Digital exhibition product display method based on deep learning

Abstract

The invention discloses a digital exhibition product display method based on deep learning, and the method comprises the steps: S1, constructing a self-evolution type dynamic feature optimization network, and optimizing the dynamic aggregation precision through feature adaptive iteration; s2, generating a unified dynamic display semantic model based on a holographic information multi-modal fusion engine; s3, the personalized display intention model is updated in real time in combination with the environment state and the display target; s4, based on an exhibition path optimization strategy, realizing real-time adjustment of an exhibit exhibition path; s5, adopting a quantization detail enhancement generation technology to generate an omnibearing exhibit detail virtual display effect; s6, realizing dynamic pushing and recommendation of multi-dimensional display content based on an interaction prediction mechanism of space-time perception; s7, constructing a lightweight quantum enhancement classification model, and realizing efficient exhibit dynamic display management; and S8, constructing an immersive multi-dimensional display effect of the exhibit in the virtual exhibition space. The system has the advantages of excellent user interaction experience and high exhibition management efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of product display, and in particular to a digital exhibition product display method based on deep learning. Background Art

[0002] With the rapid development of digital technology, digital exhibitions have gradually become an important form of displaying and disseminating product information. In the digital exhibition environment, the efficient presentation and dynamic interaction of exhibits through virtual display technology has become the key to improving user experience. However, existing technologies still face many challenges in the dynamic adaptability, multimodal fusion and personalized recommendation of digital exhibition product display.

[0003] In traditional digital exhibition display methods, most rely on static content presentation and simple rule-driven recommendations. Although these methods can meet basic display needs, they often seem powerless when faced with complex user interaction behaviors, dynamic exhibition environments, and multi-dimensional exhibit display goals. Specifically, the existing technology has obvious deficiencies in the following aspects:

[0004] 1. Lack of dynamic adaptability: Existing display methods usually use fixed display paths or content layouts, which cannot respond to user interaction and changes in the exhibition environment in real time. This static display mode is difficult to meet the personalized needs of users, resulting in a relatively monotonous and unattractive user experience.

[0005] 2. Insufficient multimodal fusion: During the exhibition process, the user's behavior data, environmental data and the multimodal features of the exhibits were not effectively integrated. The existing technology lacks an efficient multimodal feature aggregation method, resulting in insufficient semantic consistency and interactive effects of the display content, and unable to achieve an immersive display experience.

[0006] 3. Limited personalized recommendation capabilities: Traditional display methods usually recommend exhibits based on simple rules or statistical models, lacking in-depth modeling of user behavior intentions. This approach cannot accurately predict user needs, resulting in low relevance and accuracy of recommended content.

[0007] 4. Inefficient display path optimization: Existing technologies mostly rely on static algorithms in the planning and dynamic adjustment of display paths. They are difficult to adapt to complex changes in exhibition scenes and real-time interactive feedback from users, and cannot provide efficient dynamic path optimization solutions.

[0008] 5. Insufficient detail rendering: The virtual detail display of exhibits usually relies on fixed graphics rendering methods and lacks technical support for dynamic enhancement of exhibit details. This deficiency makes the detail expression of the exhibited content limited and it is difficult to fully display the full picture and value of the exhibits.

[0009] 6. Limitations in data processing capabilities: When dealing with cross-domain exhibit data, existing methods are often limited by data dimensions, computational complexity, and model generalization ability, and cannot efficiently classify and dynamically manage complex exhibit features.

[0010] Therefore, how to provide a digital exhibition product display method based on deep learning is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0011] An object of the present invention is to propose a digital exhibition product display method based on deep learning. The present invention makes full use of technologies such as dynamic feature aggregation and adaptive display optimization, multi-modal deep semantic fusion, personalized intention modeling, reinforcement learning path optimization, generative detail rendering, and virtual simulation mapping, and details the methods for realizing dynamic adaptation display of exhibits, multi-dimensional content optimization, and personalized recommendation, and has the advantages of strong dynamic response ability, high richness of display content, excellent user interaction experience, and high exhibition management efficiency.

[0012] A digital exhibition product display method based on deep learning according to an embodiment of the present invention includes the following steps:

[0013] S1. Construct a self-evolving dynamic feature optimization network, take multi-dimensional data streams as inputs, dynamically capture the changing features of exhibits in the spatio-temporal environment, and optimize the dynamic aggregation accuracy through feature adaptive iterative optimization;

[0014] S2. Based on the holographic information multi-modal fusion engine, deeply align and fuse visual holographic streams, natural language semantic streams, and tactile feedback data in the multi-modal space to generate a unified dynamic display semantic model;

[0015] S3. Through a high-dimensional interaction intention guidance mechanism, based on the multi-dimensional intention parsing and prediction of user interaction behaviors, combine the environmental state and display goals to update the personalized display intention model in real time;

[0016] S4. Based on the exhibition path optimization strategy of deep reinforcement learning, use the deep deterministic policy gradient algorithm to continuously model the user behavior and environmental feedback in the state space to realize the real-time adjustment of the exhibit display path;

[0017] S5. Adopt quantization detail enhancement generation technology, combine the exhibit feature information with the dynamic virtual environment to generate an all-round virtual display effect of the exhibit details;

[0018] S6. Through a spatio-temporal perception-based interaction prediction mechanism, realize the dynamic push and recommendation of multi-dimensional display content by real-time modeling of user interaction time series, exhibition scene dynamics, and individual behavior intentions;

[0019] S7. Build a lightweight quantum-enhanced classification model, and use self-supervised learning to accurately classify cross-domain exhibit data to achieve efficient dynamic display management of exhibits;

[0020] S8. An exhibit reshaping method based on virtual simulation mapping, which constructs an immersive multi-dimensional display effect of exhibits in a virtual exhibition space through cross-domain data generation and dynamic environment mapping technology.

[0021] Optionally, the specific content of S1 includes:

[0022] S11. Conduct distributed collection and structured preprocessing on the multi-dimensional input data of exhibits. The input data includes the static feature data X s of the exhibits, the dynamic behavior feature data X d (t), the environmental context awareness data E(t), and the user multi-modal interaction data U(t). Generate a feature data stream through feature normalization, dimension compression, and semantic mapping;

[0023] S12. In the feature extraction module, use a hierarchical neural encoder to hierarchically deconstruct the input data, and encode the static feature data X s into a high-order feature vector F s , and process the dynamic data X d (t), the environmental data E(t), and the interaction data U(t) through a recurrent neural network based on time series to generate corresponding time series features F d (t), F e (t), and F u (t);

[0024] S13. Build a multi-input multi-output feature optimization network based on a self-evolving feature optimization algorithm:

[0025] F agg (t) = W s ·F s + W d (t)·F d (t) + W e (t)·F e (t) + W u (t)·F u (t);

[0026] Among them, F agg (t) is the aggregated feature representation, and W s , W d (t), W e (t), W u (t) are the dynamic weight coefficients of static, dynamic, environmental, and interaction features respectively, and the weights are adjusted in real time through an adaptive evolution strategy combined with a feature importance function;

[0027] S14. An optimization mechanism driven by user feedback signals, which uses the real-time prediction error Δ(t) and the user response value R(t) to establish an optimization function, and updates the optimization feature weight W through gradient i (t + 1):

[0028]

[0029] where η is the dynamic learning rate and L represents the loss function;

[0030] S15. Input the optimized feature vector F opt (t) into the feature dynamic control module, and process it through a multi-layer perceptron network to realize the modeling and adaptive expression of the spatio-temporal dynamic features of the exhibition items.

[0031] Optionally, the S2 specifically includes:

[0032] S21. Perform distributed acquisition and multi-modal preprocessing on the visual holographic stream, natural language semantic stream, and tactile feedback data, and respectively generate the visual initial feature V raw (t), semantic initial feature S raw (t), and tactile initial feature H raw (t) through a feature decoder, and use an adaptive noise reduction module to eliminate redundant information to obtain the optimized features V(t), S(t), and H(t);

[0033] S22. In the multi-modal feature alignment module, construct a feature projection network based on the self-attention mechanism, project the visual feature V(t), semantic feature S(t), and tactile feature H(t) into a unified high-dimensional representation space, and achieve feature alignment through dynamic weight adjustment:

[0034] F align (t) = β v ·V′(t) + β s ·S′(t) + β h ·H′(t);

[0035] where F align (t) is the aligned multi-modal feature, V′(t), S′(t), and H′(t) are the projected feature vectors, and β v , β s , β h are the feature alignment dynamic weights, which are updated in real time by an adaptive optimization strategy;

[0036] S23. Use a deep fusion engine based on the multi-layer attention mechanism to perform feature fusion on F align (t), and calculate and generate the fusion feature F fusion (t) through the multi-layer attention mechanism:

[0037]

[0038] Among them, α i (t) is the dynamic attention weight of the i-th modality, and F′ i (t) is the enhanced representation of the multimodal features;

[0039] S24. Perform semantic decomposition and structured modeling on the fused feature F fusion (t) through a deep semantic parsing network to generate a semantic hierarchical vector M semantic (t), and extract semantic logical relationships and dynamic semantic expressions through a multi-layer semantic decoder;

[0040] S25. Establish a semantic consistency verification module to perform semantic integrity verification and logical consistency check on the generated semantic model M semantic (t), and output an optimized dynamic display semantic model.

[0041] Optionally, the specific steps of S3 are as follows:

[0042] S31. Collect multimodal user interaction behavior data, including user touch action data T m (t), voice command data V n (t), eye gaze tracking data G p (t), gesture recognition data P q (t), and interaction context variable C r (t), and combine environmental state data E k (t) and exhibit characteristic data O l (t), and perform structured preprocessing on the collected data through a multimodal feature decoder to generate an initial interaction feature F x (t);

[0043] S32. In the multi-dimensional intention parsing module, use a multimodal feature embedding network to project the initial feature F x (t) into high-dimensional features to generate time series features F y (t), and model historical interaction behaviors through a recursive network based on the self-attention mechanism to generate predicted features F z (t);

[0044] S33. Based on user behavior characteristics and environmental context information, define a real-time dynamic intention optimization function:

[0045] I u (t) = δ1·F y (t) + δ2·E k (t) + δ3·O l (t) + δ4·C r (t);

[0046] Among them, I u (t) represents the dynamic user intention model, and δ1, δ2, δ3, δ4 are weight parameters, which are dynamically updated through an optimization algorithm based on reinforcement learning;

[0047] S34. Dynamically adjust I u (t) through a multi-objective joint optimization strategy, and optimize the matching between the user preference expression and the display target in combination with the intention guidance mechanism to generate a personalized display intention model I optimal (t):

[0048] I optimal (t) = ψ1·I u (t) + ψ2·Δ(T m (t), V n (t), G p (t));

[0049] Among them, ψ1 and ψ2 are the weight factors for multi-objective joint optimization, and Δ(T m (t), V n (t), G p (t)) represents the dynamic correction function of the user interaction behavior characteristics;

[0050] S35. Input the optimized personalized display intention model I optimal (t) into the dynamic display configuration module, and optimize the display order S of the exhibits d , layout adjustment parameter L d and interaction method A d in real time.

[0051] Optionally, the S4 specifically includes:

[0052] S41. Construct a state space model for optimizing the exhibition path, and the state variable set is Σ(t) = {U h (t), E j (t), R k (t)}, where U h (t) represents the current user behavior feature vector, E j (t) represents the environmental feedback signal, and R k (t) represents the dynamic correlation feature between the exhibit and the user's interest;

[0053] S42. Based on the action space Λ(t) = {ξ1, ξ2,..., ξ m} of deep reinforcement learning, each action ξ i represents the adjustment method of the exhibit display path, including exhibit position exchange, dynamic insertion, priority reordering, and path extension;

[0054] S43. Define the reward function Ω(t) for path optimization:

[0055] Ω(t) = κ1·I u (t) + κ2·C q (t) - κ3·T m (t);

[0056] Where, I u (t) represents the user interest matching score, C q (t) represents the coherence score of the display path, T m (t) represents the time consumption of path adjustment, and κ1, κ2, κ3 are the weight factors of the reward function;

[0057] S44. Use the deep deterministic policy gradient algorithm to train the path optimization policy network:

[0058]

[0059] Where, Θ(t) are the parameters of the policy network, η is the learning rate, Q(Σ(t), Λ(t)|Θ(t)) is the expected return value of the state and action combination, and is estimated and updated using a deep neural network;

[0060] S45. In the path optimization decision module, combine the user's real-time behavior data U real (t) and the environmental dynamic feedback E dyn (t), generate the optimal path adjustment action sequence Λ opt (t) through the trained policy network, and dynamically adjust the exhibit display order P final (t).

[0061] Optionally, the specific content of S5 includes:

[0062] S51. Collect multi-dimensional feature data of exhibits, including geometric shape information G x , surface material characteristics T y , dynamic behavior sequence D z (t) and virtual environment mapping data E w (t), and perform quantization processing on the original data through the feature quantization module to generate the quantization feature vector F p (t);

[0063] S52. Use the quantization detail enhancement network to enhance the details of the quantization feature F p (t), and define the exhibit detail enhancement function:

[0064] F enhanced (t) = α1·G x +α2·T y +α3·Φ(Dz (t), E w (t));

[0065] Among them, F enhanced (t) represents the enhanced exhibit feature, α1, α2, α3 are weight parameters, Φ(D z (t), E w (t)) represents the joint correction function between the dynamic behavior and the virtual environment;

[0066] S53. Combine the virtual environment to perform dynamic rendering on the enhanced feature F enhanced (t), and use a detail enhancement generator to construct a dynamic rendering model R dynamic (t):

[0067] R dynamic (t) = β1·F enhanced (t) + β2·Ψ(E w (t));

[0068] Among them, β1, β2 are rendering parameters, Ψ(E w (t)) represents the environment mapping optimization function;

[0069] S54. Through a multi-layer detail verification module, layer-by-layer verification is performed on the geometric shape, surface texture, and dynamic behavior effect of the rendering model R dynamic (t), and the final optimized virtual display model M optimized (t) is generated;

[0070] S55. Input the optimized virtual display model M optimized (t) into the display control module, and adjust the display perspective, detail performance, and dynamic response of the exhibit according to the user's real-time feedback through the dynamic interaction engine, and generate a real-time updated virtual display effect V final (t).

[0071] Optionally, the specific content of S6 includes:

[0072] S61. Collect the multi-modal interaction data of the user, including the user interaction time series feature T x (t), the dynamic state S of the exhibition scene y (t), and the individual behavior intention B z (t), and perform normalization processing and hierarchical coding on the data through the data preprocessing module to generate a spatio-temporal perception feature set F spatio (t) = {T x (t), S y (t), B z (t)};

[0073] S62. Use a recurrent neural network based on spatio-temporal perception for Fspatio (t) Perform high-dimensional encoding to construct a user interaction prediction model:

[0074] P interaction (t) = θ1·Λ(T x (t)) + θ2·Γ(S y (t)) + θ3·Ω(B z (t));

[0075] Among them, P interaction (t) represents the user interaction prediction value, Λ(T x (t)), Γ(S y (t)), Ω(B z (t)) are the spatio-temporal encoding functions of time series features, scene dynamics, and behavior intentions respectively, and θ1, θ2, θ3 are dynamic weight parameters;

[0076] S63. Combine the historical behavior pattern H m (t) and the user's real-time interaction feature P interaction (t) to define the priority score function of the recommended content:

[0077] R priority (t) = η1·P interaction (t) + η2·H m (t) + η3·C k (t);

[0078] Among them, R priority (t) is the dynamic priority of the recommended content, η1, η2, η3 are priority weight parameters, and C k (t) represents the matching score of the current exhibit features;

[0079] S64. Sort the priority R priority (t) of the recommended content through a multi-dimensional dynamic content push engine, and combine the user preference distribution U pref (t) and the scene dynamic feedback E dyn (t) to generate a real-time recommended content set D recommend (t);

[0080] S65. Use the dynamic display optimization module to input D recommend (t) into the display logic controller, and generate a personalized dynamic display scheme V dynamic (t) by adjusting the display order, display style, and interaction method in real time.

[0081] Optionally, the S7 specifically includes:

[0082] S71. Collect cross-domain exhibit data, including the static attribute features G s of the exhibit and the dynamic behavior sequence Ht (t) and multimodal interaction information I u (t), and the original data is normalized, noise-filtered, and high-dimensional embedded through a feature preprocessing module to generate an initial feature set F init (t) = {G s , H t (t), I u (t)};

[0083] S72. Construct a lightweight quantum-enhanced classification model to perform feature encoding and quantum enhancement on F init (t):

[0084] F enh (t) = δ1·Ψ1(G s ) + δ2·Ψ2(H t (t)) + δ3·Ψ3(I u (t));

[0085] Among them, F enh (t) is the feature representation after quantum enhancement, Ψ1(G s ), Ψ2(H t (t)), Ψ3(I u (t)) are the quantum enhancement functions of static features, dynamic features, and multimodal interaction data respectively, and δ1, δ2, δ3 are feature enhancement weight parameters;

[0086] S73. Use the self-supervised learning mechanism to perform feature classification on F enh (t):

[0087]

[0088] Among them, L cls (t) is the classification loss function, Q(y j |(t)) is the predicted probability distribution of class y j , and ξ j is the class balance weight;

[0089] S74. Optimize L cls (t) through a lightweight classification network, and use the gradient descent method combined with the self-supervised learning strategy to adjust the model parameters to generate an optimized classification result R cls (t);

[0090] S75. Input the classification result R cls (t) into the exhibit dynamic management module, dynamically group and optimize the exhibits according to the classification categories, and generate a display plan set S disp (t).

[0091] Optionally, the S8 specifically includes:

[0092] S81. Collect cross - domain data of the exhibit, including the physical attribute features P a , the dynamic behavior sequence D b (t), the material detail features M c and the environmental context information E d (t), and perform normalization processing and multi - modal encoding on the data through the cross - domain data fusion module to generate the feature set F cross (t) = {P a , D b (t), M c , E d (t)};

[0093] S82. Perform dynamic feature reshaping on F cross (t) based on the virtual simulation mapping technology:

[0094] F remap (t) = κ1·P a +κ2·Ξ(D b (t), M c )+κ3·Ψ(E d (t));

[0095] Among them, F remap (t) represents the reshaped feature set, κ1, κ2, κ3 are feature mapping weight parameters, Ξ(D b (t), M c ) is the joint mapping function of dynamic behavior and material attributes, and Ψ(E d (t)) is the dynamic correction function of the environmental context;

[0096] S83. Use the multi - dimensional virtual rendering engine to perform immersive multi - dimensional rendering on F remap (t) to construct the virtual display model R immersive (t):

[0097] R immersive (t) = λ1·F remap (t)+λ2·Φ(E d (t));

[0098] Among them, λ1, λ2 are rendering parameters, and Φ(E d (t)) represents the multi - dimensional rendering function of the virtual environment;

[0099] S84. Through the multi - layer verification module, hierarchically verify the geometric shape, material details, dynamic behavior performance, and environmental interaction effects of the virtual display model R immersive (t) to generate the optimized virtual exhibit model M optimized(t);

[0100] S85. Apply M optimized (t) to the virtual exhibition space, and through the dynamic scene control module, combine the user's real-time interaction behavior data U e (t) and the changes in the exhibition environment S f (t) to dynamically adjust the display details, interaction logic, and environmental adaptation of virtual exhibits, and generate an immersive multi-dimensional display effect V final (t).

[0101] The beneficial effects of the present invention are:

[0102] (1) By combining the dynamic feature aggregation network and the adaptive display optimization algorithm, the present invention realizes the efficient feature capture and real-time adjustment of exhibits in the dynamic exhibition scene, enabling the system to accurately adapt to user behavior and environmental changes, thereby enhancing the dynamic response ability of the display content and solving the deficiency that traditional display methods cannot be adapted in real time.

[0103] (2) Through the holographic information multi-modal fusion engine, combining semantic depth alignment and feature dynamic fusion technologies, the present invention provides multi-modal deep understanding and interaction optimization of the display content of exhibits, enabling the display system to cooperate efficiently at multi-modal levels such as vision, language, and touch, and enhancing the user's immersive experience and sense of interaction.

[0104] (3) Based on the path optimization algorithm of reinforcement learning, using the deep deterministic policy gradient method, the present invention dynamically optimizes the display path of exhibits in real time, effectively shortening the path planning time and improving the display efficiency, and solving the problems of slow response and poor adaptability of traditional path optimization methods.

[0105] (4) Through the quantization detail enhancement generation technology, combining the dynamic features of exhibits and the virtual environment mapping, the present invention generates a high-precision virtual display effect of exhibit details, significantly enhancing the detail expressiveness and visual quality of exhibits, and overcoming the limitation of insufficient detail rendering in traditional methods.

[0106] (5) Using the lightweight quantum-enhanced classification model and the self-supervised learning mechanism, the present invention realizes the efficient classification and dynamic management of cross-domain exhibit data, significantly improving the accuracy and processing efficiency of exhibit classification, and providing strong support for the accurate recommendation and dynamic management of exhibition content. Description of the Drawings

[0107] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0108] Figure 1Flowchart of a digital exhibition product display method based on deep learning proposed by the present invention. Detailed implementation manners

[0109] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0110] Refer to Figure 1 , a digital exhibition product display method based on deep learning, including the following steps:

[0111] S1. Construct a self-evolving dynamic feature optimization network, take multi-dimensional data streams as inputs, dynamically capture the change features of exhibits in the spatio-temporal environment, and optimize the dynamic aggregation accuracy through feature adaptive iterative optimization;

[0112] In this embodiment, S1 specifically includes:

[0113] S11. Perform distributed acquisition and structured preprocessing on the multi-dimensional input data of the exhibits. The input data includes the static feature data X s of the exhibits, the dynamic behavior feature data X d (t), the environmental context awareness data E(t), and the user multi-modal interaction data U(t). Generate feature data streams through feature normalization, dimension compression, and semantic mapping;

[0114] S12. In the feature extraction module, use a hierarchical neural encoder to hierarchically deconstruct the input data, encode the static feature data X s into a high-order feature vector F s , and process the dynamic data X d (t), the environmental data E(t), and the interaction data U(t) through a recurrent neural network based on time series to generate corresponding time series features F d (t), F e (t), and F u (t);

[0115] S13. Construct a multi-input multi-output feature optimization network based on the self-evolving feature optimization algorithm:

[0116] F agg (t) = W s · F s + W d (t) · F d (t) + W e (t) · F e (t) + W u (t) · F u (t);

[0117] Among them, Fagg (t) is the aggregated feature representation, W s , W d (t), W e (t), W u (t) are the dynamic weight coefficients of static, dynamic, environmental, and interaction features respectively, and the weights are adjusted in real time through an adaptive evolution strategy combined with a feature importance function;

[0118] S14. An optimization mechanism driven by user feedback signals, which uses the real-time prediction error Δ(t) and the user response value R(t) to establish an optimization function, and optimizes the feature weight W i (t + 1):

[0119]

[0120] where η is the dynamic learning rate and L represents the loss function;

[0121] S15. Input the optimized feature vector F opt (t) into the feature dynamic control module, and process it through a multi-layer perceptron network to realize the modeling and adaptive expression of the spatio-temporal dynamic features of the exhibit.

[0122] S2. Based on the holographic information multi-modal fusion engine, deeply align and fuse the visual holographic stream, natural language semantic stream, and tactile feedback data in the multi-modal space to generate a unified dynamic display semantic model;

[0123] In this embodiment, S2 specifically includes:

[0124] S21. Perform distributed acquisition and multi-modal preprocessing on the visual holographic stream, natural language semantic stream, and tactile feedback data, and respectively generate visual initial features V raw (t), semantic initial features S raw (t), and tactile initial features H raw (t) through a feature decoder, and use an adaptive noise reduction module to eliminate redundant information to obtain optimized features V(t), S(t), and H(t);

[0125] S22. In the multi-modal feature alignment module, construct a feature projection network based on the self-attention mechanism, project the visual feature V(t), semantic feature S(t), and tactile feature H(t) into a unified high-dimensional representation space, and achieve feature alignment through dynamic weight adjustment:

[0126] F align (t) = β v ·V′(t) + β s ·S′(t) + β h ·H′(t);

[0127] Among them, F align (t) is the aligned multi-modal feature, V′(t), S′(t) and H′(t) are the projected feature vectors, and β v , β s , β h are the dynamic weights for feature alignment, which are updated in real time by an adaptive optimization strategy;

[0128] S23. Use a deep fusion engine based on a multi-layer attention mechanism to perform feature fusion on F align (t), and calculate and generate the fused feature F fusion (t) through the multi-layer attention mechanism:

[0129]

[0130] Among them, α i (t) is the dynamic attention weight of the i-th modality, and F′ i (t) is the enhanced representation of the multi-modal feature;

[0131] S24. Perform semantic decomposition and structured modeling on the fused feature F fusion (t) through a deep semantic parsing network to generate the semantic hierarchical vector M semantic (t), and extract semantic logical relationships and dynamic semantic expressions through a multi-layer semantic decoder;

[0132] S25. Establish a semantic consistency verification module to perform semantic integrity verification and logical consistency check on the generated semantic model M semantic (t), and output the optimized dynamic display semantic model.

[0133] S3. Through a high-dimensional interaction intention guidance mechanism, based on the multi-dimensional intention parsing and prediction of user interaction behaviors, combine the environmental state and display objectives to update the personalized display intention model in real time;

[0134] In this embodiment, S3 specifically includes:

[0135] S31. Collect multi-modal user interaction behavior data, including the touch action data T m (t), voice command data V n (t), gaze tracking data G p (t), gesture recognition data P q (t) and interaction context variable C r (t), combine the environmental state data E k (t) and exhibit characteristic data O l (t), and perform structured preprocessing on the collected data through a multi-modal feature decoder to generate the initial interaction feature F x (t);

[0136] S32. In the multi-dimensional intention parsing module, use the multi-modal feature embedding network to project the initial feature F x (t) into high-dimensional features to generate time series features F y (t), and model the historical interaction behavior through a recurrent network based on the self-attention mechanism to generate predicted features F z (t);

[0137] S33. Based on the user behavior characteristics and environmental context information, define a real-time dynamic intention optimization function:

[0138] I u (t) = δ1·F y (t) + δ2·E k (t) + δ3·O l (t) + δ4·C r (t);

[0139] where I u (t) represents the dynamic user intention model, and δ1, δ2, δ3, δ4 are weight parameters, which are dynamically updated through an optimization algorithm based on reinforcement learning;

[0140] S34. Dynamically adjust I u (t) through a multi-objective joint optimization strategy, and combine the intention guidance mechanism to optimize the matching of user preference expression and display goals, and generate a personalized display intention model I optimal (t):

[0141] I optimal (t) = ψ1·I u (t) + ψ2·Δ(T m (t), V n (t), G p (t));

[0142] where ψ1, ψ2 are weight factors for multi-objective joint optimization, and Δ(T m (t), V n (t), G p (t)) represents the dynamic correction function of user interaction behavior characteristics;

[0143] S35. Input the optimized personalized display intention model I optimal (t) into the dynamic display configuration module, and perform real-time optimization on the display order S d of the exhibits, the layout adjustment parameter L d and the interaction method A d .

[0144] S4. Exhibition path optimization strategy based on deep reinforcement learning. Using the deep deterministic policy gradient algorithm, through continuous modeling of user behavior and environmental feedback in the state space, real-time adjustment of the exhibition path of exhibits is achieved;

[0145] In this embodiment, S4 specifically includes:

[0146] S41. Construct a state space model for exhibition path optimization. The set of state variables is Σ(t) = {U h (t), E j (t), R k (t)}, where U h (t) represents the current behavior feature vector of the user, E j (t) represents the environmental feedback signal, and R k (t) represents the dynamic correlation feature between the exhibit and the user's interest;

[0147] S42. Action space Λ(t) = {ξ1, ξ2, …, ξ m} based on deep reinforcement learning. Each action ξ i represents the adjustment method of the exhibition path of the exhibit, including exhibit position exchange, dynamic insertion, priority reordering, and path extension;

[0148] S43. Define the reward function Ω(t) for path optimization:

[0149] Ω(t) = κ1·I u (t) + κ2·C q (t) - κ3·T m (t);

[0150] Among them, I u (t) represents the user interest matching score, C q (t) represents the coherence score of the display path, T m (t) represents the time consumption of path adjustment, and κ1, κ2, κ3 are the weight factors of the reward function;

[0151] S44. Use the deep deterministic policy gradient algorithm to train the path optimization policy network:

[0152]

[0153] Among them, Θ(t) are the parameters of the policy network, η is the learning rate, Q(Σ(t), Λ(t)|Θ(t)) is the expected return value of the state-action combination, and is estimated and updated using a deep neural network;

[0154] S45. In the path optimization decision module, combine the real-time behavior data U real (t) of the user and the dynamic environmental feedback E dyn(t), generate the optimal path adjustment action sequence Λ through the trained policy network opt (t), and dynamically adjust the exhibit display order P final (t).

[0155] S5. Adopt the quantization detail enhancement generation technology, combine the exhibit feature information with the dynamic virtual environment to generate the all-round virtual display effect of the exhibit details;

[0156] In this embodiment, S5 specifically includes:

[0157] S51. Collect the multi-dimensional feature data of the exhibit, including the geometric shape information G x , the surface material feature T y , the dynamic behavior sequence D z (t) and the virtual environment mapping data E w (t), and perform quantization processing on the original data through the feature quantization module to generate the quantization feature vector F p (t);

[0158] S52. Use the quantization detail enhancement network to enhance the details of the quantization feature F p (t), and define the exhibit detail enhancement function:

[0159] F enhanced (t) = α1·G x +α2·T y +α3·Φ(D z (t), E w (t));

[0160] Among them, F enhanced (t) represents the enhanced exhibit feature, α1, α2, α3 are weight parameters, and Φ(D z (t), E w (t)) represents the joint correction function between the dynamic behavior and the virtual environment;

[0161] S53. Combine the virtual environment to perform dynamic rendering on the enhanced feature F enhanced (t), and adopt the detail enhancement generator to construct the dynamic rendering model R dynamic (t):

[0162] R dynamic (t) = β1·F enhanced (t)+β2·Ψ(E w (t));

[0163] Among them, β1, β2 are rendering parameters, and Ψ(E w (t)) represents the environment mapping optimization function;

[0164] S54. Layer-by-layer verification is performed on the geometric shape, surface texture, and dynamic behavior effects of the rendered model R dynamic (t), and the final optimized virtual display model M optimized (t) of the exhibit is generated;

[0165] S55. The optimized virtual display model M optimized (t) is input into the display control module, and the display perspective, detail performance, and dynamic response of the exhibit are adjusted according to the real-time feedback of the user through the dynamic interaction engine, and the real-time updated virtual display effect V final (t) is generated.

[0166] S6. Through the interaction prediction mechanism based on spatio-temporal perception, real-time modeling is performed on the user interaction timing sequence, exhibition scene dynamics, and individual behavior intentions to achieve dynamic push and recommendation of multi-dimensional display content;

[0167] In this embodiment, S6 specifically includes:

[0168] S61. Collect multi-modal interaction data of the user, including the user interaction timing feature T x (t), the dynamic state S y (t) of the exhibition scene, and the individual behavior intention B z (t). The data is normalized and hierarchically encoded through the data preprocessing module to generate the spatio-temporal perception feature set F spatio (t) = {T x (t), S y (t), B z (t)};

[0169] S62. Use the recurrent neural network based on spatio-temporal perception to perform high-dimensional encoding on F spatio (t) to construct a user interaction prediction model:

[0170] P interaction (t) = θ1·Λ(T x (t)) + θ2·Γ(S y (t)) + θ3·Ω(B z (t));

[0171] Among them, P interaction (t) represents the user interaction prediction value, and Λ(T x (t)), Γ(S y (t)), Ω(B z (t)) are the spatio-temporal encoding functions of the timing feature, scene dynamics, and behavior intention respectively, and θ1, θ2, θ3 are dynamic weight parameters;

[0172] S63. Combine the historical behavior pattern H m(t) and the real-time interaction feature P with the user interaction (t), define the priority score function of the recommended content:

[0173] R priority (t) = η1·P interaction (t) + η2·H m (t) + η3·C k (t);

[0174] Among them, R priority (t) is the dynamic priority of the recommended content, η1, η2, η3 are priority weight parameters, and C k (t) represents the matching score of the current exhibit features;

[0175] S64. Through the multi-dimensional dynamic content push engine, sort the priority R priority (t) of the recommended content, and combine the user preference distribution U pref (t) and the scenario dynamic feedback E dyn (t) to generate the real-time recommended content set D recommend (t);

[0176] S65. Use the dynamic display optimization module to input D recommend (t) into the display logic controller, and generate a personalized dynamic display plan V dynamic (t) by adjusting the display order, display style and interaction method in real time.

[0177] S7. Build a lightweight quantum-enhanced classification model, and use self-supervised learning to accurately classify cross-domain exhibit data to achieve efficient dynamic display management of exhibits;

[0178] In this embodiment, S7 specifically includes:

[0179] S71. Collect cross-domain exhibit data, including the static attribute features G of the exhibits s , the dynamic behavior sequence H t (t) and the multi-modal interaction information I u (t), and perform normalization, noise filtering and high-dimensional embedding processing on the original data through the feature preprocessing module to generate the initial feature set F init (t) = {G s , H t (t), I u (t)};

[0180] S72. Build a lightweight quantum-enhanced classification model to perform feature encoding and quantum enhancement on F init (t):

[0181] F enh (t) = δ1·Ψ1(Gs ) + δ2·Ψ2(H t (t)) + δ3·Ψ3(I u (t));

[0182] Among them, F enh (t) is the feature representation after quantum enhancement, and Ψ1(G s ), Ψ2(H t (t)), Ψ3(I u (t)) are the quantum enhancement functions of static features, dynamic features, and multimodal interaction data respectively, and δ1, δ2, δ3 are feature enhancement weight parameters;

[0183] S73. Use the self-supervised learning mechanism to classify the features of F enh (t):

[0184]

[0185] Among them, L cls (t) is the classification loss function, Q(y j |(t)) is the predicted probability distribution of class y j , and ξ j is the class balance weight;

[0186] S74. Optimize L cls (t) through a lightweight classification network, and use the gradient descent method combined with the self-supervised learning strategy to adjust the model parameters to generate the optimized classification result R cls (t);

[0187] S75. Input the classification result R cls (t) into the exhibit dynamic management module, dynamically group and optimize the exhibits according to the classification categories, and generate the display plan set S disp (t).

[0188] S8. The exhibit reshaping method based on virtual simulation mapping constructs an immersive multi-dimensional display effect of the exhibit in the virtual exhibition space through cross-domain data generation and dynamic environment mapping technology.

[0189] In this embodiment, S8 specifically includes:

[0190] S81. Collect cross-domain data of the exhibit, including the physical attribute characteristics P a , the dynamic behavior sequence D b (t), the material detail characteristics M c , and the environmental context information E d (t). Normalize the data and perform multimodal encoding through the cross-domain data fusion module to generate the feature set F cross (t) = {Pa , D b (t), M c , E d (t)};

[0191] S82. Dynamically reshape the features of F cross (t) based on virtual simulation mapping technology:

[0192] F remap (t) = κ1·P a +κ2·Ξ(D b (t), M c ) + κ3·Ψ(E d (t));

[0193] Among them, F remap (t) represents the reshaped feature set, κ1, κ2, κ3 are feature mapping weight parameters, Ξ(D b (t), M c ) is the joint mapping function of dynamic behavior and material properties, and Ψ(E d (t)) is the dynamic correction function of the environmental context;

[0194] S83. Use a multi-dimensional virtual rendering engine to perform immersive multi-dimensional rendering on F remap (t) and construct a virtual display model R immersive (t):

[0195] R immersive (t) = λ1·F remap (t) + λ2·Φ(E d (t));

[0196] Among them, λ1, λ2 are rendering parameters, and Φ(E d (t)) represents the multi-dimensional rendering function of the virtual environment;

[0197] S84. Through a multi-layer verification module, hierarchically verify the geometric shape, material details, dynamic behavior performance, and environmental interaction effects of the virtual display model R immersive (t) to generate an optimized virtual exhibit model M optimized (t);

[0198] S85. Apply M optimized (t) to the virtual exhibition space, and dynamically adjust the display details, interaction logic, and environmental adaptation of the virtual exhibit by combining the dynamic scene control module with the user's real-time interaction behavior data U e (t) and the exhibition environment change S f (t) to generate an immersive multi-dimensional display effect V final (t).

[0199] Example 1:

[0200] To verify the feasibility of the present invention, the digital exhibition product display method based on deep learning of the present invention is applied to an international digital technology exhibition. The goal of this exhibition is to showcase various technological products in a digital way, including smart home devices, virtual reality devices, and related innovative technologies. The exhibition attracted more than 200 exhibitors and more than 5,000 visitors from all over the world. The exhibition scenario is complex, the needs of visitors are diverse, and the types of exhibits are numerous. The traditional static display method can no longer meet the personalized needs and real-time response requirements.

[0201] In this scenario, the digital exhibition system is deployed based on the method of the present invention, covering dynamic feature aggregation, path optimization, detail enhancement, multimodal fusion, and virtual simulation mapping technology. The system is comprehensively applied throughout the whole process from the collection of visitors' behaviors to the optimization of exhibit displays. The following are the specific implementation steps and actual effects:

[0202] The system first uses the dynamic feature aggregation module to collect the interactive behavior data of visitors, the exhibit environment data, and the dynamic changes of the exhibition scene in real time. These data include the line-of-sight trajectories, residence times of visitors, and touch feedback on specific exhibits, etc. In the preliminary test, the system processes more than 50GB of data every day and completes feature extraction and optimization within seconds. This function significantly improves the dynamic adaptation ability of exhibits and solves the limitation that the traditional display mode cannot quickly respond to user needs.

[0203] In terms of optimizing the display path, the system plans a dynamically adjustable visit path for visitors based on the reinforcement learning algorithm. During peak hours, the visitor flow is dense, and the path planning of the traditional method takes a long time and cannot be dynamically adjusted. The test results show that after adopting the path optimization module of the present invention, the response time of path planning is reduced from an average of 45 seconds to 3 seconds, and the proportion of visitors in congested areas is reduced.

[0204] The system also provides high-precision rendering for virtual display content through quantization detail enhancement generation technology. Taking a certain smart home device as an example, its virtual display presents real surface textures, dynamic behavior simulations, and light and shadow effects with the support of a multi-dimensional rendering engine. The residence time of visitors on this exhibit is increased from an average of 30 seconds in the traditional display to 85 seconds, and the display attractiveness is significantly improved.

[0205] In terms of personalized recommendation, the system uses a spatio-temporal perception-based interactive prediction module and a lightweight quantum-enhanced classification model to push exhibit content that matches the interests of visitors to them. One day, a visitor showed a high interest in virtual reality devices. The system captured their stay behavior and interaction patterns in real time and accurately recommended 5 related devices, with a recommendation accuracy rate of 94%. The feedback from the visitors on the recommended content shows that the relevance of the recommended content is significantly better than that of traditional rule-based recommendation methods.

[0206] In addition, the virtual simulation mapping module of the present invention supports the multi-dimensional display of exhibits in the virtual exhibition space. In the test, a drone product achieved flight dynamic simulation and environment adaptation display through this module. The interaction frequency of users in the virtual scenario increased by 120%, and the intended purchase volume of the exhibits increased by 38%. The specific effect data is shown in Table 1 below:

[0207] Table 1 Analysis of the Deployment Effect of the Digital Exhibition System

[0208]

[0209] From the above tests and practical applications, it can be seen that the present invention has significantly improved the dynamic adaptability of exhibit display, the optimization efficiency of display paths, and the accuracy of personalized content recommendation in complex digital exhibition scenarios. The system has demonstrated excellent performance in key technologies such as multi-modal fusion, real-time response, and virtual simulation display, and has solved many deficiencies of traditional display methods in dynamic adaptation ability, data processing efficiency, and user experience optimization.

[0210] Specifically, the present invention realizes the efficient adaptation of exhibits to environmental changes through dynamic feature aggregation technology, greatly shortening the response time and providing guarantee for intelligent management in complex exhibition scenarios. Through the path optimization module, the efficiency of visit path planning during peak hours of the exhibition has increased by 15 times, while significantly reducing the proportion of congested areas, improving the mobility and satisfaction of visitors.

[0211] In terms of multi-modal fusion and recommendation, the present invention accurately identifies the interest points of visitors based on a deep learning-based prediction mechanism, and provides personalized exhibit recommendations in combination with quantum-enhanced classification technology, with the recommendation accuracy rate increased to 94%, effectively meeting the personalized needs of visitors.

[0212] In addition, the application of the present invention in the field of virtual display has comprehensively improved the display effect of exhibits. Through generative detail enhancement and virtual simulation mapping technologies, the detail expressiveness and interaction experience of exhibits have reached the leading level in the industry. The interaction frequency and intended purchase rate of users have increased significantly, further verifying the practical value of the present invention in optimizing exhibition effects and promoting commercial transformation.

[0213] As described above, it is only the preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A digital exhibition product display method based on deep learning, characterized in that: The steps include: S1. Build a self-evolving dynamic feature optimization network, take multi-dimensional data stream as input, dynamically capture the changing characteristics of exhibits in the spatiotemporal environment, and optimize the dynamic aggregation accuracy through feature adaptive iteration; S2. Based on the holographic information multimodal fusion engine, the visual holographic stream, natural language semantic stream and tactile feedback data are deeply aligned and fused in the multimodal space to generate a unified dynamic display semantic model; S3, through the high-dimensional interactive intention guidance mechanism, based on the multi-dimensional intention analysis and prediction of user interactive behavior, combined with the environment status and display goals, the personalized display intention model is updated in real time; S4. Exhibition path optimization strategy based on deep reinforcement learning, using deep deterministic policy gradient algorithm, through continuous modeling of user behavior and environmental feedback in state space, to achieve real-time adjustment of exhibit display path; S5. Use quantized detail enhancement generation technology, combine exhibit feature information with dynamic virtual environment, and generate a full range of exhibit detail virtual display effects; S6. Through the interactive prediction mechanism based on time and space perception, the user interaction sequence, exhibition scene dynamics and individual behavior intention are modeled in real time to achieve dynamic push and recommendation of multi-dimensional display content; S7. Build a lightweight quantum-enhanced classification model and use self-supervised learning to accurately classify cross-domain exhibit data to achieve efficient dynamic display management of exhibits; S8. An exhibit reshaping method based on virtual simulation mapping constructs an immersive multi-dimensional display effect of exhibits in a virtual exhibition space through cross-domain data generation and dynamic environment mapping technology.

2. A digital exhibition product display method based on deep learning according to claim 1, characterized in that: The S1 specifically includes: S11. Distributed collection and structured preprocessing of the multi-dimensional input data of the exhibits. The input data includes the static feature data X s , dynamic behavior feature data X d (t), environmental context perception data E(t) and user multimodal interaction data U(t), generate feature data stream through feature normalization, dimension compression and semantic mapping; S12. In the feature extraction module, a hierarchical neural encoder is used to deconstruct the input data hierarchically, converting the static feature data X s Encoded as a high-order feature vector F s , and process dynamic data X through a time series based recurrent neural network d (t), environmental data E(t) and interaction data U(t), generate the corresponding time series features F d (t), F e (t) and F u (t); S13. Construct a multi-input and multi-output feature optimization network based on a self-evolutionary feature optimization algorithm: F agg (t)=W s ·F s +W d (t)·F d (t)+W e (t)·F e (t)+W u (t)·F u (t); Among them, F agg (t) is the aggregate feature representation, W s , W d (t), W e (t), W u (t) are the dynamic weight coefficients of static, dynamic, environmental and interactive features, respectively. The weights are adjusted in real time through adaptive evolution strategy combined with feature importance function; S14, based on the optimization mechanism driven by user feedback signals, the optimization function is established using the real-time prediction error Δ(t) and the user response value R(t), and the feature weight W is optimized through gradient update i (t+1): Among them, η is the dynamic learning rate, L represents the loss function; S15, the optimized feature vector F opt (t) The data is input to the feature dynamic control module and processed by a multi-layer perceptron network to achieve modeling and adaptive expression of the spatiotemporal dynamic features of the exhibits.

3. According to the deep learning-based digital exhibition product display method of claim 1, it is characterized in that: The S2 specifically includes: S21, perform distributed collection and multimodal preprocessing on the visual holographic stream, natural language semantic stream and tactile feedback data, and generate the visual initial features V through the feature decoder raw (t), semantic initial feature S raw (t) and the initial tactile feature H raw (t), and use the adaptive denoising module to eliminate redundant information and obtain the optimized features V(t), S(t) and H(t); S22. In the multimodal feature alignment module, a feature projection network based on the self-attention mechanism is constructed to project the visual features V(t), semantic features S(t) and tactile features H(t) into a unified high-dimensional representation space, and feature alignment is achieved through dynamic weight adjustment: F align (t)=β v ·V ′ (t)+β s ·S ′ (t)+β h ·H ′ (t); Among them, F align (t) is the aligned multimodal feature, V ′ (t), S ′ (t) and H ′ (t) is the eigenvector after projection, β v , β s , β h Dynamic weights for feature alignment are updated in real time by an adaptive optimization strategy; S23, using the deep fusion engine based on the multi-layer attention mechanism to align (t) Perform feature fusion and generate fusion feature F through multi-layer attention mechanism. fusion (t): Among them, α i (t) is the dynamic attention weight of the i-th modality, F′ i (t) is the enhanced representation of multimodal features; S24, through the deep semantic parsing network to fusion feature F fusion (t) Perform semantic decomposition and structural modeling to generate a semantic hierarchical vector M semantic (t), extracting semantic logical relations and dynamic semantic expressions through multi-layer semantic decoders; S25, establish a semantic consistency verification module to check the generated semantic model M semantic (t) Perform semantic integrity verification and logical consistency check, and output an optimized dynamic display semantic model.

4. The method for displaying digital exhibition products based on deep learning according to claim 1, characterized in that: The S3 specifically includes: S31, collecting multimodal user interaction behavior data, including user touch action data T m (t), voice command data V n (t), gaze tracking data G p (t), posture recognition data P q (t) and the interaction context variable C r (t), combined with the environmental status data E k (t) and exhibit characteristic data O l (t), the collected data is structured preprocessed through the multimodal feature decoder to generate the initial interaction feature F x (t); S32, in the multi-dimensional intent parsing module, the initial feature F is embedded using the multimodal feature embedding network x (t) Perform high-dimensional feature projection to generate time series features F y (t), model the historical interaction behavior through a recursive network based on the self-attention mechanism to generate the prediction feature F z (t); S33. Based on user behavior characteristics and environmental context information, define a real-time dynamic intent optimization function: I u (t)=δ1·F y (t)+δ2·E k (t)+δ3·O l (t)+δ4·C r (t); Among them, I u (t) represents the dynamic user intention model, δ1, δ2, δ3, δ4 are weight parameters, which are dynamically updated through the optimization algorithm based on reinforcement learning; S34, through multi-objective joint optimization strategy u (t) Dynamically adjust and optimize the matching between user preference expression and display target by combining the intention guidance mechanism to generate a personalized display intention model I optimal (t): I optimal (t)=ψ1·I u (t)+ψ2·Δ(T m (t),V n (t),G p (t)); Among them, ψ1, ψ2 are weight factors for multi-objective joint optimization, Δ(T m (t),V n (t),G p (t)) represents the dynamic correction function of the user's interactive behavior characteristics; S35, optimize the personalized display intention model I optimal (t) Input the dynamic display configuration module, and use the dynamic decision engine to determine the display order of the exhibits. d , layout adjustment parameter L d and interaction mode A d Perform real-time optimization.

5. The method for displaying digital exhibition products based on deep learning according to claim 1, characterized in that: The S4 specifically includes: S41. Construct a state space model for exhibition path optimization. The state variable set is Σ(t)={U h (t),E j (t),R k (t)},U h (t) represents the user’s current behavior feature vector, E j (t) represents the environmental feedback signal, R k (t) Dynamic correlation characteristics between exhibits and user interests; S42. Action space Λ(t) based on deep reinforcement learning = {ξ1,ξ2,…,ξ m }, each action ξ i Indicates the adjustment methods of the exhibit display path, including exhibit position exchange, dynamic insertion, priority reordering and path extension; S43. Define the reward function Ω(t) for path optimization: Ω(t)=κ1·I u (t)+κ2·C q (t)-κ3·T m (t); Among them, I u (t) represents the user interest matching score, C q (t) represents the coherence score of the display path, T m (t) represents the time consumption of path adjustment, κ1, κ2, κ3 are the weight factors of the reward function; S44. Use the deep deterministic policy gradient algorithm to train the path optimization strategy network: Where Θ(t) is the parameter of the policy network, η is the learning rate, and Q(Σ(t),Λ(t)|Θ(t)) is the expected return value of the state and action combination, which is estimated and updated using a deep neural network; S45. In the path optimization decision module, combined with the user's real-time behavior data U real (t) and environmental dynamic feedback E dyn (t), the optimal path adjustment action sequence Λ is generated through the trained strategy network opt (t), and dynamically adjust the order of exhibits display P final (t).

6. The method for displaying digital exhibition products based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51. Collect multi-dimensional feature data of exhibits, including geometric information G x , surface material characteristics T y , Dynamic Behavior Sequence D z (t) and virtual environment mapping data E w (t), the original data is quantized through the feature quantization module to generate a quantized feature vector F p (t); S52, using quantized detail enhancement network to quantize feature F p (t) Perform detail enhancement and define the exhibit detail enhancement function: F enhanced (t)=α1·G x +α2·T y +α3·Φ(D z (t),E w (t)); Among them, F enhanced (t) represents the enhanced exhibit features, α1, α2, α3 are weight parameters, Φ(D z (t),E w (t)) represents the joint correction function between dynamic behavior and virtual environment; S53, combined with the virtual environment to enhance the feature F enhanced (t) Perform dynamic rendering and use the detail enhancement generator to build a dynamic rendering model R dynamic (t): R dynamic (t)=β1·F enhanced (t)+β2·Ψ(E w (t)); Among them, β1, β2 are rendering parameters, Ψ(E w (t)) represents the environment mapping optimization function; S54, through the multi-layer detail verification module, the rendering model R dynamic The geometric shape, surface texture and dynamic behavior effect of (t) are verified layer by layer, and the final optimized exhibit virtual display model M is generated. optimized (t); S55, the optimized virtual display model M optimized (t) Input the display control module, and adjust the display perspective, detail expression and dynamic response of the exhibits according to the real-time feedback of users through the dynamic interaction engine to generate a real-time updated virtual display effect V final (t).

7. The method for displaying digital exhibition products based on deep learning according to claim 1, characterized in that: The S6 specifically includes: S61. Collect multimodal interaction data of users, including user interaction time series features T x (t), dynamic state of exhibition scene S y (t) and individual behavioral intention B z (t), the data preprocessing module normalizes and hierarchically encodes the data to generate a spatiotemporal perception feature set F spatio (t) = {T x (t),S y (t),B z (t)}; S62, using a recurrent neural network based on spatiotemporal perception to spatio (t) Perform high-dimensional encoding and build a user interaction prediction model: P interaction (t)=θ1·Λ(T x (t))+θ2·Γ(S y (t))+θ3·Ω(B z (t)); Among them, P interaction (t) represents the user interaction prediction value, Λ(T x (t)),Γ(S y (t))、Ω(B z (t)) are the spatiotemporal encoding functions of temporal features, scene dynamics, and behavioral intentions, respectively, and θ1, θ2, θ3 are dynamic weight parameters; S63, combined with historical behavior patterns H m (t) and user real-time interaction features P interaction (t), defines the priority score function of the recommended content: R priority (t)=η1·P interaction (t)+η2·H m (t)+η3·C k (t); Among them, R priority (t) is the dynamic priority of the recommended content, η1, η2, η3 are priority weight parameters, C k (t) represents the matching score of the current exhibit feature; S64, prioritize recommended content R through multi-dimensional dynamic content push engine priority (t) is sorted and combined with the user preference distribution U pref (t) and scene dynamic feedback E dyn (t), generate real-time recommended content set D recommend (t); S65, using the dynamic display optimization module to recommend (t) Input the display logic controller, and generate a personalized dynamic display solution V by adjusting the display order, display style and interaction method in real time. dynamic (t).

8. The method for displaying digital exhibition products based on deep learning according to claim 1, characterized in that: The S7 specifically includes: S71. Collect cross-domain exhibit data, including static attribute features of exhibits G s , dynamic behavior sequence H t (t) and multimodal interaction information I u (t), the original data is normalized, noise filtered and high-dimensional embedded through the feature preprocessing module to generate the initial feature set F init (t) = {G s ,H t (t),I u (t)}; S72, build a lightweight quantum enhanced classification model for F init (t) Feature encoding and quantum enhancement: F enh (t)=δ1·Ψ1(G s )+δ2·Ψ2(H t (t))+δ3·Ψ3(I u (t)); Among them, F enh (t) is the characteristic representation after quantum enhancement, Ψ1(G s ),Ψ2(H t (t))、Ψ3(I u (t)) are the quantum enhancement functions of static features, dynamic features and multimodal interaction data, respectively, and δ1, δ2, δ3 are feature enhancement weight parameters; S73, using self-supervised learning mechanism to enh (t) Perform feature classification: Among them, L cls (t) is the classification loss function, Q(y j |(t)) is the category y j The predicted probability distribution of j Balance weights for categories; S74, through the lightweight classification network cls (t) Optimize and use the gradient descent method combined with the self-supervised learning strategy to adjust the model parameters and generate the optimized classification result R cls (t); S75, the classification result R cls (t) Input the exhibit dynamic management module, dynamically group and optimize the exhibits according to the classification category, and generate a display solution set S based on user needs and scene context disp (t).

9. The method for displaying digital exhibition products based on deep learning according to claim 1, characterized in that: The S8 specifically includes: S81. Collect cross-domain data of exhibits, including physical attribute characteristics of exhibits P a , Dynamic Behavior Sequence D b (t), material detail features M c and environmental context information E d (t), the data is normalized and multimodally encoded through the cross-domain data fusion module to generate the feature set F cross (t) = {P a ,D b (t),M c ,E d (t)}; S82, based on virtual simulation mapping technology cross (t) Perform dynamic feature reshaping: F remap (t)=κ1·P a +κ2·Ξ(D b (t),M c )+κ3·Ψ(E d (t)); Among them, F remap (t) represents the reshaped feature set, κ1, κ2, κ3 are feature mapping weight parameters, Ξ(D b (t),M c ) is the joint mapping function of dynamic behavior and material properties, Ψ(E d (t)) is the dynamic correction function of the environmental context; S83, using multi-dimensional virtual rendering engine to remap (t) Perform immersive multi-dimensional rendering to build a virtual display model R immersive (t): R immersive (t)=λ1·F remap (t)+λ2·Φ(E d (t)); Among them, λ1,λ2 are rendering parameters, Φ(E d (t)) represents the multi-dimensional rendering function of the virtual environment; S84, through the multi-layer verification module, the virtual display model R immersive (t) is hierarchically verified in terms of geometric shape, material details, dynamic behavior, and environmental interaction effects, and an optimized virtual exhibit model M is generated. optimized (t); S85, M optimized (t) Applied to virtual exhibition space, the dynamic scene control module is combined with the user's real-time interactive behavior data U e (t) and exhibition environment changes S f (t) Dynamically adjust the display details, interaction logic and environment adaptation of virtual exhibits to generate immersive multi-dimensional display effects V final (t).

Images