Visual environment detection method and device, storage medium and computer device

Abstract

This application discloses a method, apparatus, storage medium, and computer device for visual environment detection. The method includes: in response to receiving twin model update information corresponding to a target park, generating a content area distribution map corresponding to each content type based on key events within a target time period, the frequency of occurrence of key events, the number of times the digital twin model corresponding to each sub-area is displayed, and key content tags under the content type; generating a prediction model corresponding to each area based on the different importance of each area in the content area distribution map, wherein each prediction model shares a common model module; determining a sequence of visual environment detection distribution maps under the content type corresponding to the content area distribution map using the prediction model set; generating a corresponding twin model based on the sequence of visual environment detection distribution maps; and in response to receiving twin model display information for at least one content type, overlaying and displaying the corresponding at least one twin model.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a visual environment detection method, apparatus, storage medium, and computer equipment. Background Technology

[0002] Currently, with the continuous development of digital twin and artificial intelligence technologies, predicting and visually modeling various events within the park has become an important development direction. This technology allows for timely adjustments based on predicted trends of events within the park, effectively preventing significant losses.

[0003] In the field of visual modeling of various events in the park, the current common approach is to first obtain the specific content of events under each type, then select the corresponding prediction model for the event content under each type, and finally output the prediction results. Different events correspond to different prediction models, and the prediction models corresponding to different types of events are trained separately.

[0004] However, this existing technology has significant drawbacks in practical applications. When not all events in the target area require extremely high prediction accuracy, deploying and training prediction models for each event separately consumes substantial server resources. Especially when the target area has a large number of events but limited server computing resources, prediction models for each event may queue for training. This not only delays predictions for some events but also significantly extends the development cycle of the corresponding twin models, failing to meet the needs of efficient park management. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method, apparatus, storage medium, and computer device for visual environment detection.

[0006] According to one aspect of this application, a method for visualizing environment detection is provided, the method comprising: In response to receiving the twin model update information corresponding to the target park, for each content type, a generation step is performed to obtain the twin model corresponding to each content type; In response to receiving twin model display information for at least one content type, the corresponding twin model is overlaid and displayed. The displayed twin model supports at least one of the following processing operations: model dragging, region search, model scaling, and visualization of the target location environment. The generation step includes: Based on the key events within the target time period, the frequency of occurrence of key events, the number of times the digital twin model corresponding to each sub-region is displayed, and the key content tags under the content type, a content area distribution map corresponding to the content type is generated. Based on the different importance of each region in the content region distribution map, a prediction model is generated for each region to form a prediction model set. Different regions use prediction models with different model structures, and each prediction model shares a model module. The closer the importance is, the more common the number of the same model modules. Using the prediction model set, determine the sequence of visual environment detection distribution maps under the content type corresponding to the content region distribution map; Based on the sequence of visual environment detection distribution maps, a corresponding twin model is generated.

[0007] According to another aspect of this application, a visual environment detection device is provided, the device comprising: An execution unit is configured to respond to receiving twin model update information corresponding to the target park, and for each content type, execute a generation step to obtain a twin model corresponding to each content type; wherein, the generation step includes: generating a content area distribution map corresponding to the content type based on key events within the target time period, the frequency of occurrence of key events, the number of times the digital twin model corresponding to each sub-region is displayed, and key content tags under the content type; generating a prediction model corresponding to each region based on the different importance levels of each region in the content area distribution map, forming a prediction model set, wherein different regions use prediction models with different model structures, and the prediction models share model modules, with the number of identical model modules corresponding to the closer the importance levels are; using the prediction model set, determining the sequence of visual environment detection distribution maps under the content type corresponding to the content area distribution map; and generating the corresponding twin model based on the sequence of visual environment detection distribution maps; The display unit is used to respond to receiving twin model display information for at least one content type, and to overlay and display the corresponding twin model. The displayed twin model supports at least one of the following processing operations: model dragging, region search, model scaling, and visualization of the target location environment.

[0008] According to another aspect of this application, a storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the above-described visual environment detection method.

[0009] According to another aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described visualization environment detection method.

[0010] By employing the above technical solutions, the visualization environment detection method, apparatus, storage medium, and computer equipment provided in this application embodiment, utilize a set of prediction models with model structure associations. With reasonable use of server resources, this enables the accurate and efficient generation of twin models for various content types within a target area, as well as the display of twin models to meet diverse display needs. Specifically, the reason for the insufficient accuracy and efficiency in generating related twin models is that, when high prediction accuracy is not required for all events in the target area, the separate deployment and training of event-specific prediction models often incurs significant server resource overhead. Furthermore, when there are many events in the target area and limited server computing resources, scenarios arise where event-specific prediction models are trained in a queue, leading to delays in event prediction and a long twin model establishment cycle. Based on this, the visualization environment detection method of some embodiments of this disclosure firstly, in response to receiving the twin model update information corresponding to the target park, performs the following generation steps for each content type: Step 1: Based on each key event within the target time period, the frequency of occurrence of key events, the number of times the digital twin model corresponding to each sub-region is displayed, and the key content tags under the content type, the content area distribution map corresponding to the content type can be accurately determined, thereby determining the distribution of each type of event within the corresponding area of ​​the target park. The prediction model corresponding to each region in the content area distribution map is determined. Based on the different importance levels of the regions, prediction models with different model structures are adopted. Each prediction model shares a common model module, and the closer the importance levels are, the more common the number of identical model modules. Here, by setting prediction models with different model structures under different importance levels, and by setting common model modules among the prediction models and the greater the number of identical model modules corresponding to closer importance levels, it is possible to achieve that each event under the same content type uses the same prediction model framework. Furthermore, for the prediction model framework under the same framework, corresponding model training can be implemented, which can greatly improve the deployment efficiency of the prediction model and reduce the computational load of model training and the selection cost of model structure. Therefore, efficient training and deployment of models corresponding to various content events under the same content type can be achieved. Thirdly, using the obtained prediction model set, the sequence of visual environment detection distribution maps representing changes in visual environment content at different times under the content region distribution map can be accurately determined, facilitating the subsequent generation of accurate twin models. Fourthly, based on the sequence of visual environment detection distribution maps, twin models corresponding to the content type can be accurately generated. Finally, in response to receiving twin model display information for at least one content type, the corresponding at least one twin model is overlaid and displayed. The displayed twin model supports the following processing operations: model dragging, region search, model scaling, and visualization of the target location's environment.Here, by setting up dynamic selection and overlay of twin models, the predicted content of events under different content types can be displayed. In summary, by implementing a common prediction model framework for various events of the same content type, and by training corresponding models within the same framework, the deployment efficiency of prediction models can be greatly improved, the computational load of model training and the selection cost of model structure can be reduced, and scenarios where prediction models for events are trained in a queue can be avoided, which could lead to delays in event prediction and a long development cycle for the corresponding twin models.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a visual environment detection method provided in an embodiment of this application is shown. Figure 2 A flowchart illustrating another visual environment detection method provided in an embodiment of this application is shown; Figure 3 This paper shows a schematic diagram of the structure of a visual environmental detection device provided in an embodiment of this application; Figure 4 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0013] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0014] This embodiment provides a visual environment detection method, such as Figure 1 As shown, the method includes: Step 101: In response to receiving the twin model update information corresponding to the target park, for each content type, perform a generation step to obtain the twin model corresponding to each content type; wherein, the generation step includes steps 1011 to 1014: Step 1011: Generate a content area distribution map corresponding to the content type based on the key events within the target time period, the frequency of occurrence of key events, the number of times the digital twin model corresponding to each sub-region is displayed, and the key content tags under the content type.

[0015] In some embodiments, the executing entity of the above-described visual environment detection method (e.g., an electronic device) can determine the content area distribution map corresponding to the content type based on key events within the target time period, the frequency of occurrence of key events, the number of times the digital twin model corresponding to each sub-area is displayed, and the key content tags under the content type. The target park can be the park to be visually detected. Visual environment detection can be a processing method that performs environmental detection on the environment of the target park and visualizes the environmental content. Twin model update information can be information on updating the model content of the digital twin model corresponding to the target park. In practice, twin model update information can be a model content update instruction in a predetermined format. In practice, for the park environment of the target park, the corresponding digital twin model supports the display of the following content types: infrastructure content type, park security content type, park fire protection content type, and application-related content type. The infrastructure content type can be the content type of work content related to infrastructure. The park security content type can be the content type of work content related to security facilities. The park fire protection content type can be the content type of work content related to fire protection facilities. The application-related content type can be the content type of the relationship between various applications. For example, if the target park is a target factory, the infrastructure can be the operational facilities within the factory. For example, if the target factory is a power plant, the corresponding infrastructure could be electrical equipment. Security facilities can be facilities related to factory security. For example, security facilities could be security cameras, access control systems, anti-theft systems, and patrol systems. Fire protection facilities could be fire alarm systems and fire elevators. Depending on the content type, there are event prediction functions for the corresponding facilities. For example, for infrastructure, there is a need for facility anomaly detection. For security facilities, there is a need for security anomaly detection. For fire protection facilities, there is a need for fire anomaly detection. Regarding the relationships between applications, there is a need for application graph structure representation. That is, by categorizing the various types of facilities within the target park, at least one content type can be obtained. The target time period can be a predetermined time period before the current time. For example, if the current time is July 15th, the corresponding target time period could be January 15th to July 15th. Key events can be important events occurring within the target park. For example, for infrastructure, a significant event could be a malfunction event or a work-related event (e.g., a workload requirement event). Each key event can be an event within a content type that is considered "important." The frequency of key events can be the number of times a key event occurs within a target time period. For example, for a content type called "park fire safety," the corresponding key events include: fires caused by welding operations, electrical faults, and violations of operating procedures. The frequency of fires caused by welding operations within the target time period is 5 times.Electrical faults and operational violations are expected to occur 3 times within the target time period. Each sub-region can be a functional division of the target park. For example, for a power plant, the corresponding sub-regions could include: main plant area, process unit area, power distribution unit area, hydropower facility area, new energy facility area, transportation and loading / unloading area, and administrative management area. The digital twin model corresponding to each sub-region can be a display model showing the events occurring in each facility within that sub-region. The number of displays refers to the number of times the digital twin model corresponding to the sub-region is clicked and expanded. Key content tags can represent the importance of the corresponding event content. For example, key content tags can be one of the following: Level A, Level B, or Level C. Key content tags can also be color-coded. For example, red tags represent the highest importance level, yellow tags represent a secondary importance level, and gray tags represent the lowest importance level. The content area distribution map can be a regional distribution map of content events under each content type in the park map corresponding to the target park. That is, the content type of the event content corresponding to a region is unique. For example, the content area distribution map can use different colors to display the area range under different content types. A content area distribution map can be a map showing the distribution of various content events within a target park. In practice, a content area distribution map can be used to display the areas where various content events occur under different content types.

[0016] As an example, the aforementioned executing entity can set the feature weights of each feature under the target time period, the frequency of occurrence of key events, and the number of times the corresponding digital twin model is displayed in each sub-region, to label the importance of content events and obtain a content region distribution map.

[0017] In some optional implementations of certain embodiments, after step 1101, the steps further include: The first step is to display the content area distribution map on the target canvas page. Then, using the various canvas processing operations supported by the target canvas page, the content area distribution map is further subdivided to obtain a re-divided distribution map. The target canvas page can be a page that displays the content area distribution map in canvas format. Canvas processing operations can be various image processing operations supported on the canvas. For example, canvas processing operations can be, but are not limited to, at least one of the following: image deletion, image addition, image modification, and tag addition. In practice, the re-divided distribution map can be determined by the operation object corresponding to the target canvas page.

[0018] The second step is to determine the re-divided distribution map as a content area distribution map.

[0019] Step 1012: Based on the different importance of each region in the content region distribution map, generate a prediction model corresponding to each region to form a prediction model set. Different regions use prediction models with different model structures, and each prediction model has a shared model module. The closer the importance is, the more model modules are corresponding to each other.

[0020] In some embodiments, the aforementioned execution entity can determine the prediction model corresponding to each region in the content area distribution map. Different prediction models with different structures are used based on the varying importance of the regions. These prediction models share common model modules, with a greater number of identical model modules corresponding to regions of similar importance. Each region corresponds to a unique content type. Each region corresponds to at least one content event. In this scenario, the prediction model can be a model that predicts the development trend of content events within the region. That is, the main content of the digital twin model of this disclosure, based on the prediction model, for visualizing the environment, is the changing trend of content events under different content types, facilitating emergency adjustments to address trend changes in various content types within the target park. For example, for the infrastructure content type, the corresponding prediction model can be a model predicting infrastructure failures or a model predicting whether there are operational irregularities in the infrastructure. For the park fire safety content type, the corresponding prediction model can be a model predicting the probability of fire incidents. For the application-related content type, the corresponding prediction model can be a model generating application association graphs. For the park security content type, the corresponding prediction model can be a model for security monitoring. In practice, the model structure of the prediction model varies depending on the content type. In other words, there is a corresponding prediction model for one type of infrastructure content. There is another type of prediction model for the park fire safety content type. For example, for infrastructure content, which often involves fault detection and operational procedure detection, a corresponding neural network model structure for fault detection and operational procedure detection would be set up. For instance, fault detection and operational procedure detection are mainly visual detection, and a neural network model based on multiple residual layers as feature extraction layers can be used to achieve image feature extraction and subsequent fault detection or operational procedure detection. As another example, for fire protection facility content, which often involves fire detection, a corresponding neural network model structure for fire detection would be set up. For park security content, which often involves security detection, a corresponding neural network model structure for security detection would be set up. The importance of a region can be the importance of the content events corresponding to that region. The importance of a region can be distinguished by labels or colors. For the same content type, the prediction model structure for content events of different importance levels is different. Furthermore, the prediction models for content events of different importance levels share the same common modules. In practice, different common modules are used depending on the content type. For example, for content events related to infrastructure, the common module could be a residual layer connected in series with predefined layers. The higher the importance of the content event, the more residual layers are included in the common module. Additional modules for park infrastructure detection scenarios, such as attention mechanism layers, can also be added to achieve accurate prediction of content events based on their importance.In practice, the closer two content events are in importance, the more common the model modules they share.

[0021] As an example, the aforementioned implementing entity can use the association table between the content type, importance, and prediction model identifier of the region to determine the prediction model corresponding to each region in the content region distribution map.

[0022] Optionally, the executing entity may determine the prediction model corresponding to each region in the content region distribution map, including the following steps: Step 10121: Perform content region segmentation on the content region distribution map to obtain a content region information set and a content region important tag set.

[0023] Step 10122: For each content region information in the content region information set, perform the first determination steps 101221~101226: Step 101221: Determine the corresponding basic model module based on the important tags of the content area corresponding to the content area information.

[0024] Step 101222: Obtain the regional application scenario corresponding to the content area information.

[0025] Step 101223: Obtain the scene output module corresponding to the regional application scenario from the module repository, and use it as the target scene output module.

[0026] Step 101224: Obtain the historical training dataset corresponding to the content region information.

[0027] Step 101225: Send the target scene output module, the basic model module, and the historical training dataset to the distributed server cluster corresponding to the target park, so as to perform module splicing on the target scene output module and the basic model module and module training on the spliced ​​model obtained by module splicing through the distributed server cluster.

[0028] Step 101226: Combine the model training parameters and the spliced ​​model sent by the distributed server cluster to obtain the prediction model, wherein the spliced ​​model is the module splicing result of the target scene output module and the basic model module.

[0029] In some embodiments, in the visualization environment detection method, the executing entity takes the following steps to generate prediction models for different content areas within the target park. First, the content area distribution map is segmented to obtain a content area information set and a content area importance tag set. The content area information set contains detailed information about each segmented area, such as location and extent; the content area importance tag set indicates the importance of each area, such as importance levels represented by color or labels. Second, for each content area information, a first determination step is performed to determine its corresponding prediction model: based on the content area importance tags corresponding to the content area information, a basic model module matching the importance level of the content area is selected. Areas of different importance levels can use basic model modules of different complexities or types. The specific application scenario of the content area is defined, such as infrastructure monitoring, security monitoring, etc. Operation content: From a predefined module repository, a scenario output module matching the application scenario of the area is selected as the target scenario output module. Relevant historical training data for the content area is collected to form a historical training dataset for subsequent model training. The target scenario output module, basic model module, and historical training dataset are sent to the distributed server cluster corresponding to the target park. The distributed server cluster is responsible for concatenating these modules to form a concatenated model, and training this model using historical training datasets. The distributed server cluster then sends the trained model parameters back to the execution entity. The execution entity combines these model parameters with the concatenated model to form the final prediction model. Here, the concatenated model is the result of concatenating the target scene output module and the base model module. Finally, through the above steps, the execution entity can generate a corresponding prediction model for each region in the content region distribution map. These prediction models not only consider the importance of the region but also incorporate the specific application scenario and historical data of the region, thus ensuring the accuracy and effectiveness of the predictions. Simultaneously, using a distributed server cluster for module concatenation and model training improves processing efficiency and scalability.

[0030] Optionally, such as Figure 2 As shown, the distributed server cluster trains the model through the following steps: Step 201: Obtain the set of spliced ​​models to be trained at the current time. The set of spliced ​​models includes the spliced ​​models obtained by the distributed server cluster splicing the target scene output module and the basic model module.

[0031] Step 202: For each content type, perform the first training steps 2021-2029: Step 2021: Obtain the splicing model subset corresponding to the content type from the splicing model set.

[0032] Step 2022: Select at least one splicing model from the subset of splicing models that represents the most important content area important tags.

[0033] Step 2023: Obtain the historical training dataset of the content region information corresponding to the at least one splicing model, and construct it as the first training dataset.

[0034] Step 2024: Combine the at least one splicing model to obtain an initial combined splicing model.

[0035] Step 2025: Determine the initial model parameters corresponding to the initial combined splicing model.

[0036] Step 2026: Based on the first training dataset, train the initial combined splicing model under the initial model parameters to obtain the combined splicing model.

[0037] Step 2027: Remove at least one splicing model from the splicing model subset to obtain a set of splicing models after removal.

[0038] Step 2028: Generate a subset of prediction models corresponding to the set of spliced ​​models after removal based on the combined splicing model.

[0039] Step 2029: Determine the subset of prediction models corresponding to the combined splicing model and the set of spliced ​​models after removal as the subset of prediction models under the content type.

[0040] Step 203: Determine the obtained subsets of prediction models as the prediction model set.

[0041] In some embodiments, the detailed steps of model training by the distributed server cluster include: First, the distributed server cluster obtains a set of spliced ​​models to be trained at the current time. These spliced ​​models are obtained by splicing the target scene output module and the basic model module previously by the distributed server cluster. Each spliced ​​model represents a model structure under a specific content type and regional application scenario. Second, for each content type, the following first training step is performed to generate a subset of prediction models for that content type: A subset of spliced ​​models matching the current content type is selected from the spliced ​​model set. Further, at least one spliced ​​model representing the most important label for the corresponding content region is selected from the spliced ​​model subset. A historical training dataset of content region information corresponding to the selected at least one spliced ​​model is obtained and constructed as a first training dataset, which will be used for subsequent model training. The selected at least one spliced ​​model is combined to form an initial combined spliced ​​model, which will serve as the basis for subsequent training. Initial model parameters are determined for the initial combined spliced ​​model. Using the first training dataset, the initial combined spliced ​​model with the initial model parameters is trained to obtain the trained combined spliced ​​model. The process begins by removing at least one trained concatenation model from the original concatenation model subset, resulting in a set of concatenation models with the removed models. This set will be used to generate other prediction models. Based on the combined concatenation models, a subset of prediction models corresponding to the set of concatenation models with the removed models is generated. The combined concatenation models and the subset of prediction models corresponding to the set of concatenation models with the removed models are then combined to determine the subset of prediction models for the current content type. Finally, the resulting subsets of prediction models (i.e., the subsets of prediction models for each content type) are combined to determine the final set of prediction models. This set will be used for subsequent visualization environment detection and twin model generation. Through these steps, the distributed server cluster can efficiently complete the model training task and generate sets of prediction models for different content types and regional application scenarios.

[0042] Optionally, the executing entity may determine the initial model parameters corresponding to the initial combined splicing model, including the following steps: Step 20251: Obtain the historical model parameters corresponding to the content region information from the model parameter cache with a communication connection, wherein the historical model parameters are the model parameters of the historical prediction model corresponding to the content region information.

[0043] Step 20252: Determine the set of identical modules between the historical prediction model and the initial combined splicing model.

[0044] Step 20253: Obtain the historical parameters corresponding to each identical module in the same module set from the historical model parameters to obtain the historical parameter set.

[0045] Step 20254: Remove the set of identical modules from the initial combined splicing model to obtain the removed module set.

[0046] Step 20255: For each dismantled module, perform the following second determination steps 202551~202552: Step 202551: Obtain the target module whose content type and module function metadata are similar to the target module in the module repository and whose information similarity probability is higher than the target probability.

[0047] Step 202552: Determine the model parameters corresponding to the target module as the initial sub-model parameters corresponding to the demolition module.

[0048] Step 20256: Construct an initial sub-model parameter set from the initial sub-model parameters corresponding to each demolition module, and generate initial model parameters corresponding to the at least one splicing model based on the initial sub-model parameter set and the historical parameter set.

[0049] In some embodiments, during model training on a distributed server cluster, the initial model parameters of the initial combined splicing model are determined as follows: First, historical model parameters corresponding to the content region information are retrieved from the model parameter cache, which has a communication connection with the execution entity. These parameters are model parameters obtained from the historical prediction model corresponding to the content region information during previous training, reflecting the characteristics and patterns of the content region over time. Second, the structures of the historical prediction model and the initial combined splicing model are compared to determine the set of identical modules between them. This refers to a set of modules that exist in both models and have the same function. These modules have been trained in the historical prediction model, so their parameters have certain reference value. Next, historical parameters corresponding to each identical module in the set of identical modules are extracted from the historical model parameters to form a set of historical parameters. These parameters will be used as the initial parameters of the identical modules in the initial combined splicing model for subsequent model training. Further, the set of identical modules is removed from the initial combined splicing model to obtain the removed module set. The removed module set refers to the set of modules in the initial combined splicing model other than those identical to the historical prediction model. These modules require new initial parameters for training. For each module in the demolition module set, the following sub-steps are performed to determine its initial sub-model parameters: First, from the module repository, based on content type and module functional metadata, search for target modules with a higher probability of similarity to the demolition module than the target probability. Target modules are those functionally similar to the demolition module and potentially applicable to the current content area and scenario. Second, determine the model parameters corresponding to the target modules as the initial sub-model parameters for the demolition modules. These parameters will serve as the initial parameters for the demolition modules in the initial combined splicing model for subsequent model training. Finally, construct an initial sub-model parameter set from the initial sub-model parameters corresponding to each demolition module, and combine it with the historical parameter set to generate at least one initial model parameter set for the splicing model. These parameters will serve as the starting point for the initial combined splicing model in subsequent training. Through continuous adjustment and optimization, a prediction model suitable for the current content area and scenario is ultimately obtained. Through these steps, the executing entity can effectively utilize historical model parameters and resources in the module repository to determine reasonable initial model parameters for the initial combined splicing model, thereby improving the efficiency and accuracy of model training.

[0050] Optionally, the executing entity can generate a subset of prediction models corresponding to the set of models after removal based on the combined splicing model, including the following steps: Step 20281: Based on the important tags of the content area corresponding to each splicing model, divide the importance of each model in the set of splicing models after removal to obtain a model set.

[0051] Step 20282: For each model group in the model group set, perform the following second training steps 202821~202827: Step 202821: Determine the basic model module and target scene output module group corresponding to the model group.

[0052] Step 202822: Select the basic module parameters corresponding to the basic model modules from the combined splicing model.

[0053] Step 202823: Determine the historical scene output module parameter group corresponding to the target scene output module group.

[0054] Step 202824: Combine the basic module parameters and the historical scene output module parameter group to obtain the combined module parameters.

[0055] Step 202825: Use the combined module parameters as the initial model parameters of the basic model module and the target scene output module group to obtain the initial prediction model corresponding to the model group.

[0056] Step 202826: Obtain the second training dataset of the corresponding order of magnitude based on the important labels of the content regions corresponding to the model group.

[0057] Step 202827: Based on the second training dataset, train the initial prediction model to obtain the prediction model corresponding to the model group.

[0058] In some embodiments, during the distributed model training process, the execution entity first analyzes the important labels of the content regions corresponding to each model in the post-concatenation model set and groups the models accordingly. For example, models in high-risk regions are grouped together to ensure that models in key regions receive more computing resources and more refined tuning strategies during subsequent training. For each model group, the execution entity accurately extracts the basic model module parameters from the fully trained combined concatenation models. These parameters have been refined through a large amount of data and possess powerful general feature extraction capabilities. Simultaneously, output module parameters that highly match the target scene of the current model group are selected from the historical model library. Through parameter transfer and combination techniques, an initial prediction model framework with both versatility and scene adaptability is quickly constructed, avoiding the high cost and long cycle of training from scratch. Subsequently, the execution entity dynamically adjusts the scale of training data according to the importance of the model group. High-importance model groups will receive more comprehensive and representative datasets, even including real-time updated dynamic data, to enhance the model's adaptability to the latest scene changes; while low-importance model groups will use simplified datasets to improve training efficiency while ensuring basic performance. Ultimately, through multiple rounds of iterative optimization and hyperparameter tuning, each model group gradually converges to the optimal state, generating a high-precision prediction model subset that corresponds one-to-one with the original spliced ​​model set. These subsets can independently meet the needs of specific scenarios, and can also be combined and spliced ​​to form a more powerful integrated model, providing comprehensive and accurate prediction support for complex business scenarios.

[0059] Step 1013: Using the prediction model set, determine the sequence of visual environment detection distribution maps under the content type corresponding to the content region distribution map.

[0060] In some embodiments, the aforementioned executing entity can utilize the obtained prediction model set to determine a sequence of visual environment detection distribution maps for the content type corresponding to the content region distribution map. The visual environment detection distribution map can be a distribution map of the visual prediction results corresponding to each region in the content region distribution map at the current time. The visual prediction results can be the prediction results output by the prediction model for the corresponding content event. The sequence of visual environment detection distribution maps can characterize the changes in the prediction results of each sub-region at different times. That is, the sequence order of the visual environment detection distribution map sequence is arranged in chronological order. The sequence of visual environment detection distribution maps includes: a visual environment detection distribution map generated based on the current content region distribution map at the current time, and a sub-sequence of visual environment detection distribution maps generated corresponding to historical time periods.

[0061] As an example, for each region in the content region distribution map, the event content corresponding to that region is input into the corresponding prediction model to obtain the prediction result. The corresponding prediction results are then added to each region in the content region distribution map to obtain the visual environment detection distribution map for the current time. Finally, the visual environment detection distribution maps for each time period are sorted chronologically to obtain a sequence of visual environment detection distribution maps.

[0062] Step 1014: Generate the corresponding twin model based on the visualized environment detection distribution map sequence.

[0063] In some embodiments, the aforementioned execution entity can generate a corresponding twin model based on the sequence of distribution maps detected by the visualized environment. The corresponding twin model can be a twin display model of each content event under the content type.

[0064] As an example, the aforementioned execution entity can generate a corresponding twin model based on the visualized environment detection distribution map sequence using the twin model generation method.

[0065] Step 102: In response to receiving twin model display information for at least one content type, overlay and display the corresponding twin model. The displayed twin model supports at least one of the following processing operations: model dragging, region search, model scaling, and visualization of the target location environment.

[0066] In some embodiments, in response to receiving twin model display information for at least one content type, the aforementioned executing entity can overlay and display the corresponding at least one twin model. The displayed twin model supports the following processing operations: model dragging, region search, model zooming, and visualization of the target location's environment.

[0067] In some optional implementations of certain embodiments, after step 102, the steps further include: Step 103: In response to receiving the operation of displaying content for the target location in the displayed twin model, a sub-twin model of the park area corresponding to the target location is displayed with the target location as the area center. The sub-twin model displays at least one content type tag corresponding to the target location. In response to selecting a target content type tag, a brief introduction of the content changes corresponding to the target content type tag is displayed in the sub-twin model.

[0068] Step 104: In response to confirming that the target content type tag has been clicked, the visualized environment detection content for the target location at the current time and the visualized environment content change information within the target time period will pop up in the displayed twin model.

[0069] By applying the technical solution of this embodiment, and utilizing a set of prediction models with a model structure association, accurate and efficient generation of twin models for various content types within the target park can be achieved, along with the display of twin models for diverse display needs, while making reasonable use of server resources. Specifically, the reason why the generation of related twin models is not accurate and efficient is that, when the target park does not require high prediction accuracy for all events, the separate deployment and training of prediction models corresponding to events often incurs significant server resource overhead. Furthermore, when there are many events in the target park and the server's computing resources are limited, scenarios arise where prediction models for events are trained in a queue, leading to delays in event prediction and a long establishment cycle for the corresponding twin models. Based on this, the visualization environment detection method of some embodiments of this disclosure firstly, in response to receiving twin model update information corresponding to the target park, performs the following generation steps for each content type: First, based on the key events within the target time period, the frequency of occurrence of key events, the number of times the digital twin model corresponding to each sub-region is displayed, and the key content tags under the content type, the content area distribution map corresponding to the content type can be accurately determined, thereby determining the distribution of various types of events within the corresponding area of ​​the target park. The first step involves determining the prediction model corresponding to each region in the content region distribution map. Based on the varying importance of the regions, different model structures are used. These prediction models share common model modules, with a greater number of identical modules corresponding to regions closer in importance. By setting different model structures for different importance levels, and ensuring that these models share common model modules and that a greater number of identical modules corresponds to regions closer in importance, a unified prediction model framework can be used for various events within the same content type. Training corresponding models within this framework significantly improves deployment efficiency and reduces computational cost and model structure selection costs. This enables efficient training and deployment of models for various content events within the same content type. The second step utilizes the obtained prediction model set to accurately determine the sequence of visual environment detection distribution maps representing changes in visual environment content at different times within the content region distribution map, facilitating the subsequent generation of accurate twin models. The third step involves accurately generating twin models corresponding to the content type based on the visual environment detection distribution map sequence. Finally, in response to receiving twin model display information for at least one content type, the system overlays and displays the corresponding twin models. The displayed twin models support the following operations: model dragging, region search, model zooming, and visualization of the target location's environment. Here, by setting the dynamic selection and overlay of twin models, the system can display predicted content for events under different content types.In summary, by implementing a common prediction model framework for various events under the same content type, and by training corresponding models for the same prediction model framework, the deployment efficiency of prediction models can be greatly improved, the computational load of model training and the selection cost of model structure can be reduced, and the scenario of queuing for training prediction models for events can be avoided, which could lead to delays in the prediction of corresponding events and a long establishment cycle for the corresponding twin models.

[0070] Furthermore, as Figure 1 To specifically implement the method, this application provides a visual environment detection device, such as... Figure 3 As shown, the device includes: An execution unit is configured to respond to receiving twin model update information corresponding to the target park, and for each content type, execute a generation step to obtain a twin model corresponding to each content type; wherein, the generation step includes: generating a content area distribution map corresponding to the content type based on key events within the target time period, the frequency of occurrence of key events, the number of times the digital twin model corresponding to each sub-region is displayed, and key content tags under the content type; generating a prediction model corresponding to each region based on the different importance levels of each region in the content area distribution map, forming a prediction model set, wherein different regions use prediction models with different model structures, and the prediction models share model modules, with the number of identical model modules corresponding to the closer the importance levels are; using the prediction model set, determining the sequence of visual environment detection distribution maps under the content type corresponding to the content area distribution map; and generating the corresponding twin model based on the sequence of visual environment detection distribution maps; The display unit is used to respond to receiving twin model display information for at least one content type, and to overlay and display the corresponding twin model. The displayed twin model supports at least one of the following processing operations: model dragging, region search, model scaling, and visualization of the target location environment.

[0071] Optionally, the display unit is further configured to: In response to receiving a content display operation for a target location in the displayed twin model, a sub-twin model of the park area corresponding to the target location is displayed with the target location as the area center. The sub-twin model displays at least one content type tag corresponding to the target location. In response to selecting a target content type tag, a brief introduction of the content changes corresponding to the target content type tag is displayed in the sub-twin model. In response to the confirmation of clicking the target content type tag, a pop-up appears in the displayed twin model showing the visual environment detection content for the target location at the current time and the visual environment content change information within the target time period.

[0072] Optionally, the execution unit is specifically used for: The content area distribution map is displayed on the target canvas page, and the content area distribution map is further divided using the various canvas processing operations supported by the target canvas page to obtain a re-divided distribution map. The re-division distribution map is determined as a content area distribution map.

[0073] Optionally, the execution unit is specifically used for: The content region distribution map is segmented to obtain a content region information set and a content region important tag set; For each content region information in the content region information set, a first determination step is performed to determine the prediction model, wherein the first determination step includes: Based on the important tags of the content area corresponding to the content area information, determine the corresponding basic model module; Obtain the regional application scenario corresponding to the content region information; Retrieve the scene output module corresponding to the regional application scenario from the module repository, and use it as the target scene output module; Obtain the historical training dataset corresponding to the content region information; The target scene output module, the basic model module, and the historical training dataset are sent to the distributed server cluster corresponding to the target park, so that the target scene output module and the basic model module can be spliced ​​together through the distributed server cluster, and the spliced ​​model obtained after module splicing can be trained. The model training parameters sent by the distributed server cluster and the spliced ​​model are combined to obtain the prediction model, wherein the spliced ​​model is the module splicing result of the target scene output module and the basic model module.

[0074] Optionally, the execution unit is specifically used for: Obtain the set of spliced ​​models to be trained at the current time. The set of spliced ​​models includes the spliced ​​models obtained by the distributed server cluster splicing the target scene output module and the basic model module. For each content type, perform the first training step to determine a subset of prediction models for each content type; Each subset of the obtained prediction models is determined as the prediction model set; The first training step includes: Obtain a subset of splicing models corresponding to the content type from the splicing model set; From the subset of splicing models, at least one splicing model that represents the most important important tags in the corresponding content area is selected; Obtain the historical training dataset of the content region information corresponding to the at least one splicing model, and construct it as the first training dataset; The at least one splicing model is combined to obtain an initial combined splicing model; Determine the initial model parameters corresponding to the initial combined splicing model; Based on the first training dataset, the initial combined splicing model under the initial model parameters is trained to obtain the combined splicing model; Remove at least one spliced ​​model from the spliced ​​model subset to obtain a set of spliced ​​models after removal; Based on the combined splicing model, a subset of prediction models corresponding to the set of spliced ​​models after removal is generated; The subset of prediction models corresponding to the combined splicing model and the set of spliced ​​models after removal are determined as the subset of prediction models under the content type.

[0075] Optionally, the execution unit is specifically used for: Obtain the historical model parameters corresponding to the content region information from the model parameter cache where there is a communication connection, wherein the historical model parameters are the model parameters of the historical prediction model corresponding to the content region information; Determine the set of identical modules between the historical prediction model and the initial combined splicing model; The historical parameters corresponding to each identical module in the same module set are obtained from the historical model parameters to obtain the historical parameter set; The identical module set is removed from the initial combined splicing model to obtain the removed module set; For each dismantling module, a second determination step is performed to determine the initial sub-model parameters corresponding to each dismantling module; The initial sub-model parameters corresponding to each demolition module are constructed into an initial sub-model parameter set. Based on the initial sub-model parameter set and the historical parameter set, the initial model parameters corresponding to the at least one splicing model are generated. The second determining step includes: Obtain a target module whose content type and module function metadata are similar to the demolition module with a higher probability than the target probability from the module repository; The model parameters corresponding to the target module are determined as the initial sub-model parameters corresponding to the demolition module.

[0076] Optionally, the execution unit is specifically used for: Based on the important tags of the content area corresponding to each splicing model, the importance of each model in the set of splicing models after removal is divided to obtain a model set; For each model group in the model group set, the following second training step is performed to determine the prediction model corresponding to each model group: Determine the basic model module and the target scene output module group corresponding to the model group; The basic module parameters corresponding to the basic model modules are selected from the combined splicing model; Determine the historical scene output module parameter group corresponding to the target scene output module group; The basic module parameters and the historical scene output module parameter group are combined accordingly to obtain the combined module parameters; The combined module parameters are used as the initial model parameters for the basic model module and the target scene output module group to obtain the initial prediction model corresponding to the model group. Based on the important labels of the content regions corresponding to the model group, obtain a second training dataset of the same order of magnitude as the important labels of the content regions corresponding to the model group. Based on the second training dataset, the initial prediction model is trained to obtain the prediction model corresponding to the model group.

[0077] It should be noted that other corresponding descriptions of the functional units involved in the visual environment detection device provided in this application embodiment can be found in the following references. Figures 1 to 2 The corresponding descriptions in the method will not be repeated here.

[0078] This application also provides a computer device, which may specifically be a personal computer, a server, a network device, etc. Figure 4 As shown, the computer device includes a bus, a processor, memory, and a communication interface, and may also include an input / output interface and a display device. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores location information. The network interface allows communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the various method embodiments.

[0079] Those skilled in the art will understand that Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0080] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, having stored thereon a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0081] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0082] It should be noted that the user personal information involved in the embodiments of this application is all authorized (with the knowledge and consent) by the relevant parties or fully authorized by all parties, and the executing entity can obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with the relevant laws and regulations of the relevant countries and regions, and do not violate public order and good morals. It should be noted that if any software tools or components other than those of this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use.

[0083] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0084] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0085] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A visual environment detection method, characterized in that, The method includes: In response to receiving the twin model update information corresponding to the target park, for each content type, a generation step is performed to obtain the twin model corresponding to each content type; In response to receiving twin model display information for at least one content type, the corresponding twin model is overlaid and displayed. The displayed twin model supports at least one of the following processing operations: model dragging, region search, model scaling, and visualization of the target location environment. The generation step includes: Based on the key events within the target time period, the frequency of occurrence of key events, the number of times the digital twin model corresponding to each sub-region is displayed, and the key content tags under the content type, a content area distribution map corresponding to the content type is generated. Based on the different importance of each region in the content region distribution map, a prediction model is generated for each region to form a prediction model set. Different regions use prediction models with different model structures, and each prediction model shares a model module. The closer the importance is, the more common the number of the same model modules. Using the prediction model set, determine the sequence of visual environment detection distribution maps under the content type corresponding to the content region distribution map; Based on the sequence of visual environment detection distribution maps, a corresponding twin model is generated.

2. The method according to claim 1, characterized in that, The method further includes: In response to receiving a content display operation for a target location in the displayed twin model, a sub-twin model of the park area corresponding to the target location is displayed with the target location as the area center. The sub-twin model displays at least one content type tag corresponding to the target location. In response to selecting a target content type tag, a brief introduction of the content changes corresponding to the target content type tag is displayed in the sub-twin model. In response to the confirmation of clicking the target content type tag, a pop-up appears in the displayed twin model showing the visual environment detection content for the target location at the current time and the visual environment content change information within the target time period.

3. The method according to claim 1, characterized in that, After determining the content area distribution map corresponding to the content type based on key events within the target time period, the frequency of occurrence of key events, the number of times the digital twin model corresponding to each sub-region is displayed, and the key content tags under the content type, the method further includes: The content area distribution map is displayed on the target canvas page, and the content area distribution map is further divided using the various canvas processing operations supported by the target canvas page to obtain a re-divided distribution map. The re-division distribution map is determined as a content area distribution map.

4. The method according to claim 1, characterized in that, The step of determining the prediction model corresponding to each region in the content region distribution map includes: The content region distribution map is segmented to obtain a content region information set and a content region important tag set; For each content region information in the content region information set, a first determination step is performed to determine the prediction model, wherein the first determination step includes: Based on the important tags of the content area corresponding to the content area information, determine the corresponding basic model module; Obtain the regional application scenario corresponding to the content region information; Retrieve the scene output module corresponding to the regional application scenario from the module repository, and use it as the target scene output module; Obtain the historical training dataset corresponding to the content region information; The target scene output module, the basic model module, and the historical training dataset are sent to the distributed server cluster corresponding to the target park, so that the target scene output module and the basic model module can be spliced ​​together through the distributed server cluster, and the spliced ​​model obtained after module splicing can be trained. The model training parameters sent by the distributed server cluster and the spliced ​​model are combined to obtain the prediction model, wherein the spliced ​​model is the module splicing result of the target scene output module and the basic model module.

5. The method according to claim 4, characterized in that, The distributed server cluster trains the model through the following steps: Obtain the set of spliced ​​models to be trained at the current time. The set of spliced ​​models includes the spliced ​​models obtained by the distributed server cluster splicing the target scene output module and the basic model module. For each content type, perform the first training step to determine a subset of prediction models for each content type; Each subset of the obtained prediction models is determined as the prediction model set; The first training step includes: Obtain a subset of splicing models corresponding to the content type from the splicing model set; From the subset of splicing models, at least one splicing model that represents the most important important tags in the corresponding content area is selected; Obtain the historical training dataset of the content region information corresponding to the at least one splicing model, and construct it as the first training dataset; The at least one splicing model is combined to obtain an initial combined splicing model; Determine the initial model parameters corresponding to the initial combined splicing model; Based on the first training dataset, the initial combined splicing model under the initial model parameters is trained to obtain the combined splicing model; Remove at least one spliced ​​model from the spliced ​​model subset to obtain a set of spliced ​​models after removal; Based on the combined splicing model, a subset of prediction models corresponding to the set of spliced ​​models after removal is generated; The subset of prediction models corresponding to the combined splicing model and the set of spliced ​​models after removal are determined as the subset of prediction models under the content type.

6. The method according to claim 5, characterized in that, The determination of the initial model parameters corresponding to the initial combined splicing model includes: Obtain the historical model parameters corresponding to the content region information from the model parameter cache where there is a communication connection, wherein the historical model parameters are the model parameters of the historical prediction model corresponding to the content region information; Determine the set of identical modules between the historical prediction model and the initial combined splicing model; The historical parameters corresponding to each identical module in the same module set are obtained from the historical model parameters to obtain the historical parameter set; The identical module set is removed from the initial combined splicing model to obtain the removed module set; For each dismantling module, a second determination step is performed to determine the initial sub-model parameters corresponding to each dismantling module; The initial sub-model parameters corresponding to each demolition module are constructed into an initial sub-model parameter set. Based on the initial sub-model parameter set and the historical parameter set, the initial model parameters corresponding to the at least one splicing model are generated. The second determining step includes: Obtain a target module whose content type and module function metadata are similar to the demolition module with a higher probability than the target probability from the module repository; The model parameters corresponding to the target module are determined as the initial sub-model parameters corresponding to the demolition module.

7. The method according to claim 5, characterized in that, The step of generating a subset of prediction models corresponding to the set of models after removal from the combined splicing model includes: Based on the important tags of the content area corresponding to each splicing model, the importance of each model in the set of splicing models after removal is divided to obtain a model set; For each model group in the model group set, the following second training step is performed to determine the prediction model corresponding to each model group: Determine the basic model module and the target scene output module group corresponding to the model group; The basic module parameters corresponding to the basic model modules are selected from the combined splicing model; Determine the historical scene output module parameter group corresponding to the target scene output module group; The basic module parameters and the historical scene output module parameter group are combined accordingly to obtain the combined module parameters; The combined module parameters are used as the initial model parameters for the basic model module and the target scene output module group to obtain the initial prediction model corresponding to the model group. Based on the important labels of the content regions corresponding to the model group, obtain a second training dataset of the same order of magnitude as the important labels of the content regions corresponding to the model group. Based on the second training dataset, the initial prediction model is trained to obtain the prediction model corresponding to the model group.

8. A visual environmental monitoring device, characterized in that, The device includes: An execution unit is configured to respond to receiving twin model update information corresponding to the target park, and for each content type, execute a generation step to obtain a twin model corresponding to each content type; wherein, the generation step includes: generating a content area distribution map corresponding to the content type based on key events within the target time period, the frequency of occurrence of key events, the number of times the digital twin model corresponding to each sub-region is displayed, and key content tags under the content type; generating a prediction model corresponding to each region based on the different importance levels of each region in the content area distribution map, forming a prediction model set, wherein different regions use prediction models with different model structures, and the prediction models share model modules, with the number of identical model modules corresponding to the closer the importance levels are; using the prediction model set, determining the sequence of visual environment detection distribution maps under the content type corresponding to the content area distribution map; and generating the corresponding twin model based on the sequence of visual environment detection distribution maps; The display unit is used to respond to receiving twin model display information for at least one content type, and to overlay and display the corresponding twin model. The displayed twin model supports at least one of the following processing operations: model dragging, region search, model scaling, and visualization of the target location environment.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

10. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

Images