1320 results about "Backpropagation" patented technology

The invention discloses an artificial intelligence-based engineering resource allocation optimization method and system, and relates to the technical field of artificial intelligence and resource management crossing. The problems of resource configuration dynamic change and resource conflict coordination are solved. And processing the multi-modal heterogeneous data in the target engineering scene through space-time alignment and semantic coding to generate a dynamic data stream. A task, resource and environment node relationship is constructed based on a dynamic heterogeneous graph network, and a node dependency weight is updated through an event-driven mechanism. Two-stage collaborative optimization is executed, in the first stage, a baseline scheme is generated through Monte Carlo sampling, and in the second stage, resource conflicts are eliminated through back propagation negotiation. The resource dynamic entropy is monitored in real time, a rebalance algorithm is triggered during local overload, and a distribution scheme is adjusted in combination with security constraints. And finally, outputting the optimization scheme to an engineering management system for execution, dynamically returning updated graph network parameters, forming a data acquisition, optimization decision and feedback closed loop, and improving the intelligence and robustness of resource allocation.

The invention relates to the technical field of artificial intelligence, in particular to an artificial intelligence-based image segmentation and dynamic target recognition method, which comprises the following steps of: accurately positioning a candidate region through multi-modal space-time fusion and dynamic confidence coefficient screening; strengthening spatial-temporal feature expression in a layering manner through a multi-level feature decoupler, and generating a multi-dimensional feature enhanced spatial-temporal candidate region; through a deformable segmentation network, a deformation convolution kernel and edge motion matching loss are combined, joint optimization of a geometric boundary and motion continuity is realized, and the segmentation robustness of a flexible target is improved; through optical flow back propagation dynamic correction and confidence coefficient propagation, high-precision segmentation masks with consistent time and space are output; and through a target trajectory re-identification and completion mechanism driven by a graph attention network, and in combination with optical flow deformation prediction, stable tracking in a shielding scene is realized.

The invention provides a computer document intelligent compliance detection system based on deep learning, and relates to the technical field of data processing, and the system comprises the steps: carrying out the topological structure analysis of an adjacent matrix through a graph neural network, so as to detect an abnormality, obtaining a structure deviation degree, and analyzing a decision path node sequence through a pre-training language model to generate semantic conflict features; fusing the structure deviation degree with the semantic conflict feature to generate a risk feature vector; generating an augmented rule set according to the risk feature vector, and updating the formalized rule base; updating parameters of the feature coding component through gradient back propagation; optimizing a differentiable logic layer judgment threshold value and the weight of an analysis module; and obtaining the updated feature coding component, the rule base version and the risk quantification parameter. According to the invention, the efficiency of document compliance detection is effectively improved.

The invention provides an automatic heuristic algorithm planning method based on a large language model, and the method comprises the following steps: carrying out the initialization and problem modeling, starting from a basic heuristic mode through guiding the large language model, generating a candidate algorithm set in combination with a plurality of cognitive perspectives, and providing diversified starting points for a search space; configuring core parameters of Monte Carlo tree search; in each iteration process, planning is started in a heuristic space by utilizing Monte Carlo tree search, and the process is composed of five core stages of selection, reflection, expansion, simulation and back propagation; and after all iterations are completed, the path with the highest average reward and the corresponding optimal heuristic algorithm are returned, and the global optimality of the final solution is ensured. According to the method, the effective experience can be automatically extracted from the heuristic strategy generated historically, and real-time feedback adjustment and strategy induction optimization of the heuristic structure are realized, so that the knowledge migration and generalization ability in the search process is remarkably enhanced.

The invention provides an intelligent drilling speed prediction method based on physical feature guidance and multi-source information fusion, and relates to the technical field of intelligent drilling speed prediction, and the method specifically comprises the following steps: collecting multi-source heterogeneous data from a drilling real-time database, a logging system, a logging system and a geological database; constructing a dual-channel deep learning prediction model, wherein the dual-channel deep learning prediction model comprises a dual-channel convolution feature extraction module, a feature fusion module, a time sequence fusion module, a time sequence modeling module and a full connection layer which are connected in sequence; obtaining a predicted drilling speed by using a dual-channel deep learning prediction model; a joint loss function is constructed by considering a data driving error and a physical constraint error, an error is calculated according to the joint loss function, and network parameters are updated through back propagation; carrying out loop iteration training until convergence; and the trained dual-channel deep learning prediction model is used for drilling speed prediction. According to the technical scheme, the problems that in the prior art, a mechanism model is insufficient in precision, and a data driving model is poor in reliability are solved.

Disclosed in the present invention is a lightweight defect detection method based on hybrid multi-scale knowledge distillation. The method comprises: constructing a dataset; constructing a teacher network model and a lightweight student network model; using the dataset to train the teacher network model, and saving a weight file of the trained teacher network model; and loading into the teacher network model the saved weight file of the teacher network model, inputting defect images in the dataset into the teacher network model and the student network model to respectively obtain first multi-scale features and second multi-scale features, respectively inputting the first multi-scale features and the second multi-scale features into a cascaded knowledge blending module to obtain final deeply fused first multi-scale features and final deeply fused second multi-scale features, then calculating a hybrid multi-scale knowledge loss, and in combination with the prediction loss of the student network model, using a backpropagation algorithm to update network parameters, so as to obtain a trained lightweight student network model for implementing defect detection of intelligent manufacturing products. The cognitive ability and recognition performance for defects of different scales are improved.

The invention discloses a rolling bearing fault diagnosis method based on a multi-scale residual attention network and an adaptive Transform encoder. The rolling bearing fault diagnosis method comprises the following steps: acquiring original vibration data in the running process of a rolling bearing; segmenting the collected original vibration data into samples with specified lengths, and dividing the samples into a training data set and a test data set; inputting the training data set into a multi-scale residual attention network to perform preliminary multi-scale feature extraction; inputting the feature information extracted by the multi-scale residual attention network into an adaptive Transform encoder to obtain time sequence features; finally obtained feature information is subjected to GAP processing and then is input into a Softmax layer for fault diagnosis; the forward propagation calculation and the back propagation calculation are repeatedly executed to optimize model parameters until the diagnosis accuracy and loss of the training data set reach a stable level; and inputting the test data set into the trained model for fault diagnosis, and determining the health condition of the rolling bearing. According to the method, the adaptability and the diagnosis accuracy in time sequence dependence scenes such as rolling bearing fault diagnosis are enhanced.

The invention relates to the field of target detection, in particular to a UMFNet-YOLO-based joint detection algorithm under a low-light condition, which adopts a multi-scale decomposition and high-frequency-low-frequency feature collaborative enhancement strategy, separates image details from global features through high / low-pass filtering, realizes multi-resolution feature complementation in combination with dynamic weight distribution, and improves the detection accuracy. Target contour and texture information degraded in severe weather can be effectively recovered; a parallel channel compression and spatial semantic enhancement module is designed, rain and fog noise interference is suppressed through cross-dimensional adaptive feature screening, and significance expression of a key target is enhanced; an image enhancement network UMFNet is cascaded with YOLOv8, and strong correlation between feature expression and a detection task is ensured by synchronously optimizing an enhancement module and detecting network parameters through back propagation. Compared with a baseline YOLOv8 detector, the method has the advantages that the average precision (mAP50) is obviously improved, and the method is obviously superior to a traditional enhancement-detection separation scheme. The method can be applied to robust target detection scenes in complex environments such as automatic driving and security monitoring.

The invention relates to a training method, a data processing method, electronic equipment and a computer readable storage medium, and relates to the technical field of computers. The training method comprises the steps that a training sample is input into a to-be-trained model, forward propagation is executed, output of the to-be-trained model is obtained, the to-be-trained model comprises a plurality of expert layers, and a specific activation value is not stored in the forward propagation process; determining the value of a loss function according to the output of the to-be-trained model; according to the value of the loss function and the loss function, back propagation is executed, in the back propagation process of each expert layer, a recalculation task of a specific activation value and a first communication task are executed in an overlapping mode, and for each expert layer, the first communication task comprises a communication task of data interaction among multiple devices corresponding to the expert layer.

The invention discloses a road safety early warning method and system based on a mixed precision quantification visual large model, and the method comprises the steps: collecting road traffic safety videos and pictures, and carrying out the preprocessing, data enhancement and marking, thereby forming a diversified data set; a pre-trained visual large model is selected as a teacher model, after fine tuning, output layer and middle layer knowledge is extracted, key features are weighted, and meanwhile, a lightweight neural network is taken as a student model, same input is received, and prediction and middle feature maps are output. And inputting data into the two models and the student model to carry out mixing precision quantification forward propagation, constructing a total loss function containing tasks, knowledge distillation and quantification learning loss, and updating parameters through back propagation. And after training is completed, exporting a quantitative model, and deploying the quantitative model to an edge computing platform to realize safety early warning. The lightweight model can realize rapid reasoning on edge equipment such as a vehicle-mounted road side, and the problem that performance and efficiency are difficult to consider in a traditional model compression method is solved.

The invention discloses a network intrusion detection method and system based on edge attention learning, and a storage medium, and the method comprises the steps: converting an original network flow into a network flow diagram, and constructing a training diagram and a test diagram under the condition that a coarse-grained label and a fine-grained label are reserved; edge embedding representation is obtained through edge feature reservation, adaptive weight distribution and multi-layer feature extraction of the training graph and the test graph; based on the edge embedding representation, performing coarse-grained detection to identify a basic attack category, and performing fine-grained classification by using multi-scale feature fusion related to global graph attributes; and adversarial training: through initializing adversarial disturbance and optimizing disturbance based on loss function gradient iteration, superposing final disturbance into the training graph, and based on loss function back propagation updating, obtaining a trained network intrusion detection model. The method provided by the invention can effectively capture the depth characteristics of the key attack and maintain the stable detection performance in the confrontation environment.

The invention discloses a cross-source data three-dimensional reconstruction method and system based on improved Gaussian sputtering, and the method comprises the steps: collecting an unmanned plane inclined image and a ground panoramic image of a target region, and constructing a time-space correlation data set; based on multi-view geometric constraints, space-time coding matching point pairs are established through an adaptive feature pyramid, intelligent incremental cross-source data sparse reconstruction is carried out, and point cloud and camera parameters are output; adopting improved Gaussian sputtering, compressing a three-dimensional Gaussian kernel into a two-dimensional Gaussian primitive through double tangent vector constraint, and fitting surface geometry to realize multi-scale reconstruction; and optimizing primitive parameters by using a differentiatable renderer, completing multi-scale fine reconstruction through gradient back propagation, and generating a high-precision three-dimensional model. According to the method, multi-scale accurate geometric prior input and accurate camera poses are provided for three-dimensional reconstruction, the dependence on professional manual operation in a traditional three-dimensional reconstruction method is greatly reduced, and meanwhile, the geometric accuracy and visual fidelity of a reconstruction result are remarkably improved.

The invention provides a lithium battery SOH estimation method based on a time sequence diagram neural network under physical constraints. The method comprises the following steps: collecting battery charge-discharge cycle data; time-related statistical characteristics and IC curve characteristics are extracted from the battery charge-discharge cycle data, and analysis characteristics are obtained after PCC analysis; calculating a time sequence distance between analysis features based on a sliding window and multi-channel collaborative dynamic time warping, and constructing a graph structure; constructing a time sequence diagram neural network, firstly extracting spatial features through diagram convolution, and then modeling a degradation law through time convolution to obtain a preliminary SOH predicted value; defining a physical residual error loss and a monotonicity regular term, and performing weighted combination on the physical residual error loss and the monotonicity regular term and data fitting loss to construct a joint optimization objective function; based on the optimization target, a final constrained SOH predicted value is obtained through a back propagation training model; therefore, in combination with a physical information constraint mechanism, high-precision and physically consistent SOH prediction is realized.

The invention discloses a sparse view angle three-dimensional reconstruction method based on 3DGS. The method comprises the following steps: providing a Gaussian field initialization model; according to the method, the FPN, the 2D U-Net and the MLP are used for forming a Gaussian parameter initialization network, the network integrates high-level and low-level semantic information, and the understanding of the network on multi-scale features is enhanced. And then a complete Gaussian field model is obtained by combining initial point cloud coordinates generated by DUSt3R. And then performing projection and rasterization under different visual angles on the 3D Gaussian ball by taking three-dimensional Gaussian splashing as a main body frame of an algorithm, performing adaptive adjustment of cloning and pruning on the 3D Gaussian ball with relatively large gradient change, and outputting an output result which is optical information and depth information of each visual angle image after rendering. Then, in gradient back propagation, pixel value absolute value deviation between the RGB image obtained by rendering and the RGB true value is supervised through L1 norm loss; the scene rendering quality is high, calculation is simple, and engineering implementation is easy.

The invention relates to the technical field of computer-generated holographic three-dimensional display and deep learning, in particular to an unsupervised deep learning method for realizing three-dimensional holographic display. The method comprises the steps of generating a depth map corresponding to a two-dimensional image; the depth image and the two-dimensional image are spliced in the channel dimension to serve as input of a hologram encoder, and a double-U-Net cascade neural network architecture serves as the hologram encoder; angular spectrum diffraction back propagation is carried out through the generated pure phase hologram to realize three-dimensional scene discretization reconstruction, a reconstructed image with a specified depth is obtained, a depth map is uniformly quantized to obtain a plurality of binary masks, loss calculation is carried out on the reconstructed image and a target image superposed with the corresponding depth binary masks, network parameter optimization is carried out, and a target image with the depth corresponding to the target image is obtained. And when the training of the double U-Net cascade neural network architecture is converged, the training stage is ended. According to the method, high-quality three-dimensional hologram reconstruction is realized through layered angular spectrum propagation, and the method has relatively high precision and detail reduction capability.

The invention provides a multi-modal-based advertisement recommendation method and a related device, and systematically solves the technical bottleneck of a traditional recommendation technology in a complex scene through multi-modal feature alignment, dynamic weight optimization and causal effect decoupling. The method comprises the following steps: firstly, respectively extracting modal features of a video, a text and a user behavior sequence, and constructing a cross-modal contrast loss function to strengthen multi-modal semantic alignment; secondly, dynamically updating each modal weight through back propagation, and realizing advertisement recall and sorting in combination with cross-modal similarity calculation; an anti-fact causal reasoning module is further introduced, interference of environment mixed variables on the recommendation effect is eliminated through tendency score estimation and anti-fact result prediction, and the real causal effect of advertisement exposure is accurately quantified. According to the scheme, representation learning, dynamic decision and causal inference are deeply fused, a generalized and anti-noise technical framework is provided for short video advertisement recommendation, and the performance boundary is remarkably superior to that of traditional collaborative filtering, matrix decomposition and other methods.

The invention provides a spatial-temporal feature fusion-based dynamic graph convolution electroencephalogram depression detection method, which comprises the following steps of: firstly, segmenting a sample into fragments with the length of 1 second, and calculating power spectral density (PSD) as an input feature by adopting a Welch method; the time sequence and spatial characteristics of the EEG signals are synchronously extracted through a double-branch architecture, wherein one branch captures the long-term time sequence dependence of the EEG signals by using a GRU; and the other branch adopts an improved TSCN (separable convolution is introduced), multi-scale spatial features from fine to rough are extracted through causal convolution and expansion convolution of residual layered stacking, after double-branch features are adaptively fused based on an attention mechanism, a dynamic graph structure is constructed, functional connection evolution of brain intervals is modeled by using a graph convolution network, and a dynamic graph structure is constructed. The topological structure of the network is optimized through a back propagation process, and finally depression identification is realized through a Softmax classifier. According to the method, the time sequence modeling capability of the GRU and the multi-scale spatial analysis capability of the TSCN are fused, the representation limitation of a single model is broken through, the dynamic change of a brain function network is adaptively captured through dynamic graph convolution, the physiological interpretability is enhanced, deep complementary fusion of EEG spatial and temporal characteristics is realized, the depression recognition accuracy is remarkably improved, and the method is suitable for popularization and application. And an efficient tool is provided for auxiliary diagnosis of mental diseases.

The invention discloses a dual-channel time-frequency fusion driving Mama main power distribution network detection and classification method, which comprises the following steps: S1, collecting fault voltage and current signal data under different main power distribution network topologies, and constructing a training set; s2, constructing a model based on dual-channel time-frequency fusion driving Mama, wherein the model is used for calculating and obtaining the fault probability and the fault category label of the kth fault; s3, the training set trains the model by using a back propagation and gradient descent method to obtain a trained main and distribution network fault detection network which is used for inputting a fault data set and then mapping a corresponding fault classification label; and S4, inputting fault voltage and current signal data of the main power distribution network by using the trained model, and executing detection classification operation. According to the method, through multi-scale feature reconstruction and cross-modal fusion, the detection robustness in a complex noise environment is remarkably improved, and the fault starting moment, the duration time and the propagation path are accurately captured.

The invention discloses a social robot identification method and system based on deep learning, and relates to the technical field of robots, and the method comprises the steps: carrying out the smooth processing of a tweet through a large language model, and generating a tweet fused with expression semantics in combination with a natural language processing model; capturing an emotion expression difference between a robot account and a real user; generating a comprehensive feature vector of global context sensing; outputting fusion features; the output fusion features are mapped to a high-dimensional space through a linear layer, and the detection probability of a robot account is output through an activation function; and based on adaptive moment estimation, gradient back propagation is carried out by using a cross entropy loss function, and parameters of the graph convolutional network model are optimized to obtain a final classification result. According to the method, the problems of sparse text expression and emotion information loss are effectively relieved, and the separability of the social robot and the real user in emotion behavior modes is enhanced.

A machine learning model may be configured for training using an associated learning technique. A model configured for end-to-end backpropagation may adapted for associated learning by introducing functions for projecting hidden vectors and labels to a shared representation space and for reconstructing labels from representation vectors. An associated learning loss may be calculated at each layer, with the resulting gradients backpropagated locally through that layer rather than all layers. A reconstruction loss may be calculated using each layer's output including the predicted label. Training by associated learning may be parallelized (e.g., layer by layer) to yield efficiency gains. In addition, associated learning training may be more robust to training label errors. The resulting model may be used to, for example, predict data sequences in an autoregressive manner in which subsequent portions of the output data sequence are predicted in part based on previous predicted portions of the output data sequence.

The invention discloses a fine tuning system of a large-scale pre-training model in a federated learning environment and application thereof. The system comprises a local disturbance gradient estimation module, a differential privacy protection module and a global model aggregation and update module. The local disturbance gradient estimation module is used for calculating a global model loss value by combining forward propagation with a zero-order optimization method so as to estimate a gradient and realize all-parameter fine tuning; the differential privacy protection module performs differential privacy protection processing on the estimated disturbance gradient to prevent gradient information from leaking user sensitive data; and the global model aggregation and update module reconstructs a disturbance vector and completes global model update based on a random seed and a scalar gradient uploaded by a client. Compared with the prior art, on the premise of not depending on back propagation, all-parameter fine tuning of a large-scale pre-training model is achieved, data privacy is guaranteed, meanwhile, calculation and memory expenses are remarkably reduced, and the method is suitable for a resource-limited distributed calculation environment.

A hybrid neural network architecture is disclosed that integrates matrix multiplication-free (MatMul-free) transformation layers with spiking neural network (SNN) layers for efficient, low-power computation. The system includes an interface module configured to convert intermediate continuous-valued data from MatMul-free layers into a spike-compatible format using encoding techniques such as rate coding, phase coding, or threshold-based conversion. The SNN layers process the spike-encoded data in an event-driven manner, enabling sparse, temporal inference. Training is supported by a hybrid optimization strategy combining backpropagation in MatMul-free components with surrogate gradient descent or spike-timing-dependent plasticity (STDP) in SNN layers. The architecture reduces computational complexity, supports real-time adaptability, and enables deployment in energy-constrained environments such as edge devices and neuromorphic platforms. The system may be implemented in hardware, software, or a co-designed pipeline optimized for dynamic sensor data, control signals, or continuous inference tasks.

A sample batch is accessed and a conditioning is computed based on given constraints. A dense relaxation is computed based on the computed conditioning and the sample batch to generate kernels for conditioning a diffusion optimization model. A sample is selected from the sample batch based on a randomly sampled timestep and random noise is sampled from a Gaussian distribution. A noisy representation of the selected sample is sampled based on the randomly sampled timestep and the sampled random noise. The diffusion optimization model is run in a forward direction using the noisy representation of the selected sample to generate a prediction. An error loss is computed to minimize an error between the prediction and a target; the target is the sampled random noise. A diffusion optimization loss is computed based on the error loss and the diffusion optimization model is updated based on the diffusion optimization loss using backpropagation.

The invention discloses an air conditioner resource allocation method and system based on a cold accumulation prediction algorithm, and relates to the field of air conditioner cold accumulation, and the method comprises the steps: collecting historical load operation data of an air conditioner, analyzing the historical load operation data through employing a grey relational analysis algorithm, and constructing a cold load prediction model in combination with the cold accumulation prediction algorithm; training the cold load prediction model by using a back propagation neural network algorithm, inputting load operation data collected in real time into the trained cold load prediction model, and outputting a cold load prediction result; and in combination with the cooling load prediction result and the electricity price information, an air conditioner resource allocation strategy is formulated, and a model prediction control algorithm is used for optimizing the allocation strategy. Through integration of the load prediction technology and the real-time electricity price information, the cold load demand of the whole day can be accurately predicted, energy waste is reduced, the energy utilization efficiency is improved, the dynamic electricity price market can be responded, and economic maximization is achieved.

The invention discloses a multi-mode bearing residual life prediction method and system based on balance optimization. The method comprises the following steps: firstly, in an offline stage, acquiring time sequence data of a vibration signal from a bearing operation data set, and generating a time-frequency image through wavelet transform to ensure that the data comprises two modes of a time sequence and an image; secondly, extracting time sequence features of the time sequence data by using a long short-term memory (LSTM) network, and extracting spatial features of a time-frequency image by using a graph convolutional network (GCN); then, features extracted by the LSTM and the GCN are spliced to generate a joint feature vector, and the joint feature vector is input into a full connection layer to output an RUL predicted value; in the training process, the loss contribution difference of the LSTM mode and the GCN mode is monitored in real time through an adaptive optimization strategy, the gradient update amplitude of each mode is adjusted through back propagation, the optimization of the dominant mode inhibiting the vulnerable mode is avoided, and the full extraction of the multi-mode features is ensured.

To create high-performance metasurface devices (110) in an inverse design process over a large design space (100), a deep neural network may be used as a surrogate model in lieu of a full physics simulation to more efficiently predict figures of merit for given input metasurface topologies during the iterative topology optimization. The neural network may also serve to efficiently compute, via backpropagation, gradients of the figures of merit with respect to design parameters, as are used to update the topology in each iteration.

The invention relates to a textile product defect identification method based on improved YOLOv11. The method comprises the following steps: acquiring a textile product defect image data set; performing pretreatment; dividing into a training set and a verification set; the method comprises the following steps: introducing MConv into a YOLOv11 backbone network, adding a CCIAP module behind a C2PSA module, and applying BiFPN in a path aggregation network; performing prediction through YOLO Head to obtain N prediction feature maps; the overall loss of the network is calculated, and network parameters are optimized through back propagation; predicting the verification set image through a network to output AP values of various categories; repeating the above steps to obtain a trained YOLOv11 network; and detecting the test image or video by using the trained detector to obtain a detection result. According to the method, the MConv is introduced into the YOLOv11 network to enlarge the receptive field, the CCIAP module is added behind the C2PSA to improve the feature extraction capability, and the BiFPN is applied to the Neck layer to enhance the feature fusion capability, so that the target detection precision is improved and the real-time detection of textile product flaws is realized under the condition that the reasoning speed is not influenced.

The invention discloses an autoregressive language model training process-oriented data leakage risk quantification method, which comprises the following steps of: executing enhanced member data division processing on an original training text set, and generating a ternary partition text set containing member, non-member and key boundary texts through an optimization function; for each text, using the target model to extract member attributes from three channels of forward reasoning, back propagation and state evolution, and fusing the member attributes into member attribute vectors; inputting the member attribute vector into a hierarchical comparison embedded network, and mapping the member attribute vector into an optimized embedded representation vector through comparison learning including similar clustering, heterogeneous separation and boundary positioning; and inputting the embedded representation vector into a dynamic reasoning classifier capable of perceiving a training stage, judging the risk membership degree of the dynamic reasoning classifier, and generating a data leakage risk quantification result. According to the invention, real-time and fine-grained dynamic quantification can be carried out on the leakage risk, and timely early warning is provided for model training.

The invention discloses a chest radiology medical report generation method based on a knowledge graph, and the method comprises the steps: obtaining a to-be-processed chest radiology image, inputting the to-be-processed chest radiology image into a trained medical report generation network, and obtaining a chest radiology medical report. The training process of the network comprises the following steps: acquiring a chest radiology medical report set; constructing a chest radiology knowledge graph; extracting text features, image features and knowledge features from the chest radiology medical report, inputting the text features, the image features and the knowledge features into a double-path knowledge enhancement gating module, aligning the text features and the image features, aligning the image features and the knowledge features, and calculating alignment loss; fusing the aligned features to obtain enhanced fusion features; inputting the enhanced fusion features into a classifier, and calculating classification loss; the fusion features are input into a decoder, a chest radiology medical report is generated, and generation loss is calculated; and network parameters are optimized through loss back propagation. According to the invention, the medical report generation quality and accuracy can be obviously improved.

A traffic anomaly detection method and system based on improved BERT integrating contrastive learning is provided, the method includes: obtaining traffic data and preprocessing the data; building an improved BERT model including an embedding layer and 12 Transformer encoder networks; performing weight sharing operation among a first 6 Transformer encoder networks and a last 6 Transformer encoder networks, respectively; building a classification network; building a total loss function based on a cross-entropy loss and a contrastive loss; performing unsupervised pre-training on the improved Bidirectional Encoder Representations from Transformers, BERT model; fine-tune training the improved BERT model; updating model parameters through backpropagation to obtain a trained improved BERT model; feed test traffic data into the trained improved BERT model; and obtain a traffic detection outcome. The present disclosure enhances the generalization ability of the model while maintaining stability and accuracy.