Method and device for identifying fire hazards through AI visual analysis technology

Abstract

The embodiments of the present application provide a method and device for identifying fire hazards through AI visual analysis technology, by innovatively constructing a multi-source data acquisition and fusion mechanism, integrating multimodal data such as image video streams, three-dimensional space, thermal imaging, and environmental perception. A feature extraction model based on transfer learning is designed, combined with a hierarchical classification structure and an attention mechanism, to achieve high-precision hazard feature identification through integrated learning. Time series analysis and spatial positioning technology are introduced to construct a hazard feature association network and evolution model to achieve hazard development trend prediction and common hazard discovery. This method effectively solves the shortcomings of traditional technologies in multimodal data processing, feature recognition, and trend prediction, and significantly improves the intelligence level and early warning capabilities of fire hazard identification.

Description

Technical Field

[0001] The present application relates to the field of data processing, and specifically to a method and device for identifying fire hazards through AI visual analysis technology. Background Art

[0002] Existing fire hazard identification methods have significant shortcomings. Traditional systems have limitations in multi-source data collection and fusion, making it difficult to effectively integrate multimodal data such as images, videos, depth information, thermal imaging, and acoustic signals, affecting the comprehensiveness of hazard identification.

[0003] Furthermore, existing technologies face bottlenecks in feature extraction and model training. Most systems fail to fully utilize transfer learning and attention mechanisms, lacking in-depth exploration of hierarchical relationships between features, resulting in low recognition accuracy.

[0004] Existing systems have technical shortcomings in hazard trend analysis. They lack the ability to predict hazards based on their time-series evolution, making it difficult to accurately predict their development trends and hindering the timeliness of preventive measures. Addressing these issues is crucial for improving fire hazard identification. Summary of the Invention

[0005] In response to the problems in the existing technology, this application provides a method and device for identifying fire hazards through AI visual analysis technology, which can effectively solve the shortcomings of traditional technology in multimodal data processing, feature recognition and trend prediction, and significantly improve the intelligence level and early warning capabilities of fire hazard identification.

[0006] In order to solve at least one of the above problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a method for identifying fire hazards using AI visual analysis technology, comprising:

[0008] Collecting image and video stream data from a security monitoring system, a mobile inspection terminal, and a fire inspection robot, acquiring scene depth information through a binocular camera to construct a three-dimensional space model, collecting infrared thermal imaging data to obtain temperature distribution features, collecting environmental perception signals to extract environmental features, preprocessing the image and video stream data, the three-dimensional space model, the temperature distribution features, and the environmental features to obtain normalized training data, extracting multimodal feature vectors from the normalized training data, annotating the multimodal feature vectors to construct training samples, selecting a basic feature extraction layer from a pre-trained model library to construct a transfer learning model, inputting the training samples into the transfer learning model, training a fire hazard feature neural network model group based on a hierarchical classification structure and an attention mechanism, and performing integrated learning on the recognition results of the fire hazard feature neural network model group to obtain a fusion recognition model;

[0009] receiving multi-source data collected in real time, performing time series decomposition and feature extraction on the multi-source data to construct a dynamic feature sequence, inputting the dynamic feature sequence into the fusion recognition model to perform multimodal feature recognition to obtain hidden danger feature data, and analyzing the time evolution trend of the hidden danger feature data based on the dynamic feature sequence to obtain trend prediction data;

[0010] The hidden danger feature data is spatially located to construct hidden danger marking data, a hidden danger feature association network is constructed based on the hidden danger marking data and the trend prediction data, a hidden danger evolution model is established using the hidden danger feature association network, hidden danger development trends are predicted through the hidden danger evolution model, a cluster analysis is performed on the hidden danger feature data to discover common hidden dangers, and the hidden danger marking data, the hidden danger development trends and the common hidden dangers are integrated into structured data and output to a monitoring terminal.

[0011] Furthermore, the method further includes: performing convolution decomposition and pooling operations on the normalized training data to extract image texture features, extracting three-dimensional spatial features through a deep image analysis module, obtaining temperature gradient features using an infrared feature extraction module, extracting acoustic features based on a spectrogram analysis module, fusing the image texture features, the three-dimensional spatial features, the temperature gradient features, and the acoustic features to obtain a multimodal feature vector, normalizing the feature components of the multimodal feature vector and adding time tags, and performing category and attribute annotation on the multimodal feature vector based on manual annotation rules to construct training samples;

[0012] Perform feature layer analysis on the models in the pre-trained model library, select a representative basic feature extraction layer according to the application scenario, copy the weight parameters and network structure of the basic feature extraction layer to the target network to build a transfer learning model, adjust the input layer and output layer of the transfer learning model to adapt them to the data format of the training samples, and input the training samples into the adjusted transfer learning model for feature mapping.

[0013] Furthermore, the method further includes: constructing a hierarchical classification structure based on a preset fire hazard classification system, dividing hazard types into multiple subcategory nodes according to physical characteristics, assigning a corresponding feature recognition model to each subcategory node, embedding a self-attention calculation unit and a cross-modal attention calculation unit in the feature recognition model, using the self-attention calculation unit to weight the importance of feature vectors, calculating the association weights between different modal features through the cross-modal attention calculation unit, inputting the training samples into each feature recognition model for back-propagation training to obtain a fire hazard feature neural network model group;

[0014] Calculate the recognition accuracy and confidence score of each model in the fire hazard characteristic neural network model group, assign a fusion weight coefficient to each model based on the recognition accuracy and confidence score, perform a weighted combination of the recognition results of the model group and the corresponding fusion weight coefficient, perform decision fusion on the weighted combination results through a voting mechanism to obtain a fusion recognition model, and iteratively optimize the fusion weight coefficient of the fusion recognition model.

[0015] Furthermore, the method further includes: receiving data streams collected by multiple sensors, dividing the data streams into time windows to obtain data segments, aligning and synchronizing the data segments according to timestamps, performing wavelet transform on the synchronized data to obtain time-frequency components, extracting amplitude features and phase features of the time-frequency components, sequentially splicing the amplitude features and phase features based on a sliding window method to construct a dynamic feature sequence, and performing normalization processing and noise reduction filtering on the dynamic feature sequence;

[0016] The dynamic feature sequence is parsed into a feature tensor according to a preset time step, the feature tensor is channel-separated to obtain multiple feature components, the feature components are input into the corresponding input ends of the fusion recognition model, the recognition result and confidence of each feature component are calculated by the fusion recognition model, the recognition results are screened and combined according to the confidence to obtain hidden danger feature data, and the hidden danger feature data is subjected to time series correlation analysis.

[0017] Furthermore, the method further includes: constructing a time series sample matrix according to the time dimension from the dynamic feature sequence and the hidden danger feature data, performing singular value decomposition on the time series sample matrix to extract principal component features, calculating the time increment and change rate of the principal component features, performing trend decomposition on the principal component features based on the change rate to obtain a periodic component and a trend component, reconstructing the periodic component and the trend component into a time series curve, and performing piecewise fitting on the time series curve to calculate fitting parameters;

[0018] A state prediction model is constructed using the fitting parameters, and historical time series data is input into the state prediction model for parameter training. The development trend of hidden danger characteristics is extrapolated based on the trained state prediction model. According to the extrapolation calculation results, a probability assessment of the hidden danger state at the predicted time point is performed to obtain trend prediction data, and the credibility of the trend prediction data is quantitatively assessed.

[0019] Furthermore, the method further includes: performing target detection and positioning on the hidden danger feature data to obtain bounding box coordinates, mapping the bounding box coordinates to a three-dimensional space model to obtain spatial position parameters, calculating the relative distance and azimuth between the spatial position parameters and a preset reference point, integrating the type label of the hidden danger feature data, the spatial position parameters, the relative distance, and the azimuth to construct hidden danger marking data, and spatially clustering the hidden danger marking data according to the relative distance;

[0020] A node feature matrix is ​​constructed based on the hidden danger marking data, a node transfer probability matrix is ​​constructed using the trend prediction data, the node feature matrix and the node transfer probability matrix are combined into a hidden danger feature association network, the association strength and propagation path between nodes in the hidden danger feature association network are calculated, a state transfer equation is constructed according to the association strength and the propagation path, and a hidden danger evolution model is established based on the state transfer equation.

[0021] Furthermore, the method further includes: numerically solving the state transition equation of the hidden danger evolution model to obtain an evolution trajectory, calculating the development direction and change rate of the evolution trajectory, extrapolating the evolution trajectory according to a preset time interval to obtain a hidden danger state prediction value, comparing the hidden danger state prediction value with a threshold value at a corresponding time point to obtain a hidden danger development trend, performing density clustering analysis on the hidden danger feature data to obtain feature clusters, calculating the center of gravity position and distribution range of the feature clusters, and identifying common hidden dangers according to the number and size of the feature clusters;

[0022] The hidden danger marking data is field mapped according to a preset data structure format, the hidden danger development trend data is time-serialized, the common hidden danger data is converted into an attribute relationship table, and the field mapping data, the time-serialized data and the attribute relationship table are integrated into structured data based on a unified data template, and the structured data is transmitted to the storage area specified by the monitoring terminal through a data interface.

[0023] In a second aspect, the present application provides a device for identifying fire hazards using AI visual analysis technology, comprising:

[0024] A model fusion module is used to collect image and video stream data from security monitoring systems, mobile inspection terminals, and fire inspection robots, obtain scene depth information through a binocular camera to construct a three-dimensional space model, collect infrared thermal imaging data to obtain temperature distribution characteristics, collect environmental perception signals to extract environmental characteristics, preprocess the image and video stream data, the three-dimensional space model, the temperature distribution characteristics, and the environmental characteristics to obtain normalized training data, extract multimodal feature vectors from the normalized training data, annotate the multimodal feature vectors to construct training samples, select a basic feature extraction layer from a pre-trained model library to construct a transfer learning model, input the training samples into the transfer learning model, train a fire hazard feature neural network model group based on a hierarchical classification structure and an attention mechanism, and perform integrated learning on the recognition results of the fire hazard feature neural network model group to obtain a fused recognition model;

[0025] a trend prediction module, configured to receive multi-source data collected in real time, perform time series decomposition and feature extraction on the multi-source data to construct a dynamic feature sequence, input the dynamic feature sequence into the fusion recognition model to perform multimodal feature recognition to obtain hidden danger feature data, and analyze the time evolution trend of the hidden danger feature data based on the dynamic feature sequence to obtain trend prediction data;

[0026] The hidden danger identification module is used to spatially locate the hidden danger feature data to construct hidden danger marking data, construct a hidden danger feature association network based on the hidden danger marking data and the trend prediction data, establish a hidden danger evolution model using the hidden danger feature association network, predict the hidden danger development trend through the hidden danger evolution model, perform cluster analysis on the hidden danger feature data to discover common hidden dangers, and integrate the hidden danger marking data, the hidden danger development trend and the common hidden dangers into structured data and output it to the monitoring terminal.

[0027] In a third aspect, the present application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the method for identifying fire hazards through AI visual analysis technology are implemented.

[0028] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for identifying fire hazards through AI visual analysis technology.

[0029] In a fifth aspect, the present application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the method for identifying fire hazards through AI visual analysis technology.

[0030] It can be seen from the above technical solution that the present application provides a method and device for identifying fire hazards through AI visual analysis technology, by innovatively constructing a multi-source data acquisition and fusion mechanism, integrating multimodal data such as image video streams, three-dimensional space, thermal imaging and acoustic signals. A feature extraction model based on transfer learning is designed, combined with a hierarchical classification structure and an attention mechanism, to achieve high-precision hazard feature identification through integrated learning. Time series analysis and spatial positioning technology are introduced to construct a hazard feature association network and evolution model to achieve hazard development trend prediction and common hazard discovery. This method effectively solves the shortcomings of traditional technologies in multimodal data processing, feature recognition and trend prediction, and significantly improves the intelligence level and early warning capabilities of fire hazard identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0032] Figure 1 This is a flow chart of a method for identifying fire hazards using AI visual analysis technology in an embodiment of the present application;

[0033] Figure 2 This is a structural diagram of a device for identifying fire hazards using AI visual analysis technology in an embodiment of the present application;

[0034] Figure 3 Schematic diagram of the structure of the electronic device in the embodiment of the present application.

[0035] Reference numerals:

[0036] Electronic device 9600, central processing unit 9100, memory 9140, communication module 9110, input unit 9120, audio processor 9130, display 9160, power supply 9170, buffer memory 9141, application / function storage unit 9142, data storage unit 9143, driver program storage unit 9144, antenna 9111, speaker 9131, microphone 9132. DETAILED DESCRIPTION

[0037] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0038] The acquisition, storage, use, and processing of data in this application's technical solution comply with relevant national laws and regulations.

[0039] Taking into account the problems existing in the prior art, the present application provides a method and device for identifying fire hazards through AI visual analysis technology, which innovatively constructs a multi-source data acquisition and fusion mechanism to integrate multimodal data such as image video streams, three-dimensional space, thermal imaging, and acoustic signals. A feature extraction model based on transfer learning is designed, combined with a hierarchical classification structure and an attention mechanism, to achieve high-precision hazard feature identification through integrated learning. Time series analysis and spatial positioning technology are introduced to construct a hazard feature association network and evolution model to achieve hazard development trend prediction and common hazard discovery. This method effectively solves the shortcomings of traditional technologies in multimodal data processing, feature recognition, and trend prediction, and significantly improves the intelligence level and early warning capabilities of fire hazard identification.

[0040] In order to effectively address the deficiencies of traditional technologies in multimodal data processing, feature recognition, and trend prediction, and significantly improve the intelligent level and early warning capabilities of fire hazard identification, this application provides an embodiment of a method for identifying fire hazards using AI visual analysis technology, see Figure 1 The method for identifying fire hazards through AI visual analysis technology specifically includes the following contents:

[0041] Step S101: collecting image and video stream data from a security monitoring system, a mobile inspection terminal, and a fire inspection robot, acquiring scene depth information through a binocular camera to construct a three-dimensional space model, collecting infrared thermal imaging data to obtain temperature distribution features, collecting environmental perception signals to extract environmental features, preprocessing the image and video stream data, the three-dimensional space model, the temperature distribution features, and the environmental features to obtain normalized training data, extracting multimodal feature vectors from the normalized training data, annotating the multimodal feature vectors to construct training samples, selecting a basic feature extraction layer from a pre-trained model library to construct a transfer learning model, inputting the training samples into the transfer learning model, training a fire hazard feature neural network model group based on a hierarchical classification structure and an attention mechanism, and performing integrated learning on the recognition results of the fire hazard feature neural network model group to obtain a fusion recognition model;

[0042] Optionally, this embodiment addresses the limitations of a single data source, incomplete feature extraction, and suboptimal model training during fire hazard identification. By doing so, a hazard identification model training solution based on multimodal data fusion has been innovatively designed. This embodiment first constructs a multi-source data acquisition network, integrating multiple acquisition terminals such as fixed security surveillance cameras, patrol personnel handheld terminals, and intelligent patrol robots to achieve comprehensive monitoring of key areas within a building. High-definition cameras capture image and video stream data, capturing the real-time status of key areas such as fire escape routes, emergency exits, and firefighting facilities.

[0043] This embodiment optimizes the method of obtaining scene depth information. A stereo image pair is captured by a binocular camera, and the three-dimensional spatial structure is reconstructed based on the parallax calculation principle:

[0044] Depth = (f B) / d

[0045] Where f is the camera focal length, B is the binocular baseline distance, and d is the parallax of the corresponding points. This method can accurately obtain the spatial position and size information of the object, providing basic data for subsequent spatial positioning of hidden dangers.

[0046] This embodiment innovatively captures temperature distribution characteristics. Infrared thermal imaging equipment is used to scan key areas and acquire temperature field distribution data. In electrical equipment inspection, this method can effectively detect potential faults such as localized overheating and poor contact. Temperature anomalies are also an important early warning indicator of fires.

[0047] This embodiment designs a method for extracting environmental perception features. For example, it extracts environmental acoustic features by collecting environmental perception signals using a high-sensitivity microphone array and extracting environmental features using spectrogram analysis. These features can reflect information such as equipment operating status and abnormal vibrations, providing important evidence for early detection of potential hazards.

[0048] This embodiment optimizes the data preprocessing process. Image data is de-noised, enhanced, and normalized; 3D model data is filtered and mesh optimized; temperature data is calibrated and smoothed; and acoustic data is subjected to noise reduction and spectral analysis. This preprocessing ensures the quality and consistency of all types of data.

[0049] This embodiment demonstrates significant advantages in multimodal feature extraction. It extracts visual features such as texture, shape, and color from images through a deep learning network; uses a three-dimensional convolutional network to extract spatial structural features; uses heat map analysis to extract temperature gradient features; and extracts acoustic features through spectrum analysis. These features describe the manifestations of fire hazards from different dimensions.

[0050] This embodiment innovatively designs a feature annotation strategy. It establishes a labeling system encompassing multiple categories, including equipment failure, channel blockage, and illegal electricity use. For each type of hidden danger, detailed attribute information such as location, type, and severity is annotated. This annotation information provides accurate supervisory signals for model training.

[0051] This embodiment optimizes the method for constructing a transfer learning model. It selects models from a library of pre-trained models that excel in tasks such as object detection and scene understanding, and extracts the weights and structure of their underlying feature layers. This approach fully leverages the advantages of pre-trained models in feature extraction and significantly reduces the need for training data.

[0052] This embodiment innovatively implements a model training strategy. It employs a hierarchical classification structure to decompose the hidden danger identification task into multiple specialized subtasks. An attention mechanism is introduced within each subtask to enhance the model's ability to perceive key features. This divide-and-conquer approach improves the model's recognition accuracy.

[0053] This embodiment significantly improves recognition performance through model ensemble learning. The prediction results of different sub-models are weighted and integrated, comprehensively considering the strengths of each model. For example, when identifying electrical fire hazards, temperature features and image features have different importance. Proper weight distribution improves recognition accuracy.

[0054] This embodiment establishes a complete model training system. From data collection to feature extraction, and then to model training and integration, a systematic training process is formed. This training system can adapt to the hidden danger identification needs of different scenarios and has good scalability.

[0055] This embodiment provides intelligent technical support for fire safety management. Through multimodal data analysis and deep learning technology, it achieves automated and intelligent identification of fire hazards. This approach not only improves the accuracy of hazard detection but also reduces the workload of manual inspections.

[0056] This embodiment achieves an intelligent upgrade for hidden danger identification. By integrating multi-source data and deep learning technology, hidden danger identification becomes more comprehensive and accurate. This intelligent identification method provides powerful technical support for fire safety management.

[0057] Step S102: receiving multi-source data collected in real time, performing time series decomposition and feature extraction on the multi-source data to construct a dynamic feature sequence, inputting the dynamic feature sequence into the fusion recognition model to perform multimodal feature recognition to obtain hidden danger feature data, and analyzing the time evolution trend of the hidden danger feature data based on the dynamic feature sequence to obtain trend prediction data;

[0058] Optionally, this embodiment addresses the challenges of capturing the dynamic evolution of fire hazards and poor real-time identification, by innovatively designing a dynamic feature recognition solution based on time series analysis. This embodiment first establishes a real-time multi-source data reception mechanism, synchronously collecting video streams, depth information, thermal imaging data, and acoustic signals via a data bus. To ensure real-time data integrity, a buffer queue management mechanism is employed to time-align and synchronize data streams at different sampling rates.

[0059] This embodiment optimizes the time series decomposition algorithm. It divides multi-source data into time windows, with the window length dynamically adjusted based on the characteristic time scales of different types of hidden dangers. For example, for electrical fire hazards, temperature anomalies evolve over a longer period, requiring a larger time window. Meanwhile, for sudden hazards like channel blockages, smaller time windows are needed to capture rapid changes.

[0060] This embodiment innovatively realizes multi-scale feature extraction. The time-frequency analysis of the signal is performed through wavelet transform:

[0061] WT(a,τ) = ∫ x(t)ψ((t-τ) / a)dt

[0062] Where x(t) is the input signal, ψ is the wavelet function, a is the scale parameter, and τ is the translation parameter. This method can effectively extract feature changes at different time scales, providing rich time-frequency features for hidden danger identification.

[0063] This embodiment designs a method for constructing dynamic feature sequences. Using a sliding window technique, features extracted within continuous time windows are sequentially concatenated. The window sliding step size is set based on actual monitoring requirements, ensuring feature continuity while avoiding the accumulation of redundant data. This method transforms discrete feature observations into a continuous time-series feature stream.

[0064] This embodiment optimizes the feature recognition strategy. By inputting a dynamic feature sequence into a trained fusion recognition model, the model's internal attention mechanism automatically captures changes in the importance of features across different modalities. For example, in nighttime scenarios, the weight of thermal imaging features is automatically increased; in equipment status monitoring, the importance of acoustic features is correspondingly increased.

[0065] This embodiment demonstrates significant advantages in real-time hidden danger identification. By dynamically integrating multimodal features, it can accurately capture early signs of hidden dangers. For example, in electrical equipment monitoring, comprehensive analysis of multi-dimensional features such as temperature anomalies, unusual sounds, and changes in equipment appearance can identify potential equipment failures in advance.

[0066] This embodiment innovatively designs a trend analysis method. It performs time series analysis on the identified hidden danger feature data to extract its evolution trend characteristics. It uses the autoregressive moving average model to model the feature sequence:

[0067] X(t) = φ1X(t-1) + ... + φpX(tp) + ε(t)

[0068] Where X(t) is the time series feature value, φi is the autoregressive coefficient, and ε(t) is the random error term. This model can be used to predict the development trend of the feature.

[0069] This embodiment optimizes the strategy for generating prediction data. Based on a trend model trained with historical data, it predicts the development trends of hidden danger characteristics. The model considers factors such as seasonal changes and cyclical fluctuations, improving prediction accuracy. For example, when predicting congestion in crowded places, the model considers daily variations in pedestrian flow.

[0070] This embodiment innovatively implements a credibility assessment for prediction results. By calculating the prediction interval and confidence level, the reliability of the prediction results is quantitatively assessed. This assessment mechanism provides a reference for management decisions and helps prioritize inspection and rectification work.

[0071] This embodiment significantly improves the foresight of hidden danger identification through dynamic feature analysis and trend prediction. By analyzing and predicting the evolution of hidden dangers, fire safety management shifts from passive response to active prevention.

[0072] This embodiment establishes a complete dynamic monitoring system. From real-time data collection to feature extraction and trend prediction, a systematic monitoring process is formed. This dynamic monitoring system can promptly identify potential risks and provide a basis for preventive maintenance.

[0073] This embodiment provides an intelligent early warning method for fire safety management. Through real-time feature recognition and trend prediction, it achieves early warning of fire hazards. This early warning mechanism not only improves the timeliness of hazard detection but also enhances proactive safety management.

[0074] This embodiment achieves an intelligent upgrade for hidden danger identification. Through dynamic feature analysis and trend prediction, hidden danger identification is made more timely and accurate. This intelligent identification method provides powerful technical support for fire safety management.

[0075] Step S103: spatially locate the hidden danger feature data to construct hidden danger marking data, construct a hidden danger feature association network based on the hidden danger marking data and the trend prediction data, establish a hidden danger evolution model using the hidden danger feature association network, predict the hidden danger development trend through the hidden danger evolution model, perform cluster analysis on the hidden danger feature data to find common hidden dangers, integrate the hidden danger marking data, the hidden danger development trend and the common hidden dangers into structured data and output it to the monitoring terminal.

[0076] Optionally, this embodiment addresses the complex spatial distribution of fire hazards, the difficulty in understanding inter-hazard correlations, and the inaccurate prediction of development trends. By innovatively designing a hazard evolution prediction solution based on spatial positioning and correlation analysis, this solution first spatially locates the identified hazard feature data and calculates the bounding box coordinates of the hazard target using a deep learning object detection algorithm. Combining this with a three-dimensional spatial model constructed using a binocular camera, the two-dimensional coordinates are mapped into three-dimensional space to obtain the spatial parameters of the hazard location.

[0077] This embodiment optimizes the spatial positioning method. A relative coordinate system is established based on a preset reference point, and the distance and azimuth between the hidden danger location and the reference point are calculated:

[0078] Position = [r cos(θ) cos(φ), r cos(θ) sin(φ), r sin(θ)]

[0079] Where r is the relative distance, θ is the elevation angle, and φ is the azimuth angle. This positioning method facilitates the description of the spatial distribution characteristics of hidden dangers.

[0080] This embodiment innovatively constructs hidden danger marker data. It integrates information such as the hidden danger's type label, spatial location, and relative distance into structured marker data. For example, for an electrical fire hazard, the marker data includes multi-dimensional information such as device type, fault characteristics, temperature anomaly level, and spatial location. This marker data provides the foundation for subsequent correlation analysis.

[0081] This embodiment designs a method for constructing a feature association network. A node feature matrix is ​​constructed based on hidden danger labeling data, with each node containing the attribute characteristics of the hidden danger. Trend prediction data is used to construct a transition probability matrix between nodes, describing the evolution of hidden danger states. By combining the feature matrix and the transition matrix, a complete hidden danger feature association network is formed.

[0082] This embodiment optimizes the strategy for establishing a hidden danger evolution model. The association strength and propagation path between nodes in the association network are calculated, and the state transition equation is constructed:

[0083] S(t+1) = T S(t) +ε(t)

[0084] Where S(t) is the state vector at time t, T is the transfer matrix, and ε(t) is the random disturbance term. This model can describe the dynamic evolution of the hidden danger state.

[0085] This embodiment demonstrates significant advantages in predicting hidden dangers. By numerically solving the evolution model, the evolution trajectory of the hidden danger state is obtained. Based on the development direction and rate of change of the evolution trajectory, the development trend of the hidden danger can be predicted. For example, in electrical equipment monitoring, the model can predict the spread and development speed of temperature anomalies.

[0086] This embodiment innovatively designs a common hidden danger discovery method. It performs density cluster analysis on hidden danger feature data and discovers feature clusters based on the distance measurement of feature space:

[0087] Cluster_density = Σ(exp(-||xi - xj||² / 2σ²))

[0088] Where xi and xj are eigenvectors, and σ is the density parameter. By analyzing the distribution characteristics of the feature clusters, common hidden danger types can be identified.

[0089] This embodiment optimizes the data integration strategy. Hidden danger mark data is mapped to fields according to a unified data structure, hidden danger development trend data is time-series processed, and common hidden danger data is converted into attribute relationship tables. This information is integrated through data templates to form standardized structured data.

[0090] This embodiment innovatively implements a data output mechanism. It configures data interface parameters and establishes a secure connection channel with the monitoring terminal. It transmits structured data to a designated storage area on the monitoring terminal via the data interface, ensuring the reliability and security of data transmission.

[0091] This embodiment significantly improves the scientific nature of hidden danger management through spatial analysis and evolutionary prediction. Spatially located hidden danger distribution analysis helps optimize inspection routes and resource allocation. The prediction results of the evolutionary model provide a basis for decision-making on hidden danger rectification.

[0092] This embodiment establishes a complete hidden danger analysis system. From spatial positioning to correlation analysis and then to trend prediction, a systematic analysis process is formed. This analysis system can fully grasp the distribution characteristics and development patterns of hidden dangers.

[0093] This embodiment provides scientific decision-making support for fire safety management. Through spatial analysis and trend prediction, it achieves precise management of fire hazards. This management approach not only improves the targeted treatment of hazard problems but also enhances the effectiveness of preventive measures.

[0094] This embodiment achieves an intelligent upgrade in hidden danger management. Through spatial positioning and evolution prediction, hidden danger management becomes more accurate and efficient. This intelligent management approach provides strong technical support for fire safety work.

[0095] From the above description, it can be seen that the method for identifying fire and firefighting hazards through AI visual analysis technology provided in the embodiment of the present application can integrate multimodal data such as image video streams, three-dimensional space, thermal imaging and acoustic signals by innovatively constructing a multi-source data acquisition and fusion mechanism. A feature extraction model based on transfer learning is designed, combined with a hierarchical classification structure and an attention mechanism, to achieve high-precision hazard feature identification through integrated learning. Time series analysis and spatial positioning technology are introduced to construct a hazard feature association network and evolution model to achieve hazard development trend prediction and common hazard discovery. This method effectively solves the shortcomings of traditional technologies in multimodal data processing, feature recognition and trend prediction, and significantly improves the intelligence level and early warning capabilities of fire and firefighting hazard identification.

[0096] In one embodiment of the method for identifying fire hazards using AI visual analysis technology of the present application, the following contents may also be specifically included:

[0097] Step S201: performing convolution decomposition and pooling operations on the normalized training data to extract image texture features, extracting three-dimensional spatial features using a deep image analysis module, obtaining temperature gradient features using an infrared feature extraction module, extracting acoustic features based on a spectrogram analysis module, fusing the image texture features, the three-dimensional spatial features, the temperature gradient features, and the acoustic features to obtain a multimodal feature vector, normalizing the feature components of the multimodal feature vector and adding time tags, and annotating the multimodal feature vector with categories and attributes based on manual annotation rules to construct training samples;

[0098] Step S202: Perform feature layer analysis on the models in the pre-trained model library, select a representative basic feature extraction layer according to the application scenario, copy the weight parameters and network structure of the basic feature extraction layer to the target network to build a transfer learning model, adjust the structure of the input layer and output layer of the transfer learning model to adapt to the data format of the training sample, and input the training sample into the adjusted transfer learning model for feature mapping.

[0099] Optionally, this embodiment addresses issues such as incomplete fire hazard feature extraction, difficulty fusing multimodal data, and low model training efficiency by innovatively designing a deep learning training solution based on multimodal feature fusion. This embodiment first performs deep feature extraction on normalized training data and then processes the image data using a multi-layer convolutional neural network. The convolution kernel design takes into account texture features at different scales, with small-sized convolution kernels used to extract local details and large-sized convolution kernels used to capture global structure.

[0100] This embodiment optimizes the 3D spatial feature extraction method. The depth information acquired by the binocular camera is processed by a depth image analysis module, and 3D reconstruction is achieved using a point cloud registration algorithm: point_cloud = f(depth_map,camera_params), where depth_map is the depth map and camera_params are the camera parameters. This method accurately reconstructs the spatial structure of potential hazards.

[0101] This embodiment innovatively realizes the extraction of temperature gradient features. The infrared feature extraction module is used to analyze thermal imaging data and calculate the spatial gradient of the temperature field:

[0102]

[0103] Where T is the temperature value, and x and y are spatial coordinates. Gradient analysis can effectively identify temperature anomaly areas and heat transfer directions.

[0104] This embodiment designs an acoustic feature extraction method. Based on the spectrogram analysis module to process the acoustic signal, the short-time Fourier transform is used to obtain the time-frequency features:

[0105] STFT(t,f) = ∫ x(τ)w(τ-t)e^(-j2πfτ)dτ

[0106] Where x(τ) is the acoustic signal, w is the window function, t is time, and f is frequency. This analysis method can capture abnormal characteristics of equipment operating sounds.

[0107] This embodiment optimizes the feature fusion strategy and performs adaptive weight fusion on features of different modalities:

[0108] Feature_fusion = Σ(wi Feature_i)

[0109] Where wi is the feature weight, and Feature_i is the feature of each modality. The weight coefficient is automatically optimized through backpropagation, so that the fused feature can better express the hidden danger characteristics.

[0110] This embodiment demonstrates significant advantages in feature extraction. Through the collaborative analysis of multimodal features, it is able to comprehensively capture the manifestations of hidden dangers. For example, in electrical equipment monitoring, combining multidimensional features such as abnormal temperature, abnormal operating sounds, and changes in appearance can accurately identify equipment failures.

[0111] This embodiment innovatively designs a labeling rule system. It establishes labeling rules encompassing multiple dimensions, including device type, fault type, and hazard level. For each feature vector, its category attributes and time information are detailed, providing complete supervision information for model training.

[0112] This embodiment optimizes the transfer learning strategy. Models that excel in tasks such as object detection and scene understanding are selected from a library of pre-trained models as the base network. Through feature layer analysis, a basic feature extraction layer with good generalization capabilities is extracted.

[0113] This embodiment innovatively implements model structure adjustment. The input and output layers of the transfer learning model are restructured to adapt to the data format of multimodal features. The input layer is designed with multiple feature channels to receive feature data from different modalities; the output layer is adjusted based on the requirements of the hidden danger classification task.

[0114] This embodiment significantly improves model training performance through feature mapping optimization. Leveraging the feature extraction capabilities of pre-trained models reduces the computational overhead of training from scratch. Furthermore, fine-tuning ensures the model is adaptable to the specific needs of fire hazard identification.

[0115] This embodiment establishes a complete feature learning system. From multimodal feature extraction to feature fusion and then to transfer learning, a systematic feature learning process is formed. This learning system can effectively extract and utilize the characteristic information of hidden dangers.

[0116] This embodiment provides an intelligent feature analysis method for fire safety management. Through deep learning technology, it achieves automated extraction and analysis of fire hazard characteristics. This analysis method not only improves the comprehensiveness of feature extraction but also enhances the efficiency of model training.

[0117] This embodiment achieves an intelligent upgrade for feature analysis. Through multimodal feature fusion and transfer learning, feature extraction becomes more accurate and efficient. This intelligent analysis method provides powerful technical support for fire safety management.

[0118] In one embodiment of the method for identifying fire hazards using AI visual analysis technology of the present application, the following contents may also be specifically included:

[0119] Step S301: Construct a hierarchical classification structure based on a preset fire hazard classification system, divide hazard types into multiple subcategory nodes according to physical characteristics, assign a corresponding feature recognition model to each subcategory node, embed a self-attention calculation unit and a cross-modal attention calculation unit in the feature recognition model, use the self-attention calculation unit to weight the importance of feature vectors, and use the cross-modal attention calculation unit to calculate the association weights between different modal features. Input the training samples into each feature recognition model for back-propagation training to obtain a fire hazard feature neural network model group;

[0120] Step S302: Calculate the recognition accuracy and confidence score of each model in the fire hazard characteristic neural network model group, assign a fusion weight coefficient to each model based on the recognition accuracy and confidence score, perform a weighted combination of the recognition results of the model group and the corresponding fusion weight coefficient, perform decision fusion on the weighted combination results through a voting mechanism to obtain a fusion recognition model, and iteratively optimize the fusion weight coefficient of the fusion recognition model.

[0121] Optionally, this embodiment addresses the complex types of fire hazards, low feature recognition accuracy, and poor model fusion effects by innovatively designing a model training scheme based on hierarchical classification and attention mechanisms. This embodiment first constructs a hierarchical hazard classification system, scientifically classifying hazards based on physical characteristics. For example, electrical fire hazards can be subdivided into subcategories such as aging lines, poor contact, and excessive loads; flammable and explosive hazards can be subdivided into subcategories such as gas leaks, illegal storage, and substandard environments.

[0122] This embodiment optimizes the allocation strategy for feature recognition models. A dedicated feature recognition model is designed for each subcategory node. The network structure of each model is optimized based on the complexity of the hidden danger characteristics to ensure appropriate feature extraction capabilities. For example, for temperature anomaly detection, the model focuses on thermal imaging features; for channel blockage detection, it emphasizes spatial feature analysis.

[0123] This embodiment innovatively implements the self-attention mechanism. The self-attention calculation unit is embedded in the feature recognition model:

[0124] Attention(Q,K,V) = softmax(QK^T / √d)V

[0125] Where Q, K, and V are the query matrix, key matrix, and value matrix, respectively, and d is the feature dimension. Through the attention mechanism, the model can adaptively focus on important features and improve recognition accuracy.

[0126] This embodiment designs a cross-modal attention calculation method. The cross-modal attention unit processes the association between different modal features:

[0127] Cross_attention = Σ(wi,j Feature_i Feature_j)

[0128] Where wi,j is the inter-modal association weight, and Feature_i and Feature_j are features of different modalities. This mechanism can capture the synergistic relationship between multimodal features.

[0129] This embodiment optimizes the model training strategy. It uses an end-to-end backpropagation training method, simultaneously optimizing multiple steps, including feature extraction, attention calculation, and classification prediction. The learning rate is dynamically adjusted during training to ensure the model converges to the optimal solution.

[0130] This embodiment demonstrates significant advantages in model organization. Through a hierarchical model structure, each sub-model can focus on identifying specific types of hidden dangers, enhancing the model's professionalism. Furthermore, the introduction of an attention mechanism enables the model to intelligently process complex feature relationships.

[0131] This embodiment innovatively designs a model fusion strategy. Calculate the recognition accuracy of each model:

[0132] Accuracy = (TP + TN) / (TP + TN + FP + FN)

[0133] Where TP is true positive, TN is true negative, FP is false positive, and FN is false negative. Initial weights are assigned to the model based on the accuracy.

[0134] This embodiment optimizes the confidence assessment method. It calculates a confidence score by analyzing the probability distribution of the model output. The higher the confidence score, the more confident the model is in its prediction, and the corresponding weighting coefficient increases. This mechanism ensures that high-quality predictions play a dominant role in the fusion process.

[0135] This embodiment innovatively implements a decision-making fusion mechanism, using weighted voting to integrate the prediction results of each model:

[0136] Final_result = argmax(Σ(wi Prediction_i))

[0137] Where wi is the model weight and Prediction_i is the prediction result of each model. In this way, the advantages of each model are fully utilized.

[0138] This embodiment improves fusion performance through iterative optimization. The performance of the fusion model is regularly evaluated, and the weight coefficients are dynamically adjusted based on the evaluation results. This adaptive optimization mechanism ensures that the fusion model can maintain good recognition results.

[0139] This embodiment establishes a complete model training and fusion system. From hierarchical classification to attention calculation and model fusion, a systematic recognition framework is formed. This framework can effectively handle complex fire hazard identification tasks.

[0140] This embodiment provides an intelligent identification tool for fire safety management. Through multi-model collaboration and intelligent fusion, it achieves accurate identification of fire hazards. This identification method not only improves recognition accuracy but also enhances the adaptability of the system.

[0141] This embodiment achieves an intelligent upgrade for hidden danger identification. Through hierarchical classification and model fusion, hidden danger identification is made more accurate and reliable. This intelligent identification method provides strong technical support for fire safety management.

[0142] In one embodiment of the method for identifying fire hazards using AI visual analysis technology of the present application, the following contents may also be specifically included:

[0143] Step S401: receiving data streams collected by multiple sensors, dividing the data streams into time windows to obtain data segments, aligning and synchronizing the data segments according to timestamps, performing wavelet transform on the synchronized data to obtain time-frequency components, extracting amplitude and phase features of the time-frequency components, sequentially concatenating the amplitude and phase features based on a sliding window method to construct a dynamic feature sequence, and performing normalization and noise reduction filtering on the dynamic feature sequence;

[0144] Step S402: Parse the dynamic feature sequence into a feature tensor according to a preset time step, perform channel separation on the feature tensor to obtain multiple feature components, input the feature components into the corresponding input end of the fusion recognition model, calculate the recognition result and confidence of each feature component through the fusion recognition model, filter and combine the recognition results according to the confidence to obtain hidden danger feature data, and perform time series correlation analysis on the hidden danger feature data.

[0145] Optionally, this embodiment addresses the challenges of real-time processing of multi-source data, discontinuous feature extraction, and unstable recognition results by innovatively designing a dynamic feature processing solution based on time-frequency analysis. This embodiment first establishes a multi-channel data reception mechanism to simultaneously receive multiple sensor data sources, including video streams, thermal imaging, depth information, and acoustic signals. A high-speed cache queue is used to manage data flows, ensuring real-time and continuous data reception.

[0146] This embodiment optimizes the time window division strategy. According to the sampling characteristics of different types of data, an adaptive window division method is designed:

[0147] Window_size = max(fs Tc, min_size)

[0148] Where fs is the sampling frequency, Tc is the feature correlation time, and min_size is the minimum window size. This method ensures that the window size can fully capture feature changes while maintaining computational efficiency.

[0149] This embodiment innovatively achieves data synchronization and alignment. Based on timestamp information, interpolation and resampling techniques are used to achieve time alignment of different data streams. For data with different sampling rates, linear interpolation is used to ensure consistency of data points on the timeline. For example, low-frame-rate thermal imaging data can be interpolated to match the timing of high-frame-rate video data.

[0150] This embodiment designs an efficient time-frequency analysis method. Continuous wavelet transform is performed on the synchronized data:

[0151] CWT(a,b) = (1 / √a)∫x(t)ψ ((tb) / a)dt

[0152] Where a is the scale parameter, b is the translation parameter, and ψ is the wavelet basis function. Through multi-scale analysis, the time domain and frequency domain characteristics of the signal can be obtained simultaneously.

[0153] This embodiment optimizes the feature extraction strategy. It extracts amplitude and phase features from the time-frequency components. Amplitude features reflect changes in signal strength, while phase features contain information about the signal structure. The combination of these two types of features can comprehensively describe the dynamic characteristics of a signal. For example, in equipment anomaly detection, amplitude anomalies in vibration signals often indicate a fault, while phase changes reflect the type of fault.

[0154] This embodiment shows significant advantages in dynamic feature construction. The sliding window method is used to sequentially splice features:

[0155] Feature_sequence = [F(tn), ..., F(t-1), F(t)]

[0156] Where F(t) is the feature vector at time t, and n is the sequence length. By continuously updating the feature sequence, real-time tracking of dynamic changes is achieved.

[0157] This embodiment innovatively designs a feature preprocessing method. It normalizes the dynamic feature sequence to make features of different dimensions comparable. It also uses a wavelet threshold denoising method to eliminate random disturbances and improve the signal-to-noise ratio of the feature.

[0158] This embodiment optimizes the feature tensor construction strategy. It reorganizes the dynamic feature sequence into a three-dimensional feature tensor at a fixed time step, maintaining the structured organization of the time, feature, and batch dimensions. This organization facilitates subsequent deep learning processing.

[0159] This embodiment innovatively implements a feature separation mechanism. It performs channel separation on the feature tensor and assigns features of different modalities to corresponding processing channels. For example, visual features, thermal features, depth features, etc. are input into dedicated input ports of the fusion recognition model.

[0160] This embodiment improves recognition reliability through confidence evaluation. The recognition result and confidence of each feature component are calculated:

[0161] Confidence = Softmax(output_logits)

[0162] Where output_logits is the raw score output by the model. Reliable recognition results are screened based on the confidence threshold.

[0163] This embodiment establishes a complete time series analysis framework. From data reception to feature extraction and then to result recognition, a continuous processing flow is formed. This framework can effectively process real-time data streams and ensure the continuity of feature extraction and recognition.

[0164] This embodiment provides real-time analysis capabilities for fire safety monitoring. Through dynamic feature processing and real-time identification, continuous monitoring of fire hazards is achieved. This monitoring method not only improves the timeliness of identification but also enhances system reliability.

[0165] This embodiment achieves an intelligent upgrade in feature processing. Through time-frequency analysis and dynamic feature construction, feature extraction is made more accurate and stable. This intelligent processing method provides powerful technical support for fire safety monitoring.

[0166] In one embodiment of the method for identifying fire hazards using AI visual analysis technology of the present application, the following contents may also be specifically included:

[0167] Step S501: constructing a time series sample matrix based on the time dimension of the dynamic feature sequence and the hidden danger feature data, performing singular value decomposition on the time series sample matrix to extract principal component features, calculating the time increment and change rate of the principal component features, performing trend decomposition on the principal component features based on the change rate to obtain periodic components and trend components, reconstructing the periodic components and trend components into a time series curve, and performing piecewise fitting on the time series curve to calculate fitting parameters;

[0168] Step S502: Use the fitting parameters to construct a state prediction model, input historical time series data into the state prediction model for parameter training, extrapolate the development trend of hidden danger characteristics based on the trained state prediction model, and perform probability evaluation on the hidden danger state at the predicted time point according to the extrapolation calculation results to obtain trend prediction data, and quantitatively evaluate the credibility of the trend prediction data.

[0169] Optionally, this embodiment addresses issues such as inaccurate fire hazard trend predictions, difficulty identifying cyclical changes, and low reliability of prediction results by innovatively designing a trend prediction solution based on time series analysis. This embodiment first aligns and integrates dynamic feature sequences and hazard feature data along the time dimension to construct a time series sample matrix containing multidimensional features. This matrix structure preserves complete information about feature evolution over time.

[0170] This embodiment optimizes the principal component feature extraction method. Perform singular value decomposition on the time series sample matrix:

[0171] X = UΣV^T

[0172] Where X is the time series sample matrix, U is the left singular matrix, Σ is the singular value matrix, and V is the right singular matrix. By selecting the eigenvectors corresponding to the important singular values, the main change patterns in the data are extracted.

[0173] This embodiment innovatively implements feature change analysis. Calculate the increments of the principal component features at adjacent time points:

[0174] ΔF(t) = F(t) - F(t-1)

[0175] Where F(t) is the eigenvalue at time t. The rate of change is calculated based on the increment, reflecting the dynamic trend of the characteristic. This method can effectively capture the speed characteristics of hidden danger development.

[0176] This embodiment designs a trend decomposition strategy. It uses seasonal decomposition to process the principal component characteristics, breaking them down into cyclical and trend components. The cyclical component reflects regular changes in the hazard characteristics, such as equipment usage cycles and ambient temperature fluctuations, while the trend component reflects long-term development trends.

[0177] This embodiment optimizes the time series reconstruction method. The decomposed periodic and trend components are recombined to construct a physically meaningful time series curve. This reconstruction method preserves the main characteristics of the data while reducing the impact of random noise.

[0178] This embodiment shows significant advantages in curve fitting. Perform segmented fitting on the reconstructed time series:

[0179] y(t) = Σ(ai φi(t))

[0180] Where φi(t) is the basis function and ai is the fitting parameter. By selecting appropriate segmentation points and basis functions, accurate description of trends at different stages can be achieved.

[0181] This embodiment innovatively designs a state prediction model. A nonlinear state space model is constructed based on fitting parameters. The model structure takes into account the historical dependence and nonlinear evolution characteristics of features. This model design makes the prediction results more consistent with the actual laws of hidden danger development.

[0182] This embodiment optimizes the model training strategy. Historical time series data is fed into the prediction model in chronological order, and a sliding window method is used for parameter training. The learning rate is dynamically adjusted during training to ensure that the model converges to the optimal solution.

[0183] This embodiment innovatively implements a trend extrapolation algorithm. Based on the trained state prediction model, the hidden danger state at a future time point is predicted:

[0184] State(t+k) = f(State(t), Parameters)

[0185] Where k is the prediction step size and f is the state transition function. The long-term trend prediction is achieved through multi-step iteration.

[0186] This embodiment improves prediction reliability through probabilistic assessment. Probabilistic modeling is performed on the prediction results to calculate the probability of occurrence of different states. This probabilistic prediction method provides more comprehensive information for risk assessment.

[0187] This embodiment establishes a complete prediction evaluation system. Construct prediction credibility evaluation indicators:

[0188] Confidence = g(prediction_error, sample_size, trend_stability)

[0189] Where g is the evaluation function, which comprehensively considers factors such as prediction error, sample size and trend stability.

[0190] This embodiment provides predictive decision support for fire safety management. Through trend prediction and credibility assessment, it enables scientific prediction of the development trends of fire hazards. This prediction method not only improves the accuracy of early warnings but also provides a basis for preventive maintenance.

[0191] This embodiment achieves an intelligent upgrade to predictive analysis. Through time series analysis and trend prediction, hidden danger warnings are more accurate and reliable. This intelligent prediction method provides powerful technical support for fire safety management.

[0192] In one embodiment of the method for identifying fire hazards using AI visual analysis technology of the present application, the following contents may also be specifically included:

[0193] Step S601: performing target detection and positioning on the hidden danger feature data to obtain bounding box coordinates, mapping the bounding box coordinates to a three-dimensional space model to obtain spatial position parameters, calculating the relative distance and azimuth between the spatial position parameters and a preset reference point, integrating the type label of the hidden danger feature data, the spatial position parameters, the relative distance, and the azimuth to construct hidden danger marker data, and spatially clustering the hidden danger marker data based on the relative distance;

[0194] Step S602: construct a node feature matrix based on the hidden danger marking data, construct a node transfer probability matrix using the trend prediction data, combine the node feature matrix and the node transfer probability matrix into a hidden danger feature association network, calculate the association strength and propagation path between nodes in the hidden danger feature association network, construct a state transfer equation according to the association strength and the propagation path, and establish a hidden danger evolution model based on the state transfer equation.

[0195] Optionally, this embodiment addresses issues such as inaccurate spatial positioning of fire hazards, difficulty in understanding inter-hazard correlations, and difficulty in predicting their evolution patterns. By innovatively designing a hazard modeling solution based on spatial analysis and network evolution, this embodiment first employs a deep learning object detection network to locate hazard feature data and obtain the bounding box coordinates of the target. The detection network employs a multi-scale feature fusion strategy to improve detection capabilities for objects of varying sizes.

[0196] This embodiment optimizes the spatial mapping method. The two-dimensional bounding box coordinates are mapped to the three-dimensional space through the camera projection model:

[0197] P_3D = R K^(-1) P_2D + T

[0198] Where P_2D is the two-dimensional coordinate, K is the camera intrinsic parameter matrix, R is the rotation matrix, and T is the translation vector. This mapping method achieves an accurate conversion from image space to physical space.

[0199] This embodiment innovatively implements spatial parameter calculation. A spatial coordinate system is established based on a preset reference point to calculate the relative distance and azimuth of the hidden danger location:

[0200] Distance = √((x-x0)² + (y-y0)² + (z-z0)²)

[0201] Angle = arctan2((y-y0), (x-x0))

[0202] Where (x0, y0, z0) is the coordinate of the reference point. This representation method is convenient for describing the spatial distribution characteristics of hidden dangers.

[0203] This embodiment designs a method for constructing hidden danger marker data. It integrates hidden danger type labels, spatial location parameters, relative distance, and azimuth into structured marker data. For example, for electrical fire hazards, the marker data includes multi-dimensional information such as device type, fault characteristics, and spatial location.

[0204] This embodiment optimizes the spatial clustering strategy and uses a density clustering algorithm to group the hidden danger mark data:

[0205] Cluster = DBSCAN(Points, eps, min_samples)

[0206] Where eps is the neighborhood radius and min_samples is the minimum number of samples. Cluster analysis is used to identify spatially close groups of hidden dangers.

[0207] This embodiment demonstrates significant advantages in constructing a feature network. A node feature matrix is ​​constructed based on hidden danger labeling data, with each node containing the hidden danger's attribute characteristics and spatial information. A node transition probability matrix is ​​constructed using trend prediction data to describe the evolution of hidden danger status.

[0208] This embodiment innovatively designs an association network structure. It combines the node feature matrix and the transition probability matrix into a hidden danger feature association network. The edge weights in the network reflect the strength of the mutual influence between nodes, while the direction indicates the propagation path of the influence.

[0209] This embodiment optimizes the method for calculating association strength. The association strength is calculated based on the similarity and spatial distance of node features:

[0210] Correlation = f(Feature_similarity, Spatial_distance)

[0211] Where f is the correlation function, which takes into account both feature similarity and spatial proximity. This calculation method can accurately describe the interactions between hidden dangers.

[0212] This embodiment innovatively implements propagation path analysis. It uses graph theory algorithms to calculate the shortest and critical paths in the network, identifying the main channels through which hidden dangers spread. This path information is of great guiding significance for formulating preventive measures.

[0213] This embodiment describes the system evolution through the state transition equation. The state transition equation is constructed based on the correlation strength and propagation path:

[0214] dS / dt = A S + B U + ε

[0215] Where S is the state vector, A is the state transfer matrix, U is the control input, and ε is the random disturbance term.

[0216] This embodiment establishes a complete evolutionary model system. From spatial positioning to network construction and then to state evolution, a systematic modeling process is formed. This model system can accurately describe the spatial distribution and evolution of fire hazards.

[0217] This embodiment provides a scientific analytical tool for fire safety management. Through spatial analysis and evolutionary modeling, it achieves precise location and development prediction of fire hazards. This analytical approach not only improves the accuracy of hazard identification but also enhances the targeted nature of preventive measures.

[0218] This embodiment achieves an intelligent upgrade for hidden danger analysis. Through spatial modeling and network analysis, hidden danger management becomes more accurate and efficient. This intelligent analysis method provides powerful technical support for fire safety management.

[0219] In one embodiment of the method for identifying fire hazards using AI visual analysis technology of the present application, the following contents may also be specifically included:

[0220] Step S701: Numerically solve the state transition equation of the hidden danger evolution model to obtain an evolution trajectory, calculate the development direction and change rate of the evolution trajectory, extrapolate the evolution trajectory according to a preset time interval to obtain a hidden danger state prediction value, compare the hidden danger state prediction value with the threshold value at the corresponding time point to obtain the hidden danger development trend, perform density clustering analysis on the hidden danger feature data to obtain feature clusters, calculate the center of gravity position and distribution range of the feature clusters, and identify common hidden dangers based on the number and size of the feature clusters;

[0221] Step S702: Map the hidden danger marking data to fields according to a preset data structure format, perform time serialization processing on the hidden danger development trend data, convert the common hidden danger data into an attribute relationship table, integrate the field mapping data, the time serialization data and the attribute relationship table into structured data based on a unified data template, and transmit the structured data to the storage area specified by the monitoring terminal through a data interface.

[0222] Optionally, this embodiment addresses the challenges of accurately predicting the evolution of fire hazards, identifying common hazards, and irregular data integration by innovatively designing a hazard assessment solution based on evolutionary analysis and data structuring. This embodiment first uses numerical integration to solve the state transition equation of the hazard evolution model, obtaining the evolution trajectory of the hazard state over time. High-precision numerical calculations are performed using the Runge-Kutta method to ensure the stability and accuracy of the solution.

[0223] This embodiment optimizes the evolution trajectory analysis method. Calculate the tangent vector and curvature of the trajectory at each time point:

[0224] Direction = dS / dt / ||dS / dt||

[0225] Rate = ||d²S / dt²|| / ||dS / dt||

[0226] Where S is the state vector and t is the time variable. These parameters reflect the direction and speed characteristics of the hidden danger development.

[0227] This embodiment innovatively implements state prediction. Extrapolation calculation based on known trajectory:

[0228] S(t+Δt) = S(t) + v(t) Δt + 0.5 a(t) Δt²

[0229] Where v(t) is the instantaneous rate of change, a(t) is the acceleration term, and Δt is the prediction interval. This method can be used to predict the state of hidden dangers at future points in time.

[0230] This embodiment designs a trend assessment strategy. The predicted value is compared with the dynamic threshold:

[0231] Risk_level = f(S_pred -S_threshold)

[0232] Where S_pred is the predicted state value, S_threshold is the threshold, and f is the evaluation function. This evaluation method can detect potential risks in a timely manner.

[0233] This embodiment optimizes the density clustering algorithm and uses the improved DBSCAN method to cluster the hidden danger feature data:

[0234] Density(p) = Σ exp(-||p-qi||² / 2σ²)

[0235] Where p is the feature point, qi is the other points in the neighborhood, and σ is the density parameter. Density analysis is used to identify clustered areas in the feature space.

[0236] This embodiment shows significant advantages in identifying common hidden dangers. Calculate the statistical features of the feature cluster, including the center of gravity position:

[0237] Centroid = (1 / n)Σxi

[0238] Where xi is the position vector of the point in the cluster, and n is the number of points. By analyzing the spatial distribution characteristics of the feature clusters, a universal hidden danger pattern is discovered.

[0239] This embodiment innovatively designs a data structuring solution. Hidden danger marking data is mapped according to a unified field structure to ensure the standardization of the data format. For example, information such as spatial location, hidden danger type, and risk level is mapped to predefined data fields.

[0240] This embodiment optimizes time series processing. Time series organization of hidden danger development trend data:

[0241] TimeSeries = {<t1,S1> ,<t2,S2> , ...,<tn,Sn>}

[0242] Where ti is the timestamp and Si is the corresponding status value. This organization method facilitates subsequent trend analysis and query.

[0243] This embodiment innovatively implements attribute relationship table conversion, converting common hidden danger data into a standard relational structure and establishing associations between hidden danger features. This structured representation facilitates data management and analysis.

[0244] This embodiment integrates various types of information through a unified data template. The template design takes into account the integrity, consistency and scalability of the data, ensuring that data from different sources can be effectively integrated.

[0245] This embodiment establishes a complete data transmission mechanism. Configure a secure data interface and establish a reliable connection with the monitoring terminal:

[0246] Connection = establish_connection(endpoint, protocol, auth_token)

[0247] Transfer structured data to a designated storage area through an encrypted channel.

[0248] This embodiment provides a systematic analysis tool for fire safety management. Through evolutionary analysis and data structuring, a comprehensive assessment of fire hazards is achieved. This analysis method not only improves the accuracy of hazard prediction but also enhances the standardization of data management.

[0249] This embodiment achieves an intelligent upgrade in hidden danger management. Through evolutionary prediction and data integration, hidden danger management becomes more scientific and efficient. This intelligent management approach provides strong technical support for fire safety work.

[0250] In order to effectively address the deficiencies of traditional technologies in multimodal data processing, feature recognition, and trend prediction, and significantly improve the intelligent level and early warning capabilities of fire hazard identification, the present application provides an embodiment of a device for identifying fire hazards using AI visual analysis technology, which is used to implement all or part of the method for identifying fire hazards using AI visual analysis technology. Figure 2 The device for identifying fire hazards through AI visual analysis technology specifically includes the following contents:

[0251] The model fusion module 10 is used to collect image and video stream data from the security monitoring system, mobile inspection terminals, and fire inspection robots, obtain scene depth information through a binocular camera to construct a three-dimensional space model, collect infrared thermal imaging data to obtain temperature distribution characteristics, collect environmental perception signals to extract environmental characteristics, preprocess the image and video stream data, the three-dimensional space model, the temperature distribution characteristics, and the environmental characteristics to obtain normalized training data, extract multimodal feature vectors from the normalized training data, annotate the multimodal feature vectors to construct training samples, select a basic feature extraction layer from a pre-trained model library to construct a transfer learning model, input the training samples into the transfer learning model, train a fire hazard feature neural network model group based on a hierarchical classification structure and an attention mechanism, and perform integrated learning on the recognition results of the fire hazard feature neural network model group to obtain a fused recognition model;

[0252] The trend prediction module 20 is configured to receive multi-source data collected in real time, perform time series decomposition and feature extraction on the multi-source data to construct a dynamic feature sequence, input the dynamic feature sequence into the fusion recognition model to perform multimodal feature recognition to obtain hidden danger feature data, and analyze the time evolution trend of the hidden danger feature data based on the dynamic feature sequence to obtain trend prediction data;

[0253] The hidden danger identification module 30 is used to spatially locate the hidden danger feature data to construct hidden danger marking data, construct a hidden danger feature association network based on the hidden danger marking data and the trend prediction data, establish a hidden danger evolution model using the hidden danger feature association network, predict the hidden danger development trend through the hidden danger evolution model, perform cluster analysis on the hidden danger feature data to discover common hidden dangers, and integrate the hidden danger marking data, the hidden danger development trend and the common hidden dangers into structured data and output them to the monitoring terminal.

[0254] From the above description, it can be seen that the device provided in the embodiment of the present application for identifying fire and fire hazards through AI visual analysis technology can integrate multimodal data such as image video streams, three-dimensional space, thermal imaging and acoustic signals by innovatively constructing a multi-source data acquisition and fusion mechanism. A feature extraction model based on transfer learning is designed, combined with a hierarchical classification structure and an attention mechanism, to achieve high-precision hazard feature identification through integrated learning. Time series analysis and spatial positioning technology are introduced to construct a hazard feature association network and evolution model to achieve hazard development trend prediction and common hazard discovery. This method effectively solves the shortcomings of traditional technologies in multimodal data processing, feature recognition and trend prediction, and significantly improves the intelligence level and early warning capabilities of fire and fire hazard identification.

[0255] From a hardware perspective, in order to effectively address the shortcomings of traditional technologies in multimodal data processing, feature recognition, and trend prediction, and significantly improve the intelligent level and early warning capabilities of fire hazard identification, this application provides an embodiment of an electronic device for implementing all or part of the method for identifying fire hazards using AI visual analysis technology. The electronic device specifically includes the following:

[0256] A processor, a memory, a communications interface, and a bus; wherein the processor, memory, and communications interface communicate with each other via the bus; the communications interface is used to transmit information between the device for identifying fire hazards using AI visual analysis technology and related equipment such as core business systems, user terminals, and related databases; the logic controller can be a desktop computer, a tablet computer, a mobile terminal, etc., but this embodiment is not limited thereto. In this embodiment, the logic controller can be implemented with reference to the embodiment of the method for identifying fire hazards using AI visual analysis technology and the embodiment of the device for identifying fire hazards using AI visual analysis technology in the embodiment, the contents of which are incorporated herein and repeated parts are not repeated.

[0257] It is understandable that the user terminal may include a smart phone, a tablet electronic device, a network set-top box, a portable computer, a desktop computer, a personal digital assistant (PDA), a vehicle-mounted device, a smart wearable device, etc. Among them, the smart wearable device may include smart glasses, a smart watch, a smart bracelet, etc.

[0258] In practical applications, part of the method for identifying fire hazards using AI visual analysis technology can be performed on the electronic device side as described above, or all operations can be completed on the client device. The specific selection can be based on the processing capabilities of the client device and the limitations of the user's usage scenario. This application does not limit this. If all operations are completed on the client device, the client device may also include a processor.

[0259] The aforementioned client device may include a communication module (i.e., a communication unit) capable of establishing a communication connection with a remote server to facilitate data transmission with the server. The server may include a server at the task scheduling center or, in other implementation scenarios, a server on an intermediate platform, such as a server on a third-party server platform that is communicatively linked to the task scheduling center server. The server may comprise a single computer device, a server cluster consisting of multiple servers, or a distributed server configuration.

[0260] Figure 3 Schematic block diagram of the system structure of the electronic device 9600 according to an embodiment of the present application. Figure 3 As shown, the electronic device 9600 may include a central processing unit 9100 and a memory 9140; the memory 9140 is coupled to the central processing unit 9100. It is worth noting that the Figure 3 is exemplary; other types of structures may also be used to supplement or replace this structure to implement telecommunication functions or other functions.

[0261] In one embodiment, the method for identifying fire hazards using AI visual analysis technology can be integrated into the central processing unit 9100. The central processing unit 9100 can be configured to perform the following control:

[0262] Step S101: collecting image and video stream data from a security monitoring system, a mobile inspection terminal, and a fire inspection robot, acquiring scene depth information through a binocular camera to construct a three-dimensional space model, collecting infrared thermal imaging data to obtain temperature distribution features, collecting environmental perception signals to extract environmental features, preprocessing the image and video stream data, the three-dimensional space model, the temperature distribution features, and the environmental features to obtain normalized training data, extracting multimodal feature vectors from the normalized training data, annotating the multimodal feature vectors to construct training samples, selecting a basic feature extraction layer from a pre-trained model library to construct a transfer learning model, inputting the training samples into the transfer learning model, training a fire hazard feature neural network model group based on a hierarchical classification structure and an attention mechanism, and performing integrated learning on the recognition results of the fire hazard feature neural network model group to obtain a fusion recognition model;

[0263] Step S102: receiving multi-source data collected in real time, performing time series decomposition and feature extraction on the multi-source data to construct a dynamic feature sequence, inputting the dynamic feature sequence into the fusion recognition model to perform multimodal feature recognition to obtain hidden danger feature data, and analyzing the time evolution trend of the hidden danger feature data based on the dynamic feature sequence to obtain trend prediction data;

[0264] Step S103: spatially locate the hidden danger feature data to construct hidden danger marking data, construct a hidden danger feature association network based on the hidden danger marking data and the trend prediction data, establish a hidden danger evolution model using the hidden danger feature association network, predict the hidden danger development trend through the hidden danger evolution model, perform cluster analysis on the hidden danger feature data to find common hidden dangers, integrate the hidden danger marking data, the hidden danger development trend and the common hidden dangers into structured data and output it to the monitoring terminal.

[0265] As can be seen from the above description, the electronic device provided in the embodiment of the present application integrates multimodal data such as image video streams, three-dimensional space, thermal imaging and acoustic signals by innovatively constructing a multi-source data acquisition and fusion mechanism. A feature extraction model based on transfer learning is designed, combined with a hierarchical classification structure and an attention mechanism, to achieve high-precision hidden danger feature identification through integrated learning. Time series analysis and spatial positioning technology are introduced to construct a hidden danger feature association network and evolution model to achieve hidden danger development trend prediction and common hidden danger discovery. This method effectively solves the shortcomings of traditional technologies in multimodal data processing, feature recognition and trend prediction, and significantly improves the intelligence level and early warning capabilities of fire and fire hazard identification.

[0266] In another embodiment, the device for identifying fire hazards through AI visual analysis technology can be configured separately from the central processing unit 9100. For example, the device for identifying fire hazards through AI visual analysis technology can be configured as a chip connected to the central processing unit 9100, and the method function of identifying fire hazards through AI visual analysis technology can be realized through the control of the central processing unit.

[0267] like Figure 3 As shown, the electronic device 9600 may further include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is worth noting that the electronic device 9600 does not necessarily have to include Figure 3 In addition, the electronic device 9600 may also include all components shown in Figure 3 For components not shown, reference may be made to the prior art.

[0268] like Figure 3 As shown, the central processing unit 9100 is sometimes also referred to as a controller or operation control, and may include a microprocessor or other processor device and / or logic device. The central processing unit 9100 receives input and controls the operation of various components of the electronic device 9600.

[0269] Memory 9140 can be, for example, one or more of a cache, flash memory, hard drive, removable media, volatile memory, non-volatile memory, or other suitable devices. It can store the aforementioned failure-related information and also store programs that execute the relevant information. The CPU 9100 can execute the programs stored in memory 9140 to implement information storage or processing.

[0270] The input unit 9120 provides input to the central processing unit 9100. The input unit 9120 may be, for example, a keypad or touch input device. The power supply 9170 is used to provide power to the electronic device 9600. The display 9160 is used to display objects such as images and text. The display may be, for example, an LCD display, but is not limited thereto.

[0271] The memory 9140 may be a solid-state memory, such as a read-only memory (ROM), random access memory (RAM), or SIM card. Alternatively, it may be a memory that retains information even when power is off, can be selectively erased, and is capable of storing additional data. Examples of such memory are sometimes referred to as EPROMs. The memory 9140 may also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 may include an application / function storage unit 9142 for storing application programs and function programs, or processes used by the central processing unit 9100 to execute operations of the electronic device 9600.

[0272] The memory 9140 may also include a data storage unit 9143 for storing data, such as contacts, digital data, images, sounds, and / or any other data used by the electronic device. The driver storage unit 9144 of the memory 9140 may include various driver programs for communication functions of the electronic device and / or for executing other functions of the electronic device (such as messaging applications, address book applications, etc.).

[0273] The communication module 9110 is a transmitter / receiver that transmits and receives signals via the antenna 9111. The communication module 9110 (transmitter / receiver) is coupled to the central processor 9100 to provide input signals and receive output signals, which may be the same as the case of a conventional mobile communication terminal.

[0274] Based on different communication technologies, multiple communication modules 9110 may be provided in the same electronic device, such as cellular network modules, Bluetooth modules, and / or wireless local area network modules. The communication module 9110 (transmitter / receiver) is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130, providing audio output via the speaker 9131 and receiving audio input from the microphone 9132, thereby implementing common telecommunication functions. The audio processor 9130 may include any suitable buffer, decoder, amplifier, etc. Furthermore, the audio processor 9130 is coupled to the central processing unit 9100, enabling local recording via the microphone 9132 and playback of stored audio via the speaker 9131.

[0275] Embodiments of the present application also provide a computer-readable storage medium capable of implementing all steps of the method for identifying fire hazards using AI visual analysis technology, where the execution subject is a server or a client, in the above-mentioned embodiment. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements all steps of the method for identifying fire hazards using AI visual analysis technology, where the execution subject is a server or a client, in the above-mentioned embodiment. For example, when the processor executes the computer program, the following steps are implemented:

[0276] Step S101: collecting image and video stream data from a security monitoring system, a mobile inspection terminal, and a fire inspection robot, acquiring scene depth information through a binocular camera to construct a three-dimensional space model, collecting infrared thermal imaging data to obtain temperature distribution features, collecting environmental perception signals to extract environmental features, preprocessing the image and video stream data, the three-dimensional space model, the temperature distribution features, and the environmental features to obtain normalized training data, extracting multimodal feature vectors from the normalized training data, annotating the multimodal feature vectors to construct training samples, selecting a basic feature extraction layer from a pre-trained model library to construct a transfer learning model, inputting the training samples into the transfer learning model, training a fire hazard feature neural network model group based on a hierarchical classification structure and an attention mechanism, and performing integrated learning on the recognition results of the fire hazard feature neural network model group to obtain a fusion recognition model;

[0277] Step S102: receiving multi-source data collected in real time, performing time series decomposition and feature extraction on the multi-source data to construct a dynamic feature sequence, inputting the dynamic feature sequence into the fusion recognition model to perform multimodal feature recognition to obtain hidden danger feature data, and analyzing the time evolution trend of the hidden danger feature data based on the dynamic feature sequence to obtain trend prediction data;

[0278] Step S103: spatially locate the hidden danger feature data to construct hidden danger marking data, construct a hidden danger feature association network based on the hidden danger marking data and the trend prediction data, establish a hidden danger evolution model using the hidden danger feature association network, predict the hidden danger development trend through the hidden danger evolution model, perform cluster analysis on the hidden danger feature data to find common hidden dangers, integrate the hidden danger marking data, the hidden danger development trend and the common hidden dangers into structured data and output it to the monitoring terminal.

[0279] As can be seen from the above description, the computer-readable storage medium provided in the embodiment of the present application integrates multimodal data such as image video streams, three-dimensional space, thermal imaging and acoustic signals by innovatively constructing a multi-source data acquisition and fusion mechanism. A feature extraction model based on transfer learning is designed, combined with a hierarchical classification structure and an attention mechanism, to achieve high-precision hidden danger feature identification through integrated learning. Time series analysis and spatial positioning technology are introduced to construct a hidden danger feature association network and evolution model to achieve hidden danger development trend prediction and common hidden danger discovery. This method effectively solves the shortcomings of traditional technologies in multimodal data processing, feature recognition and trend prediction, and significantly improves the intelligence level and early warning capabilities of fire and fire hazard identification.

[0280] The embodiments of the present application also provide a computer program product capable of implementing all steps of the method for identifying fire hazards using AI visual analysis technology in the above-mentioned embodiment, where the execution subject is a server or a client. When the computer program / instructions are executed by a processor, the steps of the method for identifying fire hazards using AI visual analysis technology are implemented. For example, the computer program / instructions implement the following steps:

[0281] Step S101: collecting image and video stream data from a security monitoring system, a mobile inspection terminal, and a fire inspection robot, acquiring scene depth information through a binocular camera to construct a three-dimensional space model, collecting infrared thermal imaging data to obtain temperature distribution features, collecting environmental perception signals to extract environmental features, preprocessing the image and video stream data, the three-dimensional space model, the temperature distribution features, and the environmental features to obtain normalized training data, extracting multimodal feature vectors from the normalized training data, annotating the multimodal feature vectors to construct training samples, selecting a basic feature extraction layer from a pre-trained model library to construct a transfer learning model, inputting the training samples into the transfer learning model, training a fire hazard feature neural network model group based on a hierarchical classification structure and an attention mechanism, and performing integrated learning on the recognition results of the fire hazard feature neural network model group to obtain a fusion recognition model;

[0282] Step S102: receiving multi-source data collected in real time, performing time series decomposition and feature extraction on the multi-source data to construct a dynamic feature sequence, inputting the dynamic feature sequence into the fusion recognition model to perform multimodal feature recognition to obtain hidden danger feature data, and analyzing the time evolution trend of the hidden danger feature data based on the dynamic feature sequence to obtain trend prediction data;

[0283] Step S103: spatially locate the hidden danger feature data to construct hidden danger marking data, construct a hidden danger feature association network based on the hidden danger marking data and the trend prediction data, establish a hidden danger evolution model using the hidden danger feature association network, predict the hidden danger development trend through the hidden danger evolution model, perform cluster analysis on the hidden danger feature data to find common hidden dangers, integrate the hidden danger marking data, the hidden danger development trend and the common hidden dangers into structured data and output it to the monitoring terminal.

[0284] As can be seen from the above description, the computer program product provided in the embodiment of the present application integrates multimodal data such as image video streams, three-dimensional space, thermal imaging and acoustic signals by innovatively constructing a multi-source data acquisition and fusion mechanism. A feature extraction model based on transfer learning is designed, combined with a hierarchical classification structure and an attention mechanism, to achieve high-precision hidden danger feature identification through integrated learning. Time series analysis and spatial positioning technology are introduced to construct a hidden danger feature association network and evolution model to achieve hidden danger development trend prediction and common hidden danger discovery. This method effectively solves the shortcomings of traditional technologies in multimodal data processing, feature recognition and trend prediction, and significantly improves the intelligence level and early warning capabilities of fire and fire hazard identification.

[0285] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, apparatuses, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0286] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (apparatus), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0287] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0288] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0289] Specific embodiments are used in the present invention to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

Claims

1. A method for identifying fire hazards using AI visual analysis technology, characterized in that: The method comprises: Collecting image and video stream data from a security monitoring system, a mobile inspection terminal, and a fire inspection robot, acquiring scene depth information through a binocular camera to construct a three-dimensional space model, collecting infrared thermal imaging data to obtain temperature distribution features, collecting environmental perception signals to extract environmental features, preprocessing the image and video stream data, the three-dimensional space model, the temperature distribution features, and the environmental features to obtain normalized training data, extracting multimodal feature vectors from the normalized training data, annotating the multimodal feature vectors to construct training samples, selecting a basic feature extraction layer from a pre-trained model library to construct a transfer learning model, inputting the training samples into the transfer learning model, training a fire hazard feature neural network model group based on a hierarchical classification structure and an attention mechanism, and performing integrated learning on the recognition results of the fire hazard feature neural network model group to obtain a fusion recognition model; receiving multi-source data collected in real time, performing time series decomposition and feature extraction on the multi-source data to construct a dynamic feature sequence, inputting the dynamic feature sequence into the fusion recognition model to perform multimodal feature recognition to obtain hidden danger feature data, and analyzing the time evolution trend of the hidden danger feature data based on the dynamic feature sequence to obtain trend prediction data; Perform target detection and positioning on the hidden danger feature data to obtain bounding box coordinates, map the bounding box coordinates to a three-dimensional space model to obtain spatial position parameters, calculate the relative distance and azimuth between the spatial position parameters and a preset reference point, integrate the type label of the hidden danger feature data, the spatial position parameters, the relative distance and the azimuth to construct hidden danger marking data, and perform spatial clustering on the hidden danger marking data according to the relative distance; construct a node feature matrix based on the hidden danger marking data, construct a node transfer probability matrix using the trend prediction data, combine the node feature matrix and the node transfer probability matrix into a hidden danger feature association network, calculate the association strength and propagation path between nodes in the hidden danger feature association network, construct a state transfer equation based on the association strength and the propagation path, establish a hidden danger evolution model based on the state transfer equation, predict the hidden danger development trend through the hidden danger evolution model, perform cluster analysis on the hidden danger feature data to discover common hidden dangers, integrate the hidden danger marking data, the hidden danger development trend and the common hidden dangers into structured data and output it to the monitoring terminal.

2. The method for identifying fire hazards using AI visual analysis technology according to claim 1, characterized in that: The step of extracting a multimodal feature vector from the normalized training data, annotating the multimodal feature vector to construct a training sample, selecting a basic feature extraction layer from a pre-trained model library to construct a transfer learning model, and inputting the training sample into the transfer learning model includes: Performing convolution decomposition and pooling operations on the normalized training data to extract image texture features, extracting three-dimensional spatial features through a deep image analysis module, obtaining temperature gradient features using an infrared feature extraction module, extracting acoustic features based on a spectrogram analysis module, fusing the image texture features, the three-dimensional spatial features, the temperature gradient features, and the acoustic features to obtain a multimodal feature vector, normalizing the feature components of the multimodal feature vector and adding time tags, and annotating the multimodal feature vector with categories and attributes based on manual annotation rules to construct training samples; Perform feature layer analysis on the models in the pre-trained model library, select a representative basic feature extraction layer according to the application scenario, copy the weight parameters and network structure of the basic feature extraction layer to the target network to build a transfer learning model, adjust the input layer and output layer of the transfer learning model to adapt them to the data format of the training samples, and input the training samples into the adjusted transfer learning model for feature mapping.

3. The method for identifying fire hazards using AI visual analysis technology according to claim 1, characterized in that: The fire hazard feature neural network model group is trained based on the hierarchical classification structure and the attention mechanism, and the recognition results of the fire hazard feature neural network model group are integrated and learned to obtain a fusion recognition model, including: A hierarchical classification structure is constructed based on a preset fire hazard classification system. Hazard types are divided into multiple subcategory nodes according to physical characteristics. A corresponding feature recognition model is assigned to each subcategory node. A self-attention calculation unit and a cross-modal attention calculation unit are embedded in the feature recognition model. The self-attention calculation unit is used to weight the importance of feature vectors. The cross-modal attention calculation unit is used to calculate the association weights between different modal features. The training samples are input into each feature recognition model for back-propagation training to obtain a fire hazard feature neural network model group. Calculate the recognition accuracy and confidence score of each model in the fire hazard characteristic neural network model group, assign a fusion weight coefficient to each model based on the recognition accuracy and confidence score, perform a weighted combination of the recognition results of the model group and the corresponding fusion weight coefficient, perform decision fusion on the weighted combination results through a voting mechanism to obtain a fusion recognition model, and iteratively optimize the fusion weight coefficient of the fusion recognition model.

4. The method for identifying fire hazards using AI visual analysis technology according to claim 1, characterized in that: The receiving of multi-source data collected in real time, performing time series decomposition and feature extraction on the multi-source data to construct a dynamic feature sequence, and inputting the dynamic feature sequence into the fusion recognition model to perform multimodal feature recognition to obtain hidden danger feature data, includes: Receive data streams collected by multiple sensors, divide the data streams into time windows to obtain data segments, align and synchronize the data segments according to timestamps, perform wavelet transform on the synchronized data to obtain time-frequency components, extract amplitude and phase features of the time-frequency components, sequentially concatenate the amplitude and phase features based on a sliding window method to construct a dynamic feature sequence, and perform normalization and noise reduction filtering on the dynamic feature sequence; The dynamic feature sequence is parsed into a feature tensor according to a preset time step, the feature tensor is channel-separated to obtain multiple feature components, the feature components are input into the corresponding input ends of the fusion recognition model, the recognition result and confidence of each feature component are calculated by the fusion recognition model, the recognition results are screened and combined according to the confidence to obtain hidden danger feature data, and the hidden danger feature data is subjected to time series correlation analysis.

5. The method for identifying fire hazards using AI visual analysis technology according to claim 1, characterized in that: The step of analyzing the time evolution trend of the hidden danger feature data based on the dynamic feature sequence to obtain trend prediction data includes: Constructing a time series sample matrix from the dynamic feature sequence and the hidden danger feature data according to the time dimension, performing singular value decomposition on the time series sample matrix to extract principal component features, calculating the time increment and change rate of the principal component features, performing trend decomposition on the principal component features based on the change rate to obtain a periodic component and a trend component, reconstructing the periodic component and the trend component into a time series curve, and performing piecewise fitting on the time series curve to calculate fitting parameters; A state prediction model is constructed using the fitting parameters, and historical time series data is input into the state prediction model for parameter training. The development trend of hidden danger characteristics is extrapolated based on the trained state prediction model. According to the extrapolation calculation results, a probability assessment of the hidden danger state at the predicted time point is performed to obtain trend prediction data, and the credibility of the trend prediction data is quantitatively assessed.

6. The method for identifying fire hazards using AI visual analysis technology according to claim 1, characterized in that: The method includes predicting the hidden danger development trend by using the hidden danger evolution model, performing cluster analysis on the hidden danger feature data to discover common hidden dangers, and integrating the hidden danger marking data, the hidden danger development trend, and the common hidden dangers into structured data and outputting it to the monitoring terminal, including: Numerically solving the state transition equation of the hidden danger evolution model to obtain an evolution trajectory, calculating the development direction and change rate of the evolution trajectory, extrapolating the evolution trajectory according to a preset time interval to obtain a hidden danger state prediction value, comparing the hidden danger state prediction value with the threshold value at the corresponding time point to obtain the hidden danger development trend, performing density cluster analysis on the hidden danger feature data to obtain feature clusters, calculating the center of gravity position and distribution range of the feature clusters, and identifying common hidden dangers based on the number and size of the feature clusters; The hidden danger marking data is field mapped according to a preset data structure format, the hidden danger development trend data is time-serialized, the common hidden danger data is converted into an attribute relationship table, and the field mapping data, the time-serialized data and the attribute relationship table are integrated into structured data based on a unified data template, and the structured data is transmitted to the storage area specified by the monitoring terminal through a data interface.

7. A device for identifying fire hazards using AI visual analysis technology, characterized in that: The device comprises: A model fusion module is used to collect image and video stream data from security monitoring systems, mobile inspection terminals, and fire inspection robots, obtain scene depth information through a binocular camera to construct a three-dimensional space model, collect infrared thermal imaging data to obtain temperature distribution characteristics, collect environmental perception signals to extract environmental characteristics, preprocess the image and video stream data, the three-dimensional space model, the temperature distribution characteristics, and the environmental characteristics to obtain normalized training data, extract multimodal feature vectors from the normalized training data, annotate the multimodal feature vectors to construct training samples, select a basic feature extraction layer from a pre-trained model library to construct a transfer learning model, input the training samples into the transfer learning model, train a fire hazard feature neural network model group based on a hierarchical classification structure and an attention mechanism, and perform integrated learning on the recognition results of the fire hazard feature neural network model group to obtain a fused recognition model; a trend prediction module, configured to receive multi-source data collected in real time, perform time series decomposition and feature extraction on the multi-source data to construct a dynamic feature sequence, input the dynamic feature sequence into the fusion recognition model to perform multimodal feature recognition to obtain hidden danger feature data, and analyze the time evolution trend of the hidden danger feature data based on the dynamic feature sequence to obtain trend prediction data; A hidden danger identification module is used to perform target detection and positioning on the hidden danger feature data to obtain bounding box coordinates, map the bounding box coordinates into a three-dimensional space model to obtain spatial position parameters, calculate the relative distance and azimuth between the spatial position parameters and a preset reference point, integrate the type label of the hidden danger feature data, the spatial position parameters, the relative distance and the azimuth to construct hidden danger marking data, and spatially cluster the hidden danger marking data according to the relative distance; construct a node feature matrix based on the hidden danger marking data, construct a node transfer probability matrix using the trend prediction data, combine the node feature matrix and the node transfer probability matrix into a hidden danger feature association network, calculate the association strength and propagation path between nodes in the hidden danger feature association network, construct a state transfer equation based on the association strength and the propagation path, establish a hidden danger evolution model based on the state transfer equation, predict the hidden danger development trend through the hidden danger evolution model, perform cluster analysis on the hidden danger feature data to discover common hidden dangers, integrate the hidden danger marking data, the hidden danger development trend and the common hidden dangers into structured data and output it to the monitoring terminal.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method for identifying fire hazards through AI visual analysis technology as described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for identifying fire hazards by using AI visual analysis technology as described in any one of claims 1 to 6 are implemented.

Images