Intelligent management system applied to exhibition center

Abstract

The invention discloses an intelligent management system applied to a convention and exhibition center, relates to the technical field of convention and exhibition intelligent management, and realizes unified and high-precision data input through protocol conversion and multi-mode denoising. Second-level alarm and topology reconstruction are supported in cooperation with anomaly detection and self-adaptive management on the edge side; the comprehensive interpretation precision of the passenger flow and the environment situation is improved by using a multi-modal fusion algorithm; then mapping the perception information into a virtual venue situation by means of a digital twinning and self-learning mechanism, and realizing continuous optimization of sensor calibration and algorithm weight; finally, multi-Agent complex network dynamics is introduced, interaction evolution of a large-scale crowd in an emergency scene is simulated, closed-loop management of active prediction and emergency layout is provided for intelligent exhibition, and leap-type improvement in multiple levels of monitoring precision, safety disposal, resource scheduling and the like is expected to be achieved so as to cope with exhibition layout adjustment.

Description

Technical Field

[0001] The present invention relates to the technical field of smart exhibition management, and in particular to a smart management system applied to exhibition centers. Background Art

[0002] In venues like super-large exhibition centers, stadiums, and smart campuses, hundreds to thousands of heterogeneous sensing devices—including high-definition cameras, infrared thermal imagers, RFID / Bluetooth positioning probes, environmental monitors (temperature, humidity, air quality, and noise), Wi-Fi probes, and facial and license plate recognition terminals—are often deployed simultaneously, connected to access control, fire protection, security robots, building automation, and energy management systems. These devices are distributed throughout exhibition halls, corridors, public areas, and backstage areas, often experiencing obstructions and blind spots due to booth reorganization, temporary partitions, weather changes, or nighttime lighting adjustments. Peak visitor density increases lead to video congestion, multipath attenuation from metal structures, and wireless signal interference. Inconsistencies in device models and communication protocols result in complex data formats and fragmented interfaces. Management expects to obtain detailed crowd heat maps, security status, and environmental indicators within seconds to support intelligent scheduling and predictive operations and maintenance, ensuring safety, enhancing the user experience, and optimizing energy consumption.

[0003] After searching, it was found that in the Chinese invention patent application publication number CN112202857A, a smart management system for exhibition centers was disclosed. It includes a system server group and several access control devices, several identity recognition devices, several self-service machines, several self-service invoice printers, several self-service inquiry machines, clients, and security network modules connected to the system server group. The present invention incorporates exhibition centers, organizers, exhibitors, buyers, exhibition-related service providers, and visitors into a unified management system through a secure data server and a sub-server separately connected thereto, thus achieving data interoperability and unified management. An integrated design was implemented at the architectural level, with a unified style, simple operation, and convenient later maintenance, which improves management efficiency, reduces labor costs, and increases corporate profits.

[0004] However, existing smart venue systems generally rely on decentralized subsystems:

[0005] CCTV video is manually monitored, and passenger flow statistics and environmental monitoring are independent, lacking a unified data aggregation and precise fusion mechanism. The core technical issue is: How can we achieve real-time data alignment and adaptive calibration across protocols and modalities in the face of dynamic changes in venue layout and multi-source heterogeneous data environments, thereby ensuring the precise consistency of multi-terminal sensory data in time, space, and numerical dimensions, and providing trusted input for subsequent AI fusion analysis and edge anomaly detection? This issue directly restricts the system's response speed to sudden security incidents, the accurate interpretation of high-density crowds, and the precise control of complex environmental parameters. It is a key bottleneck that urgently needs to be overcome in the refined management of smart exhibition venues.

[0006] To this end, the present invention provides a smart management system applied to a convention and exhibition center. Summary of the Invention

[0007] (1) Technical problems solved

[0008] In response to the shortcomings of the existing technology, the present invention provides an intelligent management system for convention and exhibition centers, which realizes unified and high-precision data input through protocol conversion and multimodal denoising; cooperates with anomaly detection and adaptive management on the edge side to support second-level alarm and topology reconstruction; and uses multimodal fusion algorithms to improve the comprehensive judgment accuracy of passenger flow and environmental conditions; then uses digital twins and self-learning mechanisms to map the perception information into virtual venue conditions to achieve continuous optimization of sensor calibration and algorithm weights; finally, multi-agent complex network dynamics is introduced to simulate the interactive evolution of large-scale crowds in emergency scenarios, providing closed-loop management of active prediction and emergency layout for smart exhibitions, in order to achieve leapfrog improvements in monitoring accuracy, safety disposal, resource scheduling and other aspects, solving the technical problems recorded in the background technology.

[0009] (2) Technical solution

[0010] To achieve the above objectives, the present invention is implemented through the following technical solutions: A smart management system applied to a convention and exhibition center includes: when massive amounts of heterogeneous sensor data are simultaneously aggregated, a multi-source perception network utilizes protocol conversion and time series alignment algorithms to perform format unification and noise filtering on the raw readings, transforming them into a multi-source data set;

[0011] After the preprocessed multi-source dataset has completed time series correction, the multimodal fusion module performs cross-modal aggregation on the feature vectors of each channel through convolutional networks and dynamic weight scheduling to generate a multimodal fusion vector;

[0012] When the multimodal fusion vector is sent to the edge node, the anomaly detection and edge adaptation engine uses anomaly scoring and perception topology reconstruction mechanisms to provide instant alerts and rapid device calibration for suspicious human traffic or extreme environmental data, generating adaptive perception configurations.

[0013] After completing the preliminary calibration of the multi-source fusion data, the digital twin and self-learning module integrates the multimodal fusion vector, anomaly score, and adaptive perception configuration into the situation vector of the virtual venue through the situation mapping function and self-learning feedback function, and dynamically updates the sensor weights.

[0014] Through multi-agent modeling and the introduction of network dynamics mechanisms in the update rules, the virtual crowd distribution is time-series calculated and the crowding effect is deduced. Simulation errors are introduced between the simulation results and the actual monitoring data to optimize the emergency evacuation plan and layout configuration.

[0015] Furthermore, the multi-source perception network configures a heterogeneous protocol conversion module on the gateway side, maps the communication protocols of various sensors to a unified interface protocol based on the sensor node identifier and network access identifier, maintains an access dynamic mapping table at the network aggregation end, and fixes the correspondence between the sensor node identifier and the perception channel.

[0016] Furthermore, the raw data from the perception channels are aligned to a unified time scale, and the data alignment transformation operation is defined based on the time series fitting method, and the time series misalignment is dynamically corrected during the alignment process; the aligned data are denoised by the weighted spectral renormalization function to obtain a multi-source data set.

[0017] Furthermore, a learnable feature mapping operator is constructed for each perception channel, and the multi-source data set collected from the perception channel is mapped into a feature vector; the feature vectors of all channels are spliced ​​or adaptively fused to obtain a multi-channel fused feature vector.

[0018] Furthermore, after introducing a learnable or adaptive weight set, the dynamic weight is calculated by the weight scheduling function to represent the weight of the perception channel: the weighted feature vector is weighted at the element level according to the obtained dynamic weight to obtain a multimodal fusion vector with dynamic collaborative characteristics. When the exhibition hall layout or the flow of people changes, the weight scheduling function updates the weight.

[0019] Furthermore, an anomaly detection function is used to measure the degree of deviation of the multimodal fusion vector in high-dimensional space and output an anomaly score. If the obtained anomaly score exceeds the preset dynamic threshold function, an anomaly event is generated, triggering edge adaptive perception and rapid calibration. The corresponding timestamp and anomaly level are written into the edge alarm queue, where the anomaly event contains anomaly intensity, timestamp, and associated channel information.

[0020] Furthermore, the decision function for edge adaptive perception is called to reallocate and activate some spare sensors, and to quickly calibrate key sensors to generate the perception network reconstruction decision set and sensor incremental calibration parameters, and synchronously update the calibration parameter matrix.

[0021] Furthermore, virtual entities corresponding to key entities are created in the digital twin space, and twin nodes are defined using sensor identification and network mapping information to obtain a node set in the digital twin space;

[0022] The situation mapping function is called to project the multimodal fusion vector, anomaly score and perception network reconstruction decision set into the digital twin to obtain the global situation vector; after the anomaly score is calculated, the anomaly score is projected into the risk intensity display of each node in the virtual venue through the situation mapping function.

[0023] Furthermore, given the global situation vector and internal simulation parameters at the current moment, a predictive scenario simulation function is introduced to perform scenario calculations for future periods to obtain twin state predictions;

[0024] The twin status prediction is presented in the digital twin visualization interface as a scene animation or trend chart, and outputs predictive indicators, including peak passenger flow periods, safety alarm probability, and environmental load.

[0025] Furthermore, the self-learning feedback function uses abnormal events, reconstructed decision sets and situation vectors to iteratively update sensor weights, calibration parameters and fusion algorithms. Among them: the self-learning feedback function automatically collects the differences between the actual scene and the twin simulation in each learning cycle, incorporates them into the learning loss function, drives the iterative correction of the sensor's calibration parameter matrix set and dynamic weights, and updates the status of the virtual entity node.

[0026] Furthermore, a multi-agent set is added to the digital twin space to construct the state vector of each agent. The state vector is used to describe its state parameters such as position, speed, target points of interest, emergency behavior tendencies, etc. in the digital twin environment, and define the interaction rules between people and scenes in the digital twin space.

[0027] Furthermore, for each agent, an update rule is defined in the digital twin space to simulate its behavioral evolution, adjacency relationships are defined in the agent set, and a network dynamics mechanism is introduced into the update rule; the status of all agents is updated in parallel in the digital twin space to generate the overall population distribution at the next moment.

[0028] Furthermore, an error measurement operator is constructed to compare the obtained virtual population distribution with the actual monitoring data. If the obtained simulation error is continuously higher than expected, the parameters are adjusted in the digital twin space;

[0029] If the simulation accuracy gradually improves, the simulation results will be used to adjust the venue layout, optimize emergency drill plans, or optimize security and evacuation strategies, generate evaluation results on the accuracy of crowd simulation and the effectiveness of strategies, and feed back key improvement suggestions or parameter corrections to the digital twin and self-learning engine.

[0030] (3) Beneficial effects

[0031] The present invention provides a smart management system for a convention and exhibition center, which has the following beneficial effects:

[0032] Step 1 implements the multi-source dataset X clean The alignment and cross-protocol unification of (t) ensure that multi-source sensor data maintains high availability and low noise in subsequent processes; its output not only filters redundant signals but also lays a solid foundation for multimodal fusion.

[0033] In the second step, under the collaboration of deep learning and dynamic weight allocation, the data of different channels are mapped into a multimodal fusion vector Y(t), which can flexibly respond to sudden changes in the exhibition hall layout, and use dynamic collaborative algorithms to maximize information complementarity and accuracy, thereby increasing the ability to make unified judgments on multi-dimensional factors such as human flow, security, and environment.

[0034] Step three uses the anomaly score δ(t) and adaptive perception network to achieve second-level processing on the edge side, closely integrating distributed AI with physical environment requirements, quickly detecting abnormal events and emergency linkage, and sharing necessary information summaries between edge nodes and the cloud, which not only reduces bandwidth load but also keeps key decisions within a controllable range, forming a pattern of local efficiency and global insight.

[0035] Step 4: Inject the situation vector U(t) into the digital twin model to establish a realistic mapping of the venue's global situation. Use the self-learning module to continuously absorb information such as the anomaly score δ(t) and dynamically update sensor parameters and fusion weights. This allows the digital twin to no longer be limited to visualization, but to form a cyclic evolution through self-learning and the actual system, pushing venue management towards deep intelligence with active prediction and iterative optimization.

[0036] Step five superimposes multi-agent crowd simulation on this foundation, aligning the situation vector U(t) with complex network dynamics to simulate crowd flow evolution in specific scenarios. The simulation results are then compared with the actual multimodal fusion vector Y(t) to refine the crowd model and layout strategy. This hybrid integration of large-scale simulation, digital twins, and edge-based adaptation enables early prediction of congestion and resource bottlenecks, enabling precise decision-making. Self-learning bridges the gap between virtual and real life, enabling rapid iteration of crowd safety, environmental control, and resource scheduling, generating comprehensive performance that is more in-depth and forward-looking than isolated technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flow chart of the smart management system for the convention and exhibition center of the present invention. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0039] See also Figure 1 The present invention provides a smart management system for a convention and exhibition center, including:

[0040] Step 1: When the data from cameras, infrared and RFID in the super-large exhibition center is reported through a unified gateway, the multi-source perception network synchronously calls the protocol conversion module and the high-order denoising algorithm to convert the original observation value X raw Perform communication alignment, format unification, and redundancy filtering operations to reconstruct it into a unified and time-synchronized multi-source dataset X clean (t);

[0041] The step 1 includes the following:

[0042] Step 101: Establishing a multi-source perception network and converting heterogeneous protocols

[0043] Define sensor node identification and network access identifier It is used to uniquely identify different sensor devices and their corresponding data stream sources; at the network level, by configuring a protocol conversion module at the main switch or edge gateway, the proprietary communication protocols of various sensors are mapped to a unified interface protocol. For example, the data of Zigbee, LoRaWAN or custom TCP / IP protocol is encapsulated into a unified format and identified as a sensor node. Manage the index.

[0044] For different data types (video stream, infrared signal stream, RFID reading, environmental monitoring value, etc.), multi-source perception channels (denoted as etc.), and maintain access dynamic mapping table (MappingTable) at the network aggregation end to identify sensor nodes Corresponding perception channel The corresponding relationship between them is fixed so that the next step can directly reference these identification information for data processing.

[0045] When in use, the sensor node is identified and network access identifier The management mode can be quickly accessed when the venue scale is expanded or new types of sensors are deployed without large-scale changes to the network structure; a dynamic mapping table (MappingTable) is used to manage sensor node identification and sensory channels The mapping relationship enables the network to be quickly remapped when the venue layout is temporarily changed or the equipment is updated, which can improve the overall adaptability.

[0046] Step 102: Data alignment and advanced denoising preprocessing

[0047] Due to different perception channels ( The data sampling frequency (e.g., the sampling frequency of the data) is different from the data distribution delay. Therefore, the original data needs to be aligned with the unified time scale t. The data alignment transformation operation Γ is defined as follows:

[0048] X align (t) = Γ(X raw (t-τ1),X raw (t-τ2),…,X raw (t-τ n );β)

[0049] Where: X raw (t-τ n ) represents the sensor n at time t-τ n The original data of τ n Data delay estimation of the nth sensing channel;

[0050] β is the internal reference parameter of the alignment operator, which is used to adjust the weight distribution of the historical window during the alignment operation. For example, when interpolating or compensating between a multi-frame rate video stream and low-frequency infrared data, β can adjust the smoothness of the interpolation function.

[0051] The Γ operation can be based on wavelet filtering, singular spectrum reconstruction (SVD-based) or other advanced time series fitting methods to dynamically correct time series misalignment during the alignment process;

[0052] To ensure the accuracy of subsequent multimodal data fusion, it is necessary to perform denoising on the aligned data. In order to enhance the anti-interference ability and take into account the characteristics of multimodality, the following weighted spectral reshaping function Θ can be defined:

[0053] X denoise (t) = Θ(X align (t),λ,γ)

[0054] Where: X denoise (t) is the aligned multi-source data vector; λ is the adaptive filter threshold adjustment factor, which is used to dynamically determine the attenuation coefficient of each frequency band; γ is the smoothness control factor of the denoising algorithm, and its value range is determined according to the requirements;

[0055] The weighted spectral reshaping function Θ can not only attenuate high-frequency noise using hybrid wavelet transform, but also retain the core characteristic signal at the low-frequency end. Through frequency domain weighting, it can effectively filter out various irregular interferences in the environment, such as local occlusion or sudden electromagnetic interference.

[0056] The sensor node identifier established in step 101 and channel identification Inherited to the result after processing in this step, we get the aligned and clean multi-source dataset X clean (t).

[0057] When used, the delay difference of multi-source data is corrected by the Γ operator to achieve more accurate time sequence alignment, avoiding confusion of character trajectory or environmental measurement caused by data dislocation during subsequent fusion. The weighted spectrum reshaping function Θ is used for weighted spectrum reshaping, which can achieve differential attenuation of the noise characteristics of each between different perception channels, and finally obtain the multi-source data set X clean (t),Introducing advanced timing reconstruction algorithms in the alignment operation can significantly reduce the data mismatch rate in high dynamic scenarios.

[0058] Step 2: When the preprocessed multi-source dataset X clean (t) When the timing correction is completed and can be used for deep mining, the multimodal fusion module uses convolutional networks and dynamic weight scheduling to calculate the feature vector F of each channel. x (t) performs embedding mapping and adaptive aggregation to generate a multimodal fusion vector Y(t), which accurately depicts the flow of people and the environment while eliminating single-source blind spots;

[0059] The second step includes the following:

[0060] Step 201: Feature extraction and latent space mapping

[0061] For each perception channel Define a learnable feature mapping operator Φ x , from the perception channel The clean data collected at time t is represented as Perception Channel The feature vector in the latent space is denoted as F x (t):

[0062]

[0063] Where: Represents the perception channel Denoised data at time t; η x For the perception channel The feature extraction hyperparameters used to control the feature mapping operator Φ x In the settings of convolution kernel size, attention gate structure or autoencoder depth, different channels can use different hyperparameters η x Φ x Use multi-layer deep networks or other advanced embedding functions, such as multi-head attention networks, to extract high-order pattern information of the modality;

[0064] In order to obtain cross-modal correlation information, the feature vectors F of all channels are x (t) Perform concatenation or adaptive fusion to obtain a unified potential representation F all (t), define the aggregation operation Used to represent the stacking and alignment of different eigenvectors:

[0065]

[0066] Among them F all (t) is the preliminary multi-channel fusion feature vector, which is still in the high-dimensional latent space and has not yet been assigned any weight or dynamic coordination strategy.

[0067] When used, through the mapping operator Φ x By performing high-order nonlinear mapping on the data of each channel, potential correlations between different modalities such as vision and infrared, which are difficult to capture with a single channel, can be found; by aggregating features, the relative independence of each channel can be retained, and the exclusive mapping operator Φ is used for different channel data. x Hyperparameter η x By customizing the optimal extraction method for each modality before fusion, we can achieve more fine-grained differentiated processing, rather than blindly processing all modalities with the same model. During latent space aggregation, instead of performing direct linear concatenation, high-dimensional attention or adaptive mechanisms can be combined to ensure that the features of different channels retain maximum plasticity during subsequent dynamic collaboration.

[0068] Step 202: Dynamic weight allocation and collaborative output

[0069] In order to achieve adaptive fusion of multiple channels, a learnable or adaptive weight set {ω x (t)}, let ω x (t) represents the sensing channel at time t The weight of the equation satisfies: 0≤ω x (t)≤1, and

[0070] Dynamic weight ω x The value of (t) can be dynamically generated by comprehensively evaluating the real-time status of the venue (such as the current crowd distribution, emergencies, and the availability of various sensors), or it can be calculated by the weighted scheduling function Ψ:

[0071] ω x (t)=Ψ(F all (t),ξ)

[0072] Where: ξ represents the set of hyperparameters within the weight scheduling function, such as attention threshold, local density threshold or adaptive threshold; Ψ can be based on the multi-head attention mechanism or graph network structure of Transformer to obtain the multi-channel fusion feature vector F all (t) learns the importance distribution of each mode at present;

[0073] According to the dynamic weight ω obtained x(t), the final multimodal fusion output Y(t) can be defined as:

[0074]

[0075] Where: ⊙ is the vector element-level weighted operation; F x (t) is the single-channel latent feature vector;

[0076] Through this weighted summation, a multimodal fusion vector Y(t) with dynamic collaborative characteristics can be obtained. When the layout of the exhibition hall or the flow of people changes, the weight scheduling function Ψ will update the weight ω x (t), so that the features of important channels account for a larger proportion in the fusion results, thereby achieving cross-modal adaptive scheduling.

[0077] When using, when some sensor data in the venue is blocked or faulty, resulting in a single-channel potential feature F x (t) When an abnormal offset occurs, the weight scheduling function Ψ can automatically lower its dynamic weight ω x (t), to avoid the mode from interfering with the overall result, based on the multi-channel fusion feature vector F all (t) and dynamic weight ω x (t) can quickly generate a multimodal fusion vector Y(t) without repeatedly going back to the original data for large-scale retraining, meeting the real-time requirements of the exhibition center;

[0078] Introducing the weight scheduling function Ψ to multi-channel fusion feature vector F all (t) Conduct context-aware dynamic analysis to achieve cross-modal importance allocation, rather than simply fixing the weight of a certain channel based on prior experience; through element-level weighted summation, the contribution of single-channel features in multimodal fusion can be refined to the dimensional level, and feature components of different dimensions can be independently amplified or attenuated.

[0079] Step 3: When the multimodal fusion vector Y(t) is sent to the edge node and needs to identify safety and passenger flow anomalies in real time, the anomaly detection and edge adaptation engine uses the high-dimensional anomaly score δ(t) and the local perception topology reconstruction algorithm to provide instant alerts and rapid equipment calibration for suspicious passenger flow or extreme environment data, generating the perception network reconstruction decision set R(t);

[0080] The step three includes the following:

[0081] Step 301: Real-time anomaly detection

[0082] Since the multimodal fusion vector Y(t) already contains multi-source information such as vision, infrared, RFID, and environmental sensors in the cross-modal dimension, the vector can more comprehensively reflect the current situation of the venue;

[0083] To measure the degree of deviation and fluctuation of the multimodal fusion vector Y(t) in high-dimensional space from time t-Δt to time t, the following integral anomaly detection function Ω is defined to output the anomaly score δ(t):

[0084]

[0085] Where: δ(t) is the anomaly score. A larger value means a stronger anomaly signal or a sharp fluctuation occurs in the time window [t-Δt, t]; Δt is the size of the lookback time window;

[0086] Y(s) is the multimodal fusion vector at time s, which contains the information after fusion of multi-source features such as vision, infrared, and RFID;

[0087] H is a symmetric semi-positive definite matrix (or positive definite matrix), which is used to perform weighted measurement on each dimension of Y(s);

[0088] is the first-order derivative (gradient) of the multimodal fusion vector Y(s) with respect to time s, which describes the rate of change of Y(s) in the time domain. represents its Euclidean norm (size);

[0089] β is the fluctuation sensitivity coefficient, which is used to balance the numerical deviation in high-dimensional space Rate of change with time Importance in anomaly detection;

[0090] If the obtained abnormal score δ(t) exceeds the preset dynamic threshold function Δv thd When Δv is Δv, it is considered that an abnormal event ε(t) occurs at time t; at this time, the subsequent edge adaptive perception and fast calibration (step 302) can be triggered; thd This is the internal hyperparameter adapted to the dynamic environment of the venue in this step. It is used to dynamically update the specific value of Δ(·) based on factors such as the current crowd density, sensor health, and historical normal data distribution.

[0091] If an abnormal event ε(t) is generated, the corresponding timestamp and abnormal level are written to the edge alarm queue. The abnormal event ε(t) contains the abnormal intensity (represented by the abnormal score δ(t)), timestamp, associated channel information, etc.

[0092] When used, it can ensure sensitive capture of anomalies in different exhibition hall layouts, crowd density and sensor status changes, avoiding the traditional method being easily interfered by simple noise, and the dynamic threshold function Δv thresholdDynamic adjustments can be made in real time based on the venue's historical data and current context, enhancing the accuracy of detecting unexpected events (such as short-term surges in passenger traffic). Using methods such as manifold comparison or local topology measurements, rare patterns or extreme data can be more flexibly identified in high-dimensional fused feature spaces. Incorporating adaptive decision thresholds into the hyperparameter v for online learning allows the anomaly detection algorithm to continuously evolve as the venue operates.

[0093] Step 302: Edge adaptive perception and fast calibration

[0094] After receiving the abnormal event ε(t), obtain the multimodal fusion vector Y(t), the dynamic mapping table of the sensor (used to locate the specific device and sensory channels

[0095] Define a decision function Π for edge adaptive sensing, reallocate and activate some backup sensors based on currently detected anomalies, sensor availability, and venue physical conditions, and quickly calibrate key sensors;

[0096] Its core execution logic is:

[0097] R(t)=Π(ε(t),Y(t),ζ)

[0098] Where: R(t) is the perception network reconstruction decision set, which records the sensor identifiers that need to be activated or deactivated and its new sampling frequency, network routing, edge computing node allocation, etc.;

[0099] ζ is a hyperparameter, which may involve the maximum allowed network load, the shortest sampling period limit, temporary layout change information, etc.

[0100] The decision function Π can make dynamic perception decisions based on the load assessment model on the edge side, physical distribution constraints, and judgment of the anomaly type (such as security anomalies or environmental anomalies).

[0101] For sensors that may experience accuracy deviations, or for specific sensing channels where noise increases suddenly in abnormal environments, fast calibration operations can be performed:

[0102]

[0103] in: Representative perception channel Previous calibration parameter matrices (such as the imaging geometric transformation matrix of the visual channel camera, the sensitivity factor of the environmental sensor, etc.);

[0104] κ is the incremental calibration value automatically calculated based on the current anomaly type and sensor health status;

[0105] It represents a nonlinear combination with the original parameter matrix and can be solved iteratively based on the comparative data of the benchmark measurement point or the adjacent sensor;

[0106] This process is performed on the edge side, which can greatly speed up the correction of sudden failures or abnormal drifts.

[0107] Step 302 finally produces the perception network reconstruction decision set R(t) and the calibration parameter matrix after calibration Synchronous updates enable multi-source data collection and fusion at subsequent moments to be performed under the new configuration.

[0108] When used, after detecting an anomaly, the edge node can quickly reconstruct the local perception network based on the abnormal event ε(t) and the multimodal fusion vector Y(t), without waiting for a long link decision from the cloud. This can significantly reduce the delay in responding to emergencies. Through a rapid calibration mechanism, sensor parameters can be adjusted promptly to maintain subsequent data quality when a local scene suddenly changes (for example, when a large device is suddenly added to an exhibition area, causing infrared sensors to saturate). Tightly binding the abnormal event ε(t) to the rapid calibration operation allows for the reconstruction of sensor topology and parameters in the shortest possible time, without relying on manual intervention or centralized update processes.

[0109] Step 4: When the edge has generated the perception network reconstruction decision set R(t) and completed the preliminary calibration of the multi-source fusion data, the digital twin and the self-learning module use the situation mapping function Φ DT and the self-learning feedback function Γ Learn , the multimodal fusion vector Y(t), the anomaly score δ(t), and the perception network reconstruction decision set R(t) are integrated into the situation vector U(t) of the virtual venue and the sensor weights are dynamically updated to form a growable twin model;

[0110] The step 4 includes the following contents:

[0111] Step 401: Digital twin model construction and real-time situation mapping

[0112] To connect with key entities in the physical venue (venue area, sensor equipment, perception channel etc.) one-to-one correspondence in the digital twin space Create virtual entities corresponding to key entities and use them with sensor identifiers and network mapping information to define the twin nodes. At this time, the digital twin space The node set of

[0113]

[0114] in Represents the twin node corresponding to a sensor entity, d Zone It can represent a virtual representation of an exhibition area or a specific functional area, etc. It represents the twin nodes that correspond one-to-one to the physical sensor entities and the twin nodes that correspond to specific sensing channels in the digital twin space;

[0115] In order to project the multimodal fusion vector Y(t) and anomaly score δ(t) obtained in the previous step into the digital twin, a situation mapping function Φ is constructed. DT , the specific form is as follows:

[0116] U(t)=Φ DT (Y(t),R(t),δ(t),χ)

[0117] U(t) is the global state vector in the digital twin space, recording the status updates of each twin node (such as temperature and humidity conditions, safety risk level, and crowd density). χ is a defined set of internal hyperparameters or rules, such as the coordinate system of the twin space, the visualization style of each node, and the connectivity topology between expansion zones.

[0118] After calculating the anomaly score δ(t), the digital twin platform uses the situation mapping function Φ DT The score is projected as a risk intensity display for each node in the virtual venue. After receiving the perception network reconstruction decision set R(t) and the risk intensity display, the parameters of the corresponding nodes are synchronously updated in the virtual model to display the changes in sensor topology and risk levels in real time.

[0119] When in use, the global situation vector U(t) graphically displays crowd heat, dangerous hotspots, security trends, etc., providing managers or upper-level business modules with a global perspective; the twin model also stores the associated information of the sensor topology (that is, the perception network reconstruction decision set R(t), the multimodal fusion vector Y(t) and the anomaly score δ(t)), breaking through the limitations of traditional digital twins that often only perform static visualization or single-dimensional parameter mapping, and forming a dynamic and interconnected venue ecological model.

[0120] Step 402: Predictive scenario simulation and strategy deduction

[0121] In the digital twin space On the global situation vector U(t), the predictive scenario simulation function Ψ is introduced Sim , given the global situation vector U(t) at the current time t, the predictive scenario simulation function Ψ Sim The scenario calculation will be performed for the future period [t, t+T]:

[0122]

[0123] Among them, U future (s) represents the twin state prediction at time s (s≤t+T);

[0124] Internal simulation parameters, such as passenger flow growth model, exhibition hall layout change assumptions, and probability distribution of possible emergencies;

[0125] Predictive scenario simulation function Ψ Sim It can be based on high-level methods such as differential equations, reinforcement learning simulation or multi-agent system simulation, rather than simple linear extrapolation;

[0126] When making predictions, different strategic assumptions can be imported simultaneously (e.g. emergency evacuation plan A, air conditioning load balancing plan B, security personnel scheduling plan C, etc.), their effects can be simulated and compared in parallel in the digital twin model, and several groups of assumptions can be generated. Combined with the objective function (such as energy consumption, safety risk, personnel congestion, etc.), it provides managers with optimization decision-making suggestions.

[0127] Twin state prediction U future (s) It can be displayed in the digital twin visualization interface as a scene animation or trend chart, and predictive indicators (such as peak passenger flow period, safety alarm probability, environmental load) can be reported to the upper-level business platform as specific values ​​or icons;

[0128] Before an event occurs in a real venue, the impact can be quickly assessed through multiple scenarios in the virtual scene, significantly shortening the decision-making time for temporary layout adjustments or emergency response. Leveraging multimodal fusion data Y(t) and anomaly information δ(t), the twin model has a more accurate input foundation, improving simulation accuracy and reliability. Through parallel multi-strategy simulation, different layouts or emergency plans can be switched in the digital twin with a single click, quantifying the impact on passenger flow, safety, and the environment in real time, surpassing the traditional approach of only rough simulation or single-path prediction. Using differential equations, reinforcement learning, or multi-agent methods to construct virtual scene dynamics can further broaden the depth of simulation and deduction.

[0129] Step 403: Self-learning closed loop and sensor weight iteration

[0130] Comprehensively utilize abnormal events δ(t), reconstructed decision set R(t), situation vector U(t), etc. to continuously iteratively update sensor weights, calibration parameters and fusion algorithms, and introduce self-learning feedback function Γ Learn :

[0131]

[0132] Where: W newIndicates the new optimal version of the calibration parameters or other sensor configurations used in the previous step of anomaly detection and edge adaptive processing, which can cover all sensors in multimodal fusion Processing weights, offsets, etc.

[0133] Represents the dynamic weight ω of each perception channel defined by the dynamic collaborative algorithm in step 2 x The optimization results of , after several rounds of learning, achieve more accurate cross-modal fusion; Hyperparameters include learning rate, historical memory decay factor, and alignment strategy between the twin model and physical data.

[0134] Self-learning feedback function Γ Learn In each learning cycle (or when an abnormal event occurs), the difference between the actual scene and the twin simulation is automatically collected and incorporated into the learning loss function to drive the calibration parameter matrix set W of the sensor and the dynamic weight ω x Iterative correction. The situation vector U(t) also updates the state of the virtual entity node after each iteration to ensure that the model gradually converges to the real situation; when the calibration parameter set W is updated new and optimized dynamic weights After confirmation, these updated parameters can be used in the next data collection and fusion process, achieving true accuracy with use.

[0135] When in use, by incorporating the deviation between real observations and twin predictions into the closed-loop learning process, dynamic correction of sensor offsets, algorithm weights, etc. is achieved. For convention and exhibition centers with frequent layout changes, this self-learning mechanism can continuously track venue structure adjustments to ensure that sensor coverage and fusion weights evolve accordingly; the digital twin prediction error (the difference between the physical reality of the venue and the twin simulation) is directly incorporated into the learning feedback link, and multiple constraints are optimized by combining multiple information such as abnormal information and edge adaptive decision R(t). That is, it can not only dynamically correct the sensor's own deviation W but also adjust the dynamic weight ω of the cross-modal fusion algorithm in real time. x , realizing dual self-learning of parameters and weights.

[0136] Step 5: After obtaining the iterative situation vector U(t), the large-scale crowd simulation component is updated by multi-agent modeling and the rule Θ crowd The network dynamics mechanism is introduced to perform time-series calculation and crowding effect deduction on the virtual crowd distribution S(t), and the simulation error ∈(t) is introduced between the simulation results and the actual monitoring data to optimize the emergency evacuation plan and layout configuration.

[0137] The step five includes the following:

[0138] Step 501: Multi-Agent Model Construction and Digital Twin Scenario Extension

[0139] In the digital twin space Add a multi-agent set representing individuals (or group units) in the Denoted as:

[0140]

[0141] Where N is the size of the virtual population, Represents the i-th Agent, usually corresponding to one or a group of visitors with similar behavioral characteristics; define the state vector s of each Agent i (t) is used to describe its state parameters such as position, speed, target points of interest, and emergency behavior tendencies in the digital twin environment. The state vector may include:

[0142]

[0143] where x i (t),y i (t) represents the coordinate position of the Agent, v i (t) can represent the speed or direction of movement, ρ i (t) can reflect characteristics such as urgency or crowd sensitivity;

[0144] Based on the venue coordinate system and situation vector U(t) in the digital twin model in the fourth step, in the digital twin space The interaction rules between people and scenes are defined in the virtual space (such as traffic constraints, queuing rules, danger zone avoidance, etc.), so that the multi-agent system in the virtual space is consistent with the actual physical layout;

[0145] When used, by defining the Agent set and state vector, the different behaviors of different visitors in terms of spatial movement, point of interest selection, emergency evacuation path, etc. can be simulated. Through the scene expansion operation, the real-time situation information (such as abnormal location, security situation) contained in the situation vector U(t) can directly affect the Agent behavior, ensuring the dynamic consistency of the physical and virtual crowds; the advanced behavioral characteristics such as urgency and crowd sensitivity ρ i (t) Incorporating the Agent state vector can present a more realistic crowd interaction model in emergencies or high-density exhibition area scenarios.

[0146] Step 502: Complex network dynamics simulation and multi-agent collaborative calculation

[0147] For each In the digital twin space Define a set of high-dimensional dynamic equations or update rules Θ in crowd, which is used to simulate the behavior evolution from time t to t+Δt, can be formalized as:

[0148] s i (t+Δt)=Θ crowd (s i (t),U(t),α i ,ζ)

[0149] Among them, s i (t) is The state vector at time t;

[0150] U(t) is derived from the digital twin potential output in the fourth step (including venue layout, security hotspots, historical passenger flow density, etc.), providing an external environment reference for the agent's behavior;

[0151] α i express Unique behavioral preferences or attributes (such as whether there is a VIP channel, whether one is extremely sensitive to crowding, etc.);

[0152] ζ is a set of independent simulation hyperparameters for this step, such as step size Δt, avoidance intensity, following distance, etc.

[0153] In order to consider the mutual influence between people, we can define the adjacency relationship (such as communication, proximity or social relationship) in the agent set and crowd The network dynamics mechanism (such as the attraction / repulsion of neighbor behavior, group orientation synchronization, and bottleneck effect) is introduced into the network, which can be represented as an adjacency matrix M (dimension N×N), where M ij >0 characterization and The interaction intensity in the digital twin space The state of all agents is updated in parallel, and the overall crowd distribution S(t+Δt) at the next moment is recorded. Numerical simulation or multi-threaded GPU parallelism can be used to accelerate this process, meeting the computational requirements of large-scale scenarios (N can reach tens of thousands or even more).

[0154] Environmental indices and abnormal alarm information in the situation vector U(t) can directly influence the path or speed chosen by the agent, reflecting the bidirectional influence between the crowd and the environment in real-world scenarios. Embedding the situation vector U(t) into the core crowd dynamics equations enables real-time coupling between the environment and the crowd. Using the adjacency matrix M in large-scale parallel computing, the system enhances group collaboration and is more algorithmically adaptable to the dynamic calculations of tens of thousands of agents.

[0155] Step 503: Comparison of simulation and actual monitoring data and strategy feedback

[0156] The virtual crowd distribution S(t) obtained after the digital twin simulation is compared with the actual monitoring data to construct the error measurement operator Π comp , of the following form:

[0157] ∈(t)=Π comp (S(t),Y(t),γ)

[0158] Where S(t) is the virtual crowd state; Y(t) contains multimodal observations such as passenger flow quantity and thermal distribution fused in the previous step; γ is the control hyperparameter (such as spatial position matching threshold, time tolerance, clustering radius, etc.); ∈(t) reflects the degree of difference between the simulation results and the actual venue passenger flow or crowd distribution;

[0159] If the obtained simulation error ∈(t) is continuously high, it indicates that the simulation model or parameter Θ crowd There is a mismatch with reality, and further research is needed in the digital twin space. If the simulation accuracy gradually improves, the simulation results can be used to adjust the venue layout, emergency drill plans, or optimize security and evacuation strategies, forming a linkage with the fourth step self-learning module.

[0160] Finally, the evaluation results of the crowd simulation accuracy and strategy effectiveness are generated, and key improvement suggestions or parameter corrections (such as the update of the adjacency matrix M, the update rule Θ) are provided. crowd The potential function improvements in the data are fed back to the digital twin and self-learning engine to further improve the system's control over real crowd dynamics.

[0161] When used, large-scale simulations conducted within the digital twin are no longer isolated, scripted exercises. Instead, they can be compared against actual sensor observations Y(t), continuously converging to a model that better reflects the real-world scenario. Dynamic tracking of simulation errors ∈(t) allows for precise assessment of the effectiveness of certain exhibition layouts or evacuation plans in real-world environments, enabling timely iterations to minimize safety risks or reduce congestion. Using the multimodal fusion output Y(t) to verify the accuracy of large-scale crowd simulations improves the depiction of realistic crowd distribution and stratified behavior.

[0162] Step 501 incorporates the Agent entity into the digital twin space to expand its refined depiction of crowd behavior; Step 502 combines the environmental situation U(t) and the network interaction between agents to complete the time-series crowd dynamics calculation; Step 503: Compare the simulation results with the real data Y(t), measure the simulation error ∈(t) and promptly feedback it to the digital twin and self-learning model to iteratively optimize the venue layout and emergency plan.

[0163] Through a comprehensive mechanism combining virtual-reality comparison, multi-agent collaboration, and complex network dynamics, the smart exhibition system can more proactively address large-scale crowd management issues, enabling more scientific and efficient resource allocation, from emergency evacuation to daily exhibition congestion control. By integrating and sharing data with the previous four steps, the crowd simulation system proposed in this step not only functions independently but also generates global benefits through overall collaboration.

[0164] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0165] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0166] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is only for some logical functions. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0167] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0168] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. The intelligent management system used in exhibition centers is characterized by: include, When massive amounts of heterogeneous sensor data are collected simultaneously, the multi-source perception network uses protocol conversion and time series alignment algorithms to perform format unification and noise filtering on the raw readings, transforming them into multi-source datasets. After the preprocessed multi-source dataset has completed time series correction, the multimodal fusion module performs cross-modal aggregation on the feature vectors of each channel through convolutional networks and dynamic weight scheduling to generate a multimodal fusion vector; After the multimodal fusion vector is sent to the edge node, the anomaly detection and edge adaptation engine uses anomaly scoring and perception topology reconstruction mechanisms to provide instant alerts and rapid device calibration for suspicious human traffic or extreme environmental data, generating adaptive perception configurations. After completing the preliminary calibration of the multi-source fusion data, the digital twin and self-learning module integrates the multimodal fusion vector, anomaly score, and adaptive perception configuration into the situation vector of the virtual venue through the situation mapping function and self-learning feedback function, and dynamically updates the sensor weights. Through multi-agent modeling and the introduction of network dynamics mechanisms in the update rules, the virtual crowd distribution is time-series calculated and the crowding effect is deduced. Simulation errors are introduced between the simulation results and the actual monitoring data to optimize the emergency evacuation plan and layout configuration.

2. The intelligent management system for a convention and exhibition center according to claim 1 is characterized by: The multi-source perception network configures a heterogeneous protocol conversion module on the gateway side, maps the communication protocols of various sensors to a unified interface protocol based on the sensor node identifier and network access identifier, maintains an access dynamic mapping table at the network aggregation end, and fixes the correspondence between the sensor node identifier and the perception channel.

3. The intelligent management system for a convention and exhibition center according to claim 2 is characterized by: The raw data from the perception channels are aligned to a unified time scale. The data alignment transformation operation is defined based on the time series fitting method. The time series misalignment is dynamically corrected during the alignment process. The aligned data are denoised by the weighted spectral reshaping function to obtain a multi-source data set.

4. The intelligent management system for a convention and exhibition center according to claim 3 is characterized by: A learnable feature mapping operator is constructed for each perception channel, mapping the multi-source dataset collected from the perception channel into a feature vector; The feature vectors of all channels are concatenated or adaptively fused to obtain a multi-channel fused feature vector.

5. The intelligent management system for a convention and exhibition center according to claim 4 is characterized in that: After introducing a learnable or adaptive weight set, the dynamic weight is calculated by the weight scheduling function to represent the weight of the perception channel: The weighted feature vector is weighted at the vector element level according to the obtained dynamic weight to obtain a multimodal fusion vector with dynamic collaborative characteristics. When the exhibition hall layout or the flow of people changes, the weight scheduling function updates the weight.

6. The intelligent management system for a convention and exhibition center according to claim 5 is characterized by: Use anomaly detection function to measure the degree of deviation of multimodal fusion vector in high-dimensional space and output anomaly score; If the obtained anomaly score exceeds the preset dynamic threshold function, an anomaly event is generated, triggering edge adaptive perception and rapid calibration, and the corresponding timestamp and anomaly level are written to the edge alarm queue. The anomaly event includes anomaly intensity, timestamp, and associated channel information.

7. The intelligent management system for a convention and exhibition center according to claim 6 is characterized in that: The decision function for edge adaptive perception is called to reallocate and activate some spare sensors, and quickly calibrate key sensors to generate the perception network reconstruction decision set and sensor incremental calibration parameters, and synchronize the updated calibration parameter matrix.

8. The intelligent management system for a convention and exhibition center according to claim 7 is characterized in that: Create virtual entities corresponding to key entities in the digital twin space, define twin nodes using sensor identification and network mapping information, and obtain the node set of the digital twin space; The situation mapping function is called to project the multimodal fusion vector, anomaly score and perception network reconstruction decision set into the digital twin to obtain the global situation vector; after the anomaly score is calculated, the anomaly score is projected into the risk intensity display of each node in the virtual venue through the situation mapping function.

9. The intelligent management system for a convention and exhibition center according to claim 8, characterized in that: Given the global situation vector and internal simulation parameters at the current moment, a predictive scenario simulation function is introduced to perform scenario calculations for future periods to obtain twin state predictions; The twin status prediction is presented in the digital twin visualization interface as a scene animation or trend chart, and outputs predictive indicators, including peak passenger flow periods, safety alarm probability, and environmental load.

10. The intelligent management system for a convention and exhibition center according to claim 9, characterized in that: The self-learning feedback function uses abnormal events, reconstructed decision sets, and situation vectors to iteratively update sensor weights, calibration parameters, and fusion algorithms, where: The self-learning feedback function automatically collects the differences between the actual scene and the twin simulation in each learning cycle, incorporates them into the learning loss function, drives the iterative correction of the sensor's calibration parameter matrix set and dynamic weights, and updates the status of the virtual entity node.

11. The intelligent management system for a convention and exhibition center according to claim 10, characterized in that: Add a multi-agent set to the digital twin space and construct the state vector of each agent. The state vector is used to describe its state parameters such as position, speed, target points of interest, emergency behavior tendencies, etc. in the digital twin environment, and define the interaction rules between people and scenes in the digital twin space.

12. The intelligent management system for a convention and exhibition center according to claim 11, characterized in that: For each agent, an update rule is defined in the digital twin space to simulate its behavioral evolution. Adjacency relationships are defined in the agent set, and a network dynamics mechanism is introduced into the update rule. The status of all agents is updated in parallel in the digital twin space to generate the overall population distribution at the next moment.

13. The intelligent management system for a convention and exhibition center according to claim 12, characterized in that: Construct an error measurement operator to compare the obtained virtual population distribution with the actual monitoring data. If the obtained simulation error is continuously higher than expected, adjust the parameters in the digital twin space. If the simulation accuracy gradually improves, the simulation results will be used to adjust the venue layout, optimize emergency drill plans, or optimize security and evacuation strategies, generate evaluation results on the accuracy of crowd simulation and the effectiveness of strategies, and feed back key improvement suggestions or parameter corrections to the digital twin and self-learning engine.

Images