148 results about "Multilayer perception" patented technology

The invention discloses a motor imagery early fusion decoding method based on EEG-fNIRS. The method comprises the steps that EEG and fNIRS signals in a motor imagery task are acquired and preprocessed; eEG and fNIRS feature extraction and alignment modules are used for extracting and aligning EEG feature information and fNIRS feature information respectively; performing deep fusion on the EEG and fNIRS time dimension alignment features by using a bidirectional cross attention module to obtain EEG-fNIRS early fusion features; the EEG-fNIRS early fusion features are input into a Transform encoder, different time step information is fused in a self-adaptive mode through an attention weighted pooling module, and EEG-fNIRS fusion features are obtained; and inputting the EEG-fNIRS fusion feature into a multi-layer perceptron to output a motor imagery task category. According to the method, space-time coupling characteristics of EEG and fNIRS signals can be fully utilized, deep fusion of cross-modal characteristics is realized, and the decoding performance of a motor imagery task is remarkably improved.

The invention relates to a polymer property prediction method and device based on substructure and knowledge enhancement. The method comprises the following steps: obtaining a polymer monomer SMILES to be predicted; adopting a rule-based BRICS algorithm to cut a polymer monomer SMILES to be predicted, and obtaining a substructure SMILES; converting the SMILES of the polymer monomer to be predicted and the SMILES of the substructure into a molecular image, and matching a property influence flag bit for the molecular image of the substructure based on a pre-constructed labeling dictionary; utilizing an improved double-flow ResNet network to respectively extract feature vectors of a polymer monomer molecular image to be predicted and a substructure molecular image to be predicted; respectively polymerizing the feature vector and the property influence flag bit of the substructure, and splicing the feature vector and the feature vector of the polymer monomer to be predicted to obtain a target feature vector to be predicted; and inputting the to-be-predicted target feature vector into a pre-trained multi-layer perceptron model to obtain the property of the to-be-predicted polymer monomer. Compared with the prior art, the prediction precision and generalization ability of polymer property prediction are remarkably improved.

The invention relates to a wind noise modeling method and system based on machine learning and application, and belongs to the field of marine acoustics and environmental noise modeling, and the method comprises the steps of data preprocessing and feature extraction, hybrid network model building, hybrid network model training and verification. According to the method, a hybrid network model of multiple linear regression and a multi-layer perceptron is constructed based on a known physical mechanism of wind noise and two main noise generation mechanisms of surface turbulence and bubble oscillation, a linear relation and a non-linear relation are modeled respectively, outputs of the two models are integrated through a weighted fusion strategy, and a multi-layer perceptron model is constructed. Smooth transition modeling from a low-wind-speed linear relation to a high-wind-speed nonlinear relation is achieved.

The invention discloses a plant three-dimensional point cloud completion method and system based on deep learning, and the method comprises the steps: inputting all to-be-completed sub-point cloud data into a plant point cloud completion model, and obtaining completed point cloud data; the plant point cloud completion model comprises an encoder and a decoder based on a CRA-PCN architecture, and an associated missing information enhancement module; and the association missing information enhancement module is arranged between the encoder and the decoder and is used for enhancing the global features and the local features output by the encoder through a multi-layer perceptron. According to the method, the global and local features of the complete point cloud are fused to guide the reconstruction of the missing part, so that the perception capability of the network to the missing structure is effectively enhanced, and the defects of large noise and structure distortion of the existing deep learning model in a multi-block missing or low-density scene are overcome.

The invention discloses an lncRNA-protein interaction prediction method based on bidirectional intention, and belongs to the technical field of bioinformatics, and the method comprises the following steps: data acquisition and preprocessing: constructing a data set; performing lncRNA feature extraction through a multilayer convolutional neural network; carrying out protein feature extraction based on an ACmix module; fusion of the lncRNA and protein bimodal features is carried out through a bidirectional intention network; performing lncRNA-protein interaction prediction on the fusion representation by using a three-layer and multi-layer perceptron network, and outputting at an output layer; and meanwhile, designing a loss function to carry out model optimization. The model is more comprehensive and accurate when the dependency relationship between the sequences is captured, so that the reliability and generalization ability of prediction are effectively improved.

The invention relates to the technical field of crowd counting, in particular to a crowd counting method and system based on WiFi and video modal cross-level attention, and the method comprises the steps: constructing a WiFi density map at a WiFi sensing side, and converting an irregular detection record into a fixed-size image representation; on a video sensing side, marking a region of interest for video frames collected by cameras with different visual angles, and cutting the region of interest to serve as a video side image; respectively carrying out feature coding on the WiFi density map and the video side image by adopting a convolutional neural network and self-attention combined mode; gradually aligning the WiFi modal feature embedded representation and the video modal feature embedded representation through multi-layer stacked cross-modal attention to obtain a cross-modal fusion feature; and inputting the cross-modal fusion features into a lightweight multilayer perceptron, and outputting crowd count. According to the invention, through hierarchical alignment and fusion of WiFi signals and video features, accurate estimation of the number of crowds in a large-scale complex scene can be realized.

The invention relates to the technical field of point cloud processing and deep learning, and discloses a point cloud semantic segmentation method based on multiple sparseness fusion sampling and local neighborhood space geometric color feature enhancement coding, and the method comprises the following steps: fusing two sampling results according to a preset proportion, and obtaining a sampling point set giving consideration to global and local distribution; the geometric and color features are uniformly input into the multilayer perceptron sharing the weight for feature mapping, an enhanced feature vector is generated, the enhanced feature vector is further fused with the center point features, a final code is output, and end-to-end training is achieved. The method has the beneficial effects that through multi-sparseness fusion sampling, the representativeness of dense and sparse point cloud regions can be dynamically balanced, and the segmentation performance of the network in a complex structure or non-uniform distribution scene is remarkably improved; local neighborhood space geometry and color features are fused, the perception ability of the network to fine structures and texture differences is enhanced, and the category confusion phenomenon is effectively reduced.

The application discloses a text and structured data classification method and system based on a cascade model and feature fusion, and the method comprises the following steps: for a descriptive text, semantic coding is performed on the descriptive text by using a semantic model to obtain a text hidden state tensor, and general semantic tensor is obtained by dimension reduction on the text hidden state tensor through an attention mask tensor; the general semantic tensor is input into a Transformer encoder, field semantic features are learned through a multi-head self-attention mechanism, and field enhanced semantic tensor is generated through the same dimension reduction mode. An adaptive gating network is constructed, the two kinds of tensors are spliced and adaptively weighted and fused to obtain a fusion feature tensor. The fusion feature tensor is input into a multilayer perception machine model, and a classification prediction potential target factor is output through training and modeling; the classification prediction potential target factor is integrated with a structured factor into a feature set, the feature set is input into a Lightgbm model for joint training, a classification prediction index is output, and classification prediction of the structured data with the descriptive text is completed. The accuracy of classification is improved.

The application discloses a kind of antenna array element displacement real-time reconstruction methods based on limited strain information, it is suitable for the electric performance compensation after phased array radar array surface deformation.This method uses data-driven mode, by combining numerical simulation and multilayer perception machine network, according to the strain value of limited monitoring point, the displacement of all target points is reconstructed.By introducing the loss function based on the assumption of Gaussian distribution and homoscedasticity uncertainty, the network architecture is optimized to improve the reconstruction accuracy.Using the data set generated by finite element analysis, the efficient reconstruction of multi-target point displacement on four-edge hinged plate is realized.Discrete strain data as input features, array element displacement data as output target values, are used to train and verify the deformation perception model based on MLP.The results show that sparse arrangement of monitoring points can realize synchronous and accurate reconstruction of displacement of multiple measuring points.It can be widely used in the field of deformation perception and performance compensation of complex structures.

The invention provides an unsupervised APT detection method based on a time sequence diagram attention network and a deviation network. The unsupervised APT detection method comprises the following steps: acquiring a system audit log of a fixed-width time window by using a sliding time window; entities related to the events and interaction relations between the entities are extracted, and a dynamic traceability graph is constructed in combination with occurrence time of the events; modeling the interaction behavior of each node adjacent to the node in the time dimension by using a time sequence diagram attention network so as to extract a node time dependent feature representation; predicting the type of an edge between adjacent node pairs by using a multi-layer perceptron to obtain a prediction error; evaluating the degree of deviation between the type of the predicted edge and the corresponding reference standard by using a deviation network; and based on the prediction error and the deviation degree, identifying the time window subjected to the APT attack and the corresponding attack chain. The method is remarkably superior to a traditional detection method in the aspects of detection precision and robustness, and an efficient and extensible solution is provided for APT detection.

The invention discloses a collaborative recommendation method based on differential privacy and depth representation, and belongs to the field of artificial intelligence recommendation systems. The recommendation method comprises the following steps: firstly, constructing a noise-added user portrait and an article portrait for protecting privacy by combining a collaborative filtering algorithm with a differential privacy technology; then performing deep feature extraction and optimization on the noise-added user portrait and the noise-added article portrait by using a deep learning multi-layer perceptron model; and finally, sequentially performing dot product on the final portrait vector of the user and the final portrait vectors of all candidate items to obtain a prediction score of each user-item, sorting all scores from high to low, and intercepting the sorted first N items to form a Top-N recommendation list. According to the method, the problems of data sparseness and cold start can be effectively relieved, and meanwhile, sensitive information leakage is prevented by applying privacy protection to the user and article portrait generation process, so that balance between recommendation system performance and privacy security is realized.

This invention relates to the fields of artificial intelligence and computer vision, and discloses a method for detecting residual manganese in manganese electrolysis plates based on multi-expert convolution. The method includes: constructing a first branch based on depthwise separable convolutional blocks combined with routing multilayer perception and convolutional layers with different kernels; constructing a second branch based on depthwise separable convolutional blocks and residual connections; constructing a third branch based on depthwise separable convolutional blocks; constructing a hybrid convolutional block by combining the first, second, and third branches with multi-head attention; constructing a detection backbone based on the hybrid convolutional block combined with multilayer convolution; constructing a multi-scale fusion structure based on multi-scale feature extraction and routing selection; constructing a single-input dual-output prediction head based on the hybrid convolutional block; and constructing a detection model based on the detection backbone, multi-scale fusion structure, and prediction head; acquiring an image of the manganese electrolysis plate and inputting it into the detection model to obtain the detection result of residual manganese. This method solves the problem that existing visual algorithms cannot accurately identify residual manganese in manganese electrolysis plates.

The invention discloses a post-blasting rock mass collapse volume prediction method based on point cloud deep learning. The method comprises the following steps: acquiring initial sparse point cloud data of a post-blasting rock mass and carrying out filtering preprocessing; acquiring a low-dimensional global feature vector by adopting a multi-layer perceptron encoder; photon signal features are generated by using the step-by-step autoregression prediction network and the neuromorphic photon feedforward matrix; high-density photon characteristics are formed through a coherent interference mechanism; high-density electric signal point cloud data are obtained through photoelectric conversion; and calculating the rock mass collapse volume through an Alpha Shape algorithm and outputting the rock mass collapse volume in real time. According to the invention, accurate and efficient prediction of the rock mass collapse volume is realized.

The application discloses a single-photon lidar point cloud denoising method based on multi-mode guidance and related equipment, and belongs to the field of lidar signal processing and computer vision. The three-dimensional point cloud of the single-photon lidar, the infrared image and the visible light image are synchronously collected; the infrared image and the visible light image are extracted and fused in parallel to generate multi-mode image guidance features; after the features of the point cloud are extracted, the features are projected onto the image plane to form spatially aligned alignment point cloud feature maps; the two features are interacted through bidirectional cross-attention calculation, the obtained features and the original features are aggregated, and then integrated through a multilayer perception machine to obtain multi-modal joint features; the joint features are decoded to output a denoising weight map and a signal enhancement map; the original point cloud is projected onto the image plane, noise points are filtered out by using the denoising weight map, and the reflection intensity of the reserved points is enhanced by using the signal enhancement map to generate a denoised three-dimensional image. The application realizes accurate noise filtering and effective signal enhancement, and improves the quality of the point cloud.

The invention discloses a multi-modal denoising method for a hyperspectral image, which is characterized by comprising the following steps of: S1, respectively performing low-rank tensor representation on the hyperspectral image and a registered multispectral image by utilizing Tucker decomposition, and extracting core tensors of the hyperspectral image and the registered multispectral image; s2, establishing a correlation model between the hyperspectral core tensor and the multispectral core tensor through model-driven linear mapping or data-driven multi-layer perceptron network; s3, iteratively solving a core tensor, a factor matrix and correlation model parameters by adopting an alternating direction multiplier method; and S4, reconstructing a denoised hyperspectral image by using the optimized hyperspectral core tensor and factor matrix. Compared with the prior art, the method has the advantages that the spectral details of the hyperspectral image and the high-signal-to-noise-ratio spatial information of the multispectral image are fully mined and utilized, the restoration precision in the mixed noise scene is effectively improved through the double-Tucker decomposition framework and the core tensor association strategy, and the high-quality hyperspectral image is obtained.

The application relates to a visual identification method and system of a liquid crystal angle color-changing material. Specifically, optical images of a material at a preset angle sequence are collected, Lab chroma and rotation-invariant uniform LBP texture features of each image are extracted, and a multi-dimensional initial feature sequence is formed. A double-branch identification model is constructed. One branch performs differential spectrum analysis on the feature sequence, and then extracts features through a multilayer perception machine. The other branch encodes contexts through a bidirectional GRU, and enhances key information by using angle-channel correlation attention, and then pools the key information into a context vector. The outputs of the two branches are concatenated, and are mapped into an encoding vector through a classifier. During training, a polarization loss function is used to drive the encoding vectors of the genuine product and the fake product to converge to two preset opposite polarization directions. During identification, the distances between the encoding vector of a material to be tested and the two polarization vectors are calculated, and the one with a shorter distance is determined as the genuine or fake product.

The application discloses a kind of based on multimodal data fusion's spherical fruit quality detection method, including mobile platform moves to the coordinate position of detected plant;Real-time identification fruit's space three-dimensional coordinates by the visual auxiliary positioning device set on mobile platform;Mechanical arm control system is controlled to move by mechanical arm, so that flexible mechanical paw moves to the space coordinates of detected fruit;Flexible mechanical paw carries out multiple directions'snatching action to detected fruit, and multiple directions'sensed fruit tactile data are dynamically collected by multiple-point array flexible film pressure sensor, and multiple directions'sensed fruit image data are collected to sensed fruit by industrial camera;By multimodal data processing unit, the deep semantic features and dynamic response characteristics of multiple directions'sensed fruit are extracted respectively, the deep semantic features and dynamic response characteristics are spliced and fused after input to MLP multilayer perception machine and are fused and analyzed, and the quality evaluation result of sensed fruit is obtained.

The invention provides a graph matching network-based vulnerability and patch semantic association evaluation method, which comprises the following steps of S1, performing standardization processing on a vulnerability report and patch description to obtain a preprocessed text; s2, constructing a word co-occurrence graph of the preprocessed text by utilizing a sliding window mechanism; s3, coding initial features of nodes and edges of the word co-occurrence graph into hidden vectors by using a multi-layer perceptron, and generating a vulnerability graph and a patch graph; s4, introducing an intra-graph attention mechanism and a cross-graph attention mechanism, and aggregating neighbor information in and between the vulnerability graph and the patch graph; and S5, respectively aggregating the vulnerability graph and the patch graph into graph-level representation vectors by using the patch correctness evaluation model, and calculating the Euclidean distance between the vulnerability graph and the patch graph. According to the method, the graph structure representation of the vulnerability report and the patch description is constructed, and the cross-graph attention mechanism is introduced to carry out joint semantic modeling, so that the problems of text structure information missing and lack of joint reasoning in the existing method are effectively solved.

The invention relates to a soil organic carbon content estimation method based on multi-feature coupling ensemble learning. The method comprises the following steps: screening a soil hyperspectral image of a target area by using a spectral index; establishing an SOC content feature library, calculating feature importance by using an LGBM model, and selecting input features based on the feature importance to construct a training data set; an SOC estimation framework model is constructed, a first layer comprises a random forest model, a multi-layer perceptron model and a K neighbor model, and a second layer adopts CatBoost as a meta learning device; and inputting the training data set into a first layer of the SOC estimation framework model, respectively generating predicted values of the SOC content, performing five-fold cross validation on each model sample to obtain a group of predicted superposed values, and training a meta-learner of a second layer to obtain a final SOC content prediction result. Compared with the prior art, the method has the advantages of high applicability, high prediction precision, excellent interpretability and the like under different space-time conditions.

The invention discloses a method and a system for predicting migration and blockage of fracture particles based on machine learning, and the method comprises the following steps: firstly, presetting fracture roughness coefficient (JRC), flow velocity, equivalent particle size of particles and particle quantity data, constructing a fracture model with a real shape, and carrying out a visual migration test; recording fracture blockage label data; inputting the parameters and the labels into a neural network dichotomy probability model taking a multi-layer perceptron as a core, and training to obtain a blockage probability prediction function; on the basis of the trained model, sensitivity analysis and feature importance evaluation are carried out, the influence sequence of the blockage probability on JRC, the flow velocity, the equivalent particle size of particles and the particle number is output, and prediction of the particle blockage event in the fracture is achieved; according to the invention, the particle blocking mechanism in the crack under the multi-factor coupling effect is researched, and a scientific basis is provided for design and construction of geotechnical engineering.

An autoencoder and a method for adaptive learned image compression with configurable encoder and decoder for variable bitrate applications are described, wherein the encoder comprises: - a pre-trained transformer-based architecture operable to convert input image data into quantized latent representations (9); - one or more low-rank adapter modules (W a,W b) integrated within said encoder (g a), each adapter module (W a,W b) being configured to adjust the encoding process for different target bitrates while maintaining pre-trained model parameters; - wherein said one or more low-rank adapter modules (W a,W b) are incorporated into fully connected layers (120,121) of the encoder's multi-layer perception modules (MLP), enabling said encoder (g a) to achieve efficient rate-distortion performance across a range of bitrates by fine-tuning only parameters of said adapter modules (W a,W b); - a mechanism adapted to merge said adapter modules (W a,W b) with pre-trained model weights following adaptation, thereby restoring the encoder's complexity to its original level.

The invention discloses an SF6 density relay leakage determination method. The method comprises the following steps: establishing an SF6 gas chamber leakage multi-physics field coupling model and generating a multi-state sample set; a residual multi-layer perceptron is adopted as a classification model, and hyper-parameter optimization is carried out in combination with grid search and ten-fold cross validation; constructing a leakage state intelligent identification model, and deploying the leakage state intelligent identification model on an edge device or a cloud platform; and finally, real-time and accurate classified output of the SF6 leakage state is realized. According to the method, 'non-leakage ', 'slight leakage' and'severe leakage 'can be efficiently and accurately distinguished only by depending on original air chamber structure parameters and temperature sensing information without additionally arranging an expensive detection device, and the reliability and economical efficiency of leakage judgment are remarkably improved.

The application discloses a kind of based on deep learning's staged tension prestressed long-term loss prediction method, belong to deep learning field, this method solves the problem that traditional prestress determination method prediction result robustness is poor, including: based on standard drawing set and bridge design specification obtains bridge original data;Based on construction environment condition and concrete age, calculate concrete material attribute;Based on bridge original data and concrete material attribute, utilize finite element software to simulate the staged tension process of prestressed reinforcement under different working conditions, generate bridge prestressed reinforcement staged tension loss data set;Based on multilayer perception machine neural network, construct prestressed loss prediction model, based on prestressed loss prediction model, bridge prestressed reinforcement staged tension loss data set is handled, and the long-term loss prediction result that prestressed reinforcement is caused by shrinkage and creep is output.The application can effectively predict the stress loss of prestressed reinforcement, improve the robustness and accuracy of prediction.

The invention discloses a human emotion world generation method based on MBTI character injection, and the method comprises the steps: carrying out the preprocessing of collected multi-source signals one by one, carrying out the weighted fusion through an attention mechanism, mapping a fusion feature to a two-dimensional space comprising the pleasure degree and the activation degree through a multi-layer perceptron, and obtaining a current emotion state coordinate; according to an emotional state coordinate sequence comprising a plurality of historical emotional state coordinates, a historical emotional trajectory is obtained through a Linear Attention network, and a personalized emotional evolution trajectory is obtained in combination with the current emotional state coordinates; according to the MBTI character code of the user, an embedded vector is obtained and serves as the hidden state of the Linear Attention network, and the personalized emotional evolution trajectory is modulated; and adopting a diffusion model to generate a plurality of emotion evolution trajectories according to the modulated personalized emotion evolution trajectory and the embedded vector. The method has the effects of improving the emotion perception robustness, realizing efficient long-range modeling, generating highly personalized prediction and supporting uncertainty decisions.

The application relates to the technical field of digital media processing, in particular to a kind of object attribute dynamic configuration method for digital media multi-platform.The scheme aims at the configuration problem caused by the difference of technical specifications when digital media is displayed on different platforms, generates platform feature vector containing composite features by real-time collection and analysis of each platform technical specification;Nonlinear conversion model of object basic attribute to platform specific attribute is established by using multilayer perception;Subscription-published architecture containing subscription management, message publishing and message distribution module is constructed, so that digital media object can subscribe to platform strategy change notification;When receiving the notification, trigger configuration recalculation, apply new attribute to object and update configuration information.The application can dynamically adapt to platform specification changes, improve digital media content management efficiency and display effect, has the advantages of strong universality, high configuration accuracy, etc., effectively meets the multi-platform digital media transmission demand.

A kind of motor imagination EEG signal decoding method based on multiscale filtering and attention mechanism, comprising: the original EEG data of subject is divided into single experiment electroencephalogram data and normalized;Training data is handled with time domain data enhancement strategy;EEG signal is input as four parallel networks, through a multiscale convolutional neural network including time convolution layer, spatial convolution layer and average pooling layer;All feature channels at each time point are used as token for Transform processing.The first residual sub-block of Transform is composed of layer normalization and attention mechanism, and the second residual sub-block is composed of layer normalization and channel multilayer perception structure;The space-time features output by four parallel networks are spliced along the channel dimension, and effective channel attention mechanism ECA is used for channel fusion;Two-dimensional convolution is used to classify test signal.The present application can improve the classification accuracy of motor imagination brain-computer interface and promote its practical application.

The invention is suitable for the field of power load prediction, and discloses an enhanced fusion load prediction method, system and device based on a parallel variable component sub-circuit and a multilayer perception neural network, and a medium, and the method comprises the steps: obtaining historical load time sequence data, and setting a historical sequence length to obtain an original sequence; decomposing the original sequence through a moving average decomposition method to obtain a trend sequence and a residual sequence; an enhanced prediction network is constructed through a variable component sub-circuit and a multi-layer perception neural network which are arranged in parallel; and respectively inputting the trend sequence and the residual sequence into an enhanced prediction network for prediction, and fusing a generated trend prediction result and a residual prediction result to obtain a power load prediction result. According to the method, after the sequence is decomposed through the moving average method, the parallel variable component sub-circuit and the multi-layer perception neural network are adopted, the power load result is obtained, and therefore the prediction precision and the generalization ability of the model are remarkably improved.

The application relates to an expandable multi-agent hierarchical navigation method and device based on a graph network, wherein the method comprises the following steps: receiving state information of all agents and position information of target points, and respectively generating an agent relationship graph of all agents and a target point relationship graph of all target points according to the state information and the position information, so as to calculate an agent-target point relationship matrix, and match a corresponding optimal target point for a current agent through the relationship matrix; fusing the state information and the position information through a subgraph fusion strategy to obtain a fusion result, and encoding the fusion result according to a multilayer perception strategy, extracting an encoding relationship between the current agent and the optimal target point, and generating an optimal path between the current agent and the optimal target point. Therefore, the problems that the existing multi-agent navigation method has large calculation amount and poor performance, the action space and the state space of multiple agents are too complex, overfitting is prone to occur, the migration ability is poor, and large-scale application is difficult are solved.

The invention discloses a cloud game video quality evaluation method based on spatial domain depth features and time domain change features. The method comprises the following steps: step 1, obtaining a to-be-evaluated distorted game video; 2, equally dividing the video into a plurality of segments and selecting key frames; 3, extracting a spatial feature vector; and 4, extracting a time domain feature vector. 5, the time domain feature vector and the space feature vector are connected in series, and a fusion feature vector is formed. And step 6, inputting the fusion feature vector into a multi-layer perceptron (MLP) quality regression module, and outputting a scalar value as a prediction quality score of the video sample. According to the method, the inter-frame change features are extracted in a multi-dimensional mode through physical modeling, the time feature vector and the space feature vector are connected in series to obtain the fusion feature vector, the multi-dimensional time information and the space saliency information of the distorted game video can be extracted more efficiently, and therefore the accuracy of game video quality evaluation is improved.

The application discloses a kind of underground coal mine mining process dynamic position intelligent inversion method, method includes: to the unwinding phase of interference figure of ununsolved is solved, after extracting the real part quantity, imaginary part quantity, mask, normalized X coordinate and normalized Y coordinate corresponding to interference figure are obtained, five-channel splicing is carried out again, and five-channel multidimensional feature tensor as the input data of model is formed;Depth learning model is constructed with Swin-Transformer as main feature extraction network, multilayer perception is as regression head, and hybrid loss function at least including basic loss and physical constraint loss is defined;Five-channel multidimensional feature tensor is used to train the model, and the model after training is obtained;Physical constraint loss is used to guide the model to converge in the solution space that meets physical consistency;Call the model after training, carry out inversion prediction to target mining area data, output the inversion parameter of coal mining;Efficiency and accuracy of dynamic position intelligent inversion in underground coal mining process are improved.