Smart display device adaptive adjustment method for digital exhibition

Abstract

This invention discloses an adaptive adjustment method for intelligent display equipment used in digital exhibitions, specifically relating to the field of digital exhibition technology. The method involves collecting multi-source data such as ambient light, visitor distribution, and sound through distributed lightweight sensor nodes; performing semantic fusion using a lightweight neural network to construct a dynamic exhibition hall context graph; instantiating an intelligent agent for each display device; generating resource scheduling proposals based on a shared reward mechanism and a multi-agent collaborative decision-making algorithm; an arbitrator detecting and resolving resource competition and effect offsetting conflicts; outputting a globally consistent set of control instructions; ensuring smooth execution of instructions by the devices; dynamically generating adapted display content based on contextual semantics; and continuously optimizing the decision-making model through online learning based on visitor behavior feedback. This invention achieves collaborative scheduling of equipment resources and adaptive content generation, improving the exhibition experience and resource utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of digital exhibition technology, and more specifically, to an adaptive adjustment method for intelligent display equipment used in digital exhibitions. Background Technology

[0002] With the rapid development of digital technology, traditional static and passive exhibition methods can no longer meet the growing demand of modern visitors for immersive and personalized experiences. Intelligent display equipment is widely used in museums, science and technology museums, corporate showrooms and other scenarios. However, how to enable these devices to sense the environment and the state of visitors, and to make intelligent and collaborative adaptive adjustments to improve the overall viewing experience and operational efficiency has become a key technical challenge facing the current digital exhibition field.

[0003] Currently, most existing technical solutions adopt centralized control methods based on preset rules or single sensors. For example, a relatively close existing technical solution is to deploy a central control server in the exhibition hall and connect it to several environmental sensors. The server presets a series of "if-then" rules (such as "if the ambient light value is higher than X lux, then adjust the brightness of the specified screen to Y%) to uniformly manage the display devices in the exhibition hall. This solution achieves a certain degree of automated adjustment by centrally collecting data, making logical judgments, and issuing instructions.

[0004] However, in actual use, it still has some shortcomings, such as relying on fixed rules, lacking flexibility, being unable to cope with complex dynamic changes during the exhibition, adopting a centralized processing architecture, having poor system scalability and the risk of single point of failure, lacking a collaborative decision-making mechanism and conflict resolution capability between devices, which can easily lead to conflicts between multiple devices in terms of resource competition or display effects, thus damaging the visitor experience, and the system cannot optimize itself based on implicit feedback from visitors, resulting in rigid strategies and difficulty in continuously adapting to the preferences of different visitor groups. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide an adaptive adjustment method for intelligent display devices for digital exhibitions, which solves the problems mentioned in the background art through the following solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an adaptive adjustment method for intelligent display equipment for digital exhibitions, including S1: distributed perception and semantic modeling, in which multiple lightweight sensor nodes asynchronously collect multi-source data, receive it through an embedded edge computing unit, and use a lightweight neural network model to fuse and label it to generate a dynamic and semantic exhibition hall context graph model;

[0007] S2: Intelligent Agent Decision-Making and Resource Scheduling. Instantiate intelligent agents for intelligent display devices, endow the devices with physical capability semantics and spatial location semantics. The agent receives the generated context graph and, based on a shared reward mechanism, generates a resource scheduling proposal for the device through an offline-trained multi-agent collaborative decision-making algorithm.

[0008] S3: Policy Conflict Resolution and Consistency Arbitration. The arbitrator receives proposals submitted by the intelligent agent and detects potential conflicts. Based on predefined priority rules and the current context graph, it uses rule and utility function evaluation methods to resolve conflicts and outputs a globally consistent and collaboratively executable set of device control instructions.

[0009] S4: Stealth execution and adaptive content generation. The intelligent display device executes the arbitrated control instructions and makes a seamless and smooth transition based on the context. The content management engine dynamically assembles and generates suitable display content fragments from the underlying media material library based on the current context semantics.

[0010] S5: Based on group implicit feedback, online learning continuously monitors the behavior response of visitors after adjustment as a group implicit feedback signal. Using the feedback signal, the decision model parameters of the intelligent agent are fine-tuned in an incremental learning manner to continuously align the strategy with the preferences of the current visitor group.

[0011] The technical effects and advantages of this invention are as follows:

[0012] 1. This invention overcomes the drawbacks of the existing technology where equipment "goes its own way" by introducing a multi-agent collaborative decision-making and arbitrator conflict resolution mechanism. It intelligently coordinates the actions of all display equipment in the exhibition hall, avoids interference caused by resource competition or effect cancellation, and ensures that the content display and sound and light effects transition smoothly in space and time and cooperate with each other, creating a highly harmonious and immersive overall exhibition environment for visitors.

[0013] 2. This invention constructs a dynamic context graph that can reflect the environmental state and visitor behavior in real time through distributed perception and semantic modeling. The content management engine dynamically assembles and generates highly relevant display content from the material library based on the semantics of the current scene, so that the display content can adaptively adjust with changes in crowd density, audience interest points, etc., which greatly enhances the attractiveness and effectiveness of the display.

[0014] 3. This invention introduces an online learning mechanism based on implicit feedback to continuously collect visitors' behavioral responses and fine-tunes the decision-making model through incremental learning. It continuously aligns with the preferences of the current visitor group and optimizes resource scheduling strategies, thereby improving the visitor experience while continuously increasing the energy efficiency of equipment and the overall utilization rate of system resources, forming a virtuous cycle system. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall structure of the present invention;

[0016] Figure 2 This is a schematic diagram illustrating the strategy conflict resolution and consistency arbitration of the present invention;

[0017] Figure 3 This is a schematic diagram of the adaptive content generation of the present invention;

[0018] Figure 4 This is a schematic diagram illustrating the fine-tuning of the online learning model of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] As attached Figure 1 Appendix Figure 2 Appendix Figure 3 and appendix Figure 4 The adaptive adjustment method for intelligent display equipment used in digital exhibitions includes S1: distributed perception and semantic modeling, in which multiple lightweight sensor nodes asynchronously collect multi-source data, receive it through an embedded edge computing unit, and use a lightweight neural network model to fuse and label it to generate a dynamic, semantic exhibition hall context graph model.

[0021] It should be specifically noted that the multi-source data includes ambient light intensity and color temperature, visitor distribution heatmap, local sound decibels and keyword fragment data, and the lightweight sensor nodes include environmental sensing nodes integrating light sensors, temperature and humidity sensors, infrared thermal imaging sensor nodes, and sound sensor nodes.

[0022] It should be further explained that in digital exhibition scenarios, key factors affecting the adjustment of intelligent display equipment include environmental conditions and visitor conditions. Ambient light intensity and color temperature directly determine the brightness, contrast, and color temperature parameter adjustment requirements of display equipment. If the light is too strong and the display brightness is insufficient, visitors will not be able to clearly see the display content. Visitor distribution heatmaps reflect the flow density of people in different areas of the exhibition hall. Display equipment in densely populated areas should prioritize ensuring operational stability and content update frequency, while avoiding equipment operating noise that may interfere with visitors in that area. Local sound decibel and keyword fragment data can determine whether there are explanation activities or hot topics of visitor discussion in the current area. If explanation sounds or specific topic keywords are detected, the display equipment needs to adjust the content playback rhythm and key display modules accordingly. Therefore, ambient light intensity and color temperature, visitor distribution heatmaps, local sound decibel and keyword fragment data are selected as multi-source sensing data.

[0023] A distributed lightweight sensor node deployment scheme is adopted, with one environmental sensing node integrating light and temperature / humidity sensors deployed in a grid area of ​​5 meters × 5 meters at every 5 meters on the top of the exhibition hall. The TSL2591 light sensor is selected to collect the ambient light intensity. At the same time, the color temperature detection function of the sensor is used to obtain the ambient color temperature data. The sampling frequency is set to 1 time / second to capture changes in ambient light in real time.

[0024] Infrared thermal imaging sensor nodes were deployed at the ground entrances and exits of the exhibition hall, the main exhibition area passages, and 2 meters in front of the exhibits. The FLIR Lepton 3.5 infrared thermal imaging module was selected, with a resolution of 160×120 and a frame rate of 27Hz, to generate a heat map of visitor distribution. The sampling frequency was set to 0.5 times / second to balance the real-time performance of data acquisition with energy consumption.

[0025] One sound sensor node is deployed within a 1.5-meter radius around each smart display device. The SGM3770 sound sensor is selected, with a frequency response range of 20Hz to 20kHz and a sensitivity of -46dB. It also integrates local voice segment acquisition and preliminary processing functions, collects local sound decibel data, and extracts a 3-second sound segment (sampling rate 16kHz, bit depth 16bit), with the sampling frequency set to 1 time / 2 seconds.

[0026] All sensor nodes adopt an asynchronous acquisition mode, independently acquiring data according to their own acquisition frequency, and transmitting it to the embedded edge computing unit via LoRa low-power wide area network. The transmission rate is set to 50kbps, and the transmission distance is 1-3 kilometers.

[0027] Multi-source data were preprocessed. Ambient light intensity and color temperature data were filtered using a sliding window method (window size set to 5) to remove outliers. If the light intensity value at a certain sampling time exceeded ±30% of the average value of the previous 5 samples, it was determined to be an outlier and replaced with the average value of the previous 5 samples. Visitor distribution heatmap data were processed using image morphology (dilation and erosion operations) to eliminate noise points. The dilation operation used a 3×3 rectangular structural element, and the erosion operation used a 2×2 rectangular structural element to ensure that the heatmap could accurately reflect the actual flow of people. Local sound decibel and keyword fragment data were first smoothed, and then the sound fragments were denoised (using spectral subtraction, with the signal-to-noise ratio improvement threshold set to 10dB). Sound features were extracted using MFCC (Mel-frequency cepstral coefficients) to obtain a 13-dimensional MFCC feature vector.

[0028] A lightweight neural network model is used to fuse and label the preprocessed multi-source data. By separating convolutional and linear bottleneck structures through deep learning, the number of parameters is significantly reduced while ensuring model performance. The preprocessed ambient light intensity and color temperature data (2D features), visitor distribution heatmap data (reduced to 10D features by dimensionality reduction, and retained 95% of the feature information using principal component analysis (PCA)), local sound decibel data (1D features), and sound segment MFCC feature vectors (13D features) are input into the MobileNetV2 model. The output layer of the model uses the softmax activation function to perform semantic annotation on the data. The annotation categories include 27 semantic categories of exhibition hall scenes such as "high / medium / low light - high / medium / low traffic - no discussion / explanation in progress / topic discussion".

[0029] Based on the labeled semantic data, a dynamic and semantic exhibition hall context graph model is constructed. The context graph is stored in the graph database Neo4j. The nodes in the graph include "environment state nodes", "visitor state nodes", and "time nodes", and the edges include "environment-visitor association edges" and "state-time association edges". The context graph model is updated every 30 seconds to ensure that the graph can reflect the dynamic changes of the exhibition hall in real time, providing accurate semantic data support for subsequent intelligent agent decision-making.

[0030] S2: Intelligent Agent Decision-Making and Resource Scheduling. Instantiate intelligent agents for intelligent display devices, endow the devices with physical capability semantics and spatial location semantics. The agent receives the generated context graph and, based on a shared reward mechanism, generates a resource scheduling proposal for the device through an offline-trained multi-agent collaborative decision-making algorithm.

[0031] It should be noted that the reward mechanism is divided into individual rewards and global rewards. Individual rewards evaluate the decision-making effect of a single intelligent agent, while global rewards evaluate the overall effect of collaborative decision-making among multiple intelligent agents.

[0032] It is necessary to further explain that the intelligent agent is endowed with physical capability semantics and location semantics: physical capability semantics is defined based on the device's hardware parameters and functional characteristics, while location semantics is defined based on the exhibition hall's spatial coordinate system. A three-dimensional coordinate system is established with the exhibition hall entrance as the origin, the X-axis along the length of the exhibition hall, the Y-axis along the width of the exhibition hall, and the Z-axis along the height. The installation location coordinates of each intelligent display device serve as the location semantic attribute of the intelligent agent. At the same time, the exhibition hall area where the device is located is marked (such as "Exhibition Area A on the first floor" or "Interactive Area B on the second floor"), which facilitates the consideration of spatial location relationships when making collaborative decisions among multiple intelligent agents.

[0033] The shared reward mechanism based on exhibition goals and visitor experience has core objectives including increasing visitor dwell time, optimizing resource utilization efficiency, and ensuring equipment coordination consistency. The reward function is divided into two parts: individual reward and global reward. The individual reward evaluates the decision-making effect of a single intelligent agent, while the global reward evaluates the overall effect of collaborative decision-making among multiple intelligent agents. The total reward value is the weighted sum of the individual reward and the global reward (with weights set to 0.4 and 0.6, respectively).

[0034] The individual reward function is designed as follows: If the visitor dwell time increases by more than 10% after the intelligent agent-controlled display equipment is adjusted, 20 points are awarded; an increase of 5%-10% is awarded; an increase of 0%-5% is awarded; and no increase or decrease is awarded. If the equipment's power consumption decreases by more than 15% after adjustment, 15 points are awarded; a decrease of 10%-15% is awarded; a decrease of 5%-10% is awarded; and no decrease or increase is awarded. If the semantic matching degree between the equipment content display and the current scene (evaluated by the subsequent online learning model) reaches more than 90%, 25 points are awarded; reaching 80%-90% is awarded; reaching 70%-80% is awarded; and below 70% is awarded.

[0035] The global reward function is designed as follows: If, after adjusting all intelligent display devices in the same exhibition area, the overall movement trajectory deviation rate (the proportion of deviation from the preset visitor route) of visitors is less than 10%, 30 points are awarded; less than 15%, 20 points are awarded; less than 20%, 10 points are awarded; and more than 20%, 0 points are awarded. If the resource utilization rate of all devices in the exhibition hall is less than 70%, 20 points are awarded; if all are less than 80%, 12 points are awarded; if some devices are above 80% but below 90%, 5 points are awarded; and if some devices are above 90%, 0 points are awarded. If the consistency score of the collaborative display effect of multiple devices (calculated by combining expert evaluation and visitor feedback) reaches 85 points or above (out of 100), 25 points are awarded; if it reaches 75-85 points, 15 points are awarded; if it reaches 65-75 points, 8 points are awarded; and if it is below 65 points, 0 points are awarded.

[0036] A multi-agent collaborative decision-making algorithm based on deep reinforcement learning is adopted. Through centralized training and distributed execution, collaborative decision-making of multiple agents is realized. A simulation environment of the exhibition hall scene is constructed. A virtual scene with a 1:1 scale to the actual exhibition hall is built based on the Unity3D engine. In the virtual scene, different environmental changes (such as dynamic changes in light intensity from 100 lux to 1000 lux and fluctuations in crowd density from 0.2 people / ㎡ to 1.5 people / ㎡) and visitor behaviors (such as stopping to watch, moving, discussing, and interacting) are simulated. 100,000 training samples are generated. The training samples include exhibition hall context semantic data (input features), intelligent agent decision-making actions (such as adjusting display brightness, switching content modules, and adjusting audio volume) and corresponding reward values ​​(labels).

[0037] The intelligent agent receives the exhibition hall context graph model generated by the embedded edge computing unit (extracting key semantic features from the graph to form a 26-dimensional input vector), inputs it into the Actor network model, and the model outputs preliminary resource scheduling action suggestions. The intelligent agent combines its own physical capability semantics and location semantics to verify the feasibility of the action suggestions. Each intelligent agent generates a resource scheduling proposal that includes device control actions (such as display parameter adjustment values, content module selection, and audio parameter settings), resource requirements (such as CPU usage requirements, network bandwidth requirements, and power consumption requirements), and expected effects (such as the expected increase in visitor dwell time and content matching degree). The proposal is then sent to the arbitrator via the MQTT protocol (Message Queuing Telemetry Transport Protocol, lightweight and low bandwidth consumption).

[0038] S3: Policy Conflict Resolution and Consistency Arbitration. The arbitrator receives proposals submitted by the intelligent agent and detects potential conflicts. Based on predefined priority rules and the current context graph, it uses rule and utility function evaluation methods to resolve conflicts and outputs a globally consistent and collaboratively executable set of device control instructions.

[0039] It should be specifically noted that the conflicts are divided into resource competition conflicts and effect offsetting conflicts. Resource competition conflict detection includes network bandwidth resources, power supply circuit power capacity, and central storage server storage space. Effect offsetting conflicts include the arbitrator constructing a semantic-equipment adjustment effect correlation matrix of the exhibition hall scene.

[0040] The conflict resolution types are divided into resource competition conflict resolution and effect offsetting conflict resolution. Resource competition conflict resolution includes predefining resource allocation priority rules, ranking proposals with resource competition based on priority rules, evaluating the resource utilization efficiency of each proposal using a utility function, and formulating a resource allocation scheme by combining priority and utility function values.

[0041] It should be further explained that the arbitrator receives resource scheduling proposals submitted by each smart agent via the MQTT protocol, sets up a proposal receiving buffer (with a capacity of 100 proposals), sorts the proposals according to the submission time, and performs format verification and integrity checks on the proposals. If a proposal is missing a device ID, control action parameters, or resource requirement information, it is determined to be an invalid proposal, and a resubmission instruction is returned to the corresponding smart agent. If the proposal format is complete and the parameters are within a reasonable range, it is determined to be a valid proposal and enters the conflict detection stage.

[0042] The arbitrator performs potential conflict detection on valid proposals based on preset conflict detection rules. Conflict types include resource competition conflicts and effect offsetting conflicts. The specific detection methods are as follows: resource competition conflicts occur when multiple intelligent agents' resource scheduling proposals have a demand for the same limited resource that exceeds the total resource capacity. The arbitrator establishes a resource status database to record the total capacity and currently occupied capacity of various shared resources in the exhibition hall in real time.

[0043] Network bandwidth resources: The exhibition hall uses gigabit Ethernet with a total bandwidth of 1000Mbps. The current network bandwidth usage of each device is recorded in real time (by collecting network switch port traffic data via SNMP protocol). Power supply circuit capacity: The exhibition hall is divided into 5 independent power supply circuits, each with a rated power capacity of 50kW. The current total power consumption of each circuit is recorded in real time (by collecting data from smart meters at a sampling frequency of 1 time / second). Central storage server storage space: The central storage server has a total storage space of 20TB. The used storage space capacity is recorded in real time (by querying the server file system).

[0044] For each resource scheduling proposal, its resource demand parameters are extracted and compared with the remaining capacity of the corresponding resource. If the sum of the demands of multiple proposals for the same resource is greater than the remaining capacity of that resource, then a resource contention conflict is determined to exist.

[0045] Effect offsetting conflict occurs when resource scheduling proposals from multiple intelligent agents cancel each other out in terms of display effect, resulting in a decrease in overall display effect. The arbitrator evaluates the display effect of the proposals and detects conflicts based on the association model between the semantics of the exhibition hall scene and the equipment adjustment effect.

[0046] The arbitrator constructs a semantic-equipment adjustment effect correlation matrix for the exhibition hall scene. The row dimension of the matrix is ​​the semantic category of the exhibition hall scene, the column dimension is the type of equipment adjustment action, and the matrix element is the expected effect score of a certain adjustment action in the scene (0-10 points, the higher the score, the better the effect). When the absolute value of the difference between the expected effect scores of the adjustment actions of two proposals in the current scene semantic is greater than 5, and the adjustment actions are aimed at the same influencing factor, it is determined that there is a conflict of effect cancellation.

[0047] For the detected resource competition conflicts and effect offsetting conflicts, the arbitrator resolves them by combining rule and utility function evaluation based on predefined priority rules and the current context graph. In the resource competition conflict resolution, predefined resource allocation priority rules prioritize resource allocation to display equipment in densely populated areas (population density > 1 person / ㎡), display equipment corresponding to core exhibits, and equipment with high energy efficiency (energy consumption per unit display effect < 5W / (nit・㎡)).

[0048] Proposals with resource competition are ranked based on priority rules, and the resource utilization efficiency of each proposal is evaluated using a utility function. The expected display effect score is calculated based on the correlation model between the current context graph and the device adjustment effect (0-10 points). The resource demand coefficient is set according to the resource type (network bandwidth resource coefficient is 1.2, power resource coefficient is 1.0, and storage resource coefficient is 0.8). The resource demand is the number of resources requested in the proposal.

[0049] By combining priority and utility function value, a resource allocation plan is formulated, prioritizing the resource needs of high-priority proposals. If the remaining resources can still satisfy some low-priority proposals, then the low-priority proposal with the highest utility function value is selected to allocate resources.

[0050] Based on the current context graph, the core requirements of the conflict scenario are determined, and conflict resolution rules are formulated. If the adjustment action of a certain proposal in the conflict proposal is more in line with the core requirements, the proposal is retained first, and the adjustment parameters of the other proposal are adjusted. If all conflict proposals partially meet the core requirements, a new adjustment scheme is generated by weighted fusion. The weights are set according to the matching degree between the proposal and the core requirements (the matching degree is calculated based on the semantic association model, 0-1.0).

[0051] After conflict resolution, the arbitrator integrates all valid proposals to generate a globally consistent and collaboratively executable set of device control instructions. Each control instruction includes a device ID, control parameters, execution time, and execution priority. The control instruction set is sent to the control modules of each intelligent display device via industrial Ethernet and stored in a central database for subsequent traceability and auditing. If the device control module does not receive the instruction within 100ms, the arbitrator automatically resends it. If three consecutive resending attempts fail, an alarm mechanism is triggered, indicating a device malfunction.

[0052] S4: Stealth execution and adaptive content generation. The intelligent display device executes the arbitrated control commands and makes a seamless transition based on the context. The content management engine dynamically assembles and generates suitable display content fragments from the underlying media material library based on the current context semantics.

[0053] It should be specifically noted that the control commands include display parameter adjustment, audio parameter adjustment, and content switching. The adaptive content generation includes the construction of the underlying media material library, dynamic content assembly and generation, wherein dynamic content assembly and generation includes semantic matching, material selection, module assembly, and effect optimization.

[0054] It should be further explained that after the control module of the intelligent display device receives the control command sent by the arbitrator, it first parses the command parameters, determines the adjustment type (display parameter adjustment, audio parameter adjustment, content switching), and uses a smooth transition algorithm to achieve invisible execution, avoiding abrupt phenomena such as screen flickering, sudden sound changes, and content jumps during the adjustment process.

[0055] For brightness, contrast, and color temperature parameters, a linear interpolation algorithm is used to achieve a smooth transition. When the change is less than 20%, the transition time is 1 second; when the change is between 20% and 50%, the transition time is 2 seconds; and when the change is greater than 50%, the transition time is 3 seconds.

[0056] For volume and sound effect parameters, an exponential interpolation algorithm is used to achieve a smooth transition. The volume adjustment transition time is fixed at 1.5 seconds. At the same time, the low-frequency, mid-frequency and high-frequency parameters are adjusted in real time through the sound effect equalizer to adapt to the current ambient noise (e.g., when the noise is >70dB, the mid-frequency volume is enhanced to improve speech clarity).

[0057] For content module switching, a fade-in / fade-out and sliding transition are used, with a transition time of 1.5 seconds, while ensuring the continuity of the content theme before and after the switch to improve the visitor experience.

[0058] During execution, the device control module collects real-time device operating status data (such as current brightness value, volume value, and content playback progress) and feeds it back to the arbitrator via the MQTT protocol. The arbitrator compares the feedback data with the command parameters. If the deviation is greater than 5%, a correction command is sent to ensure adjustment accuracy.

[0059] The content management engine is deployed on a central server. Based on the current context graph semantics, it dynamically assembles and generates suitable display content fragments from the underlying media material library, achieving precise "scene-content" matching. The underlying media material library adopts a layered and classified architecture, divided into a basic material layer, a theme material layer, and an interactive material layer. The basic material layer stores original media materials, including images (JPG / PNG format, resolution up to 8K), videos (MP4 / AVI format, resolution up to 4K, frame rate 30fps), audio (MP3 / WAV format, sampling rate 48kHz), and text (TXT / XML format, supports multiple languages). Each material is labeled with keywords. The theme material layer, which includes keywords, format parameters, and applicable scenarios, is based on the combination of basic materials to form theme modules. The theme modules include "Origin Videos," "Key Figure Images," "Technology Evolution Texts," and "Interactive Q&A Audios." Each theme module is labeled with its theme category and compatible device type. The interactive material layer includes interactive media materials, such as 3D models (GLB format, supporting gesture control), interactive games (HTML5 format, supporting touch operation), and virtual narrators (AI-driven, supporting voice interaction). Each interactive material is labeled with its interaction method (such as "gesture recognition," "voice control," and "touch operation") and resource requirements (such as CPU usage and memory requirements).

[0060] The media library employs a distributed storage architecture, storing popular media (accessed more than 10 times per minute) on edge caching servers (deployed on various floors of the exhibition hall), and less popular media on a central storage server. Fast media retrieval is achieved through a CDN (Content Delivery Network), with loading time controlled within one second. The content management engine, based on the semantic categories of the current context graph, generates content using a four-step process: semantic matching, media selection, module assembly, and effect optimization.

[0061] Semantic matching extracts key semantics from the context graph and matches topic modules in the material library using a semantic similarity algorithm. Topic modules with a similarity of ≥0.8 are included in the candidate list.

[0062] Material selection involves filtering materials from candidate theme modules based on device type and current environmental parameters. For example, for video wall displays in high-light scenarios, high-brightness images (brightness ≥ 800 nits) and high-contrast videos (contrast ratio ≥ 5000:1) are selected. For interactive projection in high-traffic scenarios, short-duration interactive materials (duration < 3 minutes) are selected to avoid visitors waiting too long.

[0063] The modules are assembled according to the logical structure of "introduction - core content - interactive extension" to form content segments. The total duration of the content is adjusted according to the density of people (5 minutes when there are many people and 10 minutes when there are few people).

[0064] Effect optimization involves optimizing the assembled content segments, including resolution adaptation, color correction, and text optimization, to ensure the best possible content display.

[0065] Once the content is generated, the content management engine pushes the content segments to the corresponding smart display devices via the HTTP / 2 protocol. After receiving the content, the devices cache it in their local storage (such as the 128GB SSD built into the video wall) to avoid repeated loading. At the same time, the engine records the content playback data (such as playback duration and number of interactions) and feeds it back to the content management engine for subsequent content optimization.

[0066] S5: Based on group implicit feedback, online learning continuously monitors the behavior response of visitors after adjustment as a group implicit feedback signal. Using the feedback signal, the decision model parameters of the intelligent agent are fine-tuned in an incremental learning manner to continuously align the strategy with the preferences of the current visitor group.

[0067] It should be noted that the visitor's behavioral responses include changes in dwell time, movement trajectory deviation rate, and changes in facial orientation and micro-expressions analyzed by non-contact sensors.

[0068] The online learning effectiveness evaluation system assesses the model fine-tuning effect from three dimensions: "decision accuracy", "experience optimization" and "resource efficiency". Specific indicators include decision accuracy, experience optimization and resource efficiency.

[0069] It should be further explained that the data collection on changes in dwell time was achieved using infrared thermal imaging sensors and cameras deployed in front of the display equipment. A target detection algorithm was employed to identify visitors, marking each visitor's ID and initial entry time. When a visitor left the equipment's detection range (distance > 5 meters), the departure time was recorded. Dwell time = departure time - initial entry time. Simultaneously, the average dwell time at the same equipment before and after the adjustment (average 10-minute dwell time before adjustment and average 10-minute dwell time after adjustment) was compared to calculate... ×100%.

[0070] The movement trajectory offset rate is collected by constructing a visual positioning system using multiple cameras deployed in the exhibition hall (one every 8 meters, for a total of 20). A multi-target tracking algorithm is used to obtain the real-time location coordinates of each visitor (updated once per second), generate movement trajectories, preset exhibition hall tour routes, and calculate the deviation distance between the visitor's actual trajectory and the preset route (the sum of deviation distances for every 10-meter segment). ×100%.

[0071] Facial orientation and micro-expression changes are captured by a camera in front of the device. A facial landmark detection algorithm is used to determine facial orientation. When the angle between the face and the device screen is less than 30°, it is considered "focused on the device"; when the angle is ≥30°, it is considered "unfocused on the device." The algorithm then calculates... .

[0072] Facial images are analyzed using micro-expression recognition algorithms to identify micro-expression types, count the proportion of each type of micro-expression, compare the proportion of micro-expressions before and after adjustment, and calculate the magnitude of change.

[0073] All feedback signals are transmitted in real time to the online learning server via a 5G network (transmission rate 1Gbps, latency <10ms). The server cleans and standardizes the data to form a feedback signal dataset.

[0074] The online learning server employs an incremental learning algorithm, using feedback signal datasets to fine-tune the decision model parameters of the intelligent agent. The specific process is as follows:

[0075] Correlation between feedback signals and decision-making models: Establish the correlation between feedback signals and decision-making model actions. For example, the rate of change in dwell time is correlated with the "content module selection" action, the movement trajectory offset rate is correlated with the "display parameter adjustment" action, and micro-expression changes are correlated with the "interactive material recommendation" action. The correlation strength is calculated using the Pearson correlation coefficient (|correlation coefficient|>0.6 is considered a strong correlation). Prioritize fine-tuning the model parameters that are strongly correlated with the feedback signals.

[0076] To avoid "catastrophic forgetting" of the model due to incremental learning (i.e., parameter fine-tuning in the new scenario overwrites the effective parameters of the original scenario), the EWC algorithm is used to impose constraints on the key parameters of the model: by calculating the importance weight of each parameter in the offline training stage (the contribution of the parameter to the model performance, the higher the contribution, the greater the weight), the update magnitude of high-weight parameters is limited when updating parameters in incremental learning (e.g., the update magnitude ≤ 5% of the original parameter value).

[0077] A multi-dimensional online learning effectiveness evaluation system was established to assess the model fine-tuning effect from three dimensions: "decision accuracy," "experience optimization," and "resource efficiency." Specific indicators and evaluation methods are as follows:

[0078] Decision accuracy indicators include content matching degree (the degree to which the actual displayed content matches the visitor's preferences) and adjustment accuracy rate (the deviation rate between the device adjustment parameters and the optimal parameters). Content matching degree is calculated by comparing the "percentage of visitors' micro-expressions of pleasure" and the "relevance of the content theme". If the percentage of pleasure is >40% and the content theme relevance is >0.8, the content matching degree is considered excellent. Adjustment accuracy rate is calculated by the deviation rate between the actual operating parameters of the device and the arbitrator command parameters. If the deviation rate is <3%, the adjustment accuracy rate is considered excellent. Decision accuracy indicators are statistically analyzed monthly. If the excellent rate is ≥85%, the online learning decision accuracy is considered to have met the standard.

[0079] Experience optimization metrics include the increase rate of average dwell time, the increase rate of attention ratio, and the reduction rate of movement trajectory deviation. By comparing the average data of 30 days before and after online learning, if the increase rate of average dwell time is ≥10%, the increase rate of attention ratio is ≥8%, and the reduction rate of movement trajectory deviation is ≥15%, the experience optimization effect is judged to be satisfactory.

[0080] Resource efficiency indicators include the average power consumption reduction rate and the resource utilization improvement rate (such as network bandwidth utilization and storage resource utilization). If the average power consumption reduction rate is ≥5% and the resource utilization improvement rate is ≥12%, the resource efficiency is considered to meet the standards.

[0081] If all three dimensions of indicators meet the standards, the fine-tuned model parameters will be used as the new baseline parameters for subsequent intelligent agent decision-making. If any indicators fail to meet the standards, the reasons will be analyzed and the online learning strategy will be adjusted. If the resource efficiency indicator fails to meet the standards, it may be due to an unreasonable setting of the resource demand coefficient in the utility function, which needs to be recalibrated based on actual resource usage data. In addition, an online learning iteration cycle is set, with incremental learning and model fine-tuning performed once every 7 days under normal circumstances. If there are significant changes in the exhibition hall scenario (such as changing core exhibits or holding temporary special exhibitions), an emergency iteration mechanism will be triggered to complete the collection of feedback signals, model fine-tuning, and effect evaluation within 24 hours to ensure that the decision-making model always matches the actual exhibition hall scenario.

[0082] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0083] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive adjustment method for intelligent display equipment used in digital exhibitions, characterized in that, include: S1: Distributed perception and semantic modeling: Multiple lightweight sensor nodes asynchronously collect multi-source data, receive it through an embedded edge computing unit, and use a lightweight neural network model to fuse and label it to generate a dynamic and semantic exhibition hall context graph model. S2: Intelligent Agent Decision-Making and Resource Scheduling. Instantiate intelligent agents for intelligent display devices, endow the devices with physical capability semantics and spatial location semantics. The agent receives the generated context graph and, based on a shared reward mechanism, generates a resource scheduling proposal for the device through an offline-trained multi-agent collaborative decision-making algorithm. S3: Policy Conflict Resolution and Consistency Arbitration. The arbitrator receives proposals submitted by the intelligent agent and detects potential conflicts. Based on predefined priority rules and the current context graph, it uses rule and utility function evaluation methods to resolve conflicts and outputs a globally consistent and collaboratively executable set of device control instructions. S4: Stealth execution and adaptive content generation. The intelligent display device executes the arbitrated control instructions and makes a seamless and smooth transition based on the context. The content management engine dynamically assembles and generates suitable display content fragments from the underlying media material library based on the current context semantics. S5: Based on group implicit feedback, online learning continuously monitors the behavior response of visitors after adjustment as a group implicit feedback signal. Using the feedback signal, the decision model parameters of the intelligent agent are fine-tuned in an incremental learning manner to continuously align the strategy with the preferences of the current visitor group.

2. The adaptive adjustment method for intelligent display equipment for digital exhibitions according to claim 1, characterized in that: The multi-source data includes ambient light intensity and color temperature, visitor distribution heatmap, local sound decibels and keyword fragment data, and the lightweight sensor nodes include environmental sensing nodes integrating light sensors, temperature and humidity sensors, infrared thermal imaging sensor nodes, and sound sensor nodes.

3. The adaptive adjustment method for intelligent display equipment for digital exhibitions according to claim 1, characterized in that: The reward mechanism is divided into individual rewards and global rewards. Individual rewards evaluate the decision-making performance of a single intelligent agent, while global rewards evaluate the overall performance of collaborative decision-making among multiple intelligent agents.

4. The adaptive adjustment method for intelligent display equipment for digital exhibitions according to claim 1, characterized in that: The conflicts are divided into resource competition conflicts and effect offsetting conflicts. Resource competition conflict detection includes network bandwidth resources, power supply circuit power capacity, and central storage server storage space. Effect offsetting conflicts include the arbitrator constructing a semantic-equipment adjustment effect correlation matrix of the exhibition hall scene.

5. The adaptive adjustment method for intelligent display equipment for digital exhibitions according to claim 1, characterized in that: The conflict resolution types are divided into resource competition conflict resolution and effect offsetting conflict resolution. Resource competition conflict resolution includes predefining resource allocation priority rules, ranking proposals with resource competition based on priority rules, evaluating the resource utilization efficiency of each proposal using a utility function, and formulating a resource allocation scheme by combining priority and utility function values.

6. The adaptive adjustment method for intelligent display equipment for digital exhibitions according to claim 1, characterized in that: The control commands include display parameter adjustment, audio parameter adjustment, and content switching. The adaptive content generation includes the construction of the underlying media material library, dynamic content assembly and generation, wherein dynamic content assembly and generation includes semantic matching, material selection, module assembly, and effect optimization.

7. The adaptive adjustment method for intelligent display equipment for digital exhibitions according to claim 1, characterized in that: Visitor behavior responses include changes in dwell time, movement trajectory deviation rate, and facial orientation and micro-expression changes analyzed by non-contact sensors.

8. The adaptive adjustment method for intelligent display equipment for digital exhibitions according to claim 1, characterized in that: The online learning effectiveness evaluation system assesses the model fine-tuning effect from three dimensions: "decision accuracy", "experience optimization" and "resource efficiency". Specific indicators include decision accuracy, experience optimization and resource efficiency indicators.

Images