An artificial intelligence driven digital exhibition hall dynamic display method

Abstract

This invention discloses an AI-driven dynamic exhibition method for digital exhibition halls, relating to the fields of digital display and machine learning technologies. The specific steps of this method are as follows: first, multi-dimensional data is collected, standardized, preprocessed, and stored locally; then, a federated learning system is built, an encrypted communication link is established, and relevant parameters are configured; next, a model is trained based on the local dataset, outputting only parameter gradients; subsequently, the gradients are uploaded and aggregated by a federated server to generate a global model; after distribution, the model is fine-tuned using real-time data to obtain a locally adapted model; finally, a personalized exhibition strategy is generated, achieving multi-exhibition hall linkage and adaptation. This invention, through the collaboration of a federated learning framework and a dynamic weight feature fusion algorithm, ensures data privacy and breaks down information silos. Combined with a loss function with exhibition adaptation constraints and a gradient confidence aggregation algorithm, it improves model adaptability and reliability, enabling multi-exhibition hall linkage and reuse of high-quality experience, enhancing exhibition accuracy and user experience, and reducing operating costs.

Description

Technical Field

[0001] This invention relates to the fields of digital display and machine learning technology, specifically to an artificial intelligence-driven dynamic exhibition method for digital exhibition halls. Background Technology

[0002] With the widespread application of digital technology in cultural display, corporate promotion, and science education, digital exhibition halls have gradually replaced some traditional exhibition halls, becoming an important carrier for information transmission and interactive experiences. These halls, relying on technologies such as high-definition projection, touch interaction, and 3D modeling, can provide visitors with immersive and diverse experiences, meeting their needs for personalized and interactive experiences. At the same time, more and more exhibition halls are adopting chain operation or cross-regional collaborative display models, hoping to integrate the resources and experience of multiple exhibition halls to improve the overall exhibition level. The maturity of artificial intelligence and big data technologies has made intelligent optimization of digital exhibition halls possible, and exhibition strategy adjustment schemes based on user behavior data are gradually being applied.

[0003] Current digital exhibition hall optimization technologies still have many shortcomings. First, most solutions rely solely on local data from a single exhibition hall to train models. Due to privacy and data security concerns, cross-exhibition hall data is difficult to share directly, resulting in limited data sample coverage for model training, insufficient generalization ability, and an inability to adapt to the different scenarios of various exhibition halls. Second, feature fusion often uses fixed weight settings, which cannot be dynamically adjusted according to the correlation between data and exhibition effects, easily leading to the neglect of some key features and irrelevant features interfering with model training. Third, when aggregating model parameters across exhibition halls, only the dataset size is often considered, without fully taking into account data quality and the stability of parameter gradients, resulting in the global model training effect being significantly affected by low-quality data. In addition, there is a lack of effective linkage mechanisms between exhibition halls, making it impossible to reuse exhibition experience. Each exhibition hall needs to invest a lot of resources in independent optimization, which increases operating costs and fails to meet visitors' expectations for a coherent and rich visiting experience. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an AI-driven dynamic exhibition method for digital exhibition halls. This method, through the collaboration of a federated learning framework and a dynamic weight feature fusion algorithm, ensures local data retention in each exhibition hall while breaking down information silos. By combining a cross-entropy loss function with exhibition adaptation constraints and a gradient confidence federated aggregation algorithm, it improves model adaptability and training quality. After local fine-tuning, it generates an exhibition optimization model adapted to the local environment, realizing the generation of personalized exhibition strategies and multi-exhibition hall linkage. This not only avoids data privacy risks but also reuses high-quality exhibition experience, thereby effectively improving the accuracy of exhibitions and the user experience, and reducing the operation and optimization costs of individual exhibition halls.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an artificial intelligence-driven dynamic exhibition method for digital exhibition halls, the specific steps of which are as follows: S1, Multi-source data acquisition and standardized preprocessing: Each exhibition hall collects local multi-dimensional data through distributed sensing units. After cleaning, normalization and feature fusion processing, a standardized local training dataset is generated. The standardized local training dataset is stored locally in the corresponding exhibition hall throughout the process.

[0006] S2, Federated Learning System Setup and Initialization: Construct a federated learning system that includes a federated server and local servers in each exhibition hall, establish encrypted communication links, and configure the basic structure, training hyperparameters, and parameter aggregation rules of the global exhibition optimization model.

[0007] S3, Local Exhibition Optimization Model Training: Based on the standardized local training dataset, the local servers of each exhibition hall perform localized training on the initial global exhibition optimization model to obtain the local exhibition optimization model; the training process only outputs the gradient of the model parameters.

[0008] S4, Cross-exhibition hall model parameter aggregation optimization: The local servers of each exhibition hall encrypt and upload the model parameter gradients to the federated server. The federated server completes the parameter gradient aggregation according to the preset aggregation rules and generates a globally shared exhibition optimization model.

[0009] S5, Global Model Localization Fine-tuning: The federated server distributes the global shared exhibition optimization model to the local servers of each exhibition hall, and combines the real-time operation data of the exhibition hall to complete the model fine-tuning, so as to obtain the final exhibition optimization model adapted to the local scene; S6, Dynamic Exhibition Strategy Generation and Linked Execution: Each exhibition hall generates a personalized exhibition strategy based on the final exhibition optimization model, and at the same time, according to the collaborative rules in the globally shared exhibition optimization model, the exhibition strategies of multiple exhibition halls are linked and adapted.

[0010] Furthermore, in step S1, the multi-dimensional data includes user behavior data, exhibition operation data, and environmental perception data. User behavior data is collected through high-definition cameras and touch-interactive terminals deployed in exhibition hall corridors and exhibit showcases, specifically including user visit trajectory coordinate sequences, single exhibit dwell time, interaction operation types, and trigger frequencies. Exhibition operation data is collected through the exhibition hall equipment monitoring unit, specifically including exhibit display resolution, playback frame rate, projection device power, and interactive terminal response latency. Environmental perception data is collected through temperature and humidity sensors and infrared crowd sensors, specifically including real-time crowd density in the exhibition hall, ambient light intensity, and temperature and humidity values.

[0011] Furthermore, the feature fusion processing in step S1 employs a dynamic weighted feature fusion algorithm to achieve semantic alignment of multi-dimensional data. The specific steps are as follows: Feature extraction was performed on user behavior data, exhibition operation data, and environmental perception data to obtain user feature vectors. Display feature vectors and environmental feature vectors ; Calculate the mutual information value between each feature and the exhibition effect. , , The formulas for calculating the dynamic weight coefficients and the fused feature vectors are as follows: in, This represents the fused feature vector, which is the final feature vector after semantic alignment of multi-dimensional data. The mutual information value represents the relationship between the user feature vector and the display effect, and is used to characterize the correlation between the influence of user behavior features on the display effect. This represents the dynamic weighting coefficient of user characteristics, which is adjusted in real time according to the mutual information value. This represents a user feature vector, which is a combination of features extracted from user behavior data; The mutual information value represents the relationship between the exhibition feature vector and the exhibition effect, and is used to characterize the correlation between the influence of the exhibition operation features and the exhibition effect. This represents the dynamic weighting coefficient of the exhibition features, which is adjusted in real time according to the mutual information value. This represents the exhibition feature vector, which is the combination of features extracted from the exhibition operation data; The mutual information value between the environmental feature vector and the exhibition effect is used to characterize the correlation between the influence of environmental perception features on the exhibition effect. This represents the dynamic weighting coefficient of environmental characteristics, which is adjusted in real time according to the mutual information value. This represents the environmental feature vector, which is a combination of features extracted from environmental perception data. express , , The summation result; This represents a vector concatenation operation, used to combine feature vectors of different dimensions into a fused feature vector of a unified dimension.

[0012] Furthermore, in step S2, the encrypted communication link of the federated learning system is constructed using the SSL / TLS protocol, specifically including three stages: identity authentication, data encryption, and message integrity verification. When the federated server establishes a connection with the local server in each exhibition hall for the first time, two-way identity authentication is completed through digital certificates. After successful authentication, the AES-256 encryption algorithm is used to encrypt the transmitted model parameter gradients and control commands. During transmission, a message digest is generated using the SHA-256 hash algorithm. The global exhibition optimization model adopts a deep learning network structure, with the input layer being the fusion feature vector X obtained in step S1, the hidden layer containing 3-5 layers of fully connected networks and 1 layer of attention mechanism, and the output layer being the exhibition strategy optimization parameter vector Y. The exhibition strategy optimization parameter vector Y includes content filtering weights, presentation format coefficients, and path planning parameters.

[0013] Furthermore, the local training process in step S3 employs a cross-entropy loss function with presentation adaptation constraints to optimize model parameters. The specific steps are as follows: Initialize the parameter matrix W and bias vector b of the global display optimization model; randomly sample batches of samples with batch size B from the standardized local training dataset, and input the fused feature vector X into the model to obtain the display strategy optimization parameter vector Ŷ predicted by the model; the cross-entropy loss function with display adaptation constraints is calculated as follows: in, This represents the loss function value, used to measure the deviation between the model's predicted values ​​and the actual values, as well as the suitability of the presentation strategy. This represents the label value of the actual display effect of the i-th sample, with a value range of 0-1; This represents the model's predicted value for the i-th sample, i.e., the predicted display effect label value; i represents the sample number; B represents the batch sample size, i.e., the number of samples input into the model in a single training iteration. represents the regularization coefficient, with a value ranging from 0.001 to 0.01, used to suppress model overfitting; W represents the parameter matrix of the global presentation optimization model; The square of the L2 norm is used to measure the magnitude of the weight matrix. This represents the exhibition adaptation constraint coefficient, with a value range of 0.01-0.05, used to enhance the adaptability between predicted values ​​and actual exhibition scenarios; This represents the baseline value for the actual exhibition scenario adaptation corresponding to the i-th sample, which is determined by the effect value corresponding to the best historical exhibition strategy in the exhibition hall. This represents the summation operation; The gradient of the model parameters ∇ is obtained by differentiating the loss function L using the stochastic gradient descent algorithm. With ∇b, update the parameter matrix W and the bias vector b through gradient descent; where ∇ Let ∇b represent the gradient vector of the parameter matrix W, and let ∇b represent the gradient vector of the bias vector b. Repeat the above two steps until the loss function L converges or the preset number of training iterations is reached, to obtain the local presentation optimization model and the corresponding model parameter gradients ∇W and ∇b.

[0014] Furthermore, the parameter aggregation rule in step S4 adopts a federated aggregation algorithm with gradient confidence, and the specific aggregation steps are as follows: Each exhibition hall's local server uploads the model parameter gradients ∇W and ∇b to the federated server via the encrypted communication link in step S2. The federated server performs integrity verification on the uploaded parameter gradients and removes abnormal gradients with missing values ​​≥5%. The dataset quality weight αk and gradient confidence βk for each exhibition hall are calculated. The formula for calculating the comprehensive weight γk for each exhibition hall is as follows: in, This represents the dataset quality weight of the k-th exhibition hall, used to characterize the contribution of the k-th exhibition hall dataset to the global model aggregation; This represents the number of samples in the standardized local training dataset for the k-th exhibition hall; The quality score of the k-th exhibition hall dataset is represented by a value ranging from 0 to 1. It is obtained by weighting data completeness and annotation accuracy, with weights of 0.6 and 0.4, respectively. This represents the gradient confidence level of the k-th exhibition hall, used to characterize the stability of the gradient of the parameters of that exhibition hall; This represents variance calculation, used to measure the gradient of the parameter of the k-th exhibition hall. The degree of fluctuation; The weight represents the overall weight of the k-th exhibition hall, which combines dataset quality and gradient confidence; k represents the k-th exhibition hall, k=1,2,...,n, where n is the total number of exhibition halls participating in the collaboration; This represents a summation operation, which sums the corresponding items across all participating collaborative exhibition halls; the formula for calculating the global parameter gradient is: in, This represents the gradient of the global parameter matrix, which is the gradient used to update the parameter matrix W of the global presentation optimization model after aggregation. The gradient of the parameter matrix represents the local exhibition optimization model for the k-th exhibition hall; This represents the gradient of the global bias vector, which is the gradient used to update the bias vector b of the global presentation optimization model after aggregation. This represents the gradient of the bias vector of the local exhibition optimization model for the k-th exhibition hall; Using global parameter gradient and Update the parameters of the initial global display optimization model to generate the global shared display optimization model.

[0015] Furthermore, the specific method for verifying the integrity of the parameter gradient is as follows: calculate the similarity between the uploaded parameter gradient and the preset gradient template. When the similarity is ≥95%, the gradient is determined to be complete. When the similarity is <95%, the proportion of missing gradient values ​​is further calculated. When the proportion of missing values ​​is <5%, the missing gradient is completed by interpolation. When the proportion of missing values ​​is ≥5%, the gradient is determined to be abnormal and removed.

[0016] Furthermore, the localization fine-tuning process in step S5 specifically includes: Each exhibition hall's local server collects real-time operational data from the past hour. After standardization processing in step S1, a fine-tuning dataset is generated. The number of samples in the fine-tuning dataset is 10%-20% of the standardized local training dataset. The parameters of the globally shared exhibition optimization model are used as initial parameters and loaded into the local training framework. Training is performed using the fine-tuning dataset. During training, the learning rate is set to half of the learning rate in step S3, and the number of iterations is set to 1 / 5 of the number in step S3. During training, the accuracy of the model's presentation effect prediction on the fine-tuning dataset is calculated in real time. When the accuracy is ≥90% or the accuracy fluctuation is ≤1% for three consecutive iterations, the fine-tuning is stopped, and the final presentation optimization model is obtained.

[0017] Furthermore, the process of generating personalized exhibition strategies in step S6 specifically involves: the final exhibition optimization model receiving user feature vectors collected in real time. With environmental feature vectors The system outputs an exhibition strategy optimization parameter vector Y; based on the content filtering weights in parameter vector Y, it selects the top 5 matching exhibits from the exhibition hall exhibit resource library; based on the presentation format coefficient, it determines the presentation method and playback parameters of each exhibit; based on the path planning parameters and combined with the real-time flow distribution of people in the exhibition hall, it generates a congestion-free personalized visit path; the exhibit resource library is associated with the local server of each exhibition hall, supporting real-time updates and access to exhibit information.

[0018] Furthermore, the multi-hall exhibition strategy linkage adaptation in step S6 specifically involves: the federated server synchronizing the output results of the final exhibition optimization model of each hall in real time; when it is detected that a user has completed a visit to a certain type of exhibit in the k-th hall, the associated exhibit information of that type of exhibit is mined through the globally shared exhibition optimization model, and the associated information is pushed to the m-th hall that the user may visit, where m≠k; the final exhibition optimization model of the m-th hall adjusts its local exhibition strategy based on the received associated information, and displays or highlights the associated exhibits in front; during the linkage process, only the exhibit association identification information is transmitted.

[0019] Compared with existing technologies, this AI-driven dynamic exhibition method for digital exhibition halls has the following beneficial effects: First, by combining a federated learning framework with a dynamic weighted feature fusion algorithm, this invention ensures that data from each exhibition hall is retained locally throughout the entire process without being transmitted externally, while also breaking down the information silos problem of traditional digital exhibition halls. The dynamic weighted feature fusion algorithm adjusts the weight coefficients of various features in real time based on the mutual information values ​​between user behavior, exhibition operation, environmental perception, and exhibition effects, making the fusion of data from different dimensions more aligned with actual exhibition needs and avoiding feature failure caused by fixed weights. The encrypted communication and parameter aggregation mechanism of federated learning not only avoids the privacy and security risks associated with cross-exhibition hall data sharing but also integrates the data value of multiple exhibition halls to collaboratively train the model. This allows the trained exhibition optimization model to adapt to the scene characteristics of a single exhibition hall while also absorbing the best practices from multiple exhibition halls, effectively improving the dynamic adjustment capability and adaptation accuracy of the exhibition strategy.

[0020] Second, this invention further enhances the practical value and operational reliability of the exhibition optimization model by combining a cross-entropy loss function with exhibition adaptation constraints and a gradient confidence federated aggregation algorithm. The introduction of exhibition adaptation constraint coefficients ensures that the model training process not only focuses on the deviation between predicted and actual values ​​but also aligns with the adaptation requirements of actual exhibition hall operation scenarios, reducing implementation obstacles caused by model overfitting. The gradient confidence federated aggregation algorithm, when calculating the comprehensive weights, considers both the scale and quality of the datasets from each exhibition hall and the stability of the parameter gradients, avoiding the impact of low-quality data or fluctuating gradients on the global model and ensuring the effectiveness of cross-exhibition hall collaborative training. Simultaneously, the multi-exhibition hall exhibition strategy linkage mechanism allows different exhibition halls to intelligently adjust their exhibition plans based on related exhibit information, achieving efficient reuse of high-quality exhibition experience. This enriches the user's visiting experience, reduces the independent optimization cost of a single exhibition hall, and improves the operational efficiency of the digital exhibition hall.

[0021] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0023] Figure 1 A flowchart illustrating the overall process of an AI-driven dynamic exhibition method for digital exhibition halls; Figure 2 A flowchart illustrating multi-dimensional data semantic alignment using a dynamic weighted feature fusion algorithm; Figure 3 The flowchart illustrates the implementation of cross-exhibition hall model parameter optimization using a federated aggregation algorithm with gradient confidence. Detailed Implementation

[0024] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the specific embodiments, structures, features, and effects of the present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. Embodiments This embodiment, based on federated learning and artificial intelligence technologies, combines a dynamic weighted feature fusion algorithm, a cross-entropy loss function with exhibition adaptation constraints, and a gradient confidence federated aggregation algorithm to achieve collaborative exhibition optimization under multi-exhibition hall data privacy protection. A high-quality training dataset is constructed through multi-source data collection and standardized preprocessing. Cross-exhibition hall model collaborative training is completed using a federated learning system. After localized fine-tuning, a final exhibition optimization model adapted to each exhibition hall scenario is generated. This ultimately achieves personalized exhibition strategy generation and multi-exhibition hall linkage adaptation, effectively solving problems such as high privacy risks associated with data sharing in traditional digital exhibition halls, insufficient model generalization ability, and poor adaptability of exhibition strategies, thereby improving the level of exhibition intelligence and the user's visiting experience.

[0025] The overall process of this method is as follows: Figure 1 As shown, the specific implementation process is as follows: S1, Multi-source data acquisition and standardized preprocessing To fully support the training of the exhibition optimization model, multi-dimensional data reflecting user needs, equipment status, and environmental conditions needs to be collected. Distributed sensing units are deployed in each exhibition hall. User behavior data is collected through high-definition cameras and touch-screen terminals in exhibition hall passageways and exhibit display cases, recording user visit trajectory coordinate sequences, duration of stay at a single exhibit, type of interaction, and trigger frequency. This data directly reflects user interests and visiting habits, serving as the core basis for personalized exhibition strategies. Exhibition operation data is collected by the exhibition hall equipment monitoring unit, including exhibit display resolution, playback frame rate, projection equipment power, and interactive terminal response latency, ensuring the adaptability of exhibition strategies to equipment operating status. Environmental sensing data is acquired through temperature and humidity sensors and infrared people flow sensors, covering real-time people density, ambient light intensity, and temperature and humidity values ​​in the exhibition hall, providing on-site environmental references for adjusting exhibition parameters.

[0026] The raw data collected contains issues such as noise and inconsistent units, requiring standardized preprocessing. The cleaning process removes outliers and duplicate data to ensure accuracy; normalization maps data from different dimensions to a unified range, preventing large numerical differences from affecting model training; the feature fusion stage employs a dynamic weighted feature fusion algorithm to achieve semantic alignment of multi-dimensional data. Features are first extracted from user behavior data, display operation data, and environmental perception data to obtain user feature vectors. Display feature vectors and environmental feature vectors Since the correlation between different features and the exhibition effect changes dynamically with the actual scene, fixed weights cannot meet the requirements of accurate fusion. Therefore, the mutual information value between each feature and the exhibition effect is calculated. , , A higher mutual information value indicates a stronger correlation between the features and the display effect. Based on the mutual information value, dynamic weight coefficients and fused feature vectors are calculated using the following formula: ,in, This represents the fused feature vector, which is the final feature vector after semantic alignment of multi-dimensional data. The mutual information value represents the relationship between the user feature vector and the display effect, and is used to characterize the correlation between the influence of user behavior features on the display effect. This represents the dynamic weighting coefficient of user characteristics, which is adjusted in real time according to the mutual information value. This represents a user feature vector, which is a combination of features extracted from user behavior data; The mutual information value represents the relationship between the exhibition feature vector and the exhibition effect, and is used to characterize the correlation between the influence of the exhibition operation features and the exhibition effect. This represents the dynamic weighting coefficient of the exhibition features, which is adjusted in real time according to the mutual information value. This represents the exhibition feature vector, which is the combination of features extracted from the exhibition operation data; The mutual information value between the environmental feature vector and the exhibition effect is used to characterize the correlation between the influence of environmental perception features on the exhibition effect. This represents the dynamic weighting coefficient of environmental characteristics, which is adjusted in real time according to the mutual information value. This represents the environmental feature vector, which is a combination of features extracted from environmental perception data. express , , The summation result; This represents a vector concatenation operation, used to combine feature vectors of different dimensions into a fused feature vector of a unified dimension. The feature fusion process is as follows: Figure 2 As shown.

[0027] S2, Federated Learning System Setup and Initialization To achieve collaborative training with data privacy protection across multiple exhibition halls, a federated learning system was constructed, comprising a federated server and local servers in each exhibition hall. The federated server is responsible for aggregating and distributing global model parameters, while the local servers in each exhibition hall handle local model training and data processing. An encrypted communication link based on the SSL / TLS protocol is established between the two, comprising three core components: authentication, data encryption, and message integrity verification. Upon initial connection between the federated server and the local servers, two-way authentication is performed using digital certificates to ensure the legitimacy of both parties. After successful authentication, the AES-256 encryption algorithm is used to encrypt the transmitted model parameter gradients and control commands to prevent data theft during transmission. During transmission, a message digest is generated using the SHA-256 hash algorithm to ensure data integrity. Simultaneously, the basic structure, training hyperparameters, and parameter aggregation rules of the global exhibition optimization model are configured. The global exhibition optimization model employs a deep learning network structure. The input layer is the fused feature vector X obtained in step one, and the hidden layers consist of 3-5 fully connected layers for in-depth feature mining, coupled with an attention mechanism layer to highlight the influence of key features. The output layer is the exhibition strategy optimization parameter vector Y, which includes content selection weights, presentation format coefficients, and path planning parameters, corresponding to the three core exhibition requirements: exhibit selection, display method determination, and visitor path planning, respectively. The training hyperparameters are configured to adapt the learning rate, initial iteration count, etc., to the training of the exhibition model, aligning with the scenario characteristics of digital exhibition hall optimization. The parameter aggregation rule adopts a federated aggregation algorithm with gradient confidence, laying the foundation for subsequent cross-exhibition hall parameter collaboration.

[0028] S3, Local Display Optimization Model Training Each exhibition hall's local server trains the initial global exhibition optimization model locally based on a standardized local training dataset. First, the parameter matrix W and bias vector b of the global exhibition optimization model are initialized. The parameter matrix W includes input-hidden layer weights W1 and hidden layer-output layer weights W2, and the bias vector b is used to adjust the model's output bias, with initial values ​​set to default values ​​adapted to the exhibition optimization scenario. Batch training is then performed by randomly selecting batches of samples from the standardized local training dataset. Batch training improves training efficiency and reduces the risk of overfitting. The fused feature vector X is input into the model to obtain the exhibition strategy optimization parameter vector predicted by the model. .

[0029] To ensure both accurate predictions and a close fit to actual exhibition scenarios during model training, a cross-entropy loss function with exhibition adaptation constraints is used to optimize model parameters. The calculation formula is as follows: Where L is the value of the loss function. The term is used to measure the difference between the model's predicted value and the actual value. deviation, For L2 regularization terms, The regularization coefficient is used to suppress model overfitting by limiting the magnitude of the parameter matrix W. To accommodate the constraints of the exhibition, To adapt the exhibition to the constraint coefficient, The baseline value for the actual exhibition scenario corresponding to the i-th sample is determined by the effect value corresponding to the best historical exhibition strategy in the exhibition hall. This constraint ensures that the model prediction results can meet the actual exhibition implementation requirements.

[0030] The stochastic gradient descent algorithm is used to differentiate the loss function L, yielding the model parameter gradients ∇W and ∇b. The gradient direction reflects the optimal direction for parameter adjustment. The parameter matrix W and bias vector b are updated according to the gradient descent rule. The process of sample extraction, model prediction, loss calculation, and parameter update is repeated until the loss function L converges or the preset number of training iterations is reached, resulting in a localized display optimization model adapted to the local scenario and the corresponding model parameter gradients ∇W and ∇b. During training, only the model parameter gradients are output, without disclosing the original data.

[0031] S4, Cross-exhibition hall model parameter aggregation optimization Each exhibition hall's local server uploads the trained model parameter gradients ∇W and ∇b to the federated server via the encrypted communication link established in step two. The federated server first verifies the integrity of the uploaded parameter gradients, removing abnormal gradients with missing values ​​≥5% to prevent low-quality gradients from affecting the overall model performance. To ensure the quality of the global model aggregation, the contribution of each exhibition hall is comprehensively considered, and a dataset quality weight for each exhibition hall is calculated. and gradient confidence , Through formula calculate, To standardize the number of samples in the local training dataset for the k-th exhibition hall, The dataset quality score is calculated by weighting data completeness and annotation accuracy, with weights of 0.6 and 0.4, respectively. It reflects the quantitative and qualitative contributions of the dataset; Through formula The calculation, where Var represents variance calculation, measures the gradient of the parameter of the k-th exhibition hall. The degree of fluctuation, It reflects the stability of the gradient; the smaller the fluctuation, the higher the gradient confidence.

[0032] based on and Through formula Calculate the overall weight of each exhibition hall The overall weighting takes into account the quality and quantity of the dataset, as well as gradient stability, so that gradients corresponding to high-quality data have a higher weight in the aggregation. This is achieved through the formula... The gradients of the model parameters in each exhibition hall are weighted and aggregated to obtain the global parameter gradient. and The parameters of the initial global exhibition optimization model are updated using this global parameter gradient to generate a globally shared exhibition optimization model, enabling collaborative mining of high-quality data value from multiple exhibition halls. The process of this federated aggregation algorithm with gradient confidence is as follows: Figure 3 As shown.

[0033] S5, Localized Fine-tuning of the Global Model The federated server distributes the globally shared exhibition optimization model to the local servers in each exhibition hall. Since the global model is trained collaboratively across multiple exhibition halls, it needs to be fine-tuned locally to suit the real-time scenarios of individual halls, ensuring the model fully adapts to local exhibition needs. Each exhibition hall's local server collects real-time operational data from the past hour, reflecting the latest scene status of the hall. After cleaning, normalization, and feature fusion processing in step one, a fine-tuning dataset is generated. The sample size of the fine-tuning dataset is 10%-20% of the standardized local training dataset, ensuring effective fine-tuning without excessive computational cost. The parameters of the globally shared exhibition optimization model are used as initial parameters and loaded into the local training framework for training using the fine-tuning dataset. To avoid damaging the high-quality parameters of the global model during fine-tuning, the learning rate is set to half of the learning rate in step three, and the number of iterations is set to one-fifth of that in step three, allowing for small adjustments to achieve model adaptation. During training, the model's prediction accuracy for the exhibition effect on the fine-tuning dataset is calculated in real time. When the accuracy is ≥90% or the accuracy fluctuation is ≤1% for three consecutive iterations, it indicates that the model has adapted to the local scene. Fine-tuning is stopped and the final exhibition optimization model is obtained. The local servers of each exhibition hall synchronize the parameters of the final exhibition optimization model after fine-tuning to the federated server for filing.

[0034] S6, Dynamic Exhibition Strategy Generation and Linked Execution Each exhibition hall generates a personalized exhibition strategy based on the final exhibition optimization model, which receives user feature vectors collected in real time. With environmental feature vectors The system outputs a parameter vector Y for optimizing the exhibition strategy. Based on the content filtering weights in Y, the top 5 matching exhibits are selected from the exhibition hall's exhibit resource library. The selection criteria are the degree of fit between user interests and exhibit characteristics. The presentation method and playback parameters for each exhibit are determined based on the presentation format coefficient. Presentation methods include 3D model displays, dynamic animation demonstrations, and static graphic descriptions. Playback parameters are adjusted in real time based on ambient light intensity and equipment power. Based on path planning parameters and combined with real-time visitor flow distribution in the exhibition hall, a congestion-free personalized visitor path is generated to reduce user waiting time. The exhibit resource library is linked to the local servers of each exhibition hall, supporting real-time updates and retrieval of exhibit information to ensure the timeliness of exhibition content.

[0035] Simultaneously, based on the collaborative rules in the globally shared exhibition optimization model, the exhibition strategy linkage and adaptation across multiple exhibition halls are achieved. The federated server synchronizes the final exhibition optimization model output results of each exhibition hall in real time. When it is detected that a user has completed a visit to a certain type of exhibit in the k-th exhibition hall, the globally shared exhibition optimization model mines the related exhibit information for that type of exhibit. The related information is generated based on dimensions such as exhibit theme and user interest extension. The related information is pushed to the m-th exhibition hall that the user may visit (m≠k). The final exhibition optimization model of the m-th exhibition hall adjusts its local exhibition strategy based on the received related information, displaying or highlighting related exhibits to facilitate the user's continued exploration. During the linkage process, only exhibit association identification information is transmitted, without involving the sharing of raw data, thus ensuring data privacy while enabling the cross-exhibition reuse of high-quality exhibition experience.

[0036] This embodiment collects multi-dimensional data through distributed sensing units, and provides high-quality feature support through cleaning, normalization, and dynamic weight feature fusion. Relying on federated learning, encrypted communication, and parameter aggregation mechanisms, it achieves collaborative value mining of data from multiple exhibition halls while ensuring local data retention. The cross-entropy loss function with exhibition adaptation constraints, gradient confidence federated aggregation algorithm, and local fine-tuning make the model fit the actual scenario and have strong adaptability. Finally, through personalized exhibition strategy generation and multi-exhibition hall linkage, it meets the personalized needs of visitors, avoids data sharing privacy risks, and improves the intelligent operation level of digital exhibition halls.

[0037] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An AI-driven dynamic exhibition method for digital exhibition halls, characterized in that, The specific steps of this method are as follows: S1, Multi-source data acquisition and standardized preprocessing: Each exhibition hall collects local multi-dimensional data through distributed sensing units. After cleaning, normalization and feature fusion processing, a standardized local training dataset is generated. The standardized local training dataset is stored locally in the corresponding exhibition hall throughout the process. S2, Federated Learning System Setup and Initialization: Construct a federated learning system including a federated server and local servers in each exhibition hall, establish encrypted communication links, and configure the basic structure, training hyperparameters, and parameter aggregation rules of the global exhibition optimization model. S3, Local Exhibition Optimization Model Training: Based on the standardized local training dataset, each exhibition hall's local server performs localized training on the initial global exhibition optimization model to obtain a local exhibition optimization model; The training process only outputs the model parameter gradients. S4, Cross-exhibition hall model parameter aggregation and optimization: The local servers of each exhibition hall encrypt and upload the model parameter gradients to the federated server. The federated server completes the parameter gradient aggregation according to the preset aggregation rules to generate a globally shared exhibition optimization model. S5, Global model localization fine-tuning: The federated server distributes the globally shared exhibition optimization model to the local servers of each exhibition hall, and completes model fine-tuning by combining the real-time operation data of the exhibition hall to obtain the final exhibition optimization model adapted to the local scenario. S6, Dynamic exhibition strategy generation and linkage execution: Each exhibition hall generates personalized exhibition strategies based on the final exhibition optimization model, and at the same time, realizes the linkage and adaptation of multi-exhibition hall exhibition strategies according to the collaborative rules in the globally shared exhibition optimization model.

2. The AI-driven dynamic exhibition method for digital exhibition halls according to claim 1, characterized in that, In step S1, the multi-dimensional data includes user behavior data, exhibition operation data, and environmental perception data. User behavior data is collected through high-definition cameras and touch-interactive terminals deployed in exhibition hall corridors and exhibit display cases, specifically including user visit trajectory coordinate sequences, single exhibit dwell time, interaction operation types, and trigger frequencies. Exhibition operation data is collected through the exhibition hall equipment monitoring unit, specifically including exhibit display resolution, playback frame rate, projection device power, and interactive terminal response latency. Environmental perception data is collected through temperature and humidity sensors and infrared crowd sensors, specifically including real-time crowd density, ambient light intensity, and temperature and humidity values ​​in the exhibition hall.

3. The AI-driven dynamic exhibition method for digital exhibition halls according to claim 1, characterized in that, The feature fusion processing in step S1 uses a dynamic weighted feature fusion algorithm to achieve semantic alignment of multi-dimensional data. The specific steps are as follows: Feature extraction was performed on user behavior data, exhibition operation data, and environmental perception data to obtain user feature vectors. Display feature vectors and environmental feature vectors ; Calculate the mutual information value between each feature and the exhibition effect. , , The formulas for calculating the dynamic weight coefficients and the fused feature vectors are as follows: in, Represents the fused feature vector; This represents the mutual information value between the user's feature vector and the display effect; This represents the dynamic weighting coefficients of user features; Represents the user feature vector; This represents the mutual information value between the exhibition feature vector and the exhibition effect; Indicates the dynamic weighting coefficient of the exhibition features; Represents the feature vector of the display; The mutual information value representing the environmental feature vector and the exhibition effect; Represents the dynamic weighting coefficients of environmental characteristics; Represents the environmental feature vector; express , , The summation result; Indicates vector concatenation operation.

4. The AI-driven dynamic exhibition method for digital exhibition halls according to claim 1, characterized in that, In step S2, the encrypted communication link of the federated learning system is constructed using the SSL / TLS protocol, specifically including three stages: identity authentication, data encryption, and message integrity verification. When the federated server establishes a connection with the local server of each exhibition hall for the first time, two-way identity authentication is completed through digital certificates. After successful authentication, the AES-256 encryption algorithm is used to encrypt the transmitted model parameter gradients and control commands. During transmission, a message digest is generated using the SHA-256 hash algorithm. The global exhibition optimization model adopts a deep learning network structure, with the input layer being the fusion feature vector X obtained in step S1, the hidden layer containing 3-5 layers of fully connected networks and 1 layer of attention mechanism, and the output layer being the exhibition strategy optimization parameter vector Y. The exhibition strategy optimization parameter vector Y includes content filtering weights, presentation format coefficients, and path planning parameters.

5. The AI-driven dynamic exhibition method for digital exhibition halls according to claim 1, characterized in that, The local training process in step S3 uses a cross-entropy loss function with presentation adaptation constraints to optimize model parameters. Specifically, the steps are as follows: initialize the parameter matrix W and bias vector b of the global presentation optimization model; randomly sample batches of samples (b) from the standardized local training dataset, with a batch size of B; input the fused feature vector X into the model to obtain the model-predicted presentation strategy optimization parameter vector Ŷ; the cross-entropy loss function with presentation adaptation constraints is calculated as follows: in, This represents the value of the loss function; This represents the label value of the actual display effect of the i-th sample, with a value range of 0-1; This represents the model's predicted value for the i-th sample; i represents the sample number; B represents the batch size. represents the regularization coefficient; W represents the parameter matrix of the global presentation optimization model; Represents the calculation of the square of the L2 norm; Indicates the display adaptation constraint coefficient; This represents the actual exhibition scene adaptation benchmark value corresponding to the i-th sample; This represents the summation operation; The gradient of the model parameters ∇ is obtained by differentiating the loss function L using the stochastic gradient descent algorithm. With ∇b, update the parameter matrix W and the bias vector b through gradient descent; where ∇ Let ∇W represent the gradient vector of the parameter matrix W, and ∇b represent the gradient vector of the bias vector b. Repeat the above two steps until the loss function L converges or reaches the preset number of training iterations to obtain the local display optimization model and the corresponding model parameter gradients ∇W and ∇b.

6. The AI-driven dynamic exhibition method for digital exhibition halls according to claim 1, characterized in that, The parameter aggregation rule in step S4 adopts a federated aggregation algorithm with gradient confidence. The specific aggregation steps are as follows: Each exhibition hall's local server uploads the model parameter gradients ∇W and ∇b to the federated server via the encrypted communication link in step S2. The federated server performs integrity verification on the uploaded parameter gradients and removes abnormal gradients with missing values ​​≥5%. The dataset quality weight αk and gradient confidence βk for each exhibition hall are calculated. The formula for calculating the comprehensive weight γk for each exhibition hall is as follows: , in This represents the dataset quality weight for the k-th exhibition hall; This represents the number of samples in the standardized local training dataset for the k-th exhibition hall; This represents the quality score of the k-th exhibition hall dataset; This represents the gradient confidence level of the k-th exhibition hall; Indicates variance calculation; This represents the overall weight of the k-th exhibition hall; k represents the k-th exhibition hall, k=1,2,...,n, where n is the total number of exhibition halls participating in the collaboration; This represents the summation operation; The formula for calculating the global parameter gradient is: in, Represents the gradient of the global parameter matrix; The gradient of the parameter matrix represents the local exhibition optimization model for the k-th exhibition hall; This represents the gradient of the global bias vector; This represents the gradient of the bias vector of the local exhibition optimization model for the k-th exhibition hall; the gradient of the global parameters is used. and Update the parameters of the initial global display optimization model to generate the global shared display optimization model.

7. The AI-driven dynamic exhibition method for digital exhibition halls according to claim 6, characterized in that, The specific method for verifying the integrity of the parameter gradient is as follows: calculate the similarity between the uploaded parameter gradient and the preset gradient template. When the similarity is ≥95%, the gradient is determined to be complete. When the similarity is <95%, the proportion of missing gradient values ​​is further calculated. When the proportion of missing values ​​is <5%, the missing gradient is completed by interpolation. When the proportion of missing values ​​is ≥5%, the gradient is determined to be abnormal and removed.

8. The AI-driven dynamic exhibition method for digital exhibition halls according to claim 1, characterized in that, The localization fine-tuning process in step S5 specifically includes: Each exhibition hall's local server collects real-time operational data from the past hour. After standardization processing in step S1, a fine-tuning dataset is generated. The number of samples in the fine-tuning dataset is 10%-20% of the standardized local training dataset. The parameters of the globally shared exhibition optimization model are used as initial parameters and loaded into the local training framework. Training is performed using the fine-tuning dataset. During training, the learning rate is set to half of the learning rate in step S3, and the number of iterations is set to 1 / 5 of the number in step S3. During training, the accuracy of the model's presentation effect prediction on the fine-tuning dataset is calculated in real time. When the accuracy is ≥90% or the accuracy fluctuation is ≤1% for three consecutive iterations, the fine-tuning is stopped, and the final optimized presentation model is obtained.

9. The AI-driven dynamic exhibition method for digital exhibition halls according to claim 1, characterized in that, The process of generating personalized display strategies in step S6 is as follows: the final display optimization model receives user feature vectors collected in real time. With environmental feature vectors Output the display strategy optimization parameter vector Y; Based on the content filtering weights in the parameter vector Y, the top 5 matching exhibits are selected from the exhibition hall exhibit resource library; based on the presentation format coefficient, the presentation method and playback parameters of each exhibit are determined. Based on path planning parameters and combined with real-time visitor flow distribution in the exhibition hall, a congestion-free personalized visit path is generated. The exhibit resource library is linked to the local servers in each exhibition hall, supporting real-time updates and access to exhibit information.

10. The AI-driven dynamic exhibition method for digital exhibition halls according to claim 1, characterized in that, The multi-hall exhibition strategy linkage adaptation in step S6 is as follows: the federated server synchronizes the output results of the final exhibition optimization model of each hall in real time; when it is detected that a user has completed the visit of a certain type of exhibit in the kth exhibition hall, the associated exhibit information of that type of exhibit is mined through the globally shared exhibition optimization model, and the associated information is pushed to the mth exhibition hall that the user may visit, where m≠k; the final exhibition optimization model of the mth exhibition hall adjusts the local exhibition strategy based on the received associated information, and displays or highlights the associated exhibits in front. During the linkage process, only the exhibit association identification information is transmitted.

Images