A park data fusion and intelligent decision method based on an intelligent operation center

Abstract

The present application relates to the technical field of big data processing, and more particularly to a park data fusion and intelligent decision-making method based on an intelligent operation center, comprising the following steps: S1, multi-modal sensing data acquisition: real-time acquisition of multi-modal sensing data through an Internet of Things sensing layer deployed in each area of the park, the multi-modal sensing data including video image data, environmental sensing data, equipment operation data and personnel and vehicle passing data; S2, multi-source heterogeneous data fusion based on semantic alignment: a multi-source heterogeneous data fusion engine is constructed. The present application maps heterogeneous data such as video images, environmental sensing and equipment operation to the same semantic space by constructing a unified semantic space model, realizes feature correlation and semantic alignment across data sources, and the closed-loop mechanism enables the park intelligent operation center to evolve from a traditional visual display tool to an intelligent decision-making core engine driving business optimization.

Description

Technical Field

[0001] This invention relates to the field of big data processing technology, and in particular to a method for park data fusion and intelligent decision-making based on an intelligent operation center. Background Technology

[0002] With the rapid development of the Internet of Things, big data, artificial intelligence, and digital twin technologies, the construction of smart parks has become an important part of urban digital transformation. As the nerve center of the park, the intelligent operation center integrates various system interfaces to achieve cross-departmental and cross-system data aggregation and coordinated control. However, current intelligent operation centers in parks still have the following shortcomings in data fusion and intelligent decision-making:

[0003] Existing park management systems typically integrate various business subsystems through interfaces. However, significant differences exist in the data formats, semantic expressions, and spatiotemporal benchmarks among these systems, making it difficult to achieve deep semantic alignment and correlation analysis after data aggregation. Most solutions only focus on data presentation, failing to build a unified data fusion model and effectively uncover implicit relationships between multi-source data, thus highlighting the persistent problem of data silos. While existing technologies include methods for training neural networks using labeled regional datasets to identify anomalous data, their data fusion depth remains limited to simple feature-level concatenation, lacking semantic space-based alignment and mapping mechanisms, making it difficult to support deep correlation reasoning across data domains.

[0004] Most existing decision support systems in industrial parks rely on preset rules and thresholds to trigger alarms, representing a passive, reactive decision-making model. By the time the system detects anomalies, the events have often already occurred, lacking the ability to proactively predict and intervene based on real-time sensing data. Furthermore, existing decision-making methods often depend on a single data source or transient observation data, ignoring the dynamic evolution and spatial correlation characteristics of multi-source data over time, leading to overly aggressive or inaccurate decision commands. In addition, decision results are typically presented in static reports or large-screen displays, failing to deeply integrate with digital twin technology and lacking a visualized, interactive, and predictable closed-loop decision verification mechanism. Summary of the Invention

[0005] The purpose of this invention is to solve the problem of the difficulty in intelligent management of parks in the prior art, and to propose a park data fusion and intelligent decision-making method based on an intelligent operation center.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for park data fusion and intelligent decision-making based on an intelligent operation center includes the following steps:

[0008] Step S1, Multimodal sensing data acquisition: Multimodal sensing data is collected in real time through the IoT sensing layer deployed in various areas of the park. The multimodal sensing data includes video image data, environmental sensing data, equipment operation data, and personnel and vehicle passage data.

[0009] Step S2, Multi-source heterogeneous data fusion based on semantic alignment: Construct a multi-source heterogeneous data fusion engine to perform data preprocessing, feature extraction and semantic alignment on the multimodal perception data, and generate a decision fusion dataset in a unified semantic space;

[0010] Step S3, Intelligent Decision Analysis Based on Fusion Data: Based on the decision fusion dataset, a multi-dimensional intelligent decision analysis model is constructed. The model includes a situational awareness model, an anomaly early warning model, and a resource optimization model to comprehensively evaluate and intelligently analyze the park's operational status.

[0011] Step S4, Decision Instruction Generation and Closed-Loop Verification: Generate a sequence of decision instructions based on the results of intelligent decision analysis, simulate and verify the decision effect through a digital twin engine, and send the verified decision instructions to the corresponding execution terminal.

[0012] Preferably, the data preprocessing in step S2 includes:

[0013] Denoising, enhancement, and format standardization are performed on video image data;

[0014] Environmental sensor data is processed by outlier removal, missing value imputation, and dimensional normalization.

[0015] The equipment operation data is processed by protocol parsing, timing alignment, and data cleaning.

[0016] Preferably, the feature extraction in step S2 includes:

[0017] A convolutional neural network is used to perform target detection and feature extraction on video image data to obtain a real-time feature image group, which includes human target features, vehicle target features and equipment target features.

[0018] An abnormal environment identification model is used to process environmental sensing data. The abnormal environment identification model is trained based on a historical labeled environmental dataset and outputs an environmental feature vector, which includes temperature features, humidity features and particulate matter concentration features.

[0019] Preferably, the semantic alignment in step S2 includes:

[0020] Construct a unified semantic space model for the park and define standardized semantic representations of park entities, including spatial entities, equipment entities, personnel entities, and event entities;

[0021] Based on the unified semantic space model, feature vectors extracted from different data sources are mapped to the same semantic space, realizing feature association and semantic alignment across data sources.

[0022] Preferably, the decision fusion dataset in step S2 is generated by: combining the semantically aligned target device data group, real-time feature images, and environmental feature data in a multi-dimensional manner to obtain decision fusion data, and summarizing the decision fusion data of all park areas to obtain the decision fusion dataset.

[0023] Preferably, the situational awareness model in step S3 is constructed based on a spatiotemporal correlation analysis algorithm, specifically including:

[0024] Acquire spatial location and timestamp information from the decision fusion dataset to construct a spatiotemporal data cube of the park's operational status;

[0025] Based on the aforementioned spatiotemporal data cube, the spatiotemporal autocorrelation function is used to analyze the spatial distribution characteristics and temporal evolution patterns of the operational status of each area in the park.

[0026] The overall operational status index of the park is calculated using the following formula:

[0027] I 态势 =α⋅I 安全 +β⋅I 效率 +γ⋅I 能耗 +δ⋅I 舒适

[0028] Among them, I 安全 For the security situation sub-index, I 效率 I is a sub-index of operational efficiency. 能耗 I is a sub-index of energy consumption status. 舒适 α is a sub-index of environmental comfort, where α, β, γ, and δ are the corresponding weighting coefficients, and α+β+γ+δ=1.

[0029] Preferably, the anomaly warning model in step S3 is constructed based on a time-series prediction algorithm, specifically including:

[0030] Historical time-series data from the decision fusion dataset are acquired, and a time-series prediction model is constructed using a long short-term memory network.

[0031] Based on the time-series prediction model, the future trends of various monitoring indicators in the park are predicted, and an anomaly warning is triggered when the predicted value exceeds a preset threshold.

[0032] The abnormal warnings include safety abnormality warnings, equipment failure warnings, environmental exceedance warnings, and crowd density warnings.

[0033] Preferably, the resource optimization model in step S3 is constructed based on a multi-objective optimization algorithm, specifically including:

[0034] Construct a multi-objective optimization function for park resource scheduling, wherein the optimization objectives include minimizing operational energy consumption, maximizing service efficiency, and minimizing response latency;

[0035] The multi-objective optimization function is solved using a particle swarm optimization algorithm or a genetic algorithm to obtain the optimal resource scheduling scheme.

[0036] Preferably, the simulation and deduction of the digital twin engine in step S4 includes:

[0037] A digital twin model of the park is constructed based on data from the park's 3D geographic information system and equipment mechanism model data.

[0038] The decision command sequence is input into the digital twin model to simulate the changes in the park's operational status after the decision commands are executed;

[0039] If the deviation exceeds a preset threshold, the decision instructions are iteratively optimized by comparing the simulation results with the expected goals until the simulation results meet the expected goals.

[0040] Preferably, the method further includes step S5: constructing a visual decision-making platform to display the decision fusion dataset and intelligent decision analysis results in a visual form. The visual decision-making platform includes a situation overview cockpit, an anomaly warning dashboard, a resource scheduling panel, and a decision simulation sandbox.

[0041] Compared with the prior art, the present invention has the following advantages:

[0042] 1. This invention constructs a unified semantic space model, mapping heterogeneous data such as video images, environmental sensing data, and device operation data to the same semantic space. This achieves feature association and semantic alignment across data sources, overcoming the shortcomings of existing technologies where data fusion remains at the format conversion level. Through a neural network-based feature extraction and semantic mapping mechanism, it solves the semantic gap problem of multi-source data, enabling the fused data to support deep-level associative reasoning and intelligent decision-making.

[0043] 2. This invention achieves proactive prediction and early warning of park operation status by constructing an LSTM-based time-series prediction model and a multi-objective optimization model. Unlike existing passive alarm mechanisms based on fixed thresholds, this invention enables proactive identification and pre-emptive action before anomalies occur. Simultaneously, the simulation and verification mechanism using a digital twin engine ensures the scientific validity and effectiveness of decision-making instructions, avoiding overly aggressive instructions based solely on a single transient observation.

[0044] 3. This invention constructs a complete closed-loop chain from data perception and fusion analysis to decision generation, simulation verification, and instruction execution. Through digital twin technology, it enables pre-event prediction and post-event traceability of decision-making effects, significantly improving the intelligence level and operational efficiency of park management. This closed-loop mechanism transforms the park's intelligent operation center from a traditional visualization tool into a core intelligent decision-making engine driving business optimization. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the structure of a park data fusion and intelligent decision-making method based on an intelligent operation center proposed in this invention. Detailed Implementation

[0046] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0047] Reference Figure 1 A method for park data fusion and intelligent decision-making based on intelligent operation center includes the following steps: Step S1, multimodal perception data collection; Step S2, multi-source heterogeneous data fusion based on semantic alignment; Step S3, intelligent decision analysis based on fused data; Step S4, decision instruction generation and closed-loop verification; Step S5, construction of visual decision platform.

[0048] Step S1, Multimodal sensing data acquisition: Multimodal sensing data is collected in real time through the IoT sensing layer deployed in various areas of the park. The multimodal sensing data includes video image data, environmental sensing data, equipment operation data, and personnel and vehicle passage data.

[0049] Specifically, multimodal sensing data is collected in real time through an IoT sensing layer deployed in various areas of the park (including high-definition cameras, environmental sensors, equipment status monitoring terminals, IoT gateways, etc.). Among these:

[0050] The high-definition cameras use the H.265 video encoding standard and collect video image data of key areas in the park at a sampling frequency of 25 frames per second. The resolution is configured as 1080p pixels, covering the park entrances and exits, main roads, building perimeters and key equipment areas.

[0051] The environmental sensor network deploys temperature and humidity sensors, particulate matter concentration sensors (PM2.5 / PM10), light sensors, and noise sensors to continuously collect environmental parameters at a sampling frequency of 1Hz.

[0052] The equipment status monitoring terminal accesses the operating parameters of various electromechanical equipment in the park (including air conditioning units, elevator systems, water supply and drainage pump sets, power distribution equipment, etc.) through multiple protocol interfaces such as Modbus TCP, OPC UA and BACnet.

[0053] Personnel and vehicle traffic data are collected through RFID readers, facial recognition terminals, and license plate recognition cameras deployed at park entrances and building access control systems.

[0054] The collected data undergoes preliminary preprocessing at edge computing nodes, which are deployed in data aggregation boxes across various areas of the park. These edge computing nodes are equipped with ARM-based embedded processors and GPU acceleration units, enabling real-time data cleaning and lightweight computing capabilities.

[0055] The video image data undergoes denoising, enhancement, and format standardization. Denoising employs a bilateral filtering algorithm and median filtering to suppress salt-and-pepper noise, preserving edge information while suppressing Gaussian noise. Image enhancement uses Adaptive Histogram Equalization (AHE) to improve image contrast under low-light conditions. Format standardization converts the proprietary video stream formats of cameras from different manufacturers into standardized H.264 encoded RTSP streams.

[0056] Environmental sensor data underwent outlier removal, missing value imputation, and dimensional normalization. Outlier removal employed the Grubbs test based on the 3σ criterion, marking sensor data as outliers and removing them when their deviation from the mean exceeded three times the standard deviation. Missing value imputation used Linear Interpolation, combining valid data from adjacent time points for filling in missing values. Dimensional normalization employed the Min-Max standardization method, mapping each environmental parameter to the [0,1] interval to eliminate the scale influence of different physical dimensions on subsequent fusion calculations.

[0057] The equipment operation data undergoes protocol parsing, timing alignment, and data cleaning. The protocol parsing module has built-in parsers for various industrial protocols such as Modbus, BACnet, and OPC UA, which parse the raw data frames reported by each device into standardized key-value pairs. Timing alignment uses the unified timestamp synchronized by the network time protocol as a benchmark to interpolate and align data from different devices and different sampling frequencies to the same time grid. Data cleaning uses a sliding window filtering method to remove jump data caused by communication interruption or equipment malfunction.

[0058] The preprocessed data is transmitted to the data fusion layer of the intelligent operation center via a high-speed network. The transmission network adopts a dual-redundancy architecture that combines industrial Ethernet and 5G private network. Data transmission uses the MQTT over TLS encryption protocol to ensure the integrity and confidentiality of the data during transmission.

[0059] Step S2, Multi-source heterogeneous data fusion based on semantic alignment: Construct a multi-source heterogeneous data fusion engine to perform data preprocessing, feature extraction and semantic alignment on multimodal perception data, and generate a decision fusion dataset in a unified semantic space.

[0060] The data preprocessing in step S2 includes:

[0061] Denoising, enhancement, and format standardization are performed on video image data;

[0062] Environmental sensor data is processed by outlier removal, missing value imputation, and dimensional normalization.

[0063] The equipment operation data is processed by protocol parsing, timing alignment, and data cleaning.

[0064] The data fusion layer receives the preprocessed multimodal data and performs three-level fusion processing:

[0065] Feature extraction: Convolutional neural networks are used to perform target detection and feature extraction on video image data, identify targets such as people, vehicles, and equipment, as well as their spatial location and behavioral state, and generate real-time feature image sets.

[0066] Specifically, the CNN backbone network adopts a modified YOLOv8 architecture, which includes three modules: the backbone network, the neck network, and the head network.

[0067] The backbone network adopts the CSPDarknet structure, which reduces the amount of computation while maintaining high detection accuracy through cross-stage local connections;

[0068] The neck network adopts a path aggregation network structure, which integrates shallow high-resolution feature maps and deep high-semantic feature maps to improve the detection capability of small-scale targets;

[0069] The head network outputs the coordinates (x, y, w, h) of the target detection bounding box, the target category confidence score, and the behavior status label.

[0070] In some implementations, the CNN model is pre-trained on a campus scene dataset containing over 50,000 labeled images, covering various lighting and weather conditions such as daytime, nighttime, rainy days, and foggy days. The output of feature extraction is a real-time feature image set, which includes features of personnel targets (position coordinates, direction of movement, velocity vector), vehicle targets (license plate number, vehicle type classification, driving trajectory), and equipment targets (equipment number, on / off status of indicator lights, etc.).

[0071] An anomaly identification model is employed to process environmental sensor data. This model is trained on a historically labeled environmental dataset and outputs an environmental feature vector containing dimensions such as temperature, humidity, and particulate matter concentration. The anomaly identification model uses a gradient boosting decision tree architecture, employing time-series data of temperature, humidity, and particulate matter concentration collected over the past 30 days as training samples and manually labeled "normal / abnormal" as supervision signals. The model iteratively adds weak learners (decision trees), with each new decision tree fitting the residual predictions from the previous round, ultimately integrating to output the environmental feature vector. The environmental feature vector includes temperature, humidity, and particulate matter concentration features, with each feature dimension represented by standardized numerical values.

[0072] Further explanation: The abnormal device identification model adopts the isolation forest algorithm. This algorithm constructs multiple isolation trees by randomly selecting features and randomly selecting segmentation values, and distinguishes normal data points from abnormal data points in terms of path length in the tree. Abnormal data points usually have shorter path lengths, so they are easier to isolate.

[0073] The model is trained on a dataset containing more than 100,000 historical records of equipment operation, covering multi-dimensional operating parameters such as equipment current, voltage, power factor, vibration frequency, and temperature. It outputs equipment feature vectors, which include equipment identifiers, operating status codes, and anomaly confidence scores.

[0074] First, semantic alignment: Construct a unified semantic space model for the park, defining standardized semantic representations for park entities (spatial entities, equipment entities, personnel entities, and event entities). The unified semantic space model is built on ontology and uses the WebOntology Language to describe various park entities and their attribute relationships. Spatial entities use the latitude-longitude coordinates of the park's 3D geographic information system as the spatial reference, defining standardized URI identifiers and polygonal descriptions of spatial units such as buildings, roads, areas, and floors; equipment entities use equipment codes from the equipment asset management system as unique identifiers, defining attributes such as equipment type, installation location, system affiliation, and rated parameters; personnel entities use personnel IDs as unique identifiers, defining attributes such as personnel roles, permission levels, and departments; event entities use event type codes and timestamps as composite identifiers, defining attributes such as event nature, urgency, and scope of impact.

[0075] Based on a unified semantic space model, feature vectors extracted from different data sources are projected onto the same semantic space through a semantic mapping function, achieving cross-data source feature association and semantic alignment. The semantic mapping function is defined as a nonlinear transformation f:X i →S, where X iLet S be the feature space of the i-th data source, and S be the target semantic space. This mapping is implemented using a multilayer perceptron, which includes an input layer, two hidden layers (each with 128 neurons and ReLU activation function), and an output layer (64-dimensional semantic embedding vector).

[0076] The MLP model employs supervised contrastive learning: semantically similar cross-modal data pairs are designated as positive sample pairs (e.g., people and access control card swipe records in the same area at the same time), while semantically dissimilar cross-modal data pairs are designated as negative sample pairs. The InfoNCE loss function is used to shorten the distance between positive sample pairs in the semantic space and widen the distance between negative sample pairs. After training, feature vectors extracted from any data source can be mapped to a unified 64-dimensional semantic embedding space through this MLP, achieving semantic-level alignment of cross-modal features.

[0077] Second, correlation fusion: The semantically aligned target device data group, real-time feature images, and environmental feature data are multi-dimensionally correlated and combined to construct decision fusion data. The decision fusion dataset is then compiled from the decision fusion data of all areas within the park. Correlation fusion employs a cross-modal fusion strategy based on an attention mechanism: the three types of feature vectors after semantic alignment are concatenated into a feature matrix.

[0078] F = [Fvision, Fenv, Fdevice] ∈ R 3×d

[0079] Where d=64 is the semantic embedding dimension.

[0080] Calculate the mutual attention weights among the three types of features using multi-head self-attention:

[0081] Attention(Q,K,V)=softmax(QK T / √d k V

[0082] Where Q, K, and V are all obtained from FF through linear transformation, d k The scaling factor is used. Multi-head attention computes eight attention heads in parallel, and the outputs of each head are concatenated and passed through a feedforward network to obtain the fused feature vector F. fusion ∈R d . Fuse feature vector F fusion That is, a single decision fusion data point is aggregated to form a decision fusion dataset by aggregating decision fusion data from all areas of the park.

[0083] D fusion ={F (1) fusion ,F (2) fusion ,…,F (N) fusion}

[0084] Where N represents the total number of monitoring areas within the park.

[0085] Comparative Experiment 1: Accuracy Comparison of Multi-Source Heterogeneous Data Fusion

[0086] To verify the superiority of the semantic alignment-based multi-source heterogeneous data fusion method of this invention, a comparative experiment was conducted with three traditional data fusion methods in the same campus scenario:

[0087] Table 1. Differences between different fusion methods

[0088] Evaluation indicators Traditional weighted average fusion Traditional features splicing and fusion Traditional post-integration (voting method) Semantic alignment + attention fusion Fusion accuracy (%) 72.3 78.6 74.1 91.7 Semantic alignment accuracy (F1 score) not applicable 0.61 0.58 0.89 Cross-modal association recall rate (%) 65.8 71.2 68.5 88.3 Data fusion completeness (%) 69.4 76.8 72.3 93.6

[0089] It can be known that:

[0090] Traditional weighted average fusion methods simply perform linear weighted combinations of features from different data sources, completely ignoring the semantic relationships between modalities, resulting in a fusion accuracy of only 72.3%. While traditional feature concatenation fusion methods retain all feature dimensions of each modality, the lack of semantic alignment leads to a severe "semantic gap" in the concatenated high-dimensional space, resulting in a cross-modal association recall rate of only 71.2%. Traditional post-fusion methods process each modality independently before voting, resulting in insufficient intermodal information interaction and the lowest fusion completeness (72.3%).

[0091] Semantic alignment + attention fusion achieves a fusion accuracy of 91.7% through cross-modal interaction of unified semantic space mapping and multi-head attention mechanism, which is 13.1 percentage points higher than the best traditional method (feature splicing fusion). The semantic alignment accuracy F1 score reaches 0.89, which fully demonstrates its significant advantage in solving the semantic gap problem of multi-source heterogeneous data.

[0092] Feature extraction in step S2 includes:

[0093] Convolutional neural networks are used to perform target detection and feature extraction on video image data to obtain real-time feature image groups, which include human target features, vehicle target features and equipment target features.

[0094] An abnormal environment identification model is used to process environmental sensing data. The abnormal environment identification model is trained based on historical labeled environmental datasets and outputs environmental feature vectors, which include temperature features, humidity features and particulate matter concentration features.

[0095] An abnormal device identification model is used to process the device operation data. The abnormal device identification model is trained based on a historical labeled device dataset and outputs device feature vectors.

[0096] Semantic alignment in step S2 includes:

[0097] Construct a unified semantic space model for the park and define standardized semantic representations of park entities, including spatial entities, equipment entities, personnel entities, and event entities;

[0098] Based on a unified semantic space model, feature vectors extracted from different data sources are mapped to the same semantic space, enabling feature association and semantic alignment across data sources.

[0099] The decision fusion dataset in step S2 is generated by multi-dimensionally associating and combining the semantically aligned target device data group, real-time feature images, and environmental feature data to obtain decision fusion data, and summarizing the decision fusion data of all park areas to obtain the decision fusion dataset.

[0100] Step S3, Intelligent Decision Analysis Based on Fusion Data: Based on the decision fusion dataset, a multi-dimensional intelligent decision analysis model is constructed, including a situational awareness model, an anomaly early warning model, and a resource optimization model, to comprehensively evaluate and intelligently analyze the park's operational status.

[0101] The intelligent decision-making layer runs three core analysis models in parallel based on the decision fusion dataset:

[0102] Situational Awareness Model: Based on spatiotemporal correlation analysis algorithms, a spatiotemporal data cube representing the park's operational status is constructed. The spatial distribution characteristics and temporal evolution patterns of the operational status in each area are analyzed using spatiotemporal autocorrelation functions. The overall operational status index I of the park is then comprehensively calculated. 态势 This enables real-time assessment and visualization of the overall operational status of the park.

[0103] Specifically, spatial location information (latitude and longitude coordinates and altitude) and timestamp information are obtained from the decision fusion dataset. A three-dimensional spatiotemporal data cube C∈RX×Y×T is constructed using a spatial grid (grid size of 10m×10m) and a time window (sliding window length of 5 minutes) as dimensions, where X and Y are the horizontal and vertical dimensions of the spatial grid, and T is the number of time windows. Based on the spatiotemporal data cube, the spatial autocorrelation of the operational status of each area of ​​the park is analyzed using the global Moran index.

[0104] I=N / ∑ N i=1 ∑ N j=1 w ij ⋅∑ N i=1 ∑ N j=1 w ij (x i -x ˉ (x) j -x ˉ ) / ∑N i=1 (x i -x ˉ ) 2

[0105] Where N is the total number of spatial grid cells, x i and x j Let x be the running state values ​​of the i-th and j-th grid cells, respectively. ˉ w is the mean. ij The spatial weight matrix is ​​determined using the inverse distance weighting method. Simultaneously, the local Getis-Ord G* statistic is used to identify hot and cold regions in the operational status.

[0106] Anomaly Early Warning Model: A time-series prediction model is constructed based on a Long Short-Term Memory (LSTM) network and trained on historical time-series data in the decision fusion dataset. The model predicts the future trends of various monitoring indicators in the park. When the predicted value exceeds a preset dynamic threshold, a tiered anomaly early warning is triggered. Warning types include safety anomalies, equipment failures, environmental exceedances, and abnormal crowd density.

[0107] Resource optimization model: A resource scheduling optimization engine is built based on a multi-objective optimization algorithm. The optimization objectives are to minimize operational energy consumption, maximize service efficiency, and minimize response latency. The multi-objective optimization function is solved using particle swarm optimization algorithm or genetic algorithm, and the optimal resource scheduling scheme is output.

[0108] In step S3, the situational awareness model is constructed based on a spatiotemporal correlation analysis algorithm, specifically including:

[0109] Acquire spatial location and timestamp information from the decision fusion dataset to construct a spatiotemporal data cube of the park's operational status;

[0110] Based on the spatiotemporal data cube, the spatiotemporal autocorrelation function is used to analyze the spatial distribution characteristics and temporal evolution of the operational status of each area in the park;

[0111] The overall operational status index of the park is calculated using the following formula:

[0112] I 态势 =α⋅I 安全 +β⋅I 效率 +γ⋅I 能耗 +δ⋅I 舒适

[0113] Among them, I 安全 For the security situation sub-index, I 效率 I is a sub-index of operational efficiency. 能耗 I is a sub-index of energy consumption status. 舒适α is a sub-index of environmental comfort, where α, β, γ, and δ are the corresponding weighting coefficients, and α+β+γ+δ=1.

[0114] The anomaly warning model in step S3 is constructed based on a time-series prediction algorithm, specifically including:

[0115] Historical time-series data from the decision fusion dataset are acquired, and a time-series prediction model is constructed using a long short-term memory network.

[0116] Based on the time series prediction model, the future trend of each monitoring indicator in the park is predicted, and an anomaly warning is triggered when the predicted value exceeds the preset threshold.

[0117] Anomaly warnings include safety anomaly warnings, equipment failure warnings, environmental exceedance warnings, and crowd density warnings.

[0118] Anomaly warning model: A time-series prediction model is constructed based on a Long Short-Term Memory (LSTM) network and trained on historical time-series data in the decision fusion dataset. The state update equation of the LSTM unit at time t is:

[0119] f t =σ(W f ⋅[h t−1 ,x t ]+b f )

[0120] i t =σ(W i ⋅[h t−1 ,x t ]+b i )

[0121] C ~ t =tanh(WC⋅[h t−1 ,x t ]+b C )

[0122] C t =f t ⊙C t−1 +i t ⊙C - t

[0123] o t =σ(W o ⋅[h t−1 ,x t ]+b o )

[0124] h t =o t ⊙tanh(C t )

[0125] Where f t i t o t C represents the activation vectors for the forget gate, input gate, and output gate, respectively. t In cellular state, h t In the hidden state, x t Let be the input feature vector at time t, W and b be the weight matrix and bias term, respectively, σ be the Sigmoid activation function, and ⊙ denote element-wise multiplication.

[0126] The resource optimization model in step S3 is constructed based on a multi-objective optimization algorithm, specifically including:

[0127] Construct a multi-objective optimization function for park resource scheduling, with optimization objectives including minimizing operational energy consumption, maximizing service efficiency, and minimizing response latency;

[0128] A resource scheduling optimization engine is built based on a multi-objective optimization algorithm. The optimization objectives are to minimize operational energy consumption, maximize service efficiency, and minimize response latency. The multi-objective optimization function is solved using particle swarm optimization algorithm or genetic algorithm to obtain the optimal resource scheduling scheme.

[0129] The model predicts the future trends of various monitoring indicators in the park, with prediction time windows set at three scales: 15 minutes, 30 minutes, and 60 minutes. A tiered anomaly warning is triggered when the predicted value exceeds a preset dynamic threshold. The dynamic threshold is adaptively adjusted using an exponentially weighted moving average method, rather than a fixed threshold. Anomaly warnings include security anomaly warnings (intrusion, crowd gathering, restricted area entry, etc.), equipment failure warnings (equipment parameters exceeding limits, performance degradation trends, etc.), environmental exceedance warnings (temperature and humidity exceeding limits, particulate matter concentration exceeding standards, etc.), and population density warnings (area population density exceeding safe capacity, etc.). Warning levels are divided into three levels: blue warning (attention level), yellow warning (alert level), and red warning (emergency level), with different handling procedures and response time limits corresponding to different levels.

[0130] Specifically, a multi-objective optimization function for park resource scheduling is constructed:

[0131] minF(x)=[f1(x),f2(x),f3(x)]T

[0132] Where x is a vector of decision variables (including the start / stop status of each device, operating power setpoint, scheduling priority, etc.), f1(x) is the operational energy consumption objective function (measured in kW·h for total energy consumption in the park), f2(x) is the service efficiency objective function (measured in the number of service requests per unit time), and f3(x) is the response latency objective function (measured in the average time from request issuance to service completion). The three objective functions are mutually constrained (e.g., reducing energy consumption may sacrifice service efficiency, and improving service efficiency may increase energy consumption), constituting a typical multi-objective optimization problem.

[0133] The PSO algorithm initializes M particles (M=100), each representing a candidate resource scheduling scheme, which moves in a D-dimensional search space at a velocity v. i Flight, position x i This represents a set of values ​​for the decision variable. The formulas for updating the velocity and position of particle i in the t-th iteration are:

[0134] v i (t+1)=ωv i (t)+c1r1[p best ,i−x i (t)]+c2r2[g best -x i (t)]x i (t+1)=x i (t)+v i (t+1)

[0135] Where w is the inertia weight (using a linear decreasing strategy, decreasing from 0.9 to 0.4), c1 and c2 are acceleration constants (both taken as 2.0), r1 and r2 are random numbers in the interval [0,1], pbest,i is the historical best position of particle i, and g best The optimal position is determined globally. In multi-objective optimization, the Pareto dominance relation is used to judge the quality of solutions: solution x1 dominates solution x2 if and only if x1 is not inferior to x2 on all objectives and is superior to x2 on at least one objective. After the algorithm converges iteratively, the Pareto optimal solution set is output, and the park manager selects the final scheduling scheme from the Pareto front according to the actual operation preference.

[0136] Comparative Experiment 2: Comparison of Timeliness and Accuracy of Anomaly Warning

[0137] To verify the superiority of the LSTM-based time-series prediction anomaly early warning model of this invention, a comparative experiment was conducted with three traditional anomaly early warning methods in the same park setting:

[0138] Table 2. Differences in accuracy among different early warning methods

[0139] Evaluation indicators Traditional fixed threshold alarm Traditional moving average forecast Traditional ARIMA forecast LSTM+ Dynamic Threshold Early warning accuracy rate (%) 68.2 73.5 76.8 94.6 Warning lead time (minutes) 0 (Post-event alert) 3.2 5.7 12.4 False alarm rate (%) 31.8 24.7 21.3 5.2 Missed report rate (%) 28.5 19.6 16.4 3.8

[0140] It can be known that:

[0141] Traditional fixed-threshold alarm methods only trigger alarms when monitored values ​​exceed preset thresholds, representing a typical reactive response mode with a zero-minute lead time. Furthermore, due to the fixed thresholds, they cannot adapt to dynamic changes in park operations, resulting in a false alarm rate as high as 31.8%. Traditional moving average forecasting methods extrapolate future trends based on the moving average of historical data, but they can only capture linear trends and have poor adaptability to nonlinear changes, with a lead time of only 3.2 minutes. While the traditional ARIMA (Autoregressive Integral Moving Average) model can handle some non-stationary time series, its linear assumptions limit its ability to model the nonlinear and multi-period coupling characteristics commonly found in park operation data, resulting in a warning accuracy rate of only 76.8%.

[0142] This invention effectively captures temporal dependencies and nonlinear evolution patterns over long periods of time through the gated memory units of LSTM networks. Combined with EWMA dynamic threshold adaptive adjustment, the early warning accuracy reaches 94.6%, which is 17.8 percentage points higher than the best traditional method (ARIMA). The early warning lead time reaches 12.4 minutes, which is more than twice that of the traditional method. The false alarm rate and false negative rate are reduced to 5.2% and 3.8%, respectively.

[0143] LSTM networks address the vanishing and exploding gradient problems in traditional recurrent neural networks when processing long sequences by introducing three gating units: an input gate, a forget gate, and an output gate.

[0144] Comparative Experiment 3: Comparison of Resource Scheduling Optimization Effects

[0145] To verify the superiority of the resource scheduling model based on multi-objective particle swarm optimization in this invention, a comparative experiment was conducted with three traditional resource scheduling methods in the same campus scenario:

[0146] Table 3. Differences in resource scheduling optimization effects among different tables

[0147] Evaluation indicators Traditional fixed time slot scheduling Traditional first-come, first-served scheduling Traditional single-objective optimization scheduling Multi-objective PSO Daily operating energy consumption (kW·h) 28470 27650 25820 23150 Average service response time (seconds) 18.6 22.3 14.2 9.8 Resource utilization rate (%) 61.2 58.7 73.4 87.6 Overall satisfaction with the scheduling plan 0.52 0.48 0.71 0.93

[0148] It can be known that:

[0149] Traditional fixed-slot scheduling methods uniformly allocate resources according to a preset schedule (such as air conditioning systems starting and stopping at fixed times), which cannot dynamically adjust according to real-time load, resulting in the highest energy consumption (28470 kW·h / day) and a resource utilization rate of only 61.2%. Traditional first-come, first-served (FFS) scheduling methods allocate resources according to the order in which requests arrive. While simple to implement, this completely ignores the urgency of tasks and optimal resource matching, resulting in the longest service response time (22.3 seconds) and the lowest resource utilization rate (58.7%). Traditional single-objective optimization scheduling methods focus solely on minimizing energy consumption, which improves energy efficiency (25820 kW·h / day) but sacrifices service efficiency (response time 14.2 seconds).

[0150] This invention constructs a multi-objective optimization function that includes three mutually constraining objectives: energy consumption, service efficiency, and response latency. It then uses the particle swarm optimization algorithm to search for the optimal compromise solution on the Pareto front, achieving synergistic optimization of the three objectives: daily average operating energy consumption is reduced by 10.3% compared to the optimal traditional method (single-objective optimization), average service response time is shortened by 31.0%, resource utilization is increased to 87.6%, and the overall satisfaction of the scheduling scheme (calculated based on the weighted TOPSIS method) reaches 0.93.

[0151] Step S4, Decision Instruction Generation and Closed-Loop Verification: Generate a sequence of decision instructions based on the results of intelligent decision analysis, simulate and verify the decision effect through a digital twin engine, and send the verified decision instructions to the corresponding execution terminal.

[0152] The simulation and deduction of the digital twin engine in step S4 includes:

[0153] A digital twin model of the park is constructed based on data from the park's 3D geographic information system and equipment mechanism models. The 3D GIS data includes building structure models (LOD3 level of detail), road network models, underground pipeline models, and topographic models, stored in CityGML format. Equipment mechanism models are constructed based on the physical working principles of each piece of equipment (e.g., thermodynamic models of air conditioning systems, dynamic models of elevator systems, optical models of lighting systems, etc.), and the input-output characteristics and state transition equations of the equipment are described using Modelica. The digital twin model is integrated with the decision fusion dataset D through a real-time data-driven interface. fusion The connection enables the virtual model to be updated synchronously with the physical campus.

[0154] The decision command sequence is input into the digital twin model to simulate the changes in the park's operational status after the execution of the decision commands. The simulation engine adopts a discrete event simulation mechanism with a simulation step size of 1 second, and extrapolates the trajectory of the changes in the status of each system in the park within the next 30 minutes after the issuance of the decision commands in the virtual environment. The simulation process comprehensively considers equipment response delay (usually 100ms-5s), system inertia (such as the lag effect of temperature adjustment in the air conditioning system), and the coupling effects between multiple systems (such as the impact of dimming the lighting system on the quality of video surveillance images).

[0155] By comparing the simulation results with the expected goals, if the deviation exceeds a preset threshold, the decision instructions are iteratively optimized until the simulation results meet the expected goals. Deviation evaluation employs a multi-indicator comprehensive assessment: Let the simulation result be Y. sim The expected target is Y target The overall deviation rate ϵ is defined as:

[0156] ϵ=1 / K∑ K k=1 | (Y sim,k -Y target,k / Y target,k |×100%

[0157] Where K is the total number of evaluation indicators. If ϵ>5%, an iterative optimization loop is triggered: the parameter configuration of the decision instruction is adjusted using the Bayesian optimization method, and the simulation is repeated until ϵ≤5%. The upper limit of iteration is set to 10 times. If the upper limit is reached and the condition is still not met, the process is transferred to manual review.

[0158] The decision-making level summarizes the analysis results from the three analytical models and generates a sequence of decision-making instructions. These instructions are first input into the digital twin engine for simulation and verification.

[0159] A digital twin model, constructed based on data from the park's 3D geographic information system and equipment mechanism model data, simulates the changes in the park's operational status after the execution of decision-making instructions.

[0160] If the deviation exceeds a preset threshold when comparing the simulation results with the expected target, the decision instructions are adjusted through an iterative optimization algorithm until the simulation results meet the expected target.

[0161] Once the decision instructions are verified, they are sent to the corresponding execution terminals (such as equipment control systems, security systems, energy management systems, etc.) through the IoT platform. At the same time, the decision fusion data and analysis results are pushed to the visual decision platform for visualization, thus completing the complete closed-loop management process of perception → fusion → decision → execution.

[0162] Comparative Experiment 4: Comparison of Decision Effectiveness

[0163] To verify the superiority of the decision verification mechanism based on digital twin simulation in this invention, a comparative experiment was conducted with three traditional decision-making methods in the same park setting:

[0164] Table 4 Differences between different decision-making methods

[0165] Evaluation indicators Traditional experience-based decision-making Traditional data dashboard decision making Traditional rule engine decision Digital twin simulation + closed-loop verification Decision accuracy rate (%) 52.6 67.3 74.8 93.2 Decision response time (seconds) 180 (Manual assessment) 95 (Manual assessment) 8.5 6.2 First-time approval rate of decision-making instructions (%) not applicable not applicable 61.4 89.7 Decision execution deviation rate (%) 34.8 26.1 18.5 4.3

[0166] It can be known that:

[0167] Traditional experience-based decision-making relies entirely on the manager's personal experience and subjective judgment, lacking data support. Its accuracy rate is only 52.6%, and the response time is as long as 180 seconds. Traditional data dashboards, while providing visual data displays, still require manual analysis and judgment, with a response time of 95 seconds, an accuracy rate of 67.3%, and decision quality significantly affected by the manager's professional level. Traditional rule-based decision-making, based on "if-then" rules for automated decision-making, has a relatively fast response time (8.5 seconds). However, because the rule base cannot cover all complex scenarios, the accuracy rate is only 74.8%, and lacking a pre-verification mechanism, the first-time pass rate for decision instructions is only 61.4%, with an execution deviation rate as high as 18.5%.

[0168] This invention uses a digital twin engine to perform pre-simulation and multi-indicator comprehensive verification of decision instructions in a virtual environment, achieving a decision accuracy rate of 93.2%, which is 18.4 percentage points higher than the best traditional method (rule engine); the decision response time is shortened to 6.2 seconds; and through an iterative optimization mechanism, the first-pass rate of decision instructions reaches 89.7%, and the execution deviation rate is reduced to 4.3%.

[0169] Step S5: Construct a visual decision-making platform: Display the decision fusion dataset and intelligent decision analysis results in a visual form. The visual decision-making platform includes a situation overview cockpit, an anomaly warning dashboard, a resource scheduling panel, and a decision simulation sandbox.

[0170] The overall situation overview cockpit uses a 3D real-view map of the park as its base map, overlaying it with the park's overall operational situation index I. 态势 The system displays real-time numerical values ​​and trend curves, and renders the status distribution of each region in the form of heat maps across four dimensions: safety, efficiency, energy consumption, and comfort, allowing managers to view the overall operational status of the entire region on a single screen.

[0171] The anomaly warning dashboard displays various warning events in both timeline and spatial distribution dimensions. Each warning record includes the event type, location, warning level, predicted trend, and suggested handling measures. It supports filtering by warning level and drilling down to view details by region.

[0172] The resource scheduling panel displays the real-time occupancy status, utilization statistics, and scheduling plan execution progress of various resources in the park. Managers can manually adjust scheduling parameters on the panel, and the system automatically reruns the multi-objective optimization algorithm to generate an updated Pareto optimal solution set for selection.

[0173] The decision simulation sandbox provides an interactive simulation environment for digital twin models. Managers can customize different decision-making schemes (such as adjusting the air conditioning set temperature, changing the lighting strategy, optimizing the frequency of shuttle bus departures, etc.). The system simulates and displays the evolution of the park's future operating status under the decision-making scheme in real time, supports parallel comparison of multiple schemes, and provides an intuitive "what-if" analysis tool for scientific decision-making.

[0174] It should be noted that the specific model and specifications of the visual decision-making platform need to be selected and determined based on the actual specifications of the device. The specific selection calculation method adopts the existing technology in this field, so it will not be elaborated here.

[0175] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for park data fusion and intelligent decision-making based on an intelligent operation center, characterized in that, Includes the following steps: Step S1, Multimodal sensing data acquisition: Multimodal sensing data is collected in real time through the IoT sensing layer deployed in various areas of the park. The multimodal sensing data includes video image data, environmental sensing data, equipment operation data, and personnel and vehicle passage data. Step S2, Multi-source heterogeneous data fusion based on semantic alignment: Construct a multi-source heterogeneous data fusion engine to perform data preprocessing, feature extraction and semantic alignment on the multimodal perception data, and generate a decision fusion dataset in a unified semantic space; Step S3, Intelligent Decision Analysis Based on Fusion Data: Based on the decision fusion dataset, a multi-dimensional intelligent decision analysis model is constructed. The model includes a situational awareness model, an anomaly early warning model, and a resource optimization model to comprehensively evaluate and intelligently analyze the park's operational status. Step S4, Decision Instruction Generation and Closed-Loop Verification: Generate a sequence of decision instructions based on the results of intelligent decision analysis, simulate and verify the decision effect through a digital twin engine, and send the verified decision instructions to the corresponding execution terminal.

2. The method for park data fusion and intelligent decision-making based on an intelligent operation center according to claim 1, characterized in that, The data preprocessing in step S2 includes: Denoising, enhancement, and format standardization are performed on video image data; Environmental sensor data is processed by outlier removal, missing value imputation, and dimensional normalization. The equipment operation data is processed by protocol parsing, timing alignment, and data cleaning.

3. The method for park data fusion and intelligent decision-making based on an intelligent operation center according to claim 2, characterized in that, The feature extraction in step S2 includes: A convolutional neural network is used to perform target detection and feature extraction on video image data to obtain a real-time feature image group, which includes human target features, vehicle target features and equipment target features. An abnormal environment identification model is used to process environmental sensing data. The abnormal environment identification model is trained based on a historical labeled environmental dataset and outputs an environmental feature vector, which includes temperature features, humidity features and particulate matter concentration features.

4. The method for park data fusion and intelligent decision-making based on an intelligent operation center according to claim 3, characterized in that, The semantic alignment in step S2 includes: Construct a unified semantic space model for the park and define standardized semantic representations of park entities, including spatial entities, equipment entities, personnel entities, and event entities; Based on the unified semantic space model, feature vectors extracted from different data sources are mapped to the same semantic space, realizing feature association and semantic alignment across data sources.

5. The method for park data fusion and intelligent decision-making based on an intelligent operation center according to claim 4, characterized in that, The decision fusion dataset in step S2 is generated by: combining the semantically aligned target device data group, real-time feature images, and environmental feature data in a multi-dimensional manner to obtain decision fusion data, and summarizing the decision fusion data of all park areas to obtain the decision fusion dataset.

6. The method for park data fusion and intelligent decision-making based on an intelligent operation center according to claim 5, characterized in that, The situational awareness model in step S3 is constructed based on a spatiotemporal correlation analysis algorithm, specifically including: Acquire spatial location and timestamp information from the decision fusion dataset to construct a spatiotemporal data cube of the park's operational status; Based on the aforementioned spatiotemporal data cube, the spatiotemporal autocorrelation function is used to analyze the spatial distribution characteristics and temporal evolution patterns of the operational status of each area in the park. The overall operational status index of the park is calculated using the following formula: I 态势 =α⋅I 安全 +β⋅I 效率 +γ⋅I 能耗 +δ⋅I 舒适 Among them, I 安全 For the security situation sub-index, I 效率 I is a sub-index of operational efficiency. 能耗 I is a sub-index of energy consumption status. 舒适 α is a sub-index of environmental comfort, where α, β, γ, and δ are the corresponding weighting coefficients, and α+β+γ+δ=1.

7. The method for park data fusion and intelligent decision-making based on an intelligent operation center according to claim 6, characterized in that, The anomaly warning model in step S3 is constructed based on a time-series prediction algorithm, specifically including: Historical time-series data from the decision fusion dataset are acquired, and a time-series prediction model is constructed using a long short-term memory network. Based on the time-series prediction model, the future trends of various monitoring indicators in the park are predicted, and an anomaly warning is triggered when the predicted value exceeds a preset threshold. The abnormal warnings include safety abnormality warnings, equipment failure warnings, environmental exceedance warnings, and crowd density warnings.

8. The method for park data fusion and intelligent decision-making based on an intelligent operation center according to claim 7, characterized in that, The resource optimization model in step S3 is constructed based on a multi-objective optimization algorithm, specifically including: Construct a multi-objective optimization function for park resource scheduling, wherein the optimization objectives include minimizing operational energy consumption, maximizing service efficiency, and minimizing response latency; The multi-objective optimization function is solved using a particle swarm optimization algorithm or a genetic algorithm to obtain the optimal resource scheduling scheme.

9. A method for park data fusion and intelligent decision-making based on an intelligent operation center according to claim 8, characterized in that, The simulation and deduction of the digital twin engine in step S4 includes: A digital twin model of the park is constructed based on data from the park's 3D geographic information system and equipment mechanism model data. The decision command sequence is input into the digital twin model to simulate the changes in the park's operational status after the decision commands are executed; If the deviation exceeds a preset threshold, the decision instructions are iteratively optimized by comparing the simulation results with the expected goals until the simulation results meet the expected goals.

10. A method for park data fusion and intelligent decision-making based on an intelligent operation center according to claim 9, characterized in that, It also includes step S5, constructing a visual decision-making platform, which displays the decision fusion dataset and intelligent decision analysis results in a visual form. The visual decision-making platform includes a situation overview cockpit, an anomaly warning dashboard, a resource scheduling panel, and a decision simulation sandbox.

Images