Power plant water island precision processing intelligent management system based on visual large model

Abstract

The invention relates to the field of power plant intelligent monitoring, and discloses a power plant water island precision processing intelligent management system based on a visual large model, which comprises an equipment sensing layer, a data layer, an AI algorithm application layer and a display layer, wherein the equipment sensing layer acquires a target image through power plant monitoring equipment and performs data acquisition, the data layer is used for storing the acquired data and providing computing resources, the AI algorithm application layer processes and analyzes image data by using a visual large model, and the display layer is used for displaying a processing result and providing a user interaction interface. Through a visual large model zero sample (Zero-shot) detection technology of an AI algorithm application layer, a pre-training model is utilized to learn cross-scene universal visual semantic features, a natural language Prompt cue word is combined to activate detection capability, environment differences of different power plants can be adapted without re-training, the problem that cross-environment migration of a traditional algorithm needs to be repeatedly trained is solved, and the method is suitable for being applied to various power plants. And the deployment cost and period are greatly reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring of power plants, in particular to a power plant water island fine treatment intelligent management system based on a visual large model. BACKGROUND

[0002] With the application of artificial intelligence technology in the field of power plants, water island fine treatment detection schemes based on visual models have gradually become popular. However, existing detection technologies based on deep neural networks face the core problem of high algorithm migration cost when deployed across power plant environments: the significant differences in equipment layout, lighting conditions, and other environmental differences between different power plants result in the need for the algorithm to collect a new labeled data set and perform training for the new environment, causing the deployment cycle to be extended and computational resources to be wasted. In addition, in traditional schemes, multiple brands of monitoring devices need to be configured separately, and the non-uniform data format leads to poor compatibility, the manual operation interface is scattered and lacks real-time decision support, making it difficult to meet the needs of safe production in power plants for efficient detection and rapid response. SUMMARY

[0003] In view of the deficiencies of the prior art, the present application provides a power plant water island fine treatment intelligent management system based on a visual large model, which solves the problems mentioned in the background.

[0004] To achieve the above purpose, the present application realizes the following technical scheme: a power plant water island fine treatment intelligent management system based on a visual large model, comprising: a device perception layer, a data layer, an AI algorithm application layer, and a display layer; The device perception layer obtains target images through power plant monitoring devices and performs data collection, the data layer is used to store collected data and provide computing resources, the AI algorithm application layer uses a visual large model to process and analyze image data, and the display layer is used to display processing results and provide a user interaction interface.

[0005] Preferably, the device perception layer interacts with existing monitoring devices through a standardized interface, automatically extracts and converts image data collected by different monitoring devices, and updates the monitoring state of the target area in real time, thereby achieving data integration of multiple devices, wherein the process of automatically extracting and converting image data collected by different monitoring devices is as follows: Access different brands and models of monitoring devices to extract raw image data; Convert non-standard format image data to a standard format through format conversion; Use median filtering and image enhancement algorithms to denoise and optimize the data; Real-time transmission of preprocessed image data to the data layer for storage and updating of the target area monitoring state.

[0006] Preferably, the AI algorithm application layer analyzes the collected image data through a visual large model algorithm and generates control instructions based on a predetermined configuration file, specifically including: Receiving image data from the device perception layer and selecting applicable visual large model algorithms based on a predetermined configuration file; Using the selected visual large model algorithm to extract features from the received image data, identify key features of the target area and potential problems; According to the results of feature extraction, combined with the control rules in the configuration file, generate control instructions, including but not limited to adjusting the operating parameters or monitoring mode of the power plant water island fine treatment system; The generated control instructions are fed back to the system control layer to adjust the working state of the power plant equipment in real time.

[0007] Preferably, the process of selecting applicable visual large model algorithms based on a predetermined configuration file is as follows: Extract the operating parameters, monitoring task requirements and environmental condition information of the power plant water island fine treatment system from the configuration file; According to the extracted information, select visual large model algorithms that match the current monitoring task; Configure the corresponding input parameters according to the selected algorithm, including but not limited to image data resolution, processing speed requirements, to ensure that the visual large model algorithm can adapt to the actual application requirements; The control rules in the configuration file include: Set the operating parameters of the power plant water island fine treatment system, including processing efficiency, detection accuracy, alarm threshold; Define the relationship between image analysis results and system response, including automatically adjusting the device operating mode or starting the warning mechanism when detecting abnormal targets.

[0008] Preferably, the data layer provides computing power required for algorithm operation through efficient storage devices and computing resources, wherein the computing resource configuration includes: The data layer configures at least one GPU server supporting parallel computing, each server is configured with at least 8 GPUs, and each GPU has a memory of not less than 24GB to ensure efficient processing of large-scale image data; To support fast transmission and processing of large-scale image data, the data layer uses high-speed network switching equipment to connect with GPU servers to ensure low-latency data transmission between different modules; The data layer dynamically allocates computing resources according to the requirements of processing tasks, reasonably allocates GPU resources through load balancing algorithms to avoid resource overload and ensure efficient use of computing power; High-speed SSD hard drives are used in the storage devices of the data layer to ensure fast reading and writing of data and meet the storage requirements of large-scale image data.

[0009] Preferably, the AI algorithm application layer solves the migration problem of the system in different power plant environments based on the Zero-shot zero sample capability of the visual large model, specifically including: The visual large model algorithm learns a large amount of image data in different power plant environments through a pre-training process to obtain rich feature representation capabilities; In actual application, the AI algorithm application layer uses shared features in the pre-trained model for inference through a zero-shot learning method, automatically adapting to new power plant environments without the need for re-labeling and training new data sets; When the system is deployed to a new power plant, the AI algorithm application layer adjusts the environmental parameters in the configuration file, enabling the pre-trained visual large model to recognize specific features in the new environment, ensuring that the model can quickly adapt and perform accurate image processing and target detection; To ensure the adaptability of the model, an online learning mechanism is used to dynamically adjust the inference strategy of the model based on real-time collected data, enabling the system to operate stably in various power plant environments.

[0010] Preferably, the hierarchical architecture design of the system includes: The device perception layer realizes hardware decoupling through an abstract interface, supporting plug-and-play access to monitoring devices of different brands and protocols; The AI algorithm application layer adopts a plug-in architecture, supporting dynamic loading / replacement of visual large model algorithms without modifying the underlying system; The data layer and the display layer interact through a unified data protocol, ensuring the compatibility and stability of cross-layer data transmission.

[0011] Preferably, the Zero-shot zero sample capability includes: The visual large model learns cross-scene general visual semantic features during the pre-training phase, forming a generalization capability; During deployment, the model's corresponding detection capability is activated through natural language Prompt prompt words, without the need for specific environment training data; Based on the built-in scene understanding capability of the model, it automatically adapts to environmental differences such as device layout and lighting conditions in different power plants.

[0012] Preferably, the display layer further includes integrated functions, specifically as follows: Detection result visualization module: real-time video stream with target labeling box; Data management module: count target frequency by time / place, generate interactive trend charts; Algorithm configuration module: support user-defined detection area ROI, sampling rate, alarm threshold; Historical data module: provides video segment retrieval, playback and detection record export functions.

[0013] Preferably, the visual large model algorithm carried by the AI ​​algorithm application layer adopts the following detection method: After feature extraction of the input image data, the Transformer architecture is used to achieve visual-language cross-modal semantic alignment; Use the attention mechanism to focus on the target area and output structured detection results including location coordinates and category labels; It supports multi-target parallel detection, and a single inference can identify various scenarios such as hot work and equipment anomalies.

[0014] The present invention provides an intelligent management system for water island polishing in power plants based on a large visual model. It has the following beneficial effects: 1. This invention uses the zero-shot detection technology of a large-scale visual model at the AI ​​algorithm application layer, utilizes a pre-trained model to learn common visual semantic features across scenarios, and combines it with natural language prompt word activation detection capabilities. This allows it to adapt to the environmental differences of different power plants without retraining, solving the problem of traditional algorithms requiring repeated training for cross-environment migration and significantly reducing deployment costs and cycles.

[0015] 2. The present invention supports plug-and-play access to monitoring devices of different brands and protocols through the standardized interface and abstract architecture design of the device perception layer, automatically completes video stream format conversion and image preprocessing, realizes unified collection and real-time integration of multi-brand device data, and avoids the problems of complex device management and poor compatibility in traditional solutions.

[0016] 3. Through the integrated functional design of the display layer, the present invention integrates functions such as detection result visualization, data statistics, and parameter configuration into a unified interface. Operators can dynamically adjust detection strategies and obtain analysis results in real time, solving the problems of fragmented interfaces and cumbersome operations in traditional systems and improving decision-making efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of the present invention; Figure 2 This is a workflow diagram of the AI ​​algorithm application layer of the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solution of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention. Example

[0019] See the attached Figure 1 -attached Figure 2 The embodiment of the present application provides a power plant water island fine treatment intelligent management system based on a visual large model, comprising: a device perception layer, a data layer, an AI algorithm application layer and a display layer. Among them, the device perception layer obtains target images through power plant monitoring equipment and carries out data collection, the data layer is used for storing collected data and providing computing resources, the AI algorithm application layer uses a visual large model to process and analyze image data, and the display layer is used for displaying processing results and providing a user interaction interface.

[0020] Specifically, the system adopts a hierarchical architecture design, the device perception layer collects target image data through existing monitoring equipment (such as visible light / thermal imaging camera) of the power plant, and transmits the data to the data layer for storage and providing computing resources; the AI algorithm application layer loads a visual large model to process and analyze the data, and generates detection results; the display layer is responsible for visualizing the results and providing an interactive interface. Each layer works cooperatively through a data link to form a complete closed loop from data collection to decision output.

[0021] Through hierarchical architecture, functional decoupling is realized, the device perception layer reuses existing hardware of the power plant to reduce deployment cost, the data layer centrally manages computing resources to improve processing efficiency, the AI algorithm application layer uses the semantic understanding ability of the visual large model to realize accurate detection, and the interactive design of the display layer simplifies the operation process, so that an integrated intelligent management system is finally formed to meet the automatic monitoring needs of the power plant water island fine treatment scene.

[0022] The device perception layer interacts with existing monitoring equipment through a standardized interface, automatically extracts and converts image data collected by different monitoring equipment, and updates the monitoring state of the target area in real time, so as to realize data integration of multiple devices, wherein the process of automatically extracting and converting image data collected by different monitoring equipment is as follows: Access different brands and models of monitoring equipment to extract original image data; Convert non-standard format image data to a standard format through format conversion; Adopt median filter and image enhancement algorithm to denoise and optimize the data; Real-time transmission of preprocessed image data to the data layer for storage, and updating the monitoring state of the target area.

[0023] Specifically, the device perception layer adapts to different brands and models of monitoring equipment through a standardized interface, extracts original image data, and then uniformly formats and pre-processes the data, finally transmits high-quality data to the data layer. This process realizes the standardized integration of multi-source data through protocol analysis and algorithm processing.

[0024] The AI algorithm application layer analyzes the collected image data through a visual large model algorithm and generates control instructions based on a predetermined configuration file. Specifically, it includes: Receives image data from the device perception layer and selects applicable visual large model algorithms based on a predetermined configuration file; Uses the selected visual large model algorithm to extract features from the received image data, identifying key features and potential problems in the target area; Based on the results of feature extraction, combined with the control rules in the configuration file, generate control instructions, including but not limited to adjusting the operating parameters or monitoring mode of the power plant water island fine treatment system; The generated control instructions are fed back to the system control layer to adjust the working state of the power plant equipment in real time.

[0025] Specifically, the dynamic algorithm selection mechanism based on the configuration file enables the system to adapt to different detection tasks (such as hot work and equipment anomalies); the feature extraction capability of the visual large model can accurately capture key information in complex scenarios, improving detection accuracy; the real-time control instruction generation and feedback mechanism realizes an automated closed loop from image analysis to device regulation, shortens the response time to abnormalities, and improves the safety of power plant production.

[0026] The process of selecting applicable visual large model algorithms based on a predetermined configuration file is as follows: Extract the operating parameters, monitoring task requirements, and environmental condition information of the power plant water island fine treatment system from the configuration file; Based on the extracted information, select a visual large model algorithm that matches the current monitoring task; Configure the corresponding input parameters based on the selected algorithm, including but not limited to image data resolution, processing speed requirements, to ensure that the visual large model algorithm can adapt to actual application requirements; The control rules in the configuration file include: Set the operating parameters of the power plant water island fine treatment system, including processing efficiency, detection accuracy, and alarm thresholds; Define the relationship between image analysis results and system responses, including automatically adjusting device operating mode or starting the warning mechanism when detecting abnormal targets.

[0027] Specifically, the parameter-driven algorithm selection and configuration mechanism avoids the limitations of traditional "one-size-fits-all" algorithms, improving the matching degree of models and scenarios; standardized control rule design enables detection results to be directly converted into executable production control strategies, reducing manual intervention, improving decision-making efficiency and consistency, and adapting to different power plant safety management needs through dynamic threshold setting.

[0028] The data layer provides the computing power required for algorithm operation through efficient storage devices and computing resources. The computing resource configuration includes: The data layer is equipped with at least one GPU server that supports parallel computing. A single server is equipped with at least 8 GPUs, and each GPU has a memory of at least 24GB to ensure efficient processing of large-scale image data. To support the rapid transmission and processing of large-scale image data, the data layer uses high-speed network switching equipment to connect to the GPU server, ensuring low-latency data transmission between different modules; The data layer dynamically allocates computing resources based on the needs of processing tasks and reasonably allocates GPU resources through load balancing algorithms to avoid resource overload and ensure efficient use of computing power; High-speed SSD hard drives are used in the data layer storage devices to ensure fast data reading and writing and meet the storage needs of large-scale image data.

[0029] Specifically, the parallel computing architecture significantly improves image processing speed, meeting the stringent requirements of real-time monitoring for computing efficiency; high-speed storage and transmission technology reduces data latency, ensuring the timeliness of detection results; the dynamic resource allocation mechanism avoids hardware overload, improves system stability and resource utilization, and provides underlying computing power guarantees for the efficient operation of large visual models.

[0030] The AI ​​algorithm application layer, based on the zero-shot capability of the large visual model, solves the system migration problem in different power plant environments. Specifically, it includes: The large visual model algorithm learns a large amount of image data from different power plant environments through a pre-training process to obtain rich feature representation capabilities; In actual application, the AI ​​algorithm application layer uses zero-shot learning methods to reason using shared features in pre-trained models, automatically adapting to new power plant environments without the need to re-label or train new datasets. When the system is deployed to a new power plant, the AI ​​algorithm application layer adjusts the environmental parameters in the configuration file so that the pre-trained visual model can recognize specific features in the new environment, ensuring that the model can quickly adapt and perform accurate image processing and object detection. To ensure the adaptability of the model, the model's reasoning strategy is dynamically adjusted according to real-time collected data through an online learning mechanism, so that the system can operate stably in a variety of power plant environments.

[0031] Specifically, the zero-shot detection technology completely breaks away from the dependence of traditional algorithms on training data in specific environments, and can adapt to new power plant scenarios without re-labeling or training, shortening the cross-environment deployment cycle from "several weeks" to "minute-level parameter adjustment". The online learning mechanism enables the model to continuously absorb new data to optimize the detection strategy, improve long-term generalization ability, and reduce maintenance costs and technical barriers.

[0032] The hierarchical architecture design of the system includes: The device perception layer realizes hardware decoupling through an abstract interface, supporting plug-and-play access to monitoring devices of different brands and protocols; The AI algorithm application layer adopts a plug-in architecture, supporting dynamic loading / replacement of visual large model algorithms without modifying the underlying system; The data layer and the display layer interact through a unified data protocol, ensuring the compatibility and stability of cross-layer data transmission.

[0033] Specifically, the hardware decoupling design reduces the complexity of device access and improves system scalability; the plug-in algorithm management mode supports "hot swapping" model updates, allowing the system to adapt to algorithm iteration or task changes without modifying the underlying architecture; the cross-layer data protocol ensures compatibility, avoiding integration issues caused by interface differences, and enhancing system stability and maintainability.

[0034] The Zero-shot zero-shot capability includes: The visual large model learns cross-scene general visual semantic features during the pre-training phase, forming a generalization capability; At deployment, the model's corresponding detection capability is activated through natural language Prompt prompts, without the need for specific environment training data; Based on the built-in scene understanding capability of the model, it automatically adapts to environmental differences such as device layout and lighting conditions in different power plants.

[0035] Specifically, the pre-training + Prompt mechanism enables "one training, multiple scene applications", significantly reducing algorithm development and deployment costs; the semantic-driven detection mode enables the model to understand complex detection requirements (such as multi-target identification), improving detection flexibility and accuracy, and is particularly suitable for application scenarios with diverse power plant equipment and complex environments.

[0036] The display layer further includes integrated functions, specifically as follows: Detection result visualization module: real-time video stream superimposed with target bounding boxes; Data management module: count target frequency by time / place, generate interactive trend charts; Algorithm configuration module: support user-defined detection area ROI, sampling rate, and alarm threshold; Historical data module: provides video segmentation retrieval, playback, and detection record export functions.

[0037] Specifically, the visual mark and statistical chart help the operator to quickly grasp the real-time and historical detection situation, reduce the information understanding cost; the parameter configuration module gives the user the ability to flexibly adjust the detection strategy, adapts to different monitoring priorities; the historical data backtracking function provides data support for safety accident analysis and production process optimization, and overall improves the system usability and management efficiency.

[0038] The visual large model algorithm applied by the AI algorithm application layer adopts the following detection method: After the input image data is extracted, the visual-linguistic cross-modal semantic alignment is realized through the Transformer architecture; The attention mechanism is used to focus on the target area, and the structured detection result containing the position coordinates and the category label is outputted; Supporting multi-target parallel detection, a single inference can identify multiple scenes such as hot work operation and equipment abnormality.

[0039] Specifically, the cross-modal semantic alignment technology enhances the model's understanding ability of complex scenes, the attention mechanism improves the target positioning accuracy; the multi-target parallel detection capability reduces repeated calculation, improves the detection efficiency, so that the system can monitor multiple safety hazards at the same time, adapt to the multi-dimensional safety management needs of the power plant, and reduce the risk of missed detection.

[0040] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. The intelligent management system for water island polishing in power plants based on visual large-scale models is characterized by: include: Device perception layer, data layer, AI algorithm application layer, and display layer; Among them, the device perception layer obtains the target image and collects data through the power plant monitoring equipment. The data layer is used to store the collected data and provide computing resources. The AI ​​algorithm application layer uses the visual big model to process and analyze the image data. The display layer is used to display the processing results and provide a user interaction interface.

2. The intelligent management system for water island polishing in power plants based on visual large-scale model according to claim 1 is characterized in that: The device perception layer interacts with existing monitoring devices through standardized interfaces, automatically extracts and converts image data collected by different monitoring devices, and updates the monitoring status of the target area in real time, thereby achieving data integration of multiple devices. The process of automatically extracting and converting image data collected by different monitoring devices is as follows: Connect to monitoring devices of different brands and models to extract original image data; Unify non-standard format image data into standard format through format conversion; Median filtering and image enhancement algorithms are used to denoise and optimize the data; The pre-processed image data is transmitted to the data layer storage in real time, and the monitoring status of the target area is updated.

3. The intelligent management system for water island polishing in power plants based on visual large-scale model according to claim 1 is characterized in that: The AI ​​algorithm application layer analyzes the collected image data through the visual large model algorithm and generates control instructions according to the predetermined configuration file, specifically including: Receive image data from the device perception layer and select the applicable visual model algorithm based on the predetermined configuration file; Use the selected visual big model algorithm to extract features from the received image data and identify key features and potential problems in the target area; Based on the feature extraction results and in combination with the control rules in the configuration file, control instructions are generated, including but not limited to adjusting the operating parameters or monitoring mode of the water island polishing system of the power plant; The generated control instructions are fed back to the system control layer to adjust the working status of the power plant equipment in real time.

4. The intelligent management system for water island polishing in power plants based on visual large-scale model according to claim 3 is characterized in that: The process of selecting an applicable visual large model algorithm according to a predetermined configuration file is as follows: Extract the operating parameters, monitoring task requirements and environmental condition information of the power plant water island polishing system from the configuration file; Based on the extracted information, select the visual large model algorithm that matches the current monitoring task; Configure the corresponding input parameters according to the selected algorithm, including but not limited to the resolution and processing speed requirements of the image data, to ensure that the large-scale visual model algorithm can adapt to actual application needs; The control rules in the configuration file include: Set the operating parameters of the power plant water island polishing system, including treatment efficiency, detection accuracy, and alarm thresholds; Define the relationship between image analysis results and system responses, including automatically adjusting equipment operating modes or initiating early warning mechanisms when abnormal targets are detected.

5. The intelligent management system for water island polishing in power plants based on visual large-scale model according to claim 1 is characterized in that: The data layer provides the computing power required for algorithm operation through efficient storage devices and computing resources, where the computing resource configuration includes: The data layer is equipped with at least one GPU server that supports parallel computing. A single server is equipped with at least 8 GPUs, and each GPU has a memory of at least 24GB to ensure efficient processing of large-scale image data. To support the rapid transmission and processing of large-scale image data, the data layer uses high-speed network switching equipment to connect to the GPU server, ensuring low-latency data transmission between different modules; The data layer dynamically allocates computing resources based on the needs of processing tasks and reasonably allocates GPU resources through load balancing algorithms to avoid resource overload and ensure efficient use of computing power; High-speed SSD hard drives are used in the data layer storage devices to ensure fast data reading and writing and meet the storage needs of large-scale image data.

6. The intelligent management system for water island polishing in power plants based on visual large-scale model according to claim 1 is characterized in that: The AI ​​algorithm application layer, based on the zero-shot capability of the large visual model, solves the system migration problem in different power plant environments. Specifically, it includes: The large visual model algorithm learns a large amount of image data from different power plant environments through a pre-training process to obtain rich feature representation capabilities; In actual application, the AI ​​algorithm application layer uses zero-shot learning methods to reason using shared features in pre-trained models, automatically adapting to new power plant environments without the need to re-label or train new datasets. When the system is deployed to a new power plant, the AI ​​algorithm application layer adjusts the environmental parameters in the configuration file so that the pre-trained visual model can recognize specific features in the new environment, ensuring that the model can quickly adapt and perform accurate image processing and object detection. To ensure the adaptability of the model, the model's reasoning strategy is dynamically adjusted according to real-time collected data through an online learning mechanism, so that the system can operate stably in a variety of power plant environments.

7. The intelligent management system for water island polishing in power plants based on visual large-scale model according to claim 1 is characterized in that: The system's layered architecture design includes: The device perception layer achieves hardware decoupling through abstract interfaces, supporting plug-and-play access to monitoring devices of different brands and protocols; The AI ​​algorithm application layer adopts a plug-in architecture, which supports dynamic loading / replacement of large visual model algorithms without modifying the underlying system. The data layer and presentation layer interact through a unified data protocol to ensure the compatibility and stability of cross-layer data transmission.

8. The intelligent management system for water island polishing in power plants based on visual large-scale model according to claim 1 is characterized in that: The Zero-shot capability includes: The pre-training phase of the large visual model learns common visual semantic features across scenes to develop generalization capabilities; During deployment, the model's corresponding detection capabilities are activated through natural language prompts, without the need for specific environment training data; Based on the model's built-in scenario understanding capabilities, it automatically adapts to environmental differences such as equipment layout and lighting conditions in different power plants.

9. The intelligent management system for water island polishing in power plants based on visual large-scale model according to claim 1 is characterized in that: The presentation layer further includes integrated functions, as follows: Detection result visualization module: real-time video stream superimposed with target annotation box; Data management module: count the frequency of target occurrences by time / location and generate interactive trend charts; Algorithm configuration module: supports user-defined detection area ROI, sampling rate, and alarm threshold; Historical data module: provides video segment retrieval, playback and detection record export functions.

10. The intelligent management system for water island polishing in power plants based on visual large-scale model according to claim 3 is characterized in that: The visual large model algorithm carried by the AI ​​algorithm application layer adopts the following detection method: After feature extraction of the input image data, the Transformer architecture is used to achieve visual-language cross-modal semantic alignment; Use the attention mechanism to focus on the target area and output structured detection results including location coordinates and category labels; It supports multi-target parallel detection, and a single inference can identify various scenarios such as hot work and equipment anomalies.

Images