and an image registration device, which are applied to the technical field of image processing. Comprising the following steps: acquiring a target reference image and a first target to-be-registered image of a to-be-detected area containing FOD; constructing an optimal affine transformation model according to the target reference image and the N preset pitch angles, and further performing linear transformation on the first target to-be-registered image to obtain a second target to-be-registered image; according to a Gaussian difference pyramid of a second target to-be-registered image, screening an extreme point of which the principal curvature does not exceed a preset principal curvature threshold as a first feature point, and further screening a first feature point of which the re-projection error and the sampling probability meet a preset condition as a second feature point; and solving a homogeneous linear equation set of the pixel point pairs matched with the second feature points by adopting an SVD method to obtain a transformation homography matrix, and then registering the first target to-be-registered image to obtain a target image. The contradictory closed loop of low efficiency-insufficient precision is broken through, and the aviation operation safety is guaranteed.

The invention relates to the technical field of infrared and visible light image fusion, in particular to an infrared and visible light image fusion-oriented lightweight network design method, which comprises the following steps of: introducing an SE module, channel shuffling and residual connection on the basis of deep convolution and point-by-point convolution, and combining linear transformation and channel dimension splicing, so as to obtain a lightweight network; jointly constructing a lightweight convolution module; respectively constructing an encoder and a decoder based on the lightweight convolution module, and forming an auto-encoder network based on the encoder and the decoder; iteratively training the auto-encoder network, and updating parameters of an encoder and a decoder by taking a total loss function in a segmented form based on the number of training rounds, pixel loss and SSIM loss as guidance to obtain the trained auto-encoder network; and inserting a fusion module into the trained auto-encoder network to form a fusion model. According to the method, the parameter quantity and the calculation quantity of the fusion model can be reduced and the image fusion speed can be improved under the condition of ensuring the image fusion quality.

The invention relates to a scratch detection method, device and equipment for a transparent film and a medium, and the method comprises the steps: firstly carrying out image multiplication processing and dynamic range mapping on a transparent film gray level image, and obtaining a target gray level image through threshold segmentation; calculating a pixel gradient magnitude based on the image, synthesizing an edge gradient image, and generating a first salient image through Gaussian filtering and gray linear transformation; meanwhile, logarithmic transformation is carried out on the original target image to obtain a second salient image; secondly, respectively calculating gray average values of the two salient images, determining a self-adaptive segmentation threshold by combining a preset threshold, extracting a defect region through double-image threshold segmentation, and obtaining an intersection to obtain an initial scratch region; and finally, screening according to a preset area condition to obtain a final scratch area. According to the method, through the multi-feature fusion and self-adaptive threshold technology, the problems of low contrast, uneven illumination and the like in scratch detection of the transparent film are effectively solved, the detection efficiency and accuracy are remarkably improved, and the method has high engineering application value.

The invention relates to an agricultural machine safety control method and device and a computer readable storage medium. The method comprises the following steps: synchronously acquiring agricultural machine operation data through multiple sensors, extracting terrain, earth surface and motion features, constructing feature vectors, and identifying a geographic position category label of an agricultural machine by using a pre-training model; loading a corresponding exclusive safety control rule according to the label, wherein the rule comprises a specific sensor combination, a risk mapping matrix and a basic operation parameter; screening real-time data according to a rule to calculate a risk state vector, and linearly transforming the risk state vector into a control instruction correction through a mapping matrix; and finally, a control instruction is generated in combination with the basic parameters and is issued to an actuator, so that safety control based on environment category accurate perception and dynamic rule self-adaption is realized, and the operation safety, the environment adaptability and the operation efficiency of the agricultural machinery under complex terrains are effectively improved.

The invention relates to the technical field of unmanned aerial vehicle communication encryption, and discloses a data encryption transmission system based on unmanned aerial vehicle communication. The system comprises an airspace situation awareness module, a channel feature coding module, a hierarchical encryption engine module, a key dynamic derivation module and an anti-interference relay module. The airspace situation awareness module collects electromagnetic spectrum characteristics through a multispectral sensor array to generate a dynamic spectrum fingerprint spectrum; a channel feature coding module extracts multi-dimensional channel parameters, and generates a matched chaotic mapping sequence in combination with a quantum random number generator; the hierarchical encryption engine module segments the original data stream and executes differential round hierarchical nonlinear transformation; the key dynamic derivation module generates a composite session key drifting along with time according to the flight path coordinates; the anti-interference relay module is embedded with a frequency spectrum fingerprint watermark, and a multi-hop relay link is established through adaptive beam forming. The system can adapt to the dynamic flight scene of the unmanned aerial vehicle, and guarantees the safety and stability of communication data transmission.

The invention relates to a multi-modal power data retrieval method and device and a medium, and belongs to the technical field of data retrieval, and the method comprises the following steps: obtaining and preprocessing power data, inputting a text and an image into a multi-layer perceptron to obtain an initial feature vector, and outputting a text feature vector and an image feature vector through a routing capsule algorithm; mapping to a public space by adopting linear transformation to obtain a projection vector; and inputting the projection vector into the additive attention model, and outputting a fusion feature vector. And constructing a hash function, calculating an electric power data index based on the fused feature vector, and storing data with the same index into a corresponding hash table bucket. And obtaining a user query, calculating a query fusion feature vector and an index, calculating the similarity of the query fusion feature vector and the index with each power data fusion feature vector in the corresponding bucket, and returning R power data with the highest similarity score as a retrieval result. According to the method, multi-modal information is integrated, the importance is adjusted by using an attention mechanism, and the retrieval effect is effectively improved.

The invention relates to a safety monitoring method for an irrigation and drainage project operation state and an intelligent system, and belongs to the technical field of artificial intelligence. The method comprises the following steps: collecting and marking irrigation and drainage project operation monitoring data through a sensor, and constructing a training data set; completing data normalization by combining quantile and median robust scaling with adaptive nonlinear transformation, and mining and screening high-order interaction features by combining a mutual information theory and a gradient boosting decision tree; constructing a deep classification network fusing physical prior and adaptive feature interaction, introducing physical constraint and multi-scale feature fusion, and optimizing a model through adaptive marginal classification loss and physical feature manifold alignment loss; real-time data is preprocessed and then input into the model, and operation state grade classification and graded alarm are achieved. According to the method, data noise can be inhibited, a multi-index coupling relationship can be mined, the interpretability and robustness of the model can be improved by integrating a physical rule, irrigation and drainage project abnormity can be accurately identified and early warned, and the method is suitable for intelligent safety monitoring of an irrigation area.

The invention discloses a scene perception non-intrusive load monitoring method based on a large language model. The method comprises the following steps: firstly, acquiring a total load sequence, segmenting a total electrical load sequence into a plurality of total load segments, and performing vector embedding to obtain a total load segment embedding vector; thirdly, constructing a comprehensive cue word and carrying out vector embedding to obtain a comprehensive cue word embedding vector; then, splicing the total load fragment embedding vector and the comprehensive cue word embedding vector to obtain a total load and comprehensive cue word fusion embedding vector, inputting the total load and comprehensive cue word fusion embedding vector into a trained large language model, and outputting a load estimation vector of each target device; mapping the load estimation vector of each target device through a linear transformation layer to obtain a load estimation value of each target device; and finally, correcting the load estimation value by using the load probability distribution of the target equipment in the current operation state. The power consumption scene information and priori knowledge are fully utilized, and the load monitoring accuracy of the model under the complex power consumption condition is remarkably improved.

The present disclosure relates to the technical field of artificial intelligence algorithm hardware acceleration, and in particular to a Transformer accelerator architecture based on monolithic three-dimensional integration, comprising a silicon-based logic circuit, used for controlling a current flowing direction of Transformer for logic operation; an RRAM-CIM array, used for executing a linear transformation operation; and a CFET 2T0C-CIM array, used for executing a matrix multiplication operation.