Multi-modal property management scene risk point detection system based on artificial intelligence

Abstract

The invention relates to the technical field of industrial intelligent inspection driven by artificial intelligence, and discloses a multi-modal property management scene risk point detection system based on artificial intelligence, and the system comprises a multi-modal data collection module which is used for synchronously obtaining visible light images, infrared thermal imaging and Internet of Things sensor data in building facilities; the domain self-adaptive defect generation module is based on a decoupling type generator and a StyleGAN2-ADA framework; a multi-modal feature fusion module; an attention enhancement detection module; a lightweight model compression frame; a real-time detection module; and a continuous learning module. High-fidelity defect samples are generated through the decoupling generator and the StyleGAN2-ADA, the problems of sample scarcity and class imbalance are solved, samples are generated through the cyclic generative adversarial network to cover multiple defect types, the acquisition period of the defect samples is shortened, and the requirement for training real samples is reduced.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence-driven industrial intelligent inspection technology, and specifically to an artificial intelligence-based multimodal property management scenario risk point detection system. Background Art

[0002] At present, the management of property risk points in scenarios has built a full-chain intelligent system covering data collection, defect simulation, feature analysis and decision execution by integrating cutting-edge technologies such as multimodal perception, generative adversarial networks, real-time reasoning and incremental learning. It provides an innovative paradigm for the maintenance of smart city infrastructure, breaking through the inefficiency and lag of traditional manual inspections, realizing active perception of the status of property facilities and abnormal early warning of scenario risk points, promoting the development of property operation and maintenance towards efficiency, precision and sustainability, and providing technical support for the safe operation and refined management of urban infrastructure. Traditional inspections mostly rely on the accumulation of real defect samples.

[0003] At present, collecting real defect samples requires a data collection and labeling cycle of several months or even years, which seriously restricts the development efficiency of detection models. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a multimodal property management scenario risk point detection system based on artificial intelligence to solve the problem that the real defect sample collection cycle is too long, which limits the development efficiency of the detection model.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a multimodal property management scenario risk point detection system based on artificial intelligence, comprising: Multimodal data acquisition module for synchronous acquisition of visible light images, infrared thermal imaging, and IoT sensor data in building facilities; The domain-adaptive defect generation module, based on a decoupled generator and the StyleGAN2-ADA (Style Generative Adversarial Network 2-Adaptive Discriminator Enhancement) architecture, generates high-fidelity defect samples with a resolution ≥ 1024×1024 and uses CycleGAN (Cycle Generative Adversarial Network) to transfer style across device scenarios. Multimodal feature fusion module, which is used to fuse the feature map level of visible light and infrared data through the attention-guided fusion module; Attention-enhanced detection module, which integrates spatial attention and channel attention mechanisms to locate defect areas; A lightweight model compression framework for mobile deployment through knowledge distillation, mixed-precision quantization, and hardware adaptation optimization; Real-time detection module, used for asynchronous pipeline parallel architecture and dynamic resolution adjustment, and supports industrial-grade real-time detection; The continuous learning module is used to realize online model updates through incremental learning algorithms and prototype memory libraries to adapt to the needs of identifying new defect types.

[0006] Preferably, the multimodal data acquisition module collects building facility and equipment data through a camera and infrared thermal imager bound to the drone, and adds a timestamp for data synchronization.

[0007] Preferably, the domain adaptive defect generation module includes: Defect attribute decoupling unit, used to separate the shape, texture and position parameters of the defect, and decouple the different attributes of the defect by learning and analyzing normal samples; The generator unit is used to generate 1024×1024 high-resolution defect samples through StyleGAN2-ADA, and to generate different types of defect samples by adjusting the parameters and structure of the generator; The style transfer unit is used to adapt to cross-domain differences in lighting, angles, and devices through CycleGAN. By learning the mapping relationship between different domains, the generated defect samples are transferred to different scenes.

[0008] Preferably, the multimodal feature fusion module includes: The visible light analysis unit is used to combine ResNet50 (a residual network consisting of 50 convolutional layers and fully connected layers) with visible light data, using the outputs of different levels of convolutional layers to form multi-scale feature maps covering a 16x-32x receptive field; The infrared light analysis unit is used to focus on the spatial distribution characteristics of temperature risk point areas through convolutional layers and activation functions, and uses Fourier transform to extract 0.1-5Hz heat diffusion patterns to capture dynamic temperature changes in building facilities; The fusion unit dynamically assigns fusion weights through normalized weight allocation calculation, and then generates a spatial attention map to focus on complementary regions in multimodal data, including the overlapping areas between cracks in visible light and temperature gradients in infrared.

[0009] Preferably, the attention enhancement detection module includes: The spatial attention unit is used to input the original building facility image and the annotated defect mask into the DRL-Net (deep reinforcement learning network) model, output an attention mask of the same size as the feature map, and focus high response values ​​on defect edges and areas with significant texture features; The channel attention unit is used to compress the spatial dimension of the feature map through global average pooling to generate a channel description vector. Then, two fully connected layers are used to learn the nonlinear relationship between channels and normalize it to a confidence score through the Sigmoid function. The cross-layer fusion strategy unit is used to build a bidirectional attention transfer path through a hierarchical attention transfer mechanism to fuse shallow and deep features, where the shallow layer includes the edges and textures of building facilities, and the deep layer includes the overall shape and defect types of building facilities.

[0010] Preferably, the lightweight model compression framework performs knowledge distillation through a teacher-student architecture, and adapts the system to run on mobile devices through mixed precision quantization technology.

[0011] Preferably, the real-time detection module includes: Double buffer queue unit, used for parallel processing of collected data and detection processing through two buffers; The dynamic resolution adjustment unit evaluates the complexity of the image through image processing technology and feature extraction algorithm, and adjusts the trade-off between image quality and processing speed based on the image complexity through image downsampling technology.

[0012] Preferably, the real-time detection module further includes an intelligent alarm unit, which is configured to establish three levels of alarm thresholds, including: Suspected alarm: When the detection model outputs a defect confidence level between 50% and 70%, a manual review process is triggered, and the alarm information is simultaneously pushed to the inspection terminal and associated with the device's geographic coordinates. Confirmation alarm: When the detection model outputs a defect confidence level between 70% and 90%, the system links the equipment maintenance record database. If the frequency of similar defects in historical equipment is ≥3 times, the equipment shutdown protocol is triggered. At the same time, the equipment location is obtained through geo-fencing, and a work order is generated and dispatched to the nearest operation and maintenance personnel's mobile terminal. Emergency Alarm: When the detection model outputs a defect confidence level greater than 90%, the emergency center is directly linked for processing, and drone inspections are initiated based on the geographic fence coordinates, with the captured images transmitted back in real time.

[0013] Preferably, the geo-fence uses multi-source positioning technology and spatial analysis algorithms to bind the detected risk point location to the actual device and confirm the device location.

[0014] Preferably, the continuous learning module includes: The incremental learning and defect adaptation unit collects 30-50 new defect samples, keeps the main parameters of the pre-trained model unchanged, and only unfreezes the last fully connected layer for low learning rate fine-tuning to generate the model file to be updated; The model update unit is used to transfer the model file to be updated to the edge device and automatically back up the current version before updating.

[0015] The present invention provides a multimodal property management scenario risk point detection system based on artificial intelligence. It has the following beneficial effects: 1. This invention generates high-fidelity defect samples through a decoupled generator and StyleGAN2-ADA, solving the problems of sample scarcity and category imbalance. It also generates samples covering a variety of defect types through a cyclic generative adversarial network, reducing the defect sample acquisition cycle and the demand for real training samples.

[0016] 2. The present invention can focus on the key areas in the image by weighting the spatial position in the feature map, and then focus on the defect-related feature channels based on channel attention to improve the key feature response. The combination of the two can achieve automatic focusing on the key defect area, effectively improving the accuracy of early crack recognition and the recognition accuracy of small defects.

[0017] 3. This invention transfers the knowledge of a large teacher model to a lightweight student model through knowledge distillation in a teacher-student architecture, reducing the number of parameters while maintaining high accuracy. Combined with mixed-precision quantization technology, the model parameters are represented with low precision, reducing computational complexity and storage overhead, improving inference speed and reducing power consumption, ensuring that the model runs efficiently on resource-constrained mobile devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is an architectural diagram of the domain-adaptive defect generation module of the present invention; Figure 3 This is an architecture diagram of the multimodal feature fusion module of the present invention; Figure 4 This is an architectural diagram of the attention enhancement detection module of the present invention; Figure 5 This is an architectural diagram of the real-time detection module of the present invention; Figure 6 This is an architectural diagram of the continuous learning module of the present invention. DETAILED DESCRIPTION

[0019] 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.

[0020] Please see the attached Figure 1 - Attachment Figure 6 , an embodiment of the present invention provides a multimodal property management scenario risk point detection system based on artificial intelligence, including: Multimodal data acquisition module for synchronous acquisition of visible light images, infrared thermal imaging, and IoT sensor data in building facilities; The domain-adaptive defect generation module, based on the decoupled generator and StyleGAN2-ADA architecture, generates high-fidelity defect samples with a resolution of ≥1024×1024 and performs cross-device style transfer through CycleGAN; Multimodal feature fusion module, which is used to fuse the feature map level of visible light and infrared data through the attention-guided fusion module; Attention-enhanced detection module, which integrates spatial attention and channel attention mechanisms to locate defect areas; A lightweight model compression framework for mobile deployment through knowledge distillation, mixed-precision quantization, and hardware adaptation optimization; Real-time detection module, used for asynchronous pipeline parallel architecture and dynamic resolution adjustment, and supports industrial-grade real-time detection; The continuous learning module is used to realize online model updates through incremental learning algorithms and prototype memory libraries to adapt to the needs of identifying new defect types.

[0021] Specifically, the multimodal data acquisition module simultaneously collects data from different sources, including visible light images, infrared thermal imaging images, and IoT sensor data in building facilities (such as temperature, humidity, vibration, noise, etc.), thereby comprehensively understanding the status of building facilities and providing a multi-dimensional information source for subsequent risk point detection and analysis. This ensures the effective identification of various types of risk points, including structural cracks, equipment failures, and temperature anomalies. The domain-adaptive defect generation module, based on a coupled generator and the StyleGAN2-ADA architecture, generates high-quality defect samples, ensuring high fidelity. CycleGAN also enables style transfer between different devices or scenarios, ensuring that the generated defect samples can adapt to different environments or device characteristics. This improves the diversity and quality of training data, addresses sample scarcity and class imbalance, and enhances sample generation quality. The multimodal feature fusion module fuses the features of visible light images and infrared thermal imaging data, effectively combining the advantages of visible light and infrared images, improving the detection ability of subtle defects and the early crack recognition rate. It also effectively suppresses interference from lighting and angles, and enhances the robustness of risk point location detection in complex scenarios. The attention-enhanced detection module is used to accurately locate defect areas. The spatial attention mechanism weights the spatial positions in the feature map, allowing the model to focus on key areas in the image. The channel attention mechanism focuses on feature channels related to defects, improving the response of key features. By combining these two attention mechanisms, the module can automatically focus on key defect areas and improve the recognition accuracy of small defects. The model is compressed and optimized through a lightweight model compression framework to ensure that it can run efficiently on mobile and embedded devices. Knowledge distillation reduces the model size through the teacher-student model architecture, mixed precision quantization reduces the demand for computing resources, and hardware adaptation optimization enables the model to better utilize the hardware acceleration of mobile devices.

[0022] The asynchronous pipeline parallel architecture in the real-time detection module, combined with dynamic resolution adjustment technology, ensures that the system can perform real-time anomaly detection in industrial environments. The pipeline architecture improves system throughput and processing speed by asynchronously executing tasks at different stages. Dynamic resolution adjustment automatically adjusts the image resolution based on the complexity of the scene, balancing detection speed and accuracy, ensuring detection speed without sacrificing accuracy, and improving adaptability in scenarios of varying complexity. By continuously updating the model online through the learning module, self-evolution and optimization can be achieved, timely adapting to new detection requirements and scenario changes, enhancing the long-term stability and accuracy of the model, and avoiding performance degradation caused by environmental changes.

[0023] The multimodal data acquisition module collects building facility and equipment data through the camera and infrared thermal imager attached to the drone, and adds timestamps for data synchronization.

[0024] Specifically, the visible light camera, infrared thermal imager and IoT sensor carried by the drone are used to synchronously collect high-definition images, temperature distribution and equipment operation status data of building facilities. Hardware-level timestamp synchronization (PTP protocol, error ±50μs) is used to ensure the spatiotemporal alignment of multi-source data. Automatic white balance, non-uniformity correction and sensor data filtering preprocessing are used to eliminate environmental interference and achieve pixel-level spatial alignment, providing a high-precision and high-consistency input basis for subsequent multimodal fusion analysis and defect detection.

[0025] The domain-adaptive defect generation module includes: Defect attribute decoupling unit, used to separate the shape, texture and position parameters of the defect, and decouple the different attributes of the defect by learning and analyzing normal samples; The generator unit is used to generate 1024×1024 high-resolution defect samples through StyleGAN2-ADA, and to generate different types of defect samples by adjusting the parameters and structure of the generator; The style transfer unit is used to adapt to cross-domain differences in lighting, angles, and devices through CycleGAN. By learning the mapping relationship between different domains, the generated defect samples are transferred to different scenes.

[0026] Specifically, the defect attribute unwrapping unit, based on spatial orthogonal constraints and hierarchical decoupling technology, decomposes defects into three independent attributes: shape, texture, and position. This enables controllable and combinatorial flexibility in defect generation. After decoupling, each attribute can be edited individually (for example, adjusting crack length without changing texture). This increases the diversity of generated samples by three times, reducing the defect sample acquisition cycle. The generator unit generates 1024×1024 high-resolution defect samples based on the StyleGAN2-ADA architecture, alleviates overfitting in small sample training through adaptive data augmentation, and uses hierarchical latent vectors to control the generation of details (such as crack branching morphology), so that the generated samples are closer to the real data quality.

[0027] Through the style transfer unit, CycleGAN is used to achieve unsupervised cross-domain adaptation, learn the mapping relationship between the source domain (laboratory environment) and the target domain (actual scene), adjust the lighting, angle and device imaging style while retaining the semantics of the defect, improve the diversity and adaptability of the generated samples, and make the defect samples applicable to various scenarios and devices, reducing the impact of environmental changes or device differences, thereby improving the subsequent anomaly detection capabilities under various conditions.

[0028] The multimodal feature fusion module includes: The visible light analysis unit is used to generate multi-scale feature maps covering 16x-32x receptive fields by combining ResNet50 with visible light data and using the outputs of different levels of convolutional layers. The infrared light analysis unit is used to focus on the spatial distribution characteristics of temperature anomalies through convolutional layers and activation functions, and uses Fourier transform to extract 0.1-5Hz heat diffusion patterns to capture dynamic temperature changes in building facilities; The fusion unit dynamically assigns fusion weights through normalized weight allocation calculation, and then generates a spatial attention map to focus on complementary regions in multimodal data, including the overlapping areas between cracks in visible light and temperature gradients in infrared.

[0029] Specifically, the visible light analysis unit uses ResNet50 to extract multi-scale visual features from the surface of building facilities, covering detection requirements from local details to global structures. It uses residual structures and multi-scale feature pyramids to effectively capture the locations of risk points from 0.1mm cracks to meter-level structural risks. In complex lighting scenarios, it reduces false detection rates and significantly improves the visible light modality's ability to detect surface defects. The infrared light analysis unit enhances temperature gradient features through convolutional layers and combines Fourier transforms to extract 0.1-5 Hz frequency domain thermal fluctuations (such as intermittent temperature rises caused by pump valve leakage). This improves the accuracy of infrared modality detection of internal defects, thereby enhancing the overall detection capabilities of multimodal data. The fusion unit dynamically allocates the fusion weights of visible light and infrared data through normalized weight distribution calculations, and then generates a spatial attention map to focus on the complementary areas in the multimodal data. The attention unit ensures that the fused feature map can focus on important areas rather than redundant information, effectively fusing the complementary information of visible light and infrared images, optimizing the performance of the feature map, and thus improving the accuracy and comprehensiveness of defect detection.

[0030] The attention enhancement detection module includes: The spatial attention unit is used to input the original building facility image and the annotated defect mask into the DRL-Net model, output an attention mask of the same size as the feature map, and focus high response values ​​on defect edges and texture feature-significant areas; The channel attention unit is used to compress the spatial dimension of the feature map through global average pooling to generate a channel description vector. Then, two fully connected layers are used to learn the nonlinear relationship between channels and normalize it to a confidence score through the Sigmoid function. The cross-layer fusion strategy unit is used to build a bidirectional attention transfer path through a hierarchical attention transfer mechanism to fuse shallow and deep features, where the shallow layer includes the edges and textures of building facilities, and the deep layer includes the overall shape and defect types of building facilities.

[0031] Specifically, the spatial attention unit inputs the original building facility image and the annotated defect mask, and the DRL-Net model generates an attention mask of the same size as the feature map. This attention mask helps the network focus on the most relevant areas in the image by focusing on the defect edges and areas with significant texture features, thereby improving the detection ability of detailed parts (such as cracks and damaged areas), significantly improving the positioning and recognition accuracy of subtle defects, and reducing the interference of background noise; The spatial dimensions of the feature map are compressed through global average pooling of the channel attention unit to generate a channel description vector. The nonlinear relationship between channels is then learned through two fully connected layers and normalized into confidence scores using the Sigmoid function. This effectively enhances the network's responsiveness across different channels, especially for feature channels related to defects. By optimizing channel attention allocation, the model can more accurately extract important features related to defects, improving detection accuracy. A bidirectional attention transfer path is constructed through the cross-layer fusion strategy unit, which effectively integrates the edge and texture features of the shallow layer with the overall shape and defect type of the deep layer, ensuring the complementarity of shallow and deep information, improving the fusion effect of features at different levels, and enhancing the model's ability to recognize complex defects, especially in structured images and detail processing.

[0032] The lightweight model compression framework performs knowledge distillation through a teacher-student architecture and adapts the system to run on mobile devices through mixed-precision quantization technology.

[0033] Specifically, the lightweight model compression framework performs knowledge distillation through a teacher-student architecture, transferring the knowledge of a large teacher model to a smaller student model, thereby reducing the number of model parameters while maintaining high accuracy. At the same time, through mixed-precision quantization technology, the model parameters are represented with low precision, reducing computing and storage overhead, improving inference speed and reducing power consumption, ensuring that the model can run efficiently and adapt to resource-constrained mobile devices.

[0034] The real-time detection module includes: Double buffer queue unit, used for parallel processing of collected data and detection processing through two buffers; The dynamic resolution adjustment unit evaluates the complexity of the image through image processing technology and feature extraction algorithm, and adjusts the trade-off between image quality and processing speed based on the image complexity through image downsampling technology.

[0035] Specifically, the double-buffered queue unit uses two buffers to achieve parallel processing of data collection and detection processing. One buffer is used to store the collected data, and the other buffer is used for data processing and anomaly detection, thereby avoiding the waiting time between data collection and processing and improving the system throughput and processing efficiency; The dynamic resolution adjustment unit evaluates the complexity of the image in real time and uses image downsampling technology to dynamically adjust the image resolution based on the complexity of the image, thereby finding the best balance between image quality and processing speed. The image resolution is dynamically optimized to maintain high-precision detection in complex scenes and improve processing speed in simple scenes, thereby performing efficient real-time detection and reducing the waste of system resources. The real-time detection module also includes an intelligent alarm unit, which is used to establish three levels of alarm thresholds, including: Suspected alarm: When the detection model outputs a defect confidence level between 50% and 70%, a manual review process is triggered, and the alarm information is simultaneously pushed to the inspection terminal and associated with the device's geographic coordinates. Confirmation alarm: When the detection model outputs a defect confidence level between 70% and 90%, the system links the equipment maintenance record database. If the frequency of similar defects in historical equipment is ≥3 times, the equipment shutdown protocol is triggered. At the same time, the equipment location is obtained through geo-fencing, and a work order is generated and dispatched to the nearest operation and maintenance personnel's mobile terminal. Emergency Alarm: When the detection model outputs a defect confidence level greater than 90%, the emergency center is directly linked for processing, and drone inspections are initiated based on the geographic fence coordinates, with the captured images transmitted back in real time.

[0036] Specifically, the intelligent alarm unit establishes a three-level alarm threshold and automatically classifies and processes risk points based on the defect confidence of the detection model. Its multi-level alarm mechanism can improve the processing flow of defects at different levels, thereby ensuring timely response and rapid processing.

[0037] Geofencing uses multi-source positioning technology and spatial analysis algorithms to bind the detected risk point locations to the actual devices and confirm the device location.

[0038] Specifically, geo-fencing integrates multi-source positioning technologies (such as GPS, Wi-Fi, Bluetooth, etc.) and spatial analysis algorithms to obtain the location information of the device in real time, and binds the detected risk point location with the actual device to accurately confirm the geographic location of the device, providing accurate positioning support for subsequent risk point anomaly handling, inspection tasks and alarm responses, effectively improving the accuracy and efficiency of risk point detection and emergency response.

[0039] Continuing Learning modules include: The incremental learning and defect adaptation unit collects 30-50 new defect samples, keeps the main parameters of the pre-trained model unchanged, and only unfreezes the last fully connected layer for low learning rate fine-tuning to generate the model file to be updated; The model update unit is used to transfer the model file to be updated to the edge device and automatically back up the current version before updating.

[0040] Specifically, the incremental learning and defect adaptation unit minimizes the consumption of computing resources and enables online model updates, improving adaptability to new defects without compromising existing detection capabilities, thus maintaining the model's efficiency and robustness. The model update unit is responsible for transmitting the updated model file after incremental learning to the edge device, and automatically backing up the current version before updating, ensuring that if any problems occur during the update process, the system can roll back to the previous stable version, avoiding the risks brought by update failures, and can quickly restore the normal operation of the system, ensuring a smooth transition of continuous learning and model updates.

[0041] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. The multimodal property management scenario risk point detection system based on artificial intelligence is characterized by: include: Multimodal data acquisition module for synchronous acquisition of visible light images, infrared thermal imaging, and IoT sensor data in building facilities; The domain-adaptive defect generation module, based on the decoupled generator and StyleGAN2-ADA architecture, generates high-fidelity defect samples with a resolution of ≥1024×1024 and performs cross-device style transfer through CycleGAN; Multimodal feature fusion module, which is used to fuse the feature map level of visible light and infrared data through the attention-guided fusion module; Attention-enhanced detection module, which integrates spatial attention and channel attention mechanisms to locate and detect defect areas; A lightweight model compression framework for mobile deployment through knowledge distillation, mixed-precision quantization, and hardware adaptation optimization; Real-time detection module, used for asynchronous pipeline parallel architecture and dynamic resolution adjustment, and supports industrial-grade real-time detection; The continuous learning module is used to realize online model updates through incremental learning algorithms and prototype memory libraries to adapt to the needs of identifying new defect types.

2. The multimodal property management scenario risk point detection system based on artificial intelligence according to claim 1 is characterized in that: The multimodal data acquisition module collects building facility and equipment data through the camera and infrared thermal imager bound to the drone, and adds timestamps for data synchronization.

3. The multimodal property management scenario risk point detection system based on artificial intelligence according to claim 1 is characterized in that: The domain adaptive defect generation module includes: Defect attribute decoupling unit, used to separate the shape, texture and position parameters of the defect, and decouple the different attributes of the defect by learning and analyzing normal samples; The generator unit is used to generate 1024×1024 high-resolution defect samples through StyleGAN2-ADA, and to generate different types of defect samples by adjusting the parameters and structure of the generator; The style transfer unit is used to adapt to cross-domain differences in lighting, angles, and devices through CycleGAN. By learning the mapping relationship between different domains, the generated defect samples are transferred to different scenes.

4. The multimodal property management scenario risk point detection system based on artificial intelligence according to claim 1 is characterized in that: The multimodal feature fusion module includes: The visible light analysis unit is used to generate multi-scale feature maps covering 16x-32x receptive fields by combining ResNet50 with visible light data and using the outputs of different levels of convolutional layers. The infrared light analysis unit is used to focus on the spatial distribution characteristics of temperature risk point areas through convolutional layers and activation functions, and uses Fourier transform to extract 0.1-5Hz heat diffusion patterns to capture dynamic temperature changes in building facilities; The fusion unit dynamically assigns fusion weights through normalized weight allocation calculation, and then generates a spatial attention map to focus on complementary regions in multimodal data, including the overlapping areas between cracks in visible light and temperature gradients in infrared.

5. The multimodal property management scenario risk point detection system based on artificial intelligence according to claim 1 is characterized in that: The attention enhancement detection module includes: The spatial attention unit is used to input the original building facility image and the annotated defect mask into the DRL-Net model, output an attention mask of the same size as the feature map, and focus high response values ​​on defect edges and texture feature-significant areas; The channel attention unit is used to compress the spatial dimension of the feature map through global average pooling to generate a channel description vector. Then, two fully connected layers are used to learn the nonlinear relationship between channels and normalize it to a confidence score through the Sigmoid function. The cross-layer fusion strategy unit is used to build a bidirectional attention transfer path through a hierarchical attention transfer mechanism to fuse shallow and deep features, where the shallow layer includes the edges and textures of building facilities, and the deep layer includes the overall shape and defect types of building facilities.

6. The multimodal property management scenario risk point detection system based on artificial intelligence according to claim 1 is characterized in that: The lightweight model compression framework performs knowledge distillation through a teacher-student architecture and adapts the system to run on mobile devices through mixed precision quantization technology.

7. The multimodal property management scenario risk point detection system based on artificial intelligence according to claim 1 is characterized in that: The real-time detection module includes: Double buffer queue unit, used for parallel processing of collected data and detection processing through two buffers; The dynamic resolution adjustment unit evaluates the complexity of the image through image processing technology and feature extraction algorithm, and adjusts the trade-off between image quality and processing speed based on the image complexity through image downsampling technology.

8. The multimodal property management scenario risk point detection system based on artificial intelligence according to claim 1 is characterized in that: The real-time detection module also includes an intelligent alarm unit, which is used to establish three levels of alarm thresholds, including: Suspected alarm: When the detection model outputs a defect confidence level between 50% and 70%, a manual review process is triggered, and the alarm information is simultaneously pushed to the inspection terminal and associated with the device's geographic coordinates. Confirmation alarm: When the detection model outputs a defect confidence level between 70% and 90%, the system links the equipment maintenance record database. If the frequency of similar defects in historical equipment is ≥3 times, the equipment shutdown protocol is triggered. At the same time, the equipment location is obtained through geo-fencing, and a work order is generated and dispatched to the nearest operation and maintenance personnel's mobile terminal. Emergency Alarm: When the detection model outputs a defect confidence level greater than 90%, the emergency center is directly linked for processing, and drone inspections are initiated based on the geographic fence coordinates, with the captured images transmitted back in real time.

9. The multimodal property management scenario risk point detection system based on artificial intelligence according to claim 8 is characterized in that: The geo-fence uses multi-source positioning technology and spatial analysis algorithms to bind the detected risk point location to the actual device and confirm the device location.

10. The multimodal property management scenario risk point detection system based on artificial intelligence according to claim 1 is characterized in that: The continuous learning module includes: The incremental learning and defect adaptation unit collects 30-50 new defect samples, keeps the main parameters of the pre-trained model unchanged, and only unfreezes the last fully connected layer for low learning rate fine-tuning to generate the model file to be updated; The model update unit is used to transfer the model file to be updated to the edge device and automatically back up the current version before updating.

Images