Monitoring system and method for smart park

Abstract

The invention relates to the technical field of park monitoring, in particular to a monitoring system and method for a smart park, and the method comprises the steps: collecting a building information model, geographic information system data and point cloud data of the park, and constructing a digital twinborn model through a deep learning driven multi-modal fusion method; the method comprises the following steps: acquiring the position and operation data of equipment through a digital twinborn model, processing the operation data of the equipment by utilizing a Transform-XL neural network, identifying a potential fault mode and carrying out early warning in advance, and displaying the fault position and the influence range through a visual interface of the digital twinborn model; a termite nest self-organizing mechanism is adopted for dynamic task allocation and resource scheduling, the termite nest self-organizing mechanism in the nature is simulated, cooperation between devices is optimized, and task allocation and resource scheduling are achieved. According to the invention, comprehensive, efficient and intelligent management of the smart park is realized.

Description

Technical Field

[0001] The present invention relates to the field of park monitoring technology, and in particular to a monitoring system and method for a smart park. Background Art

[0002] With the acceleration of urbanization and the rapid development of information technology, smart parks, as an important part of new urban development, are receiving increasing attention for their construction and management. Smart parks integrate advanced technologies such as the Internet of Things, big data, and cloud computing to achieve intelligent management and operation of various resources within the park. However, existing technologies still have many shortcomings in monitoring smart parks: Limitations of data integration and processing: There are various devices in smart parks, and data comes from a wide range of sources. The existing park monitoring technology mainly uses IoT sensor data, video surveillance data, etc. to form the basis of the digital twin model. Existing technologies often find it difficult to effectively integrate these multi-source heterogeneous data, resulting in serious data silos and the inability to achieve comprehensive data sharing and efficient use. At the same time, the data processing and analysis capabilities are also limited, making it difficult to meet the needs of real-time monitoring and intelligent decision-making.

[0003] Lag in monitoring and early warning: Traditional manual monitoring methods are not only inefficient, but also difficult to detect and handle potential faults in a timely manner. Even if some automated monitoring methods are adopted, due to the lag in data processing and analysis, effective early warning and intervention are often not possible before a fault occurs, thus affecting the normal operation and safety of the park.

[0004] Insufficient equipment collaborative control: There are many devices in the smart park, and how to achieve efficient collaborative control between devices is an important issue. Existing technologies often lack effective intelligent decision-making algorithms and collaborative control mechanisms, making it difficult to achieve effective information sharing and collaborative work between devices, thus affecting the overall operational efficiency and management level of the park. Summary of the invention

[0005] The present invention discloses a monitoring system and method for a smart park, aiming to achieve comprehensive, efficient and intelligent management of the smart park.

[0006] In order to achieve the purpose of the present invention, the technical solution adopted is: a monitoring method for a smart park, comprising: S1. Collect the building information model, geographic information system data and point cloud data of the park, and use the multimodal fusion method driven by deep learning to build a digital twin model; S2. Obtain the location and operation data of the equipment through the digital twin model, use the Transformer-XL neural network to process the equipment operation data, identify potential failure modes and provide early warnings, and display the failure location and impact range through the visual interface of the digital twin model; S3. Adopt the termite nest self-organizing mechanism to optimize the collaboration between devices and realize dynamic task allocation and resource scheduling.

[0007] As an optimization solution of the present invention, in S1, the point cloud data is the captured and recorded information on the structure of the campus buildings, the status of the equipment and the environmental characteristics.

[0008] As an optimization solution of the present invention, S1 includes: S1-1. After collecting the building information model, geographic information system data and point cloud data of the park, convert them to the same coordinate system WGS84 through Helmert transformation; convert BIM files into OBJ or PLY format, convert GIS data into raster images or vector graphics, and convert point cloud data into voxel grids or depth maps; S1-2. For BIM data, CNNs are used to extract building structural features; for GIS data, geographic information features are extracted through spatial coordinate transformation and feature encoding technology; for point cloud data, PointNet is used to extract 3D shape features; S1-3, multimodal feature fusion, combines the features extracted from the three data sources. Each branch is responsible for processing one type of input data, and finally the outputs of each branch are merged through the fully connected layer.

[0009] As an optimization solution of the present invention, S1 also includes: S1-4. Model training and optimization, including: 41) Use spatial consistency loss function to ensure that the results generated by the digital twin model are consistent in space;

[0010] Where: L spatial is the spatial consistency loss function, Indicates traversing and summing all point pairs (i, j); for the i-th data point x in the BIM data BIM , i ,f(x BIM,i ) represents the representation of the data point in the feature space; for the jth data point x in the BIM data BIM , j ,f(x BIM,j ) represents the representation of the data point in the feature space; ||•||2 represents the Euclidean distance, g(x BIM , i) represents the spatial coordinates of the i-th data point in the BIM data; g(x BIM , j ) represents the spatial coordinates of the jth data point in the BIM data; 42) Select the Adam optimization algorithm to minimize the spatial consistency loss function; 43) Train and evaluate the model: Initialize the model parameters during training, then perform forward propagation, calculate loss, back propagation, and update parameters for each batch of data; repeat forward propagation, calculate loss, back propagation, and update parameters until the predetermined number of epochs is reached, use K-fold cross-validation to evaluate model performance, use grid search to find the best hyperparameter combination, evaluate model performance on an independent test set, and build a digital twin model.

[0011] As an optimization solution of the present invention, in S2, it includes: S2-1. According to the operation characteristics of the equipment and the fault occurrence cycle, the time series data is divided into windows of fixed length, and the data of each window is preprocessed after windowing; S2-2. Use the preprocessed data to train the Transformer-XL model to learn the normal state and various failure modes of the equipment.

[0012] As an optimization solution of the present invention, in S2-2, it includes: S2-2-1. Dataset division: divide the preprocessed data into training set, validation set and test set; S2-2-2. Model definition: Define a classification model based on Transformer-XL; including the following steps: 1) Import the Python libraries required for model definition and training; 2) Configure the parameters of the Transformer-XL model according to the task requirements, create a TransfoXLConfig object, and set the model parameters including the dimension, number of layers, number of heads, and hidden layer dimension of the feedforward network; 3) Define a classification model class: Define a class and implement the forward propagation logic. The model will contain Transformer-XL and fully connected layers for the final classification task. The Transformer-XL model is used to extract features, and the fully connected layer is used for the final classification. Implement the forward propagation logic, including calling the Transformer-XL model to obtain the hidden state, performing pooling operations, and classifying through the fully connected layer; S2-2-3. Loss function and optimizer; the loss function uses cross entropy loss and selects Adam optimizer, which can adaptively adjust the learning rate of each parameter, and then set the hyperparameter learning rate and momentum to implement the loss function and optimizer; S2-2-4. Train the model through multiple iterations and verify the monitoring performance on the validation set; S2-2-5. Model evaluation: evaluate the generalization ability of the model on the test set.

[0013] As an optimization solution of the present invention, in S2-2-1, the data set is divided using a sliding window method and a segmented sampling method to ensure the time dependence and distribution uniformity of the time series data.

[0014] As an optimization solution of the present invention, in step S3, S3-1. Data collection and preprocessing: Use the digital twin model to obtain the real-time operation data and task requirement data of the equipment; clean and standardize the collected data to ensure that the data format is consistent and suitable for subsequent processing; S3-2. Task and resource modeling: Task modeling: Each task T i Represented as a four-group ,in: and are the start time and end time of the task respectively; i is the amount of resources required for the task, p i is the priority of the task; Resource Modeling: Each resource R j Represented as a two-tuple (c j ,l j ), where: c j is the capacity of the resource, l j is the current load of the resource; S3-3, using a swarm intelligence algorithm based on the self-organizing mechanism of termites’ nests to simulate the division of labor and cooperation behavior of termites; Individual representation: Each individual (termite) is represented as an agent responsible for performing task allocation and resource scheduling; Pheromone mechanism: Introduce a pheromone mechanism to record the allocation history of tasks and resources. τ ij (t) = τ ij (t-1) + Δτ ij ; Among them, τ ij (t) represents the pheromone concentration between task i and resource j at time t, Δτ ij represents the pheromone increment between task i and resource j from time t−1 to time t; Task allocation rule: Each agent selects tasks and resources based on pheromone concentration and task priority:

[0015] Among them, Pij (t) represents the probability of task i being assigned to resource j at time t, η ij represents the heuristic information of task i and resource j, α and β are parameters that control the weights of pheromone and heuristic information, and k is an index variable used to traverse all resources assigned to task i; S3-4, dynamic scheduling and adjustment; Real-time scheduling: Based on the current status of tasks and resources, each agent selects the optimal tasks and resources according to the task allocation rules; Dynamic adjustment: According to the changes in tasks and resources, the pheromone concentration is updated in real time and the task allocation probability is recalculated; S3-5. Optimization objective function: Objective function: Minimize the total task completion time and the imbalance of resource utilization;

[0016] Where n is the number of tasks, m is the number of resources, is the average load of resources, and λ is the weight coefficient used to balance task completion time and resource utilization.

[0017] In order to achieve the purpose of the present invention, the technical solution adopted is: a monitoring system for a smart park, including a digital twin module, a fault prediction module and a group intelligent decision-making module; Digital twin module: By combining the park's building information model, geographic information system data and point cloud data, a digital twin model is constructed using a deep learning-driven multimodal fusion method; Fault prediction module: Using the Transformer-XL neural network’s ability to process long sequence data, the module analyzes the equipment operation data in the digital twin module, identifies potential fault modes, issues early warnings, and displays the fault location and impact range through the digital twin module’s visual interface; Swarm intelligent decision-making module: Based on the digital twin module, it imitates the self-organizing characteristics of termite nests in nature to achieve efficient collaborative control among equipment within the park.

[0018] The present invention has positive effects: 1) The present invention collects building information model (BIM), geographic information system (GIS) and point cloud data, and uses a multimodal fusion method driven by deep learning to build a digital twin model, which overcomes the limitations of data integration and processing. The model can provide high-precision three-dimensional visualization, reflect the location and status of equipment in the park in real time, and ensure the accuracy and consistency of the data. It not only improves the accuracy of monitoring, but also provides a solid data foundation for subsequent fault prediction and resource scheduling; 2) The present invention uses the Transformer-XL neural network to process equipment operation data, identify potential failure modes and issue early warnings. By analyzing long sequence data, it can detect potential failure modes of equipment earlier and generate early warning information in a timely manner. This helps operation and maintenance personnel respond quickly, reduce equipment downtime, improve equipment reliability and safety, and reduce maintenance costs; 3) The present invention adopts a swarm intelligence algorithm based on the self-organizing mechanism of termite nests to optimize the collaboration between devices, realize efficient task allocation and resource scheduling, and can dynamically adjust the task allocation and resource scheduling strategies according to the real-time status of the equipment and task requirements, maximize resource utilization and minimize task completion time, thereby improving the operational efficiency of the entire park and reducing resource waste and idleness; 4) The present invention is not only applicable to the daily monitoring and management of smart parks, but also provides strong support for the construction of smart cities. By realizing comprehensive, efficient and intelligent management of parks, it helps to promote the sustainable development of smart parks. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0020] Figure 1 is a flow chart of the method of the present invention; Figure 2 It is a functional block diagram of the system of the present invention. DETAILED DESCRIPTION

[0021] The method of the present invention is further described below with reference to the accompanying drawings and specific embodiments.

[0022] Embodiment 1, as Figure 1 As shown, the monitoring method for a smart park described in this embodiment includes the following steps: S1. Collect the building information model, geographic information system data and point cloud data of the park, and use the multimodal fusion method driven by deep learning to build a digital twin model; In S1, point cloud data is the captured and recorded information about the structure of campus buildings, the status of equipment, and environmental characteristics. Point cloud data is captured and recorded including: 1) Use ground laser scanners to regularly scan buildings to detect whether there are structural deformations, cracks, tilts, displacements, etc. 2) Use a mobile laser scanning system (MLS) to scan the equipment in the park to obtain the geometric shape, position and surface condition of the equipment to ensure that wear or damage of the equipment can be discovered in time; 3) Use the LiDAR system carried by drones to regularly scan the park’s vegetation, terrain and surface features to help monitor changes in the natural environment, such as vegetation growth and terrain changes; use ultrasonic sensors to measure the water level of the park’s landscape lake, and use gas sensors to measure point cloud data such as air quality.

[0023] After collecting the park's building information model (BIM), geographic information system (GIS) data and point cloud data, proceed as follows: (1) Standardization and format conversion First, standardization is performed (ensuring that all data are aligned in the same coordinate system). BIM data (IFC format, using the local coordinate system), GIS data (Shapefile, using the WGS84 coordinate system), and point cloud data (LAS format, using the local coordinate system) are converted to the same coordinate system WGS84. The coordinate transformation method includes the Helmert transformation. After the coordinates are converted, the data is converted into a format suitable for deep learning processing.

[0024] Convert the data into a format suitable for deep learning processing. The steps include: 1) BIM data conversion: Convert BIM files (IFC format) into formats suitable for deep learning processing (OBJ, PLY).

[0025] 2) GIS data conversion: Convert GIS data (Shapefile, GeoJSON) into raster images or vector graphics.

[0026] 3) Point cloud data conversion: Convert point cloud data (LAS, LAZ format) into voxel grids or depth maps. Use point cloud processing tools to convert point cloud data in LAS or LAZ format into voxel grids or depth maps.

[0027] Standardization followed by format conversion ensures that all data is aligned in the same coordinate system, avoiding errors and inconsistencies in subsequent processing. Format conversion in the correct coordinate system ensures that the generated data has high accuracy and reliability. This sequence ensures that the data can be seamlessly integrated in subsequent processing, and the generated digital twin model is more accurate and reliable.

[0028] (2) Feature extraction: extract useful features from the data after format conversion, including: BIM data feature extraction: Analyze BIM data in OBJ or PLY format to extract building structural information (walls, door and window locations, etc.). Use convolutional neural networks (CNNs) to automatically identify these features.

[0029] GIS data feature extraction: Through spatial coordinate transformation and feature encoding technology, the geographic information in raster images (TIFF) or vector graphics (SVG) is converted into feature vectors that can be used for training.

[0030] Point cloud data feature extraction: 3D convolutional neural networks (3DCNNs) are used to extract physical state information (object shape, surface details, etc.) from point cloud data represented by voxel grids or depth maps.

[0031] (3) Multimodal feature fusion: features extracted from different data sources are fused together. Each branch is responsible for processing one type of input data, and finally the outputs of each branch are merged through a fully connected layer.

[0032] BIM branch: Use convolutional neural networks (CNNs) to process geometric data in OBJ or PLY format.

[0033] GIS branch: Use recurrent neural networks (RNNs) to process raster images (TIFF) or vector graphics (SVG).

[0034] Point cloud branch: Process voxel grids or depth maps using 3D convolutional neural networks (3D CNNs) or PointNet.

[0035] Fully connected layer (FC): The output of each branch is fused through the fully connected layer.

[0036] Z = FC([y BIM , y GIS , y PoitCloud ]), where y BIM Feature vector extracted from BIM data, y GIS Feature vector extracted from GIS data, y PoitCloud Feature vectors extracted from point cloud data.

[0037] (4) Model training and optimization 41) Use spatial consistency loss function to ensure that the results generated by the digital twin model are consistent in space;

[0038] Where: L spatial is the spatial consistency loss function, Indicates traversing and summing all point pairs (i, j); for the i-th data point x in the BIM data BIM , i ,f(x BIM,i ) represents the representation of the data point in the feature space; ||•||2 represents the Euclidean distance, g(x BIM , i) represents the spatial coordinates of the i-th data point in the BIM data.

[0039] 42) The Adam optimization algorithm is selected to minimize the spatial consistency loss function to ensure that the results generated by the digital twin model are spatially consistent. This optimization method is not only computationally efficient, but also has good stability and robustness, and is suitable for complex spatial consistency tasks.

[0040] 43) Train and evaluate the model: Initialize the model parameters during training, then perform forward propagation, calculate loss, back propagation, and update parameters for each batch of data; repeat forward propagation, calculate loss, back propagation, and update parameters until the predetermined number of epochs is reached, use K-fold cross-validation to evaluate model performance, use grid search to find the best hyperparameter combination, evaluate model performance on an independent test set, and build a digital twin model.

[0041] Collecting the park's building information model (BIM), geographic information system (GIS) and point cloud data can ensure that the generated data has high accuracy and reliability, providing a solid foundation for subsequent feature extraction and multimodal fusion, while edge computing sacrifices the model's accuracy and complex problem-solving capabilities.

[0042] S2. Through the digital twin model, the location and operation data of the equipment are obtained, and the Transformer-XL neural network is used to process the equipment operation data, identify potential failure modes and provide early warnings, and display the failure location and impact range through the visual interface of the digital twin model.

[0043] S2-1. According to the operation characteristics of the equipment and the fault occurrence cycle, the time series data is divided into windows of fixed length, and the data of each window is preprocessed after windowing; Preprocessing includes normalization, filling missing values ​​and removing outliers. Normalization uses Z-score standardization, filling missing values ​​is done through interpolation, and removing outliers is done through statistical methods to ensure data quality.

[0044] S2-2. Use the preprocessed data to train the Transformer-XL model to learn the normal state and various failure modes of the equipment. The specific implementation includes the following steps: S2-2-1. Dataset division: Divide the preprocessed data into training set, validation set and test set; specifically include the following: 1) Time series segmentation strategy: Sliding window method: Considering the time dependence of device operation data, the sliding window method is adopted for data segmentation. Assume that the size of each window is L and the sliding step is S (S < L), which can ensure partial overlap between adjacent windows, thus capturing more temporal information.

[0045] Piecewise sampling method: To avoid data deviation in certain specific time periods (such as the device frequently malfunctioning or operating normally in a certain period), windows can be randomly sampled in different time periods. For example, segment by day or week, and then randomly select a certain proportion of data from each time period as the training set, validation set, and test set.

[0046] 2) Dataset division ratio: Training set: 70% of the data is used for model training.

[0047] Validation set: 15% of the data is used for hyperparameter tuning and model performance monitoring.

[0048] Test set: 15% of the data is used for finally evaluating the generalization ability of the model.

[0049] 3) Time distribution uniformity: Uniform time sampling: To ensure that the training set, validation set, and test set contain the operating states of the device in different time periods, the method of uniform time sampling is adopted to avoid model deviation caused by time period differences. Specifically, according to the timestamps of the data, the entire time range is divided into several equally spaced time periods, and then data is sampled proportionally from each time period.

[0050] In summary, the time series segmentation strategy provides a suitable data format for subsequent steps. The dataset division ratio determines how to use these formatted data. The time distribution uniformity ensures the uniform distribution of data in time and enhances the generalization ability of the model.

[0051] S2-2-2, Model definition: Define a classification model based on Transformer-XL; the specific implementation includes the following steps: 1) Import the Python libraries required for model definition and training; 2) Configure the parameters of the Transformer-XL model according to the task requirements, create a TransfoXLConfig object, and set parameters such as the dimension, number of layers, number of heads, and hidden layer dimension of the feed-forward network of the model; 3) Define a classification model class: Define a class and implement the forward propagation logic. The model will contain Transformer-XL and fully connected layers for the final classification task. The Transformer-XL model is used to extract features, and the fully connected layer is used for the final classification. Implement the forward propagation logic, including calling the Transformer-XL model to obtain the hidden state, performing pooling operations, and performing classification through the fully connected layer.

[0052] S2-2-3, loss function and optimizer; the loss function uses cross entropy loss. For multi-category classification problems, cross entropy loss can effectively measure the difference between the probability distribution predicted by the model and the true label. Select the Adam optimizer, which can adaptively adjust the learning rate of each parameter, and then set the hyperparameter learning rate and momentum to implement the loss function and optimizer.

[0053] S2-2-4. Train the model through multiple iterations and monitor the performance on the validation set. 1) Initialize the training loop: Set the total number of training rounds (epochs), which determines how many times the model will traverse the entire training set. Each round of training will traverse all samples in the training set.

[0054] 2) Batch data loading: Use DataLoader to extract data from the training set in small batches. This not only reduces the amount of computation per iteration, but also uses GPU to accelerate the training process. Batch size is an important hyperparameter that needs to be adjusted.

[0055] 3) Forward propagation: For each batch of data, a forward propagation process is performed, that is, the input data is passed through the Transformer-XL model and the subsequent fully connected layer to obtain the prediction results.

[0056] 4) Calculate loss: Use the cross entropy loss function to compare the difference between the model output and the true label to obtain the loss value of the current batch.

[0057] 5) Back propagation and optimization: Based on the calculated loss value, the back propagation algorithm is executed to calculate the gradient, and then the Adam optimizer is used to update the model parameters. In this process, hyperparameters such as the learning rate are adaptively adjusted to accelerate convergence and avoid local optimality.

[0058] 6) Regularly evaluate the validation set: Evaluate the performance of the model on the validation set after a certain number of training steps or after each epoch. This can monitor whether the model is overfitting and adjust the model's hyperparameters accordingly.

[0059] 7) Record and adjust: Based on the performance on the validation set, record the changes in important indicators such as loss value and accuracy. If you find that the model performance is no longer improving or even decreases, you may need to adjust the hyperparameters, such as adjusting the learning rate, increasing the regularization strength, and other measures.

[0060] 8) Save the best model: During the entire training process, keep track of the best performance on the validation set and save the model weights in the corresponding state.

[0061] Through the above steps, the Transformer-XL model can be effectively trained and ensured to have good generalization ability. At the same time, monitoring the validation set can timely detect and solve possible problems such as overfitting or underfitting.

[0062] S2-2-5. Model evaluation: Evaluate the generalization ability of the model on the test set; 1) Load the best model: The best model weights saved during the training process (usually the model with the best performance on the validation set) need to be loaded back. Make sure to use the best performing model for the final test set evaluation.

[0063] 2) Prepare test data to ensure the consistency of input data format.

[0064] 3) Load test data in batches: Use the same data loader as in the training and validation phases to load the test data into the model in batches.

[0065] 4) Perform forward propagation: For each batch of test data, perform forward propagation through the loaded best model to obtain the model's prediction results. There is no need to perform backpropagation or parameter update here.

[0066] 5) Calculate the evaluation index: Use the mean square error (MSE) as the evaluation index to compare the performance of the model on the test set with the performance on the training and validation sets to check whether there is a significant difference. The specific implementation process of learning the normal state and various failure modes of the device is: (1) Convert the equipment operation data into an input format suitable for the Transformer-XL model. Assume that each input data is a time series X, where X=[x1,x2,..x t .,x T ], x t The feature vector at time point t represents the operating status of the device over time. The embedding layer is used to map the original feature vector to a higher dimensional space, E(X)=[E(x1),E(x2),...,E(x T )], where: E is the embedding matrix, which is used to convert the original feature vector into a high-dimensional vector.

[0067] (2) After the input passes through the embedding layer, it is passed to the Transformer-XL model for feature extraction, H = Transformer-XL (E (X)), where H is the hidden state of the Transformer-XL output. The hidden state is pooled (average pooling) to obtain a fixed-length feature representation:

[0068] The pooled features are input into the fully connected layer for classification: y=FC(h), where y is the predicted category probability distribution.

[0069] (3) Use the cross entropy loss function to calculate the difference between the model prediction result and the true label:

[0070] Where: C is the number of categories, y i is the probability distribution of the true labels, is the probability distribution predicted by the model.

[0071] (4) Use the Adam optimizer to perform backpropagation and update model parameters:

[0072] Where: θ is the model parameter, η is the learning rate, is the gradient of the loss function with respect to the parameters.

[0073] Through the above steps, the preprocessed data can be effectively used to train the Transformer-XL model and learn the normal state and various failure modes of the equipment.

[0074] Based on the output of the Transformer-XL model, the equipment operation status is monitored in real time, and once an abnormal pattern is detected, an early warning message is immediately generated. The early warning message should include detailed fault type, occurrence time, and recommended response measures to ensure that operation and maintenance personnel can respond quickly.

[0075] The fault location and impact range are displayed through the visualization interface of the digital twin model, and the fault location and impact range of the equipment are intuitively displayed using the three-dimensional visualization interface provided by the digital twin model. Including: 1) Mark the specific location of the faulty device in the 3D view.

[0076] 2) Display the current status of the device (normal, warning, fault) through color coding or animation effects.

[0077] 3) Provide interactive functions, allowing users to click on the device to view detailed operating data and history. Based on the actual operating conditions and feedback information, the Transformer-XL model and digital twin model are continuously optimized to improve the accuracy and reliability of the system.

[0078] S3. Dynamic task allocation and resource scheduling: The goal of dynamic task allocation and resource scheduling is to reasonably allocate resources and tasks according to the real-time status of the equipment and task requirements, so as to maximize resource utilization and minimize task completion time. It is a key step to ensure efficient operation of equipment and fault warning. The swarm intelligence algorithm (termite nest self-organization mechanism) is used to imitate the self-organization mechanism of termite nests in nature, optimize the collaboration between equipment, and achieve efficient task allocation and resource scheduling.

[0079] The specific steps of dynamic task allocation and resource scheduling are as follows: S3-1. Data collection and preprocessing: Data collection: The digital twin model is used to obtain the real-time operation data (location, status, load, etc.) of the equipment and the task requirement data (task type, priority, deadline, etc.).

[0080] Data preprocessing: Clean and standardize the collected data to ensure that the data format is consistent and suitable for subsequent processing.

[0081] S3-2, Task and Resource Modeling: Task modeling: Each task T i Represented as a four-group ,in: and are the start time and end time of the task respectively; i is the amount of resources required for the task, p i is the priority of the task; Resource Modeling: Each resource R j It can be represented as a two-tuple (c j ,l j ), where: c j is the capacity of the resource, l j is the current load of the resource; S3.3. Use a swarm intelligence algorithm based on the self-organizing mechanism of termites’ nests to simulate the division of labor and cooperation behavior of termites.

[0082] Individual representation: Each individual (termite) is represented as an agent responsible for performing task allocation and resource scheduling.

[0083] Pheromone mechanism: Introduce the pheromone mechanism to record the allocation history of tasks and resources. The higher the pheromone concentration, the greater the probability that the task or resource will be allocated.ij (t) = τ ij (t-1) + Δτ ij ; where τ ij (t) represents the pheromone concentration between task i and resource j at time t, Δτ ij represents the pheromone increment between task i and resource j from time t−1 to time t.

[0084] Task allocation rule: Each agent selects tasks and resources based on pheromone concentration and task priority:

[0085] Among them, P ij (t) represents the probability of task i being assigned to resource j, η ij represents the heuristic information (task priority and resource load) of task i and resource j, and α and β are parameters that control the weights of pheromone and heuristic information.

[0086] S3.4, dynamic scheduling and adjustment; Real-time scheduling: Based on the current status of tasks and resources, each agent selects the optimal tasks and resources according to the task allocation rules.

[0087] Dynamic adjustment: According to the changes in tasks and resources, the pheromone concentration is updated in real time and the task allocation probability is recalculated.

[0088] S3.5. Optimization objective function Objective function: Minimize the imbalance between total task completion time and resource utilization.

[0089]

[0090] Where n is the number of tasks, m is the number of resources, is the average load of resources, and λ is the weight coefficient used to balance task completion time and resource utilization.

[0091] Through the dynamic task allocation and resource scheduling of the swarm intelligence algorithm, the average task completion time is significantly reduced. Resource utilization is more balanced, avoiding resource overload or idleness. It can quickly respond to changes in tasks and resources, improving overall operating efficiency. The swarm intelligence algorithm can dynamically adjust task allocation according to environmental changes, and has strong robustness and adaptability.

[0092] By introducing swarm intelligence algorithms (simulating the self-organizing mechanism of termite nests), dynamic task allocation and resource scheduling can effectively improve the operating efficiency of smart parks and ensure efficient operation of equipment and fault warning. This method combines the division of labor and cooperation mechanism in nature with optimization algorithms, can adapt to real-time changes in task and resource status, and has high practicality and scalability.

[0093] Step 1: Initial state, task T1 = (10:00, 12:00, 5, 1), Indicates that the task starts at 10:00. Indicates that the task ends at 12:00. r1=5 means that the task requires 5 units of resources, and p1=1 means that the priority of the task is 1 (high priority).

[0094] Step 2: Resource R1 = (10,3), c1 = 10: the capacity of the resource is 10 units; l1 = 3: the current load of the resource is 3 units.

[0095] Step 2: Calculate pheromone and heuristic information; Assuming the initial pheromone concentration τ 11 (0) = 0.5, heuristic information η 11 It can be calculated based on task priority and resource load:

[0096] Step 3: Calculate the task assignment probability; Assume that only resource R1 is available, therefore:

[0097] Step 4: Task Assignment: Since P 11 (t)=1, task T1 will be assigned to resource R1, Step 5: Update resource status and pheromones Update the load of resource R1: l1 = 3 + 5 = 8; Update pheromone concentration: τ 11 (t +1) = τ 11 (t) + Δτ 11 ; Assume Δτ 11 = 0.1, then: τ 11 (t +1) = 0.5+0.1 = 0.6.

[0098] By introducing the termite nest self-organization mechanism, the intelligent and dynamic task allocation and resource scheduling are realized. Reasonable allocation decisions can be made according to the current task requirements and resource status. After the task T1 is assigned, the status of resource R1 is updated (increased from 3 to 8), which shows that the system has the ability to dynamically adjust resources according to the task execution situation; by updating the pheromone concentration, the successful path selection (i.e., the allocation of tasks to resources) is strengthened, which helps to make better decisions in the future. The increment of pheromone reflects this positive feedback mechanism. As more tasks are successfully assigned, the corresponding path selection will become more preferred. This method has achieved remarkable results in reducing task completion time, balancing resource utilization, improving system response speed and enhancing adaptability.

[0099] Embodiment 2, as Figure 2 As shown, the monitoring system for a smart park described in this embodiment includes a digital twin module, a fault prediction module and a group intelligent decision-making module; Digital Twin Module: Creates a high-precision, dynamically updated digital twin model by combining building information modeling (BIM), geographic information system, and real-time point cloud data; Fault prediction module: Using the Transformer-XL neural network’s ability to process long sequence data, the digital twin module analyzes the equipment operation data in the park, identifies potential fault modes, and issues early warnings; Swarm intelligent decision-making module: Based on the digital twin module, it imitates the self-organizing characteristics of termite nests in nature to achieve efficient collaborative control among equipment within the park.

[0100] 1) Digital Twin Module Input: Building Information Model (BIM), Geographic Information System (GIS) data, and real-time point cloud data.

[0101] Output: A highly accurate, dynamically updated digital twin model that provides real-time location, status, and environmental information of the device.

[0102] Function: As the data foundation of the entire system, it provides accurate equipment and environmental information support for the other two modules.

[0103] 2) Fault prediction module Input: Equipment operation data from the digital twin module.

[0104] Output: prediction results and warning information of equipment operating status.

[0105] Function: By analyzing the equipment operation data, potential failure modes can be identified and early warnings can be issued to operation and maintenance personnel in a timely manner to improve the reliability and safety of the equipment.

[0106] 3) Group intelligent decision-making module: Input: Real-time equipment operation data and task requirement data from the digital twin module.

[0107] Output: optimal task allocation plan and resource scheduling strategy.

[0108] Function: By simulating the division of labor and cooperation mechanism in nature, dynamically adjusting task allocation and resource scheduling, ensuring maximum resource utilization and minimization of task completion time, and improving the overall operation efficiency of the park.

[0109] The digital twin module provides real-time operation data and environmental information of the equipment. These data are used by the fault prediction module for monitoring the health status of the equipment and fault warning, and are also used by the group intelligent decision-making module for task allocation and resource scheduling. The fault prediction module relies on the high-precision data provided by the digital twin module for fault prediction, while the group intelligent decision-making module uses these data to perform efficient resource management and task allocation to ensure efficient operation and fault warning of equipment in the park. The warning information of the fault prediction module can trigger the dynamic adjustment of the group intelligent decision-making module and reallocate resources to deal with sudden failures, thereby forming a closed-loop control system and further improving the response speed and reliability of the system. This close cooperation between modules not only improves the management efficiency of the smart park, but also enhances the reliability and adaptability of the system, ensuring the safe and stable operation of equipment in the park.

[0110] As shown in Table 1, the difference in early warning time between the method of the present invention and the traditional method is shown. The specific data are as follows: Table 1 Comparison of fault prediction timeliness

[0111] It can be concluded from Table 1 that, by using the Transformer-XL neural network, the present invention can identify potential failure modes earlier and provide a longer warning time, which helps operation and maintenance personnel to respond in a timely manner.

[0112] As shown in Table 2, the differences in task completion time and resource utilization between the method of the present invention and the traditional method are shown. The specific data are as follows: Table 2 Task completion time and resource utilization

[0113] From Table 2, we can conclude that by introducing the termite nest self-organization mechanism, dynamic task allocation and resource scheduling can significantly reduce task completion time and balance resource utilization, which is significantly better than traditional methods in terms of task completion time. The more tasks there are, the more obvious the advantage. The imbalance of resource utilization is also significantly reduced, which can more evenly allocate resources and avoid resource overload or idleness.

[0114] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A monitoring method for a smart park, characterized in that: include: S1. Collect the building information model, geographic information system data and point cloud data of the park, and use the multimodal fusion method driven by deep learning to build a digital twin model; S2. Obtain the location and operation data of the equipment through the digital twin model, use the Transformer-XL neural network to process the equipment operation data, identify potential failure modes and provide early warnings, and display the failure location and impact range through the visual interface of the digital twin model; S3. Adopt the termite nest self-organizing mechanism to optimize the collaboration between devices and realize dynamic task allocation and resource scheduling.

2. A monitoring method for a smart park according to claim 1, characterized in that: In S1, the point cloud data is the captured and recorded information about the structure of the campus buildings, the status of the equipment and the environmental characteristics.

3. A monitoring method for a smart park according to claim 2, characterized in that: S1 includes: S1-1. After collecting the building information model, geographic information system data and point cloud data of the park, convert them to the same coordinate system WGS84 through Helmert transformation; convert BIM files into OBJ or PLY format, convert GIS data into raster images or vector graphics, and convert point cloud data into voxel grids or depth maps; S1-2. For BIM data, CNNs are used to extract building structural features; for GIS data, geographic information features are extracted through spatial coordinate transformation and feature encoding technology; for point cloud data, PointNet is used to extract 3D shape features; S1-3, multimodal feature fusion, combines the features extracted from the three data sources. Each branch is responsible for processing one type of input data, and finally the outputs of each branch are merged through the fully connected layer.

4. The monitoring method for a smart park according to claim 3, characterized in that: S1 also includes: S1-4. Model training and optimization, including: 41) Use spatial consistency loss function to ensure that the results generated by the digital twin model are consistent in space; ； Where: L spatial is the spatial consistency loss function, Indicates traversing and summing all point pairs (i, j); for the i-th data point x in the BIM data BIM , i ,f(x BIM,i ) represents the representation of the data point in the feature space; for the jth data point x in the BIM data BIM , j ,f(x BIM,j ) represents the representation of the data point in the feature space; ||•||2 represents the Euclidean distance; g(x BIM , i ) represents the spatial coordinates of the i-th data point in the BIM data; g(x BIM , j ) represents the spatial coordinates of the jth data point in the BIM data; 42) Select the Adam optimization algorithm to minimize the spatial consistency loss function; 43) Train and evaluate the model: Initialize the model parameters during training, then perform forward propagation, calculate loss, back propagation, and update parameters for each batch of data; repeat forward propagation, calculate loss, back propagation, and update parameters until the predetermined number of epochs is reached, use K-fold cross-validation to evaluate model performance, use grid search to find the best hyperparameter combination, evaluate model performance on an independent test set, and build a digital twin model.

5. A monitoring method for a smart park according to claim 4, characterized in that: In S2, it includes: S2-1. According to the operation characteristics of the equipment and the fault occurrence cycle, the time series data is divided into windows of fixed length, and the data of each window is preprocessed after windowing; S2-2. Use the preprocessed data to train the Transformer-XL model to learn the normal state and various failure modes of the equipment.

6. A monitoring method for a smart park according to claim 5, characterized in that: In S2-2, it includes: S2-2-1. Dataset division: divide the preprocessed data into training set, validation set and test set; S2-2-2. Model definition: Define a classification model based on Transformer-XL; including the following steps: 1) Import the Python libraries required for model definition and training; 2) Configure the parameters of the Transformer-XL model according to the task requirements, create a TransfoXLConfig object, and set the model parameters including the dimension, number of layers, number of heads, and hidden layer dimension of the feedforward network; 3) Define a classification model class: Define a class and implement the forward propagation logic. The model will contain Transformer-XL and fully connected layers for the final classification task. The Transformer-XL model is used to extract features, and the fully connected layer is used for the final classification. Implement the forward propagation logic, including calling the Transformer-XL model to obtain the hidden state, performing pooling operations, and classifying through the fully connected layer; S2-2-3, loss function and optimizer; the loss function uses cross entropy loss, selects Adam optimizer, adaptively adjusts the learning rate of each parameter, and then sets the hyperparameter learning rate and momentum to implement the loss function and optimizer; S2-2-4. Train the model through multiple iterations and verify the monitoring performance on the validation set; S2-2-5. Model evaluation: evaluate the generalization ability of the model on the test set.

7. A monitoring method for a smart park according to claim 6, characterized in that: In S2-2-1, the data set is divided using the sliding window method and segmented sampling method to ensure the time dependence and distribution uniformity of the time series data.

8. A monitoring method for a smart park according to claim 7, characterized in that: In step S3, S3-1. Data collection and preprocessing: Use the digital twin model to obtain the real-time operation data and task requirement data of the equipment; clean and standardize the collected data to ensure that the data format is consistent and suitable for subsequent processing; S3-2, Task and Resource Modeling: Task modeling: Each task T i Represented as a four-tuple ,in: and are the start time and end time of the task respectively; i is the amount of resources required for the task, p i is the priority of the task; Resource Modeling: Each resource R j Represented as a two-tuple (c j ,l j ), where: c j is the capacity of the resource, l j is the current load of the resource; S3-3, using a swarm intelligence algorithm based on the self-organizing mechanism of termites’ nests to simulate the division of labor and cooperation behavior of termites; Individual representation: Each individual termite is represented as an agent, responsible for performing task allocation and resource scheduling; Pheromone mechanism: Introduce a pheromone mechanism to record the allocation history of tasks and resources. t ij (t) = τ ij (t-1) + Δt ij ; Among them, τ ij (t) represents the pheromone concentration between task i and resource j at time t, Δτ ij represents the pheromone increment between task i and resource j from time t−1 to time t; Task allocation rule: Each agent selects tasks and resources based on pheromone concentration and task priority: ； Among them, P ij (t) represents the probability of task i being assigned to resource j at time t, η ij represents the heuristic information of task i and resource j, α is the parameter controlling the pheromone weight, and β is the parameter controlling the heuristic information weight; S3-4, dynamic scheduling and adjustment; Real-time scheduling: Based on the current status of tasks and resources, each agent selects the optimal tasks and resources according to the task allocation rules; Dynamic adjustment: According to the changes in tasks and resources, the pheromone concentration is updated in real time and the task allocation probability is recalculated; S3-5. Optimization objective function: Objective function: Minimize the total task completion time and the imbalance of resource utilization; ； Where n is the number of tasks, m is the number of resources, is the average load of resources, and λ is the weight coefficient used to balance task completion time and resource utilization.

9. A monitoring system for a smart park, used to implement a monitoring method for a smart park as described in any one of claims 1 to 8, characterized in that: Includes digital twin module, fault prediction module and swarm intelligence decision-making module; Digital twin module: By combining the park's building information model, geographic information system data and point cloud data, a digital twin model is constructed using a deep learning-driven multimodal fusion method; Fault prediction module: Using the Transformer-XL neural network’s ability to process long sequence data, the module analyzes the equipment operation data in the digital twin module, identifies potential fault modes, issues early warnings, and displays the fault location and impact range through the digital twin module’s visual interface; Swarm intelligent decision-making module: Based on the digital twin module, it imitates the self-organizing characteristics of termite nests in nature to achieve efficient collaborative control among equipment within the park.

Images