Indoor intelligent exhibition appliance digital twin management system based on big data driving

Abstract

The invention discloses an indoor intelligent exhibition appliance digital twinborn management system based on big data driving, and the system comprises a behavior obtaining module which is used for deploying a multi-source heterogeneous sensing device to collect the moving track and interactive behavior original data of an exhibition viewer, and obtaining the behavior event log data of each exhibition appliance through data cleaning fusion and spatial correlation analysis; the behavior analysis module is used for converting the behavior event log data into analog load time sequence data by utilizing an exhibition appliance digital prototype based on the behavior event log data, training a physical information neural network according to the analog load time sequence data, and reasoning an exhibition appliance hidden physical state vector; and the preventive maintenance module is used for performing feature discretization and mutual information filtering on the hidden physical state vector of the exhibition appliance, constructing an optimized decision tree model fusing information gain and coefficient double splitting criteria to obtain a preventive maintenance work order, and realizing closed-loop digital twinborn management of the indoor exhibition appliance.

Description

Technical Field

[0001] This invention relates to the field of next-generation information technology, and in particular to a big data-driven digital twin management system for indoor intelligent exhibition equipment. Background Technology

[0002] Indoor exhibition equipment management technology has evolved from early manual inspections and single-point sensor monitoring to centralized monitoring based on the Internet of Things (IoT). More advanced methods have introduced the concept of digital twins, which directly collect status data by deploying various physical sensors (such as stress, displacement, and vibration sensors) on the exhibition equipment, build its virtual model, realize status visualization and alarms based on fixed thresholds, and provide some support for preventive maintenance.

[0003] Existing digital twin management systems for indoor intelligent exhibition equipment still have significant limitations. First, they rely heavily on the dense deployment of physical sensors on the exhibition equipment itself, resulting in high deployment costs, complex modifications, and potential impacts on the appearance and user experience. Second, their digital twins primarily monitor rather than predict, with passive and delayed state perception, failing to effectively detect indirect and measurable degradation processes such as structural fatigue accumulation and hidden wear of components. Finally, maintenance decisions are usually based on simple static thresholds, unable to generate dynamic and accurate maintenance strategies based on the actual load history and future usage plans of the exhibition equipment, making it difficult to achieve true intelligence and resource optimization. Therefore, this paper proposes a big data-driven digital twin management system for indoor intelligent exhibition equipment. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above objectives, the present invention proposes the following technical solution: A big data-driven digital twin management system for indoor intelligent exhibition equipment includes: Behavior acquisition module: Deploy multi-source heterogeneous sensing devices to collect raw data on visitors' movement trajectories and interactive behaviors, and obtain behavioral event log data for each exhibit through data cleaning, fusion, and spatial correlation analysis; Behavior Analysis Module: Based on behavior event log data, the digital prototype of the exhibit is used to convert it into simulated load time series data, and a physical information neural network is trained based on the simulated load time series data to infer the implicit physical state vector of the exhibit. Prevention and maintenance module: Performs feature discretization and mutual information filtering on the implicit physical state vector of the exhibit, and constructs a fusion information gain and... An optimized decision tree model based on the coefficient bi-split criterion is used to obtain preventive maintenance work orders; Digital Management Module: Based on preventive maintenance work orders, the digital twin is used to perform visualized early warning and work order dispatch, realizing closed-loop digital twin management of indoor exhibition equipment.

[0005] The process of obtaining behavioral event log data is as follows: Obtain raw data of the movement trajectory of exhibition visitors Raw data of interactive behavior Outliers in the movement trajectory data are removed by spatial neighborhood filtering rules, and the interaction behavior data is time-aligned. Multi-source data within the same time window are then fused to obtain a fused trajectory point set. ; Based on the fused trajectory point set Constructing exhibition space The partitioning model divides the exhibition space into sub-areas corresponding to each exhibit, and for each visitor's trajectory point... Calculate the spatial distance between the trajectory point and each exhibit to determine the exhibit to which the trajectory point belongs. Subregion: Match the trajectory data with the corresponding display equipment Related; At the same time, combined with interactive behavior Exhibition signage Generate behavioral event log data for each exhibit. .

[0006] The conversion process of simulated load time series data is as follows: Constructing the dynamic equations of the exhibition fixtures: ; in, The quality matrix of the exhibition equipment. Here is the damping matrix. Here is the stiffness matrix. For generalized coordinate vectors, The external excitation vector; The behavior event logs generated by the behavior acquisition module Input the dynamic equations of the display fixture, based on the behavior type Duration Determine the corresponding external incentives The dynamic equations of the display fixture were solved to obtain the simulated load time series data of the display fixture. .

[0007] The process of training a physical information neural network based on simulated load time-series data is as follows: The physical information neural network adopts a two-branch architecture with an encoding branch and a decoding branch. The encoding branch uses one-dimensional convolutional layers and gated recurrent units. In the stacked structure, the decoding branch maps the encoded features to the physical state space of the exhibit through a fully connected layer and embeds the exhibit dynamics equations as hard constraints; To simulate load time series data Physical condition monitoring data of corresponding exhibition equipment The physical state monitoring data of the exhibit serves as a training sample. This includes component wear data, fatigue accumulation data, and structural stress data; The load data is processed in the time domain, the physical condition monitoring data of the exhibition equipment is normalized, and the simulated load time series data is generated based on the time domain processed data. Compared with normalized physical condition monitoring data of exhibition equipment The loss function of the physical information neural network is constructed and validated to obtain the trained physical information neural network. .

[0008] The process of reasoning out the implicit physical state vector of the exhibit is as follows: Use the current simulated load time series data Real-time input of trained physical information neural network The temporal features of the current load are extracted through the encoding branch, and then mapped to the physical state space through the decoding branch to output the implicit physical state vector. .

[0009] The process of feature discretization and mutual information filtering of the implicit physical state vector of the exhibit is as follows: To display the implicit physical state vector To input the feature set, we integrate all historical maintenance data and extract maintenance type, maintenance timing, and maintenance effect as maintenance labels to form a labeled training dataset. ; The continuous implicit physical parameters are divided into 5 intervals by equal frequency discretization. The numerical features are transformed into categorical features. The mutual information value between each feature and the maintenance decision label is calculated, and redundant features with mutual information values ​​lower than 0.1 are filtered out.

[0010] The process of obtaining the optimized decision tree model is as follows: Based on the traditional decision tree, the information gain criterion is used to split the tree at the top three nodes of the traditional decision tree, and the splitting is performed at the bottom nodes from the fourth level down. The coefficient criterion is split; The implicit physical state vector of the exhibit is obtained through feature discretization and mutual information filtering. Selecting candidate splitting features for decision tree splitting Based on candidate splitting features The dynamic splitting criterion score is calculated to automatically select the optimal criterion, expressed as: ; in, Current candidate splitting features The training dataset at the current node The split score on the top, For information gain, for coefficient; Set a splitting stopping threshold, and select all candidate splitting features. Splitting stops when all splitting scores are less than the splitting stopping threshold, and the minimum description length is used. The pruning strategy calculates the combined value of the encoding length and data fitting error of the decision tree, pruning redundant branches. The optimized decision tree is then validated using time series cross-validation. After validation, the optimized decision tree model is obtained.

[0011] The process of visual early warning and work order dispatch based on a digital twin driven by preventive maintenance work orders is as follows: Adaptive preventive maintenance work orders After generation, the structured data in the work order is first displayed with identifiers. Maintenance type Maintenance priority Maintenance execution time The data is synchronized to the digital twin of the corresponding exhibit via a data interface. The digital twin is based on the exhibit's identifier. Complete the binding of physical exhibits with virtual twins to trigger visual alerts and work order dispatch processes.

[0012] The present invention has the following beneficial effects: 1. By employing mutual information feature filtering technology, core features strongly correlated with decision-making are extracted and maintained from the implicit physical state vector of the exhibit, effectively avoiding interference from redundant information. Furthermore, this is combined with information gain and... The optimized decision tree model with dynamic coefficient switching can quickly locate high-risk exhibit groups at the top level of the decision tree, and improve the accuracy of fault sample identification at the bottom level, so as to accurately quantify the hidden states such as internal stress accumulation and component wear and predict fault trends.

[0013] 2. By indirectly inverting the physical state of exhibits through visitor behavior data, the reliance on traditional physical sensors is reduced, achieving low-cost, non-invasive state perception. Combined with the constructed physical information neural network, data-driven and physical laws are deeply integrated, enabling high-precision inference of implicit physical states (such as structural fatigue accumulation) that are difficult to measure directly using traditional methods, thus achieving a leap from passive monitoring to active prediction.

[0014] 3. Based on the preventive maintenance work orders generated by the optimized decision tree model, the real-time status, historical data and external context are comprehensively considered, making the maintenance strategy more forward-looking and accurate, effectively optimizing the allocation of operation and maintenance resources, and finally realizing the closed-loop intelligent management of indoor smart exhibits from status perception and anomaly prediction to maintenance execution, which significantly improves operation and maintenance efficiency, exhibit reliability and user experience. Attached Figure Description

[0015] Figure 1 This is a system block diagram of the big data-driven digital twin management system for indoor intelligent exhibition equipment proposed in this invention. Detailed Implementation

[0016] 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.

[0017] Example 1: As Figure 1 As shown, the big data-driven digital twin management system for indoor intelligent exhibition equipment proposed in this invention includes: Behavior acquisition module: Deploy multi-source heterogeneous sensing devices to collect raw data on visitors' movement trajectories and interactive behaviors, and obtain behavioral event log data for each exhibit through data cleaning, fusion, and spatial correlation analysis; Deploy multi-source heterogeneous sensing devices to acquire raw data on the movement trajectories of visitors within the exhibition space. Raw data of interactive behavior ; Among them, the raw data of the movement trajectory , indicating the first The coordinates of the viewer in the two-dimensional exhibition space at each sampling time. With corresponding timestamp ; Raw data of interactive behavior , representing the The type of behavior corresponding to this interaction. Functional display equipment signage with the timestamp of occurrence , For exhibition purposes; Furthermore, the multi-source heterogeneous sensing devices include visual recognition, infrared sensing, and pressure sensing. Data acquisition is completed through the multi-source heterogeneous sensing devices, providing a basic data source for subsequent processing. First, data cleaning is performed to remove outliers caused by device vibration in the motion trajectory data, using spatial neighborhood filtering rules: ,in, for The average coordinates of the trajectory points within the neighborhood. The preset spatial distance threshold, The processed trajectory point data; Subsequently, the interaction behavior data was time-aligned, and multi-source data within the same time window were fused to obtain a fused set of trajectory points. This ensures data consistency over time, providing a unified data foundation for spatial correlation analysis. Then, spatial correlation analysis is performed based on the fused trajectory point set. Constructing exhibition space The partitioning model divides the exhibition space into sub-areas corresponding to each exhibit, and for each visitor's trajectory point... The exhibit to which the trajectory point belongs is determined by calculating its spatial distance from each exhibit. Subregion: The trajectory data is compared with the corresponding k-th display piece. Related, Indicates the k-th exhibit; At the same time, combined with interactive behavior Exhibition signage Generate behavioral event log data for each exhibit. ,in, The time when the event occurred. For behavior type, Duration of the behavior.

[0018] Behavior Analysis Module: Based on behavior event log data, the digital prototype of the exhibit is used to convert it into simulated load time series data, and a physical information neural network is trained based on the simulated load time series data to infer the implicit physical state vector of the exhibit. The process of converting digital prototypes of the exhibit into analog load timing data is as follows: For each exhibit, a digital prototype model is built. The model includes the physical structural parameters, kinematic characteristics, and mechanical response characteristics of the exhibit, and is represented by an exhibit dynamic equation: ; in, The quality matrix of the exhibition equipment. Here is the damping matrix. Here is the stiffness matrix. For generalized coordinate vectors, As the external excitation vector, this model establishes a mapping relationship between the physical state of the fixture and the external excitation, providing a physical basis for the generation of subsequent simulated loads; The behavior event logs generated by the behavior acquisition module Input the dynamic equations of the display fixture, based on the behavior type Duration Determine the corresponding external incentives ,For example: When the behavior type is touch interaction, the touch pressure and duration are converted into force excitation in the model; when the behavior type is observation, the human body radiant heat and observation time are converted into thermal excitation in the model. Then, by solving the dynamic equations of the display fixture, the simulated load time series data of the fixture are obtained. ,in For the first The equivalent load value of the exhibit at each time point is used to transform abstract behavioral events into load data that can be used for physical state analysis, thus realizing the transformation from behavioral information to physical information. Construct a physical information neural network trained with simulated load time-series data. Its network structure adopts an encoder-decoder dual-branch architecture: Encoding branch: Employs one-dimensional convolutional layers and gated recurrent units. Stacked structure for extracting simulated load timing data The temporal characteristics and nonlinear relationships in the data; Decoding branch: The encoded features are mapped to the physical state space of the exhibit through a fully connected layer, and the dynamic equations of the exhibit are embedded as hard constraints to achieve a deep integration of data-driven and physical priors; Specifically, the core innovation of physical information neural networks lies in the dynamic embedding mechanism of physical constraints, which differs from traditional methods. A fixed-weighted approach to physical losses; To simulate load time series data Physical condition monitoring data of corresponding exhibition equipment For training samples; Among them, the physical condition monitoring data of the exhibition equipment include: Structural stress data: By attaching fiber optic grating sensors to key load-bearing structures of the exhibit, such as load-bearing brackets and connecting hinges, the internal stress values ​​of the exhibit under different loads are collected, with the unit being megapascals. This data directly reflects the stress fatigue degree of the exhibit structure. Component wear data: The wear amount and vibration frequency offset of the components are collected by displacement sensors and vibration sensors installed on the moving parts of the exhibition equipment, such as rotating shafts and sliding guide rails. The wear amount is in millimeters and the vibration frequency is in Hertz. This data reflects the aging and wear status of the moving parts. Fatigue cumulative data: Based on stress monitoring data, the cumulative fatigue damage value of the structure is calculated using the rainflow counting method, with a range of values ​​including [values ​​to be filled in]. The closer the value is to 1, the closer the structure of the exhibit is to the fatigue failure threshold. First, the load data is processed in the time domain: ; in, This is the simulated load timing data after time-domain processing; Simultaneously, data from the physical condition monitoring of the exhibition equipment... Feature mapping is performed to unify heterogeneous parameters such as internal stress and component wear into a unified mapping. Intervals are used to eliminate the interference of dimensional differences on model training. ; Based on the simulated load time series data after time-domain processing Compared with normalized physical condition monitoring data of exhibition equipment The loss function of the physical information neural network is constructed, which includes data loss and physical loss, wherein: Data loss The mean square error between the predicted state and the actual monitored state is defined as: ; in, For the sample size, These are the normalized physical state monitoring values. This is the simulated load time series data after time-domain processing. For a physical information neural network, θ is the set of learnable parameters of the model; Physical loss : To output the constraint error that satisfies the dynamic equations of the exhibit, substitute the physical state vector predicted by the model into the dynamic equations of the exhibit. And calculate the residuals of the equation. Norm, to ensure that the model output conforms to physical laws; The training process uses an adaptive learning rate optimizer, with the initial learning rate set to... When the validation set loss is 10 consecutive When there is no decrease, the learning rate automatically decays to 1 / 5 of its original value. Training is iterated until the loss function converges, resulting in a trained physical information neural network. ; Use the current simulated load time series data Real-time input of trained physical information neural network The temporal features of the current load are extracted through the encoding branch, then mapped to the physical state space through the decoding branch, and finally output as a latent physical state vector. k is the index. This represents the implicit physical state vector of the kth exhibit. This vector contains implicit parameters such as internal stress, component wear degree, and structural fatigue accumulation value, which directly serve as the decision basis for the preventive maintenance module, realizing deep reasoning from load data to physical state.

[0019] Prevention and maintenance module: Performs feature discretization and mutual information filtering on the implicit physical state vector of the exhibit, and constructs a fusion information gain and... An optimized decision tree model based on the coefficient bi-split criterion is used to obtain preventive maintenance work orders; The optimized decision tree model improves upon the traditional decision tree's single information gain splitting method by fusing information gain with... The coefficient double splitting criterion calculates the dynamic splitting criterion score to automatically select the optimal criterion. The implicit physical state vector of the display fixture output by the behavior analysis module Using the input feature set and integrating the entire historical maintenance data, maintenance type, maintenance timing, and maintenance effect are extracted as maintenance label samples to form a labeled training dataset. ; Among them, the maintenance types include component replacement, accuracy calibration, maintenance timing includes within 72 hours and within 168 hours, and maintenance effects include fault elimination and condition improvement. Then, feature discretization and mutual information filtering are performed. First, the continuous latent physical parameters are divided into 5 intervals through equal frequency discretization, and the numerical features are transformed into class features to reduce the computational complexity when splitting the decision tree. Then, the mutual information value between each feature and the maintenance label is calculated. Redundant features with mutual information values ​​below 0.1 are filtered out to ensure that the input feature set retains only parameters strongly correlated with maintenance decisions. The mutual information calculation formula is as follows: ,in, As characteristic variables, To maintain the decision label variables, For joint probability distribution, , It represents a marginal probability distribution; Furthermore, only mutual information In this context, only features with high discriminative power for maintaining decision-making are allowed to enter the candidate pool. This avoids redundant features interfering with the splitting accuracy of the decision tree, and the feature variables... Corresponding display fixture implicit physical state vector A single implicit physical parameter in the training dataset Maintenance decision tags in the middle; Based on the traditional decision tree as the basic model architecture, this paper integrates information gain and... The optimal criterion is selected using the coefficient double splitting criterion. The information gain criterion is used to split the first three layers (top node) of the traditional decision tree, prioritizing the feature with the highest discriminative power for maintaining the decision. Switch to the decision tree at level 4 and below (bottom nodes) The coefficient criterion is split to improve the accuracy of fault sample identification and avoid the bias of information gain in the case of skewed sample distribution. By using feature discretization and mutual information filtering, the implicit physical state vector of the exhibit is obtained. The core features that are strongly correlated with the decision labels are selected, and these selected features form the candidate splitting features for subsequent decision tree splitting. ; Then based on candidate splitting features During the splitting process, each layer of nodes automatically selects the optimal criterion by calculating the dynamic splitting criterion score. The core logic formula is expressed as follows: ; in, Current candidate splitting features The training dataset at the current node The split score on the top, Information gain is used for the first three layers of splitting in a decision tree. , for Coefficients are used for splits at level 4 and below in the decision tree. The training dataset for the current node. As candidate splitting features, This represents a specific value of the candidate splitting feature. To obtain from the training dataset All candidate splitting features selected from the data have a value of [value]. A subset consisting of samples Indicating candidate splitting features The set of all possible values; Set a splitting stopping threshold, and select all candidate splitting features. Splitting stops when all splitting scores are less than the splitting stopping threshold, and the minimum description length is used. The pruning strategy automatically prunes redundant branches by calculating the combined value of the decision tree's encoding length and the data fitting error. Specifically, minimum description length The pruning strategy calculates the description length of the current decision tree, which consists of two parts: the encoding length and the fitting error. Encoding length: The number of bits in the encoding of a decision tree structure, reflecting the model complexity; Fitting error: The deviation between the leaf node samples and the true labels, reflecting the model fitting accuracy; Then, backtracking from the leaf nodes, we evaluate the change in description length after pruning one by one. If the description length decreases after pruning a branch, the pruning operation is retained. Repeat the backtracking evaluation until the description length cannot be further reduced through pruning, thus achieving the minimum description length. Pruning strategies; Finally, time series cross-validation was used to validate the optimized decision tree, and the optimized decision tree model was obtained after the validation was completed. The implicit physical state vector of the exhibit output by the behavior analysis module Input the optimized decision tree model, traverse the decision tree node path, locate the corresponding leaf node, and directly output the maintenance type of the current display fixture. Maintenance priority (High / Medium / Low) and Maintenance Execution Time ,For example: When the implicit physical state vector shows that the structural fatigue attenuation rate of a certain exhibit reaches 0.8, the optimization decision tree model will output the structural reinforcement, maintenance type, high maintenance priority, and maintenance execution time within 72 hours. Based on the above decision results, the system automatically generates preventive maintenance work orders. ,in, For exhibition equipment signage, For maintenance types, such as component replacement, precision calibration, etc. To maintain priority, it is divided into three levels: high, medium, and low. To maintain the execution time, e.g., within 72 hours or within 168 hours; Specifically, when maintenance work orders for multiple exhibits conflict with resources, the optimization decision tree model will dynamically schedule resources based on maintenance priority and the rate of deterioration of implicit physical state, prioritizing the allocation of resources to exhibits with faster state deterioration. The generated work orders will directly serve as input to the digital management module, driving subsequent visual early warning and work order dispatch, achieving seamless connection from physical state perception to maintenance execution.

[0020] Digital Management Module: Based on preventive maintenance work orders, the digital twin is used for visualized early warning and work order dispatch, realizing closed-loop digital twin management of indoor exhibition equipment; Adaptive preventive maintenance work orders After generation, the system first processes the structured data (exhibition equipment identifier) ​​in the work order. Maintenance type Maintenance priority Maintenance execution time The data is synchronized to the digital twin of the corresponding exhibit via a data interface. The digital twin is based on the exhibit's identifier. Complete the binding of physical exhibits with virtual twins to trigger visual alerts and work order dispatch processes; Digital twin After receiving the work order data, it is processed according to the maintenance priority. Generate tiered visual alerts, with preset high thresholds. and a low threshold : If the maintenance priority exceeds the preset high threshold, a red dynamic highlight mark will be superimposed on the 3D model of the corresponding exhibit in the twin, and the work order will be displayed at the top of the maintenance dashboard in the twin interface. If the maintenance priority is lower than the preset high threshold or higher than the preset low threshold, a yellow highlight will be added and displayed in the center of the maintenance dashboard. If the maintenance priority is lower than the preset low threshold, a blue highlight will be added and it will be displayed normally in the operation and maintenance dashboard; Meanwhile, maintenance types are embedded in the early warning information. Maintenance execution time ,For example: Exhibition equipment The system allows maintenance personnel to directly access complete maintenance decision information through a twin interface, specifying maintenance type, structural reinforcement, execution time, and timeframe within 72 hours. Based on the early warning information from the digital twin, the system automatically triggers the work order dispatch process, according to the maintenance type. Matching with operations and maintenance personnel possessing the corresponding skills, for example: "Structural reinforcement" work orders are assigned to personnel with structural repair qualifications; Combined with maintenance execution time Based on the current task load of the operations and maintenance personnel, generate the optimal dispatch plan; Finally, the work order details, including the location of the exhibit, maintenance content, priority and time requirements, are pushed to the mobile terminals of the maintenance personnel. The dispatch order is dynamically adjusted to ensure that the maintenance tasks of high-risk exhibits are executed first.

[0021] In the application, several formulas are calculated by removing dimensions and taking their numerical values. The formulas are established by collecting a large amount of data and simulating the most recent real situation. Some coefficients or weights in the formulas are set by those skilled in the art according to the actual situation, so they will not be elaborated here.

[0022] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples 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 implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0023] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A big data-driven digital twin management system for indoor intelligent exhibition equipment, characterized in that: include: Behavior acquisition module: Deploy multi-source heterogeneous sensing devices to collect raw data on visitors' movement trajectories and interactive behaviors, and obtain behavioral event log data for each exhibit through data cleaning, fusion, and spatial correlation analysis; Behavior Analysis Module: Based on behavior event log data, the digital prototype of the exhibit is used to convert it into simulated load time series data, and a physical information neural network is trained based on the simulated load time series data to infer the implicit physical state vector of the exhibit. Prevention and maintenance module: Performs feature discretization and mutual information filtering on the implicit physical state vector of the exhibit, and constructs a fusion information gain and... An optimized decision tree model based on the coefficient bi-split criterion is used to obtain preventive maintenance work orders; Digital Management Module: Based on preventive maintenance work orders, the digital twin is used to perform visualized early warning and work order dispatch, realizing closed-loop digital twin management of indoor exhibition equipment.

2. The big data-driven digital twin management system for indoor intelligent exhibition equipment according to claim 1, characterized in that, The process of obtaining behavioral event log data is as follows: Obtain raw data of the movement trajectory of exhibition visitors Raw data of interactive behavior Outliers in the movement trajectory data are removed by spatial neighborhood filtering rules, and the interaction behavior data is time-aligned. Multi-source data within the same time window are then fused to obtain a fused trajectory point set. ; Based on the fused trajectory point set Constructing exhibition space The partitioning model divides the exhibition space into sub-areas corresponding to each exhibit, and for each visitor's trajectory point... Calculate the spatial distance between the trajectory point and each exhibit to determine the exhibit to which the trajectory point belongs. Subregion: The trajectory data is compared with the corresponding k-th display piece. Related; At the same time, combined with interactive behavior Exhibition signage Generate behavioral event log data for each exhibit. .

3. The big data-driven digital twin management system for indoor intelligent exhibition equipment according to claim 2, characterized in that, The conversion process of simulated load time series data is as follows: Constructing the dynamic equations of the exhibition fixtures: ; in, The quality matrix of the exhibition equipment. Here is the damping matrix. Here is the stiffness matrix. For generalized coordinate vectors, The external excitation vector; The behavior event logs generated by the behavior acquisition module Input the dynamic equations of the display fixture, based on the behavior type Duration Determine the corresponding external incentives The dynamic equations of the display fixture were solved to obtain the simulated load time series data of the display fixture. .

4. The big data-driven digital twin management system for indoor intelligent exhibition equipment according to claim 3, characterized in that, The process of training a physical information neural network based on simulated load time-series data is as follows: The physical information neural network adopts a two-branch architecture with an encoding branch and a decoding branch. The encoding branch uses one-dimensional convolutional layers and gated recurrent units. In the stacked structure, the decoding branch maps the encoded features to the physical state space of the exhibit through a fully connected layer and embeds the exhibit dynamics equations as hard constraints; To simulate load time series data Physical condition monitoring data of corresponding exhibition equipment The physical state monitoring data of the exhibit serves as a training sample. This includes component wear data, fatigue accumulation data, and structural stress data; The load data is processed in the time domain, the physical condition monitoring data of the exhibition equipment is normalized, and the simulated load time series data is generated based on the time domain processed data. Compared with normalized physical condition monitoring data of exhibition equipment The loss function of the physical information neural network is constructed and validated to obtain the trained physical information neural network. .

5. The big data-driven digital twin management system for indoor intelligent exhibition equipment according to claim 4, characterized in that, The process of reasoning out the implicit physical state vector of the exhibit is as follows: Use the current simulated load time series data Real-time input of trained physical information neural network The temporal features of the current load are extracted through the encoding branch, and then mapped to the physical state space through the decoding branch to output the implicit physical state vector. .

6. The big data-driven digital twin management system for indoor intelligent exhibition equipment according to claim 5, characterized in that, The process of feature discretization and mutual information filtering of the implicit physical state vector of the exhibit is as follows: To display the implicit physical state vector To input the feature set, we integrate all historical maintenance data and extract maintenance type, maintenance timing, and maintenance effect as maintenance labels to form a labeled training dataset. ; The continuous implicit physical parameters are divided into 5 intervals by equal frequency discretization. The numerical features are transformed into categorical features. The mutual information value between each feature and the maintenance decision label is calculated, and redundant features with mutual information values ​​lower than 0.1 are filtered out.

7. The big data-driven digital twin management system for indoor intelligent exhibition equipment according to claim 6, characterized in that, The process of obtaining the optimized decision tree model is as follows: Based on the traditional decision tree, the information gain criterion is used to split the tree at the top three nodes of the traditional decision tree, and the splitting is performed at the bottom nodes from the fourth level down. The coefficient criterion is split; The implicit physical state vector of the exhibit is obtained through feature discretization and mutual information filtering. Selecting candidate splitting features for decision tree splitting Based on candidate splitting features The dynamic splitting criterion score is calculated to automatically select the optimal criterion, expressed as: ； in, Current candidate splitting features The training dataset at the current node The split score on the top, For information gain, for coefficient; Set a splitting stopping threshold, and select all candidate splitting features. Splitting stops when all splitting scores are less than the splitting stopping threshold, and the minimum description length is used. The pruning strategy calculates the combined value of the encoding length and data fitting error of the decision tree, pruning redundant branches. The optimized decision tree is then validated using time series cross-validation. After validation, the optimized decision tree model is obtained.

8. The big data-driven digital twin management system for indoor intelligent exhibition equipment according to claim 7, characterized in that, The process of visual early warning and work order dispatch based on a digital twin driven by preventive maintenance work orders is as follows: Adaptive preventive maintenance work orders After generation, the structured data in the work order is first displayed with identifiers. Maintenance type Maintenance priority Maintenance execution time The data is synchronized to the digital twin of the corresponding exhibit via a data interface. The digital twin is based on the exhibit's identifier. Complete the binding of physical exhibits with virtual twins to trigger visual alerts and work order dispatch processes.

Images