91results about How to "Stable training" patented technology

The invention belongs to the technical field of data asset management, and discloses an enterprise data asset auditing method and system based on big data, and the method comprises the steps: collecting financial information data, business operation data, management decision data and external industry data of an enterprise; preprocessing the collected financial information data, business operation data, management decision data and external industry data to obtain a financial feature data set, a business operation feature data set, a management decision feature data set and an external industry feature data set; performing feature extraction and weighted fusion on the financial feature data set, the business operation feature data set, the management decision feature data set and the external industry feature data set to obtain a comprehensive feature data set; constructing an audit risk prediction model, inputting the comprehensive feature data set into the audit risk prediction model, and predicting to obtain an audit risk coefficient of the enterprise data assets; and a more comprehensive view angle is provided, and the data asset risk of the enterprise can be evaluated more accurately.

The invention discloses an automatic model tuning system based on an AI training intelligent workbench, and relates to the technical field of artificial intelligence, and the system comprises the steps: setting a data collection instrument and an environment monitoring point, collecting model real-time data in a model training process, and carrying out the preprocessing; performing model performance analysis according to the model real-time data and the model original data to obtain final parameter configuration; and performing model training by utilizing the final parameter configuration, dividing the model training process into different stages, dynamically adjusting an intelligent control strategy based on model performance requirements, and performing optimization on the different stages of the model. According to the method, through multi-dimensional performance evaluation, parameter configuration is optimized, and the robustness and generalization ability of the model are improved. Through dynamic adjustment of an intelligent control strategy, fine management of a training process is realized, efficiency and resource utilization are improved, model convergence is accelerated, and finally an efficient, stable and high-adaptability model training effect is achieved.

The invention relates to the technical field of model lightweight processing, and discloses a digital twin model lightweight processing method, and the method comprises the steps: obtaining a digital twin model, constructing a multi-dimensional twin model parameter importance evaluation objective function, and evaluating the digital twin model parameters obtained through training solving; carrying out lightweight processing on the digital twinborn model parameters based on an importance evaluation result to obtain a lightweight digital twinborn model; and carrying out fine tuning on the lightweight digital twinborn model to obtain the lightweight digital twinborn model after fine tuning. Gradient evaluation and channel gain evaluation are carried out according to the gradient of the digital twin model parameters in the training process and the final training result, the digital twin model parameters with high contribution degree to the output result are selected, redundant parameters are filtered, lightweight processing of the digital twin model is achieved, and the training efficiency is improved. And performing fine tuning processing to retain the performance of the original model to obtain the digital twinborn model which reduces the consumption of computing resources and the storage cost.

The invention relates to an intelligent control method for a high-speed aircraft, and the method comprises the steps: building an aircraft control architecture, carrying out the decision-making of a neural network model, training a decision-making neural network, and carrying out the real-time compensation of a control instruction according to the flight state of the aircraft, so as to achieve the control of a variable-configuration high-speed aircraft. An aircraft controller adopts a model + learning double-layer architecture, a decision neural network is designed to be of a Double Network structure, overload errors of pitching, yawing and rolling channels are considered in a reward function design, training of a decision neural network model is based on a double-delay deterministic strategy gradient algorithm, a training process is optimized by adopting a prior experience playback technology and a target strategy smoothing mechanism, and a decision neural network model is obtained. The control system can achieve real-time intelligent adjustment of control instructions, overload quantities and attitude angles of three channels of the aircraft and corresponding rudder deflection instructions output by the controller at the last moment are obtained from the environment, the intelligent body outputs control instructions of pitching, yawing and rolling channels, and therefore intelligent control over the variable-configuration high-speed aircraft is achieved.

The invention relates to a visual attention mobility anti-attack method based on self-adaptive token fine tuning, which comprises the following steps of: constructing a feature important matrix by utilizing a proxy model middle layer, and generating a semantic-guided sparse mask; performing token-level point multiplication on the sparse mask and the initialized disturbance, and performing linear superposition on the sparse mask and the initialized disturbance and an original image to serve as initial input data; inputting the input data into the agent model for iterative training; in each round of iterative training, sparse masks are continuously updated, the number of discarded tokens is gradually increased, and gradient variance between layers in an agent module is reduced through gradient back propagation by using a gradient penalty factor and a balance factor. Meanwhile, self-adaptive truncation factor sets are designed for different modules of the agent model to weaken the attention mechanism of the model, and through T times of iterative training, a target loss function is maximized to obtain optimal mobility disturbance. Experiments prove that the method has better attack effect and mobility on the attack of the target model.

The invention discloses a multi-robot path planning method based on course deep reinforcement learning, and the method comprises the steps: designing each stage of course learning through the thought of course learning, and changing the number of robots in each stage, the size of a map, and the number of obstacles, thereby achieving the task from easiness to difficulty; building a reinforcement learning environment, defining an observation space and an action space of the robot in the environment and a reward function of the environment, and formulating an operation rule of the robot in the environment; a network structure is designed to adapt to the number of robots changing in the environment, a network model is trained through the Qatten algorithm, the multi-robot path planning problem in the continuous space can be solved, and the method has the advantages that compared with a traditional algorithm, path planning can be achieved faster and better, and the path planning efficiency is improved.

The invention provides a method for converting a multi-stage non-iodine contrast CT image into a coronary artery CTA image, and the method comprises the steps: obtaining an NCCT volume and a CTA volume which are matched, carrying out the data preprocessing, and constructing a training data set; respectively constructing an NCCT encoder-decoder network and a CTA encoder-decoder network, and carrying out pre-training based on the training data set; a mark reconstruction network is constructed and trained, a masked low-resolution CTA mark is reconstructed under the condition of an NCCT token obtained by a trained encoder, and then a low-resolution CTA volume is reconstructed by a trained decoder; a super-resolution network is constructed and trained, and a high-resolution CTA image is generated from the low-resolution CTA image; a high resolution CTA volume is generated from the low resolution NCCT volume using the trained multi-stage model. The method has the beneficial effects that the problem of posterior crash is avoided, the decoding speed during reasoning is higher, and the training process is more stable; in the low-resolution stage, the model can learn multi-scale context information, and in the high-resolution stage, the model focuses on recovering details, more fine structures are reserved in final output, and clinical requirements can be met.

The invention discloses a Chinese text grammar error correction method combining syntactic information and a pre-training language model, and the Chinese grammar error correction capability can be effectively improved by fusing the syntactic information into the pre-training language model Chine BART. The method comprises the following steps: importing a data set and converting the data set into a target format; syntactic information is generated by utilizing HanLP, part-of-speech tags and dependency relationships are converted into vectors, and the vectors are fused with default word embedding; in the encoding stage, a syntax-enhanced multi-head self-attention mechanism and a residual network are used; in the decoding stage, a self-attention mechanism and an encoder-decoder attention mechanism are combined with syntactic information for generation, and a decoder is gradually optimized in combination with a residual network; a plurality of candidate sentences are generated by using Beam Search, and the optimal candidate sentence is selected according to syntactic consistency and semantic rationality. According to the method, the syntactic information and the pre-training language model are combined, so that the model learns semantic information of sentences and pays attention to the syntactic structure, and the Chinese grammar error correction capability of the model can be improved.

The invention discloses a V2X cooperative sensing method and system based on a lightweight domain adapter. According to the invention, a lightweight domain adapter network is placed in a feature map input branch of each cooperative vehicle of a first V2X cooperative sensing network carried by an own vehicle to form a second V2X cooperative sensing network, and cooperative sensing of the own vehicle and the cooperative vehicles is realized by using the trained second V2X cooperative sensing network. According to the invention, by introducing the domain adapter network, collaborative 3D target detection under the condition of heterogeneous feature maps of the own vehicle and the collaborative vehicle is successfully realized. According to the method, under the condition that the whole basic cooperative detection network does not need to be retrained, the domain adapters corresponding to different feature maps are inserted, so that the domain difference between heterogeneous feature maps is effectively reduced, the detection performance and scene adaptability of cooperative 3D target detection based on V2X are greatly improved, and the detection efficiency is improved. The method has a high practical value for improving the collaborative perception performance of automatic driving.

The invention relates to the technical field of radar weak target detection, in particular to a radar weak target detection method and device based on deep learning. According to the technical scheme, the method comprises the following steps: data preprocessing and enhancement: receiving an original echo signal of a radar, generating three-dimensional time frequency-wavelet coefficient spatial data by combining time frequency analysis with wavelet transform, and expanding a data set in multiple modes; and feature extraction network construction: constructing a 3D-CNN and LSTM fused network, wherein the 3D-CNN has a convolution operation with a variable convolution kernel size. According to the method, the data are richer and more comprehensive through time-frequency analysis in combination with wavelet transform and data enhancement, and multi-dimensional features can be accurately captured through 3D-CNN and LSTM fusion and an adaptive convolution kernel; and the training effect is ensured through the adaptive learning rate and the early stop method, and the accuracy and adaptability of weak target detection are jointly improved.

The invention discloses a low signal-to-noise ratio DOA estimation method based on an auto-encoder and a deep neural network, comprising a low signal-to-noise ratio DOA estimation method and a DOA estimation method combining DAE and CNN, the low signal-to-noise ratio DOA estimation method comprises the following steps: giving parameters such as array radar receiving antenna array element number M, array element position dm, signal-to-noise ratio SNR, input signal sin and the like; obtaining an array receiving data covariance matrix R; taking a real part and an imaginary part of a received signal covariance matrix R in an incidence angle [-60 degrees, 60 degrees] range as input Y1 of the neural network, constructing a data set, and distinguishing a training set and a test set according to a proportion of 8: 2; dAE network parameters are initialized, data in the training set are imported, forward propagation is completed, meanwhile, back propagation is conducted on errors, and network weights are synchronously updated; a DAE network structure is further introduced to optimize signal processing, the DOA estimation accuracy is improved, and the requirement for complex computing resources is lowered.

The invention discloses a multi-task reinforcement learning method based on an emergence hybrid expert model, and the method comprises the following steps: constructing a multi-task reinforcement learning network which comprises a plurality of feedforward neural networks, and connecting a routing network in front of each feedforward neural network; performing clustering analysis on each feedforward neural network, and equally dividing the feedforward neural network into a plurality of expert modules; training each routing network based on a strategy gradient descent method, and updating parameters of corresponding expert modules; selecting experiment tasks in the multi-task learning environment; and for each experiment task, activating a corresponding expert module based on the trained routing network to perform decoupling processing, and completing multi-task reinforcement learning. According to the method, an efficient and stable optimization strategy is provided for multi-task learning, and the problem of mutual interference between parameter sharing and tasks in multi-task learning is effectively relieved.

The invention discloses a multi-part rPPG fusion physiological signal monitoring method and device and a storage medium. The method comprises the following steps: acquiring a multi-part skin video of a patient and preprocessing; editing physiological part prompt information; multi-part skin videos and physiological part prompt information are input into the model, the model can accurately extract spatial features in the skin videos of different parts under the guidance of a dynamic convolution mechanism, an asynchronous distillation technology and the physiological part prompt information, the rPPG pulse waveform features of all the parts are obtained through time sequence analysis and fusion, and then the rPPG pulse waveform features of all the parts are obtained through weighted fusion. A comprehensive rPPG waveform signal is generated; and finally, post-processing and calculating the comprehensive rPPG waveform signal to obtain a plurality of physiological signal values, and outputting a comprehensive report. The physiological state of the patient is comprehensively monitored through extraction and fusion of the rPPG signals of multiple parts such as the four limbs, and the problem that the rPPG signals cannot be extracted due to the fact that the face and the upper body are shielded is effectively solved.

The invention discloses a block sensing mask evolution method for efficient NM sparse training, and relates to the technical field of neural network compression and acceleration. The method comprises the following steps: for an N: M sparse mode, dividing a weight vector into weight blocks with a fixed size of M; a weight index is calculated for each weight block, and then a corresponding binary mask is adaptively obtained according to the weight index; calculating a sparse architecture difference for each group of weight blocks to obtain a block sensing score, and then selecting a high-frequency mask block; the binary mask and the high-frequency mask block are combined to realize joint optimization, a new mask is obtained, the original mask is updated through the new mask, pruning and weight recovery are performed, and the training loss is reduced. According to the block perception mask evolution method for efficient NM sparse training, sparse forward and backward propagation is executed in the whole training process, and the training cost is greatly reduced.

The invention discloses a model quantification method combined with efficient parameter transfer learning, which comprises the following steps: constructing an M1 model based on a pre-trained ViT model, and adding a low-rank adaptive matrix in an attention layer of the M1 model to obtain an M2 model added with a LoRA module; a plurality of sub-modules in the M2 model are converted into quantization modules, and a quantized M3 model is obtained; setting a weight quantization bit number and an activation quantization bit number, and calibrating the model M3 combined with the preset transfer learning mode to obtain each calibrated model; respectively training each calibrated model according to training logic by utilizing a preset training set; and evaluating the trained QATL-M3 model, the trained PTLQ-M3 model and the trained GQTL-M3 model by using a preset verification set to obtain an evaluation result. The invention relates to the field of deep learning, and solves the technical problems that an existing model quantification method is high in model calculation cost and prone to overfitting under the condition of low-bit quantification.

The invention provides an end-to-end automatic driving model training method and device, and the method comprises the steps: collecting the operation parameters of a vehicle, building an automatic driving model based on an actor-commentator frame, employing a DDPG algorithm fusing adaptive parameter space noise, and carrying out the training of an end-to-end automatic driving model. Initializing parameters of a commentator network and an actor network in the DDPG algorithm; an Adam optimizer is selected, an experience playback pool is initialized, interaction experience of the intelligent agent and the environment is stored, an initial learning rate is set, and the size of a batch is determined; updating parameters of the commentator network and the actor network by using the experience in the experience playback pool; according to the end-to-end automatic driving model training method and device provided by the invention, the learning rate and the batch size are adjusted to be proper, the convergence speed of the model is accelerated, and the learning rate and the batch size in hyper-parameters are adjusted, so that the model is more stable in the training process.

The invention discloses a lithium battery charging and discharging data generation method based on a hierarchical convolution generative adversarial network, and the method comprises the steps: collecting the basic parameter information of the charging and discharging process data of a target lithium battery, and building a real lithium battery charging and discharging data set; features of an input data set are extracted by adopting a convolutional layer, the number of parameters and the calculation amount are reduced through a pooling layer, and finally classification is performed through a full connection layer; constructing a generator and a discriminator of a generative adversarial network (GAN) by using a convolutional neural network (CNN), and optimizing data through adversarial training; the generator learns to generate realistic data samples from random noise, and the discriminator learns to distinguish the realistic data samples from the samples generated by the generator. According to the method, the hierarchical convolution generative adversarial network is adopted to act on lithium battery charging and discharging data generation, so that the model can better capture a complex structure of the data, and real lithium battery charging and discharging data can be better collected.

The invention discloses a polarization degree and light intensity image fusion method based on bidirectional cross attention. The method belongs to the technical field of polarization imaging, computational imaging and image processing. The technical problems that in the prior art, dynamic balance between structural detail keeping and polarization saliency enhancing is generally difficult to achieve, and particularly under the condition of a complex background or a low signal-to-noise ratio, texture missing or noise artifacts are prone to occurring in a fusion result are solved. A fusion mechanism capable of establishing a two-way information interaction relationship between a light intensity image and a polarization degree image is provided, and further processing is carried out aiming at a noise problem, so that more stable and more reliable polarization fusion imaging is realized.

The invention relates to the technical field of image processing, in particular to a training method of a semantic segmentation model and a semantic segmentation method and device. Comprising the following steps: acquiring and preprocessing tree-containing images, and constructing an image set; carrying out image segmentation based on a U-Net architecture, wherein the U-Net architecture comprises an encoder, a decoder and a jump connection; based on Dice coefficient loss and weighted cross entropy loss, determining a total loss function to optimize image background and tree category imbalance; the learning rate is adjusted in each training step through an Adam optimizer in combination with gradient first-order moment estimation and gradient second-order moment estimation, and learning rate attenuation adjustment of the Adam optimizer is achieved in combination with cosine annealing scheduling; and training the semantic segmentation model based on the configured Adam optimizer and cosine annealing scheduling until the semantic segmentation model is obtained through convergence. Therefore, based on the obtained semantic segmentation model, subsequent image segmentation processing application in an education robot can be facilitated.

The invention provides a semi-supervised fine-grained action recognition method based on time sequence disturbance and learning stabilization, and provides a semi-supervised learning fine-grained action recognition algorithm framework SeFAR based on a FixMatch semi-supervised learning system for a highly challenging task of semi-supervised fine-grained action recognition. According to the method, a two-stage time sequence element sampling method is provided, fine-granularity time sequence elements and time sequence contexts are combined, and multi-granularity time sequence information can be effectively captured, so that the capability of distinguishing subtle actions in a video is enhanced; the method comprises the following steps of: obtaining fine-grained time sequence elements, considering the importance of time sequence dynamics and designing a new'strong enhancement 'strategy, introducing proper time sequence disturbance into the obtained fine-grained time sequence elements, and keeping the time sequence of context elements, so that a weak-strong enhancement comparison data pair more suitable for a fine-grained action recognition task is constructed; besides, a self-adaptive learning adjustment strategy is further designed in the method, and back propagation loss in training is adjusted by calculating a self-adaptive coefficient so as to stabilize the training process.

The invention discloses a hydroelectric generating set fault diagnosis data enhancement method based on a diffusion model and generative adversarial training, and belongs to the field of industrial equipment fault diagnosis. According to the method, the data enhancement model combining the diffusion model and the generative adversarial network is constructed, a multi-modal non-Gaussian distribution mechanism and a latent variable control generation process are introduced, the limitation of a single Gaussian hypothesis of a traditional diffusion model is broken through, and richer and more real sample generation is realized. And a condition discriminator and a time sequence modeling mechanism are introduced, so that the training of the model under different noise levels is more stable, and the authenticity judgment capability of the discriminator on the generated sample is enhanced. The problems that an existing hydroelectric generating set fault diagnosis system is insufficient in sample, low in generated sample quality and unreal in sample distribution are solved, and the diversity and quality of data are remarkably improved.

The invention provides a small target road vehicle detection method and device and a storage medium. The method comprises the following steps: S1, constructing a CP6-YOLOv8 model; s2, constructing a data set and training a model; and S3, performing vehicle detection by using the trained model. The problems that in an existing method, precision is not high, and the number of missed detection is large are solved.

The local super-resolution reconstruction algorithm based on the transmission line insulator infrared image includes the following steps: Step 1, image recognition; Step 2, image segmentation; Step 3, super-resolution reconstruction; By using the improved adversarial network structure, unsupervised reconstruction is adopted, the RRDB network structure is adopted, when the network structure is deepened, the stability of training can be maintained, the calculation amount and the use of memory are reduced, and the reconstruction effect is improved, the SeNET attention mechanism is added, from the feature channel level, according to the importance, the useful features such as the outline and the details of the insulator are improved, and the features which are not useful for the program are suppressed. The resolution of the current transmission line insulator infrared image is low, and when the insulator detection is carried out, the workload is large and the problems such as false detection and missed detection occur.

The invention discloses a trajectory prediction system based on convolution and self-attention. The trajectory prediction system comprises an overall network architecture, a training method and a reasoning output algorithm. The whole network architecture takes a road grating map, a vehicle motion state and target information as inputs, the inputs are processed through a DenseNet layer, a self-attention coding layer, a full-connection coding layer and a decoding output layer in sequence, and a prediction track is output; according to the method, a road grating map containing information such as a collision box and a road marking and a vehicle motion state are used as input, road environment constraints are associated with a self-attention mechanism, a polar coordinate track description mode fitting vehicle kinematics characteristics is adopted, a track parameter value range is limited, common errors and abnormal tracks are effectively avoided, and the method is suitable for popularization and application. Traffic rules and driving safety requirements are met, and the reasonability and safety of formed track prediction are prominent; based on 16000-frame multi-dimensional training data collected by a Carla simulator, errors are reduced in a verification set after optimization training, and errors in a crossing turning scene are reduced.

The present application belongs to the technical field of computer vision three-dimensional reconstruction, and specifically relates to a semi-supervised single-view 3D object reconstruction method. A small amount of labeled samples are used to train a neural network, and then the trained neural network is used to generate pseudo labels for unlabeled samples and guide the training of the unlabeled samples. At the same time, the present application proposes that a discriminator scores the quality of the generated pseudo labels, and limits the bias of low-quality pseudo labels on model training. In addition, the present application proposes a prototype shape prior module based on an attention mechanism, which serves as a bridge between images and 3D shapes, reduces the difference between the two modalities, and provides shape prior for the neural network, ensuring that the reconstructed 3D shape is natural. Compared with the current mainstream method in the industry, the present application is more accurate in three-dimensional reconstruction of a single image under the condition that the amount of labeled data is very small, and the generated 3D shape is more realistic and natural.

The invention discloses an encrypted traffic classification method based on feature fusion and a packet dilated convolutional network, which identifies network traffic with high precision and ensures privacy and computational efficiency at the same time, and comprises the following steps: data preprocessing: only retaining 26 bytes to protect privacy through segmentation, filtering and truncation; feature extraction: adopting a double-sliding window mechanism and a designed hash function, generating concrete and abstract features in cooperation with an embedded layer, and realizing joint feature representation; in the model design, firstly, lightweight convolutional network design is carried out, a receptive field is expanded through a lightweight grouping dilated convolutional structure, and channel mixing and point-by-point convolution are added to increase interchannel communication; and in a classification output stage, a cosine annealing mechanism is used to dynamically adjust a learning rate so as to improve a training effect and a convergence speed. The method is superior to an existing lightweight method in the aspects of accuracy, stability, processing speed and resource occupation, is suitable for efficient traffic classification application of resource-constrained equipment, and has practical and promotional values.

The invention discloses a reinforcement learning intelligent traffic signal control method and system based on state-driven retrieval enhancement. The method comprises the steps of collecting real-time traffic flow state information of a traffic environment and constructing a global traffic state vector, retrieving a knowledge base to obtain related knowledge fragments and performing weighted aggregation to obtain a comprehensive knowledge representation vector, and constructing a fusion decision model fusing retrieval enhancement and a Markov decision mechanism, training convergence of the strategy network is completed after the enhanced state vector is generated, traffic signal control is achieved based on the converged strategy network, the generalization adaptation capacity, the anti-interference capacity and the operation stability of the signal control strategy to various traffic scenes are improved, random fluctuation of traffic flow and sudden traffic conditions can be effectively coped with, and the traffic signal control method is suitable for being popularized and applied. Through knowledge retrieval and prior knowledge fusion, traceability and interpretability of a control decision are realized, the efficiency of traffic signal control is effectively improved, the occurrence rate of intersection traffic delay and congestion is reduced, and reliable support is provided for intelligent collaborative management and control of a road network.

This invention discloses a rapid calculation method for the broadband electromagnetic performance of microstrip lines and microstrip-like two-port structures. It relates to the fields of electromagnetic calculation and electromagnetic compatibility-aided design technology, and includes: acquiring and preprocessing the layout data of the microstrip port conductor to be tested, generating a binary layout image; inputting the binary layout image into a pre-trained convolutional neural network, outputting a real-valued encoded prediction vector of the complex mode eigenvalues ​​of the dominant mode at a preset sampling frequency; denormalizing and decoding the prediction vector to obtain complex mode eigenvalue samples; identifying the complex mode eigenvalue samples based on a second-order analytical model and analytically extending them to obtain the broadband modal impedance of the target frequency band; obtaining the port impedance matrix according to the reconstruction relationship, and converting and outputting broadband scattering parameters and electromagnetic performance results. This invention can significantly reduce the computational overhead of full-wave frequency sweep and improve the simulation efficiency of microstrip lines and the iteration speed of EMC design.

The present invention relates to the field of chlorophyll a concentration prediction, and discloses a subsurface chlorophyll a concentration inversion method based on 1D-DSCAM, comprising the following steps: obtaining satellite remote sensing sea surface temperature and sea surface chlorophyll a concentration data and subsurface chlorophyll a concentration, subsurface temperature and salinity profile data of biogeochemical buoys, and preprocessing, constructing a data set, and inputting the training set into a 1D-DSCAM-GRU deep learning model for model training; the model comprises an Embedding module, a 1D-DSCAM module and a multi-layer GRU module; after preprocessing the sea surface temperature and sea surface chlorophyll a concentration data obtained by satellite remote sensing of the sea area to be predicted and the subsurface temperature and salinity profile data obtained by the biogeochemical buoys, the data are input into a qualified model for subsurface chlorophyll a concentration prediction. The method disclosed by the present invention can improve the accuracy of prediction.

The application discloses a traffic signal timing method, device and equipment based on deep reinforcement learning, relates to the field of traffic control technology, and effectively reduces the overall waiting time of vehicles at the intersection by pre-allocating the green light time of each phase according to the queue length of the current intersection lane. The application further adjusts the green light duration of the phase by the good or bad feedback result of the interaction with the traffic condition before the start of the green light time of each phase, so as to realize real-time prediction of the optimal green light time of the current phase. Further, in order to better adapt to the uncertainty and complexity of the traffic flow, the mechanism of the phase is added to the design of the state, so as to avoid the case that different intersections are in different phase green light time when the number of vehicles at different intersections is the same, thereby making the training more stable, and effectively adapting to the reasonable allocation of the green light duration of the traffic signal under the condition of the complex changing traffic flow, so as to reduce the waiting time of the vehicles at the intersection.