109 results about "Neural network modeling" patented technology

The invention discloses a power equipment defect detection method based on image recognition, and belongs to the technical field of power equipment defect detection, and the method comprises the steps: carrying out the defect simulation based on physical mechanism driving according to an equipment three-dimensional model and physical field simulation parameters, and obtaining a defect simulation data set; according to the defect simulation data set and the real inspection data, training a cross-modal deep learning network based on physical law constraint to obtain a defect identification model; performing time-space diagram neural network modeling according to the historical time sequence inspection data and the defect identification model to obtain a state evolution model; and inputting inspection data acquired in real time into the equipment health state evolution model, and performing online reasoning to obtain a defect detection result. The problems that an existing electrical equipment defect detection method excessively depends on scarce real defect samples, the generalization ability for complex working conditions is weak, and the defect evolution trend prediction ability is lacked are solved.

The invention relates to the technical field of ship design optimization, and discloses a hull shape optimization method based on neural network modeling. In the data acquisition stage of the method, initial appearance parameters and hydrodynamic performance data of a ship body are obtained, the appearance parameters comprise geometric dimensions and shape features, and the performance data comprise resistance coefficients and wave-making resistance values. In the neural network construction stage, a neural network model with a multi-layer perceptron structure is trained by using collected data, weights are updated through a back propagation algorithm, and a nonlinear mapping relation between appearance parameters and hydrodynamic performance indexes is established. In the shape optimization stage, the trained neural network model is used for carrying out iterative adjustment on the shape of the ship body, fluid dynamic performance indexes are recalculated through the model after each adjustment until preset convergence conditions are met, and finally optimized ship body shape data are output. According to the method, partial complex calculation is replaced by the neural network, and intelligent optimization of the hull appearance is realized.

The invention belongs to the cross technical field of artificial intelligence and image recognition, and discloses a cursive script image sequence recognition method for modeling by fusing a stroke structure, and the method comprises the steps: preprocessing and normalizing a cursive script image, eliminating interference information, normalizing the form of the cursive script image, and guaranteeing the stability and effectiveness of a subsequent processing flow; based on stroke decomposition and feature coding of structural analysis, stroke-level structural features are extracted through character skeleton modeling, and cursive writing characters are converted into time series data; modeling a cursive script recognition neural network fused with structural semantics, and recognizing a stroke time sequence by adopting a BiLSTM neural network; an identification output and result optimization module is designed to ensure the final identification quality; through application of scene integration and platform deployment interface design, practical deployment of the model and butt joint of a front-end platform are realized, and a complete cursive script recognition ecology is constructed. According to the method, not only is the structured representation capability of the cursive script image improved, but also the recognition precision and generalization performance of the model on complex handwriting are remarkably enhanced.

The invention specifically relates to a photovoltaic building design and control system based on a multi-modal neural network, and relates to the technical field of building energy saving, renewable energy utilization and artificial intelligence application. Comprising a multi-modal data acquisition and fusion module, a multi-modal neural network modeling and optimization design module, a real-time control strategy generation and execution module and a digital twin platform and continuous learning module. According to the method, global optimization is designed, powerful nonlinear fitting and feature fusion capabilities of the multi-modal neural network are utilized, multi-dimensional complex factors such as climate, buildings, users and a power grid are comprehensively considered, the optimal BIPV integration scheme which is high in power generation efficiency, small in influence on building performance and good in economical efficiency is rapidly generated, and the design efficiency and the scheme quality are remarkably improved.

The invention discloses an error compensation method of a strain type six-dimensional force sensor based on multi-source sensing information fusion, and belongs to the technical field of intelligent sensors. The error compensation method comprises the following steps: data acquisition; preprocessing and fusing data; constructing an error prediction model by adopting a time sequence neural network; training data construction and model training; performing online error compensation and feedback; and verifying, adjusting and optimizing the system. Based on fusion of multiple sensors (IMU, a temperature module and a timer) and time sequence neural network modeling, combined online compensation of errors caused by gravity, temperature drift and time drift of the six-dimensional force sensor is achieved, the method has the advantages of being high in compensation precision, high in response speed and high in adaptive capacity, and the application requirements of various high-precision measurement and control systems are met.

The invention relates to the technical field of industrial artificial intelligence and intelligent manufacturing, and provides a defective product root cause analysis method and system based on a multi-modal space-time diagram neural network. The method comprises the following steps: S1, constructing a dynamic factory knowledge graph; s2, multi-modal space-time diagram neural network modeling and risk prediction are carried out; s3, root cause interpretation based on an attention mechanism; and S4, reinforcement learning parameter optimization of man-machine cooperation. According to the method, the whole medicine production system is abstracted into a dynamic'factory knowledge graph ', and unified modeling is performed on complex space-time association in the graph by using a multi-modal space-time graph neural network; on the basis, autonomous and safe optimization of key process parameters is realized through a reinforcement learning module fused with expert knowledge feedback, so that a complete closed loop from sensing, diagnosis, decision making and optimization is formed.

The invention discloses an injection molding defect reason traceability system based on a knowledge graph, and the system comprises a multi-source data collection and processing module which is used for collecting and processing data in an injection molding production process; the knowledge graph construction module is used for extracting entities and relationships to generate a time-varying heterogeneous knowledge graph; the two-channel heterogeneous graph neural network modeling module is used for constructing a two-channel heterogeneous graph neural network on the time-varying heterogeneous knowledge graph; the defect traceability path searching and scoring module is used for executing traceability path searching and scoring and sorting candidate paths; the evidence fusion and cause sorting module is used for calculating a defect cause confidence coefficient and outputting a defect cause sorting result; the process adjustment scheme module is used for outputting a process adjustment scheme based on the defect cause sorting result; and the knowledge graph dynamic updating module is used for collecting an execution result of the process adjustment scheme and writing the execution result back to the time-varying heterogeneous knowledge graph. According to the invention, knowledge graph reasoning is adopted to realize injection molding defect cause intelligent traceability.

The invention discloses a spiking neural network driven robot dynamic adaptive control system, which relates to the technical field of robot control, and comprises seven modules: an environment sensing module which integrates various sensors and collects and transmits environment, attitude and interaction information; the signal preprocessing module processes data through composite filtering and feature extraction; the spiking neural network modeling module constructs a three-layer structure and performs training based on a fusion learning rule; the dynamic decision output module converts the pulse signal into a control instruction and adjusts gain; the actuating mechanism driving module drives the actuator to act; the state feedback monitoring module monitors and feeds back motion parameters and system states; and the adaptive optimization module optimizes the network and module parameters based on feedback data, and dynamically matches the environment. The control precision and the response speed of the robot in a complex environment are improved, the adaptive capacity is enhanced, the operation reliability and safety are guaranteed through multi-module cooperation, and the application scene is expanded.

The invention discloses a dynamic routing optimization transmission method based on a neural network. The method comprises the following steps: S1, collecting original state data in a network environment; s2, preprocessing the original state data; s3, constructing a causal graph based on the historical abnormal event log, and performing vectorization representation on the causal graph; s4, modeling is carried out on the state features and the topological features, state and structure fusion processing is carried out, and a performance prediction tensor is output through a multi-layer perceptron; s5, constructing a joint optimization target based on the embedded vector and the performance prediction tensor, and screening an optimal path set; and S6, generating a routing control instruction according to the optimal path set, and issuing the routing control instruction to the network forwarding equipment. According to the method, causal graph embedding and neural network modeling are fused, dynamic path optimization is realized, and the method has the advantages of high intelligence, high stability and high adaptability.

The invention discloses a natural gas production yield prediction method and system based on deep learning, and the method comprises the steps: obtaining dynamic production data, geological parameters and environmental factor data, carrying out the preprocessing of the collected data, and carrying out the multi-scale decomposition through EMD (Empirical Mode Decomposition), and obtaining the multi-scale feature data; capturing instantaneous fluctuation features based on an LSTM (Long Short-Term Memory) network, extracting a periodic trend through a CNN (Convolutional Neural Network), modeling spatial relevance between gas wells by utilizing a GNN graph neural network, and introducing an attention mechanism to dynamically weight and fuse each branch feature to obtain a mixed neural network model; and inputting the multi-scale feature data into the target hybrid neural network model for prediction, and outputting a yield prediction result. And data can be comprehensively understood, so that the yield prediction precision is improved.

The invention relates to the field of digital marketing, and discloses a video understanding-oriented hierarchical spatio-temporal semantic extraction method, which analyzes video content through a three-stage progressive analysis framework: firstly, detecting an object instance in a key frame sequence and constructing a spatio-temporal trajectory to ensure object identity continuity; modeling an object state evolution law by using a recurrent neural network, and quantifying an interaction relationship between objects; and finally, constructing an event causal graph based on the state mutation points, and generating a directed causal chain through a statistical causal test algorithm. Video content is abstracted into event causal logic from pixel features, a semantic path of'track-state-causal 'is formed, structured semantic description of decisions can be directly driven, video semantics and user behavior data are coupled, user preference portraits are dynamically constructed, and personalized material generation is guided.

The invention discloses a system call behavior modeling method based on a graph neural network, and the method comprises the following steps: S1, collecting system call event data, and carrying out the standardization processing of a resource identifier to generate a uniform resource identifier; s2, constructing a double-layer hypergraph model based on a uniform resource identifier, and jointly establishing a calling-resource layer and a constraint layer; s3, executing graph neural network modeling on the double-layer hypergraph model, and fusing two layers of embedding to generate a unified representation; s4, calculating joint loss and optimizing parameters of the double-layer hypergraph model by utilizing unified representation in a training stage; s5, adopting a reversible sliding window in a reasoning stage, outputting an abnormal score and generating a structure reconstruction plan; and S6, generating an anti-fact explanation based on the structure reconstruction plan, and outputting a calling set and a related identification sequence. According to the method, the double-layer hypergraph is constructed, and the graph neural network is combined for modeling, so that accurate detection and interpretable analysis of the system calling behavior are realized.

The invention discloses a dynamic graph distribution external detection method and system based on spectrum sensing enhanced evidence learning, and belongs to the field of graph data mining and anomaly detection. In order to solve the problem that distribution outside samples in a dynamic graph structure are difficult to accurately identify, a virtual negative sample generation method based on graph spectrum disturbance is mainly adopted, Dirichlet posterior distribution is modeled in combination with a dynamic graph encoder and an evidence neural network, and the distribution outside samples are discriminated by using uncertainty measurement. According to the method, the distributed external detection performance in a dynamic graph environment can be effectively improved, and the method has relatively high robustness and generalization ability.

The invention belongs to the technical field of power system data management and intelligent scheduling, and discloses a power grid data consistency detection method combining a knowledge graph and an AI algorithm, which realizes consistency intelligent detection of power grid multi-source heterogeneous data by fusing the knowledge graph and a graph neural network algorithm. The method has the advantages that the problems that semantics are not uniform and structural conflicts are difficult to recognize in a traditional method are effectively solved, and the accuracy and expansibility of data fusion are improved; secondly, the graph neural network is used for modeling power grid topology and state characteristics, so that the structural anomaly recognition capability and the detection precision are remarkably enhanced; thirdly, an interpretability mechanism is introduced, so that an abnormal result has a tracing path and a logic description, and data management and operation and maintenance decision making are facilitated; and 4, the maintenance cost of the rule engine is reduced, the self-adaption and intelligent upgrading of the model is realized, and the automation and intelligent level of power grid data management is comprehensively improved.

The embodiment of the invention provides a low-emission and high-efficiency boiler combustion control and management system, relates to the technical field of low-emission and high-efficiency boiler combustion control and management systems, and aims to solve the problems that the combustion efficiency is reduced and the emission of nitrogen oxides is increased due to poor air-coal matching synergism. The system comprises a sensor array, a data acquisition and preprocessing module, a multivariable coupling analysis engine, a combustion state collaborative decision center and an actuator driving array, and an optimal air-coal ratio set point is generated in real time through deep neural network modeling and a multi-target dynamic optimization algorithm. And the actuating mechanism is driven to cooperatively adjust air supply, fuel and flue gas recirculation, so that cooperative control of high efficiency and low emission is realized.

The invention relates to the technical field of thermal energy engineering and industrial artificial intelligence crossing, in particular to a boiler combustion multi-target cooperative control method based on PINN and reinforcement learning, which comprises the following steps: collecting boiler combustion related data through a multi-source sensor network, and fusing edge data; based on physical information neural network modeling training, predicting a key physical field in the boiler; constructing and training a reinforcement learning model based on PINN state embedding; and a multi-target cost function is constructed based on the reinforcement learning model, the multi-target cost function is continuously evaluated through the reinforcement learning model, the weight is automatically adjusted, and multi-target cooperative control over boiler combustion is achieved. The method is obviously superior to a traditional scheme in the aspects of physical consistency, adaptive capacity, real-time performance and multi-target cooperation, and a reproducible theory-engineering integrated new normal form is provided for efficient, clean and flexible operation of a coal-fired power plant boiler combustion system under the double-carbon background.

The invention discloses a multi-fidelity physical information neural network modeling method for spectrum separation learning, and relates to the field of multi-fidelity agent model modeling, and the method comprises the steps: S1, obtaining low-precision data corresponding to a corresponding flow field or physical field through simple numerical solution or CFD software rapid simulation calculation; s2, constructing a neural network containing a Fourier feature embedding layer, carrying out first-stage training by adopting low-precision data, and screening to obtain a high-frequency Fourier basis; s3, constructing a multi-fidelity neural network, migrating training parameters of the neural network to the multi-fidelity neural network, and obtaining a double-branch network architecture of spectrum separation learning by constructing a high-frequency branch and a low-frequency branch; and S4, performing second-stage training on the multi-fidelity neural network by combining low-precision data with corresponding physical information. According to the method, the dependence on high-fidelity data with high calculation cost and high acquisition difficulty can be greatly reduced, and the data and calculation cost in engineering application is reduced.

The invention discloses an intelligent glasses double-target speech enhancement method and system based on a multi-microphone array, terminal equipment and a medium, and relates to the technical field of speech enhancement. The method is applied to the intelligent glasses, and comprises the following steps: collecting noise-containing multichannel original time domain voice signals through a multi-microphone array of the intelligent glasses, and performing time-frequency domain conversion to obtain a frequency spectrum set; frequency domain complex spectrum amplitude-phase characteristics of each channel are extracted, a local time sequence context is introduced, and multi-channel space acoustic characteristics of amplitude-phase decoupling are constructed; and regulating and controlling the double-tone-area complex spectrum through pre-trained deep neural network modeling to obtain a target complex spectrum, and performing inverse time-frequency domain conversion and reconstruction to obtain an enhanced independent wearer voice signal and a preset direction target voice signal. According to the method, double-target voice parallel enhancement and non-target interference suppression are realized, the power consumption is low, the real-time performance is high, the method adapts to the harsh performance constraint of the end side of the intelligent glasses, and the robustness of voice enhancement in a complex noise scene is excellent.

The invention relates to the technical field of battery management systems, in particular to a method for monitoring the health state of a mining emergency lithium battery pack, a battery cell electrolyte contains 0.08%-0.15% of nano cerium oxide, 0.15%-0.25% of lithium borate and 0.03%-0.08% of amino trimethylene phosphonic acid, and a diaphragm is coated with 0.8%-1.2% of a graphene-carbon nanotube-molybdenum sulfide composite coating. The method comprises the following steps: S1, deploying a multi-dimensional sensor, collecting voltage, current and other data, and dynamically adjusting the sampling frequency; s2, carrying out noise reduction, abnormal value elimination and interpolation completion on the data; s3, extracting characteristic parameters such as capacity attenuation rate; s4, detecting the content of POF3, and associating the capacity fading rate; s5, modeling by using an improved BP neural network, and outputting an SOH value; and S6, triggering grading early warning according to the SOH value and the POF3 content. According to the method, through material innovation and multi-dimensional monitoring, SOH evaluation precision is improved, faults are early warned in advance, environmental adaptability and safety prevention and control are enhanced, the defects of a traditional method are overcome, and mining emergency power supply safety is guaranteed.

The present application relates to a kind of industrial chip transformer commonly used deep neural network modeling method, belong to on-chip transformer modeling technical field, the method includes: S1, the structure of transformer, working frequency, geometric parameter and the performance of simulation are determined, and multiple geometric parameters of transformer are defined as vector X, multiple performances to be solved are defined as vector Y;S2, in the set interval, a large number of geometric parameters X are obtained using suitable sampling method, then the corresponding performance index Y is calculated by electromagnetic simulation, and sample data set (X, Y) is formed;S3, sample data set (X, Y) is normalized;S4, the deep neural network model of on-chip transformer is established using machine learning algorithm;S5, the deep neural network model is tested and verified.The method provided by the present application can establish high-precision deep model under limited modeling sample, improve the sample efficiency required when modeling and avoid the overfitting problem of machine learning.

The present application relates to a kind of intelligent interview evaluation and feedback system based on multi-modal data fusion, specifically relates to data processing field, by meta-learning mechanism dynamic perception interview scene and generate initial fusion weight, subsequently utilize the complex interaction relationship between modalities modeled by graph neural network to carry out fine-grained correction to weight, so as to significantly improve the accuracy and scene adaptability of multi-modal evaluation, further introduce reinforcement learning, link evaluation decision and long-term performance of talents, continuously optimize weight generation strategy, ensure that evaluation standard and business goal are aligned, finally, through closed-loop iteration mechanism, make the whole scheme can be updated automatically according to new data and performance feedback, continuous evolution, with strong self-optimizing ability and long-term robustness, realize the fundamental change from static rule to dynamic intelligent decision.

The application provides a pre-training language model construction method, system and device, relates to the fields of artificial intelligence technology and semantic processing, and mainly comprises the following steps: based on a knowledge graph, a subgraph with a triple as a node unit is constructed; based on entity and relation description text, both are encoded through a conventional pre-training language model to obtain entity representation vectors and relation representation vectors; for the entity representation vectors and the relation representation vectors corresponding to the subgraph, modeling relation information is carried out based on a graph self-attention neural network, and the parameters of the graph neural network and the entity vectors are updated through a prediction link task; the parameters of the model are updated based on a contrast learning technology; and the parameters of the model are fine-tuned for a specific task. The scheme overcomes the problem that entity in the equipment field is sparse and a conventional pre-training language model cannot sufficiently learn entity semantics, and the updated pre-training language model has better understanding and reasoning capabilities in the application of the equipment field.

The invention discloses a Longboer lens reverse engineering method based on electromagnetic inverse scattering, and relates to the technical field of detection imaging of a Longboer lens, and the method comprises the following steps: S1, data collection: arranging an electromagnetic wave transmitter and a receiver outside a lens, obtaining electromagnetic information of the lens through transmitting and receiving, and obtaining the electromagnetic information of the lens; the target area is divided into a plurality of grids, a predicted scattered field is obtained through calculation and accumulation, electromagnetic wave scattering measurement and implicit neural network modeling are combined, high-precision reconstruction of the three-dimensional structure in the luneberg lens is achieved, lens scattered field data are obtained through multi-angle and multi-frequency electromagnetic wave excitation and receiving, and the reconstruction precision of the three-dimensional structure in the luneberg lens is improved. The continuous dielectric constant distribution in the lens is modeled by means of the implicit neural network, the three-dimensional structure information of the lens can be obtained more freely and accurately, and compared with analysis of a traditional Longbour lens structure, the method is more convenient and accurate.

The application discloses a kind of cross-version vulnerability identification systems based on depth map neural network, comprising: multi-version code metadata acquisition module, for collecting multiple versions of software source code and parsing generation multi-version code metadata set;Multi-version code graph generation module, for constructing multi-version code graph set;Characteristic representation matrix generation module, for generating the characteristic representation matrix of multi-version code graph;Depth map neural network modeling module, for using GLEM model, obtains the embedding matrix of each node;Potential vulnerability node identification module, for identifying the potential vulnerability node set across version;Result output module, for outputting cross-version vulnerability identification result and generating detection report.The application adopts depth map neural network and cross-version modeling method, realizes multi-version vulnerability automatic identification, with the advantages of strong intelligence, high adaptability, detection precision.

The invention relates to the field of microwave device modeling, and discloses a microwave device neural network modeling method based on a convolutional auto-encoder and accompanying sensitivity analysis. On the basis of an auto-encoder framework, a convolution auto-encoder is adopted to extract transmission function parameters, and electromagnetic accompanying sensitivity analysis is combined to construct a novel artificial neural network agent model. In the model, the convolution auto-encoder captures local correlation characteristics of an input signal by using a convolution kernel to realize efficient compression and stable reconstruction of high-dimensional data, so that the problems of discontinuity and instability in a traditional method are solved. Meanwhile, electromagnetic accompanying sensitivity analysis is introduced, gradient information of electromagnetic response to geometric parameters is obtained with low calculation cost, the gradient information is fused into the neural network training process, and the precision and generalization ability of the model are remarkably improved.

The invention discloses a metro station emergency evacuation route planning method based on a brain-like pulse neural network, and belongs to the crossing field of traffic transportation and intelligent calculation. According to the method, an evacuation environment map structure is constructed on the basis of a brain-like spiking neural network (BSNN), crowd heterogeneity modeling and a personalized path objective function are combined, and the multi-objective path planning method for the subway station large passenger flow scene is provided. According to the method, neural network modeling based on a topological graph, target function design of multi-type passenger demands, a pulse-driven rapid optimization mechanism and a congestion prediction algorithm based on membrane potential change are fused, and adaptive path adjustment and congestion avoidance and diversion among different demand personnel are realized. The system dynamically generates a personalized path through a time sequence control and path state iteration updating mechanism, and obviously improves the evacuation efficiency and the road utilization rate in a high-density people stream environment. Simulation experiments show that the method can shorten the overall evacuation time by about 48% and improve the path utilization rate by 27% in the subway peak period scene, has good response speed and self-adjusting ability, and is suitable for an intelligent decision support system of a multi-scene emergency evacuation task.

The invention provides a method and a system for quantifying a forest crown space structure of a sand-fixing forest based on a foundation and an unmanned aerial vehicle laser radar, and is applied to the technical field of data processing. According to the method, the fixed sarin quantized multi-source core data is obtained, and the standardized fusion point cloud data set is generated through point cloud preprocessing and registration correction. The method comprises the following steps: performing individual tree and forest stand scale layering on a data set, constructing ternary association nodes, aggregating topological information, and generating a forest crown structure feature association graph; then, core features are extracted through neural network modeling, high-dimensional feature embedded vectors are generated, a quantitative basic model is constructed through combination of hierarchical quantization engine iterative optimization and actual measurement calibration, finally, precision verification and dynamic correction are completed through combination of actual measurement and growth monitoring data, and multi-index requirements are balanced; and outputting a forest crown space structure quantification result adaptive to the Sarmonin full-monitoring scene.

The invention relates to the technical field of power device transient junction temperature evaluation, and discloses a neural network modeling method for IGBT transient junction temperature evaluation based on historical temperature rise characteristics, which comprises the following steps: collecting chip loss, thermal resistance aging grade, chip junction temperature and environment temperature data of an IGBT device under different working conditions, and preprocessing the collected data; constructing a transient junction temperature evaluation model by using an artificial neural network, introducing a Bayesian optimization algorithm to optimize hyper-parameters trained by a neural network model, training the model by using preprocessed data, and establishing a mapping relationship between historical temperature rise characteristic related data and future junction temperature; performing transient junction temperature evaluation on the IGBT device by using the trained artificial neural network model, evaluating the performance of the model, and performing optimization and improvement on the model according to an evaluation result; the technical problem that an existing transient junction temperature evaluation method cannot deal with a multi-chip packaging scene is solved.

The application discloses a tokamak discharge modeling system based on a bidirectional long short-term memory neural network, which is composed of multiple modules, and the modules are low-coupling functional units.The modules at least include a data transfer module, a Batch data access input module, a training self-defined parameter module, a bidirectional long short-term memory neural network modeling module and a data visualization module.The whole system architecture is mainly divided into two paths, namely data training and data modeling.Using the visualization technology, the tokamak discharge experiment personnel can refer to the discharge modeling results and visualize the model training process.The application can one-key type in the experiment proposal stage to model the whole process of the tokamak discharge curve in advance, so that the experiment personnel and the proposal design personnel can check the validity and rationality of the proposal.The application can also be used for assisting data discovery after the experiment.

The invention discloses an underground water flow field physical information neural network modeling method considering spatial heterogeneity, and belongs to the technical field of underground water flow fields, and the method comprises the steps: 1, selecting a two-dimensional underground water unstable flow equation as a physical law, collecting hydrogeological parameters, rainfall, underground water exploitation amount and geographic coordinate data, and carrying out the normalization preprocessing; 2, constructing a PINNs model based on the preprocessed data, dividing a data set according to 7: 3, dynamically adjusting the weight of a loss function by means of an NTK technology, and updating parameters through an optimization algorithm; 3, extracting a simulation residual error of the PINNs model, inputting the simulation residual error into a GTWR model for fitting, performing Kriging interpolation to obtain residual error space distribution, and superposing PINNs simulation results to obtain a final underground water flow field simulation result; according to the method, the weight of the loss function of the PINNs is dynamically adjusted by fusing the neural tangent kernel technology, the spatial heterogeneity characteristics of the hydrogeological parameters can be adapted, the weight imbalance of physical constraints and data information is avoided, and the model is guaranteed to meet the physical law and fit the actual data distribution.