829 results about "Activation function" patented technology

The invention relates to a wellbore multiphase flow model numerical solution and gas-liquid distribution state inversion method and system, and belongs to the technical field of petroleum engineering, and the method comprises the steps: 1, constructing and training a physical information neural network for drilling wellbore multiphase flow dynamic simulation and overflow gas distribution state inversion; determining input and output of the physical information neural network; determining a loss function of the physical information neural network; training a physical information neural network; 2, designing an adaptive optimization algorithm, optimizing the final solution precision and convergence speed of the physical information neural network, and obtaining an adaptive physical information neural network; designing an adaptive activation function; designing a self-adaptive sampling mechanism based on residual errors; 3, based on the self-adaptive physical information neural network, numerical solution and gas-liquid distribution state inversion of the shaft multiphase flow model are achieved. According to the method, the problem that a traditional numerical method usually needs high-precision grid division and a large number of computing resources is effectively solved.

Some embodiments provide a neural network inference circuit for executing a neural network that includes computation nodes. Each respective computation node of a set of the computation nodes includes (i) a respective linear function that includes a respective dot product of input values for the computation node and weight values for the computation node and (ii) a respective non-linear activation function. The neural network inference circuit includes a set of dot product circuits to compute the dot product for a computation node and a post-processing circuit to compute (i) a result of the linear function for the computation node based on the dot product for the computation node and (ii) an output for the computation node by applying a piecewise linear function to the result of the linear function for the computation node to apply the non-linear activation function for the computation node.

The invention provides a robot arm motion control strategy network training method and device based on deep reinforcement learning, and relates to the technical field of sensors and robots. The method comprises the following steps: in a deep reinforcement learning training environment, obtaining a jacobian sub-matrix corresponding to a robot arm; calculating an operability index under the current attitude based on a product between the Jacobi sub-matrix and a transpose matrix of the Jacobi sub-matrix; inputting the operability index into the activation function to obtain a penalty factor; and scaling the difference value between the current action output by the policy network and the action at the last moment based on the penalty factor to obtain a penalty term, and using the penalty term as a part of a reward function for training the policy network. According to the method, the self-adaptive punishment mechanism based on the operability is introduced, so that the strategy network can recognize and actively avoid the singular postures in the training process, and the stability of robot arm motion control is improved.

The invention belongs to the field of bioinformatics, and relates to an improved integrated deep learning cell communication ligand-receptor interaction prediction method. The method comprises the following steps: firstly, carrying out extraction and dimensionality reduction on biological sequence features of a ligand and a receptor, and constructing multi-modal feature input; secondly, constructing an improved deep neural network branch, introducing a batch normalization layer and a Leaky ReLU activation function, solving the problems of gradient disappearance and neuronal necrosis, and improving regularization strength to prevent overfitting; meanwhile, an enhanced heterogeneous graph auto-encoder branch is constructed, the graph embedding dimension is remarkably expanded to improve the feature capacity, and full convergence of the model is ensured by increasing training rounds; thirdly, fusing the improved deep network with the prediction probability of a heterogeneous graph auto-encoder by adopting a weighted integration strategy; and finally, outputting a potential interaction relationship based on the fusion probability. By optimizing the architecture and the strategy, the prediction accuracy and robustness are remarkably improved, and a reliable tool is provided for analyzing a complex cell communication network.

The invention relates to the technical field of robot control, and discloses a mechanical arm trajectory planning control method and system based on a BAFARNN model. The method comprises the steps that a mechanical arm kinematics model is established, and a trajectory tracking problem is converted into a time-varying equation; designing a bounded adaptive function to activate a recurrent neural network model, defining an error function and constructing a dynamic equation; designing a piecewise adaptive coefficient function, and dynamically adjusting the gain according to an error norm and time; setting a Lissajous curve as an expected trajectory, and initializing a simulation environment; the joint speed is solved in real time through an ODE numerical method, and the mechanical arm is driven to move; actual motion data is collected and compared with an instruction, and closed-loop feedback control is triggered when the actual motion data exceed a threshold value. According to the method, rapid convergence is achieved through the piecewise adaptive coefficient function, the bounded activation function and the negative feedback mechanism are adopted to suppress noise, and high-precision and real-time trajectory tracking of the mechanical arm in the dynamic environment is achieved.

The invention discloses a vehicle collision detection, description and early warning system and method based on a driving video, and belongs to the technical field of artificial intelligence and intelligent traffic safety. According to the system, on the basis of a vision-language model, vehicle-mounted videos such as an automobile data recorder are automatically analyzed, and detection, severity grading, natural language description generation and real-time early warning of vehicle collision events are achieved. The system extracts high-dimensional semantic features of video frames through a CLIP model, focuses key information through an attention weighting network, inputs the key information into a deep classification network composed of a plurality of full connection layers, a batch normalization layer and an activation function, and outputs a multi-level accident severity classification result. Integrating target detection and environment information, and generating a structured accident description text by using a fine tuning BART model; a sliding window and time sequence modeling mechanism is adopted, and recognition and early warning of the pre-collision state are achieved. The method effectively solves the technical problems of lack of semantic understanding, inaccurate severity judgment, incapability of early warning and the like of a traditional method, has high accuracy, high interpretability and real-time response capability, and is suitable for intelligent driving assistance and traffic safety monitoring scenes.

The invention discloses a cascade system online monitoring and prediction method based on a sparse self-attention mechanism, and the method comprises the steps: S1, obtaining and preprocessing multi-working-condition data of a cascade system, and constructing a time series data set; s2, performing embedded conversion and position coding on the data, and capturing sequence position information; s3, pre-fusion of adjacent time step information is realized through one-dimensional convolution; s4, a multi-head sparse attention mechanism is introduced, and a ReLU2 activation function is adopted to replace softmax so as to reduce calculation overhead; s5, completing information fusion through one-dimensional convolution, ELU activation and maximum pooling; s6, constructing an encoder containing a multi-head sparse attention mechanism; s7, designing an autoregressive decoder, and combining self-attention with cross attention; s8, adopting a HuberLoss loss function to train the model; and S9, carrying out reverse normalization on the model output to obtain a final prediction value. According to the invention, by optimizing the Transform architecture, 60 s effective prediction of the key parameters of the cascade system is realized, the prediction error is significantly reduced, and the intelligent early warning capability and the operation stability of the system are improved.

The invention relates to a landslide identification method based on a channel and space mixed attention mechanism. The method comprises the following steps: acquiring remote sensing image data, and constructing a landslide identification model comprising an encoder, an improved jump connection and feature fusion module, a mixed attention module, a decoder and an output layer. The encoder performs layer-by-layer convolution down-sampling on the remote sensing image and outputs features of each layer; the improved module carries out convolution normalization, splicing and weighted fusion on the features to obtain fusion features; in the fusion stage, a mixed attention module is embedded, EMA generates channel weights and enhances features, PPA generates space weights and enhances features, and the EMA and the PPA are spliced and subjected to convolution activation to obtain final enhanced features. After receiving, the decoder performs up-sampling fusion, and an output layer obtains a landslide segmentation result through an activation function; and finally, the model is trained by using a preset loss function, landslide identification is realized, a landslide area can be accurately extracted, and the problem of class imbalance is relieved.

The invention relates to a video snapshot compression imaging reconstruction method and system. The method comprises the following steps: inputting a video frame sequence and a time-varying mask set thereof into a measurement model to obtain initial estimation; constructing a reconstruction network which comprises a feature extraction module, a gating residual network module and a video reconstruction module; the feature extraction module comprises two three-dimensional convolution layers, each three-dimensional convolution layer is connected with an activation function, and the feature extraction module extracts initial features from the initial estimation; inputting the initial features into a gating residual network module, and outputting reconstruction information features; and the video reconstruction module fuses the reconstruction information features, and performs up-sampling and detail refining to reconstruct a video sequence. According to the method, on the premise that parameters and computing power are hardly increased, ghosting and flickering are effectively restrained, the stability of long-time reconstruction is improved, and an effective scheme is provided for SCI reconstruction with the high compression ratio, the super-definition resolution ratio and the long sequence.

The invention discloses a three-dimensional point cloud geometric information compression method based on implicit neural representation, and the method comprises the steps: constructing a trunk structure of an implicit neural network through a plurality of sine representation network layers which are connected in series, and introducing a variable-scale position coding mechanism on this basis, the method enables a network to obtain higher geometric reduction precision while keeping a compression ratio, and comprises the following steps: (1) inputting space coordinates of divided voxels into a position coding module with adjustable scale parameters; (2) feeding a coding result into an implicit neural network constructed by a network layer based on sine representation, and outputting the occupancy probability of the voxel through an activation function; (3) in a training stage, the model continuously optimizes parameters, so that the output probability distribution is highly consistent with a real occupied label; (4) after model training is completed, a method of combining an AdaRound second-order quantization optimization strategy and quantization perception training is introduced, network weight is finely adjusted, and quantization errors are reduced; (5) in a reasoning stage, judging whether the voxel is occupied or not according to a preset threshold value, and when the prediction probability exceeds the threshold value, regarding the voxel as occupied; and (6) all voxels judged to be occupied are aggregated, and reconstruction of the geometric structure of the point cloud is completed.

The invention belongs to the technical field of image data processing, and particularly relates to a super-resolution remote sensing image reconstruction method, system and device based on frequency domain enhancement, and the method comprises the steps: S1, extracting the shallow features of a low-resolution remote sensing image; s2, inputting the shallow features into a plurality of cascaded frequencies for interactive processing, performing double-branch processing on the input features, performing inverse transformation after radial weighting on different frequency components in a frequency domain to obtain first features, and obtaining second features through depth separable convolution, an activation function, a selection scanning module and layer normalization; fusing the two branch features according to the weight, performing jump connection with the original features, performing enhancement through a feedforward network, performing repeated execution for a set number of times, and performing convolution and residual connection to obtain deep features; and S3, fusing the deep features and the shallow features, and outputting a super-resolution image through convolution and pixel rearrangement up-sampling. According to the method, texture details and edge contours of the remote sensing image can be more accurately reconstructed while the structural consistency is kept.

The invention relates to the technical field of wind power, and provides a short-term wind power prediction method and system, and the method comprises the steps: obtaining the power data and meteorological data of a wind power station at a historical moment; the method comprises the following steps of: after normalizing wind power plant data, filling by adopting a plurality of filling modes, screening a plurality of features with the strongest correlation with power data, and sorting based on the screened features and the filled data to obtain input data; on the basis of the input data, through a KAN and TCN cascade prediction model, power at a future moment is obtained; wherein the KAN uses a linear combination of a B spline and a SiLU function as an activation function, and L1 regularization is carried out on the weight. And wind power reliability prediction is realized.

The invention discloses an SAR ship image detection method based on an improved YOLOv11 model, an improved C2PSA module C2DyMoETAttn is introduced, the core innovation point is that a PSABlock module is replaced by a DyMoETAttnBlock module, the DyMoETAttnBlock module fuses a Dynamic Tanh activation function, a Mona module, a TSSA attention mechanism and a frequency domain enhancement feedforward network (EDFFN), multi-dimensional modeling and robust enhancement of features are realized, and the detection accuracy is improved. The feature expression capability and the noise suppression performance under the background of small targets and complex sea clutters are effectively improved; in the deep feature fusion stage, a C3k2 module of YOLOv11 is optimized, an ScConv structure is introduced, adaptive fusion of space and channel features is realized through a joint reweighting mechanism of SRU and CRU, and the multi-scale target discrimination capability and feature selectivity are enhanced; on the bounding box regression layer, a Focaler-MPDIOU loss function is provided, and a Focaler-IoU sample difficulty adaptive mechanism is combined with MPDIOU positioning matching constraint, so that the learning ability of the model for small targets and shielded targets is enhanced, and the bounding box positioning precision and convergence stability are improved.

The invention discloses a high-dimension state space dimension reduction method and system for large-scale electric vehicle charging. The method comprises the following steps: constructing input data for electric vehicle charging control, wherein the input data comprises vehicle information Xi and scheduling information Dt; for Xi, adopting MLP as a vehicle-by-vehicle encoder to realize local nonlinear mapping of vehicle-level features; inputting the same group of MLP weights into all vehicle states in sequence, and completing dimensionality reduction coding through multilayer affine transformation and a nonlinear activation function; after the vehicle-by-vehicle sharing encoder completes local feature mapping, a low-dimensional representation set of all accessed vehicles at the moment t is obtained; and splicing and mapping the aggregation result and the global scheduling information into a compression vector. According to the scheme, through a layered structure of local coding and global aggregation, key difference information between vehicles is effectively reserved, the problem of state explosion is remarkably relieved, and compact input representation with sufficient information is provided for subsequent reinforcement learning strategy optimization.

There is provided a volume preserving trainable artificial neural network and a system and a method for building a volume preserving trainable artificial neural network. In an aspect, an artificial neural network including: an input layer to receive input data; one or more sequentially connected hidden layers, the first hidden layer connected to the input layer, to perform operations on the input data, each hidden layer including: one or more volume-preserving rotation sublayers; one or more volume-preserving permutation sublayers; one or more volume-preserving diagonal sublayers; and an activation sublayer; and a downsizing output layer connected to the activation sublayer of the last hidden layer. In some cases, the activation sublayer includes a grouped activation function acting on a grouping of input variables to the activation sublayer.

A device for deep learning acceleration with mixed precision may include matrix-vector (MV) components that each include vector-vector (VV) components that are each configured to generate a respective VV output based on an input precision mode, an output precision mode, and an accumulation of products. The accumulation of products may be calculated by adding products based on the input precision mode. Each product may be calculated by multiplying, based on the input precision mode, a map data segment and a kernel data segment. Each MV component may include one or more components configured to concatenate VV outputs to generate a concatenated VV output. The device may include activation function components that are each configured to receive a corresponding concatenated VV output, generate an activation function output based on the corresponding concatenated VV output and the output precision mode, and output the activation function output.

The invention discloses a skin stretch forming residual stress rapid calculation method which comprises the following steps: randomly generating a combination of a pre-stretching amount, a coating elongation rate and a friction coefficient, submitting the combination to finite element analysis software to execute batch simulation, and generating a result file named by process parameters; traversing all the result files, extracting node numbers and residual stress values, and storing the node numbers and the residual stress values as a text format data set corresponding to the process parameters; a full-connection neural network model is constructed, an input layer receives the three process parameters of the pre-stretching amount, the coating elongation and the friction coefficient, a hidden layer comprises multiple layers of neurons and adopts a ReLU activation function, and an output layer generates residual stress values of all nodes; training the neural network model by using the data set, and adjusting the network weight through an optimizer; and inputting target process parameters to the trained neural network model, and outputting residual stress calculation results of all nodes of the skin. The technical purposes of rapidness, high efficiency and low cost are achieved.

The invention discloses a training method of a blood glucose control model and a blood glucose control method and system, and belongs to the technical field of blood glucose control. According to the method, a framework combining a blood glucose control model and a reinforcement learning model is constructed; wherein the blood glucose control model is a full-connection neural network with two layers; the number of neurons in the first layer of the network is 3, and an adopted activation function is a linear function; the number of neurons in the second layer of the network is 1, the adopted activation function is a nonlinear activation function, the output value domain of the nonlinear activation function is bounded, the upper bound is a positive number, and the lower bound is a negative number. The blood glucose control model is a model evolved after a classic PID algorithm is improved based on a blood glucose control task, the parameter space of the model comprises the parameter space of the classic PID algorithm, and the model has a wider range, can extract features with higher expression ability, is more adaptive to a reinforcement learning process, and can simply and efficiently realize accurate control of blood glucose.

The invention discloses a fault tree Boolean function equivalent mapping method based on an untrained neural network, and relates to the field of fault tree analysis. In order to solve the problems that in the prior art, a Boolean function mapping structure is not beneficial to parallel expansion, the calculation efficiency is limited, and the Boolean function mapping structure is difficult to efficiently realize on high-parallel platforms such as a GPU, the invention provides a method for generating topological structure data by analyzing a fault tree model; the basic events, the intermediate events and the top events are mapped into neurons of an input layer, a hidden layer and an output layer respectively, a feedforward network with fixed weight and bias is constructed, and a logic activation function is defined in nodes to realize Boolean logic propagation. The input layer receives a basic event state vector, outputs a top event result through forward propagation, and can realize large-scale Boolean function mapping on a parallel platform through batch input matrixes. The method is suitable for reliability analysis, minimum cut set simplification, top event probability calculation, parallelization fault tree solving and the like of a large-scale complex system.

The invention relates to the technical field of camera detection, and discloses a camera detection method and system based on a twin neural network, and the method comprises the steps: collecting the conductive detection data of a to-be-detected camera, generating a conductive feature vector of a fixed dimension, and processing the conductive feature vector based on the twin neural network, the twin neural network comprises a first neural network branch and a second neural network branch, each neural network branch input layer is used for receiving a conductive feature vector, fixing dimensions and serving as an input signal of a subsequent neural network, the hidden layer is composed of a plurality of full-connection layers, each full-connection layer is connected with the activation function layer, and the activation function layer is connected with the second neural network branch. And the embedding output layer is used for mapping the extracted high-dimensional features into a low-dimensional embedding vector, and calculating a similarity value of the two embedding vectors to judge whether the to-be-detected camera is qualified or not. Whether the to-be-detected camera meets the qualified characteristic standard or not can be automatically judged, and fuzziness of human experience judgment is avoided.

The application relates to a hyperspectral remote sensing image quality grade evaluation method, which comprises the following steps: dividing an image into fixed-size blocks, then constructing a quality evaluation model, obtaining a feature map with spectral weight through a spatial-spectral attention module, and inputting the feature map into a frequency spectrum integration embedding module to generate an embedding feature map of local adjacent spectral bands; after linear projection, the embedding feature map is input into an encoder, and a skip connection mechanism is set to fuse the outputs of different encoders; the output of the last encoder is input into a linear layer and an activation function layer to obtain a quality evaluation result of the image; the quality evaluation model is trained, and the trained model is used for image quality grade evaluation. The application can perceive the good and bad degrees of a hyperspectral remote sensing image, and can dynamically detect and adjust the image output by an image processing system in a hyperspectral imager according to the quality grade of the image, thereby providing a more effective basis for parameter optimization of a real-time hyperspectral imager system.

Disclosed in the present invention is a traffic sign detection method, comprising: acquiring a traffic road image to be detected; preprocessing the traffic road image; inputting the preprocessed traffic road image into a pre-trained traffic sign detection model to obtain a classification result output by the model; marking a detected traffic sign on the basis of the classification result from the traffic sign detection model, and outputting an image of the marked traffic sign. The traffic sign detection model uses an improved RT-DETR model in which on the basis of an RT-DETR model, layers 5-7 are replaced with three downsampling feature extraction layers of a feature learning fusion module DualBlocks, wherein each replaced layer consists of a DualConv module, an average pooling module and an ReLu linear activation function. Layers 11 and 16 of an RT-DETR network model are separately replaced with a dynamic upsampling layer of a dynamic upsampling operator Dysample, wherein the dynamic upsampling layer consists of a sampling point generator, a sampling apparatus and an interpolation function. The present invention can reduce the number of parameters while improving the detection accuracy.

The invention discloses an electric power price prediction method and device based on time dimension, equipment and a storage medium, and relates to the technical field of electric power systems, and the method comprises the steps: carrying out the preprocessing of a multi-source data set comprising historical electric power transaction data, a load demand curve, energy production data and fuel cost, performing feature extraction on the obtained processed data set to obtain each data feature, and determining a training set and a verification set based on the data features by using Bernoulli distribution; constructing an initial power price prediction model based on a multi-head attention mechanism, a convolutional layer, a preset activation function, a bidirectional long-short-term memory network and integration of a low-rank adaptation technology, and training and verifying the initial power price prediction model by using the training set and the verification set respectively; and performing multi-scale prediction by using the obtained target power price prediction model to obtain prediction results of the power price in different time dimensions. And the real-time, day-ahead and weekly prediction accuracy of the electric power price is improved.

The invention relates to an intelligent SDN network abnormal flow detection method and system, and the method comprises the steps: collecting network flow data through an SDN switch, carrying out the preprocessing, and generating a standardized time sequence feature vector; performing multi-scale feature extraction on the standardized time sequence feature vector based on a multi-scale one-dimensional expansion convolutional neural network and a channel attention mechanism, capturing a context dependency relationship of abnormal traffic in combination with a bidirectional long-short-term memory network, and outputting a high-dimensional abnormal feature vector; and adopting a full connection layer and a Softmax activation function to classify the high-dimensional abnormal feature vector, optimizing the model performance through a composite loss function, and generating the classification probability of the abnormal traffic. According to the method, the detection precision of a complex attack mode is improved through multi-scale expansion convolution and a channel attention mechanism, and the robustness is enhanced by capturing time sequence dependence in combination with a bidirectional long-short-term memory network; the classification performance is optimized by adopting a composite loss function, and the false alarm rate is reduced; and realizing millisecond attack blocking based on a dynamic flow table rule of a classification probability.

To provide a software fault localization method based on feature intersection and KAN network. The method first extracts different types of features from bug reports and source code files, then uses a crossover layer to crossover and correlate the features and extract hidden relationships between them. Furthermore, by utilizing the KAN network's learning of parameterized nonlinear activation functions, the process of fitting a polynomial function using the feature crossover network is transformed into a process of fitting multiple univariate functions, thereby overcoming the "curse of dimensionality" and more accurately capturing and adapting to complex function changes and complex high-order feature interaction information in defect localization, thereby improving defect localization performance. Finally, the output of the KAN network is input to a fully connected layer to calculate a final similarity score, which is then sorted in descending order based on the final similarity score to obtain the defect localization results.

Generally, the present disclosure is directed to the training of robust neural network models by using smooth activation functions. Systems and methods according to the present disclosure may generate and / or train neural network models with improved robustness without incurring a substantial accuracy penalty and / or increased computational cost, or without any such penalty at all. For instance, in some examples, the accuracy may improve. A smooth activation function may replace an original activation function in a machine-learned model when backpropagating a loss function through the model. Optionally, one activation function may be used in the model at inference time, and a replacement activation function may be used when backpropagating a loss function through the model. The replacement activation function may be used to update learnable parameters of the model and / or to generate adversarial examples for training the model.

The invention discloses a method for determining a parameter lookup table of a neural network activation function, a method and system for calculating the neural network activation function based on the parameter lookup table and a storage medium, and belongs to the technical field of neural network accelerators, and the method comprises the steps: dividing the neural network activation function into a linear region and a nonlinear interpolation region; dividing the nonlinear interpolation region into a plurality of sub interpolation regions; performing linear fitting on the multiple segments of sub-interpolation regions to obtain slopes and intercept of multiple linear functions in one-to-one correspondence with the multiple segments of sub-interpolation regions; and determining a parameter lookup table according to the input variable value range of the neural network activation function corresponding to the multiple segments of sub-interpolation regions and the corresponding relationship between the slopes and intercept of the multiple linear functions, and storing the parameter lookup table in the memory. According to the method, the activation function is divided into the linear region and the nonlinear interpolation region, the activation function of each partition is approximately calculated through the linear function, the parameter lookup table is determined, and the activation function calculation is realized.

Ophthalmic medical image segmentation method includes dividing the medical image data into a training set and a test set according to an autonomously set proportion; constructing a convolutional neural network model adopting a U-shaped encoding and decoding structure based on an attention mechanism and a weighted loss function, and performing training; transmitting a to-be-segmented medical image to obtain a segmentation result, wherein the attention mechanism is introduced into the U-shaped encoding and decoding structure: a superficial layer feature map ILE of an encoder is subjected to convolution to obtain ILE-1, and a deep layer feature map IHD of a decoder is subjected to up-sampling and convolution to obtain IHD-1; the ILE-1 and the IHD-1 are multiplied to obtain IMul; the IMul and the IHD-1 are summed, and ISum is then output through an activation function; and the IMul and the ISum are spliced, and then output to a target layer.

The invention relates to an intelligent measurement laboratory environment monitoring and evaluation method and system, and belongs to the technical field of intelligent laboratory monitoring, and the system comprises a physical entity layer which collects physical perception data streams in real time through an Internet of Things sensor array; and the virtual platform layer is used for constructing a five-dimensional virtual mapping model based on a digital twinning technology and comprises the steps that the structural layer performs data standardization preprocessing and multi-source information feature extraction, and the intelligent layer uses an improved SwinTransform model to generate a personnel behavior classification probability matrix and a device state recognition probability matrix. The improved model optimizes feature extraction through a deep convolution kernel, a GeLU activation function and a simplified regularization structure; and the decision-making layer is used for calculating a laboratory operation risk index based on the probability matrix by adopting a decision-making level fusion algorithm, and triggering graded early warning when the index exceeds a preset threshold value. According to the method, the problems of metering laboratory data island, environmental response lag and weak risk early warning are solved, the personnel behavior recognition accuracy is improved, and the model convergence speed is increased.

The invention provides an equipment energy consumption dynamic evaluation method based on multi-source data fusion, and relates to the field of energy consumption evaluation, and the method specifically comprises the steps: constructing an energy consumption evaluation data set, carrying out the statistics of the missing rate, jump rate and straightness rate of an energy consumption time sequence, and generating dimension-by-dimension credibility gating. Calculating the deviation degree of the multi-source data in the energy consumption time sequence based on the median vector to obtain the dynamic fluctuation intensity, differentiating the multi-source data and the median vector, normalizing the multi-source data and the median vector in combination with the robust scale vector and the dynamic fluctuation intensity, and performing element-by-element combination operation with dimension-by-dimension credibility gating to obtain normalized output, a soft division probability is obtained through a gating circulation unit, linear transformation and a Softmax activation function in sequence, weighted fusion is carried out on normalized output, global feature representation is formed through one-dimensional convolution and summary average, dynamic energy consumption evaluation values of the main equipment and the auxiliary equipment are obtained through mapping by combining fusion representation vectors, and the dynamic energy consumption evaluation values of the main equipment and the auxiliary equipment are fused to form a dynamic energy consumption evaluation value of the equipment. Accurate and dynamic evaluation of the energy consumption state of the equipment is realized.