AI-based three-light image fire detector

Abstract

The present invention provides an AI-based three-light image type fire detector, which relates to the technical field of data processing and includes: a flame spectrum acquisition module, which is used to obtain multi-spectral image data of a monitoring area; a preprocessing module, which is used to preprocess the multi-spectral image data and generate multi-spectral data; a feature extraction module, which is used to generate multi-dimensional feature data according to the multi-spectral data; an environmental condition acquisition module, which is used to collect environmental condition data of the monitoring area; an AI classification module, which is used to perform layer-by-layer analysis and classification processing on the multi-dimensional feature data by a multi-layer neural network model to generate a preliminary fire identification result; a background adaptation module, which is used to receive the preliminary fire identification result and the environmental condition data, and process the preliminary fire identification result to obtain fire identification data; an alarm triggering module, which is used to receive the fire identification data and determine whether the preset fire alarm conditions are met; the present invention has a good detection effect.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to an AI-based three-light image-type fire detector. Background Art

[0002] Among existing fire detection technologies, image-based fire detectors usually identify fires by analyzing image features such as flames, smoke, and temperature captured by cameras. This type of detector mainly relies on image processing algorithms and simple threshold judgment models to establish preset standards for different types of fire features. When image features matching the shape of flames or smoke appear in the monitoring screen, an alarm is triggered. This type of detection method is widely used in indoor and some outdoor fire prevention places, but when dealing with complex scenes in multiple environments, false alarms are often generated due to insufficient adaptability to dynamic environmental changes. For example, in a factory workshop, false alarms may occur due to the similar shapes of smoke and flames generated by equipment heating or workers operating.

[0003] In addition, traditional fire detection technology is weak in identifying different types of fires in complex scenarios. Most traditional image-based detectors lack the fusion processing function of multi-spectral images and can only capture the apparent characteristics of flames or smoke under a single spectrum. This may result in the characteristics of flames not being obvious when the smoke is light or obscured, limiting the detection effect. Summary of the invention

[0004] The purpose of the present invention is to provide an AI-based three-light image-type fire detector, aiming to solve the problems mentioned in the background technology.

[0005] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0006] AI-based three-light image-based fire detector, including:

[0007] The flame spectrum acquisition module is used to obtain multi-spectral image data of the monitoring area, wherein the multi-spectral image data includes infrared spectrum, ultraviolet spectrum and visible spectrum data;

[0008] An image preprocessing module is used to perform denoising, edge enhancement and light intensity smoothing on multispectral image data to generate multispectral data;

[0009] A feature extraction module is used to receive multispectral data, divide it into several analysis areas, and independently extract feature information in each analysis area to generate partition feature data, integrate the partition feature data to obtain multidimensional feature data, and the multidimensional feature data includes the shape, edge, color and texture features of flame and smoke;

[0010] Environmental condition acquisition module, used to collect environmental condition data of the monitored area, including temperature, humidity, air flow velocity and air composition;

[0011] AI classification module, which uses a pre-trained multi-layer neural network model to perform layer-by-layer analysis and classification of multi-dimensional feature data, identify multiple feature combinations of flames or smoke, and generate preliminary fire identification results;

[0012] The background adaptation module is used to receive the preliminary fire identification results and environmental condition data, and perform background noise elimination and feature optimization processing on the preliminary fire identification results to obtain fire identification data; the environmental condition data includes real-time acquisition of temperature, humidity, air flow velocity and air composition change data in the monitoring area;

[0013] The alarm trigger module is used to receive fire identification data and determine whether the preset fire alarm conditions are met; when the fire identification data meets the preset fire alarm conditions, an alarm signal is output.

[0014] Preferably, the flame spectrum acquisition module further includes:

[0015] A spectral image acquisition unit, used to acquire infrared spectrum, ultraviolet spectrum and visible spectrum data of the monitored area;

[0016] A light source identification unit is used to identify the type of light source in the monitoring area by detecting the wavelength and spectral characteristics of the light source in the environment, and generate light source classification data;

[0017] A background analysis unit is used to detect background spectrum changes of the multi-spectral image data at different time intervals, generate spectrum change trend data, and generate a spectrum adjustment signal according to the spectrum change trend data;

[0018] The spectrum adjustment unit is used to adjust the acquisition parameters of infrared spectrum, ultraviolet spectrum and visible spectrum data in real time according to the light source classification data, the spectrum change trend data and the spectrum adjustment signal to obtain multi-spectral image data. The adjustment formula of the multi-spectral image data is as follows:

[0019] + + ,

[0020] in, is the adjusted multispectral image data, , , To represent infrared, ultraviolet and visible light image data respectively, is the wavelength shift of the infrared spectrum, is the weight parameter, is the wavelength shift of the ultraviolet spectrum, To adjust the constant of the UV spectrum sensitivity to wavelength shift, is the wavelength shift of the visible spectrum, is the maximum wavelength of the visible light spectrum, is the circumference constant, is the light source brightness, , , , , , is a constant parameter.

[0021] Preferably, the image preprocessing module further includes:

[0022] The multidimensional noise filtering unit is used to suppress the dynamic noise of the multispectral image data by using the spatiotemporal filtering algorithm and combining the time series characteristics of the image to generate time series image data; the denoising formula is as follows:

[0023] ,

[0024] in, is the pixel value after denoising, For the time series The pixel value of the frame, is the Gaussian spatiotemporal filter weight function, Controls the width of the filter, is the total number of filtering frames;

[0025] The edge saliency enhancement unit is used to identify and enhance the edge saliency features of flames and smoke in different spectra according to the time series image data, and generate saliency image data; the edge saliency feature enhancement formula is as follows:

[0026] ,

[0027] in, is the image after edge enhancement, is the time series image data, is the Laplace operator;

[0028] The adaptive light intensity smoothing unit performs adaptive smoothing adjustment on the light intensity change of the saliency image according to the contrast difference between the background light intensity of the monitoring area and the target spectral data to generate multi-spectral data; the smoothing formula is as follows:

[0029] ,

[0030] in, is the pixel value after light intensity smoothing, For location and location The spatial distance For the location The pixel value of the image after edge enhancement at and are the control coefficients of spatial distance and pixel difference respectively.

[0031] Preferably, the feature extraction module comprises:

[0032] A partition mapping unit is used to perform spatial mapping partition processing on multispectral data, divide the image into multiple spatial analysis areas, and generate corresponding partition mapping data for each area;

[0033] The morphology distribution extraction unit is used to analyze the flame and smoke morphology in each analysis area, calculate its edge curvature, shape contour and structural complexity, generate regional morphological feature data, and integrate them to obtain global morphological distribution data; the edge curvature, shape contour and structural complexity are calculated by the following formulas:

[0034] ,in, is the edge curvature, which is used to measure the morphological changes of the edge. is the symbol of partial derivative;

[0035] ,in, is the fractal dimension, which indicates the complexity of the shape outline. The minimum size required to cover the edge is The number of units, The scale for edge measurement;

[0036] ,in, is the structural complexity, which indicates the complexity inside the shape. For the The probability of occurrence of a structural feature, is the total number of structural features;

[0037] A feature comparison and analysis unit is used to compare and analyze the color difference, edge clarity and texture characteristics of flames and smoke in each analysis area according to the partition mapping data, and generate a regional feature matrix;

[0038] The feature integration unit is used to integrate the morphological distribution, color difference, edge clarity and texture characteristics of flames and smoke according to the global morphological distribution data and the regional feature matrix to generate multi-dimensional feature data.

[0039] Preferably, the AI ​​classification module further includes:

[0040] A feature data hierarchical processing unit, used for receiving multi-dimensional feature data and inputting the multi-dimensional feature data layer by layer into a pre-trained multi-layer neural network model for feature analysis;

[0041] An adaptive feature enhancement unit, used to dynamically adjust the weight of the intermediate classification data generated by the feature data hierarchical processing unit according to the environmental condition data, so as to obtain enhanced feature data;

[0042] The model iteration tuning unit is used to adjust the learning rate and bias value of each layer of the neural network according to the enhanced feature data.

[0043] Preferably, the adaptive feature enhancement unit comprises:

[0044] The multi-layer feature input unit is used to receive multi-dimensional feature data and input it layer by layer into the first layer of the multi-layer neural network model; the multi-layer feature input unit performs preliminary normalization processing on the input data;

[0045] The layer-by-layer feature extraction unit is used to extract the morphological distribution, color difference, edge definition and texture features of flames and smoke in each layer of the multi-layer neural network model; the layer-by-layer feature extraction unit generates intermediate feature data according to the input features of the current layer, and identifies the key feature combination related to flames and smoke in each layer to generate inter-layer correlation data;

[0046] The inter-layer feature association unit is used to transfer and integrate inter-layer association data between layers of the multi-layer neural network; the inter-layer feature association unit aggregates the association feature data according to the performance differences of flame and smoke features at different levels;

[0047] The hierarchical feature output unit is used to integrate the finally generated multi-layer associated feature data into comprehensive intermediate classification data and output it to the adaptive feature enhancement unit.

[0048] Preferably, the adaptive feature enhancement unit comprises:

[0049] An environmental condition analysis unit, used to receive environmental condition data, analyze the changing trends of temperature, humidity, air flow velocity and air composition, and generate environmental impact parameters;

[0050] A feature weight adjustment unit, used to adjust the weight of the multi-dimensional feature data of flame and smoke according to the environmental impact parameters;

[0051] ,in, After adjustment feature weights, For the original feature weights, is the temperature, For humidity, , , , , , is the weight coefficient;

[0052] A dynamic weight updating unit, used to monitor the output of the feature weight adjustment unit in real time and dynamically update the feature weight;

[0053] The weight fusion unit is used to apply the dynamically updated weights to the multi-dimensional feature data, distribute the adjusted feature weights to the feature layers such as morphology, color, edge and texture, and generate enhanced feature data.

[0054] Preferably, the background adaptation module comprises:

[0055] A multi-dimensional background elimination unit is used to eliminate background noise data in the preliminary fire identification result according to the environmental condition data set to generate fire characteristic data;

[0056] The characteristic dynamic balancing unit is used to adjust the weight of fire characteristic data under different environmental conditions and generate fire identification data.

[0057] Preferably, the alarm triggering module includes:

[0058] A credibility evaluation unit is used to calculate the confidence of the fire identification data and use the confidence as the basis for determining the trigger priority;

[0059] A multi-level trigger condition determination unit is used to set trigger conditions of different priorities. When the fire identification data meets the priority conditions, an alarm signal of the corresponding level is generated;

[0060] The multi-channel alarm output unit is used to output alarm signals through sound, light, SMS, and network channels according to the alarm level when the trigger conditions are met.

[0061] Preferably, the multi-level trigger condition determination unit further includes:

[0062] A low priority determination unit, used to determine whether to generate a low priority alarm signal according to the frequency and duration of changes in flame or smoke characteristics when the confidence level of fire identification data is lower than a preset threshold range;

[0063] A medium priority determination unit, used to generate a medium priority alarm signal according to the composite characteristics of flame and smoke features under multiple spectra when the confidence level of the fire identification data is within a preset threshold range;

[0064] The high priority determination unit is used to generate a high priority alarm signal and trigger a global alarm device when the confidence level of the fire identification data is higher than a preset threshold range.

[0065] The above solution of the present invention includes at least the following beneficial effects:

[0066] First, by integrating multispectral image data of infrared, ultraviolet and visible spectra, it can provide more comprehensive and accurate fire characteristic information than traditional single-spectrum or dual-spectrum fire detectors. The infrared spectrum can capture the thermal radiation characteristics of flames, the ultraviolet spectrum is particularly sensitive to the detection of short-wave flame characteristics, and the visible spectrum can provide the color and morphological characteristics of flames or smoke. The combined application of multiple spectra ensures that fire detectors can obtain stable and comprehensive fire characteristic data under various light sources and environmental conditions. This multispectral detection method effectively reduces false alarms caused by ambient light interference or changes in light source types under complex light sources, thereby greatly improving the accuracy and stability of fire detection.

[0067] Secondly, this fire detector ensures the clarity and stability of image features through the multiple processing functions of the image preprocessing module, including denoising, edge enhancement and light intensity smoothing. Traditional fire detectors often cause false alarms due to noise interference, especially in dynamic environments such as factories or warehouses, where background noise is often mistaken for fire features. This device effectively removes random noise in the monitored area through denoising, improving the recognition of flames and smoke; the edge enhancement function ensures the clarity of the flame boundary, making the fire features more prominent against a complex background; and light intensity smoothing can cope with the uneven intensity caused by changes in illumination, ensuring that the image maintains consistent performance under different lighting conditions such as strong light, weak light or shadow. The combination of these preprocessing functions makes fire detection more reliable in complex environments and can effectively improve the recognition accuracy of fire features.

[0068] In addition, the feature extraction module performs partition processing and multidimensional feature analysis on multispectral data, allowing the fire detector to extract multidimensional features of flames and smoke from different areas, including morphology, edges, colors, and textures. This multidimensional feature extraction method far exceeds the traditional simple morphology or color analysis, and can more accurately build a fire feature model, thereby improving the comprehensiveness and accuracy of fire identification. Especially when the flame features are not obvious or the smoke is thin, this module improves the sensitivity and recognition effect of detection through the detailed analysis of multidimensional features.

[0069] The environmental condition acquisition module further enhances the adaptability of fire detectors in complex environments. By collecting data such as temperature, humidity, air velocity and air composition in real time, the system can dynamically adjust the detection sensitivity to avoid detection deviations caused by changes in environmental conditions.

[0070] Finally, the combination of the AI ​​classification module and the background adaptation module realizes the fire feature classification and background optimization based on a multi-layer neural network. In traditional fire detection, simple threshold judgment or rule recognition is difficult to cope with the diverse features of complex scenes, while this device analyzes fire features layer by layer through a deep learning model and has the ability of adaptive learning. In addition, the background adaptation module uses real-time environmental data to remove background noise and optimize features of the preliminary recognition results, effectively reducing false alarms and ensuring the accuracy and stability of fire recognition. This multi-module combined fire detection method significantly improves the robustness and reliability of fire recognition, can adapt to changing environments, and achieve more accurate early warning of fires. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is an architecture diagram of an AI-based three-light image-type fire detector provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0072] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0073] like Figure 1 As shown, an embodiment of the present invention proposes an AI-based three-light image type fire detector, the fire detector comprising:

[0074] The flame spectrum acquisition module is used to obtain multi-spectral image data of the monitoring area, wherein the multi-spectral image data includes infrared spectrum, ultraviolet spectrum and visible spectrum data;

[0075] An image preprocessing module is used to perform denoising, edge enhancement and light intensity smoothing on multispectral image data to generate multispectral data;

[0076] A feature extraction module is used to receive multispectral data, divide it into several analysis areas, and independently extract feature information in each analysis area to generate partition feature data, integrate the partition feature data to obtain multidimensional feature data, and the multidimensional feature data includes the shape, edge, color and texture features of flame and smoke;

[0077] Environmental condition acquisition module, used to collect environmental condition data of the monitored area, including temperature, humidity, air flow velocity and air composition;

[0078] AI classification module, which uses a pre-trained multi-layer neural network model to perform layer-by-layer analysis and classification of multi-dimensional feature data, identify multiple feature combinations of flames or smoke, and generate preliminary fire identification results;

[0079] The background adaptation module is used to receive the preliminary fire identification results and environmental condition data, and perform background noise elimination and feature optimization processing on the preliminary fire identification results to obtain fire identification data; the environmental condition data includes real-time acquisition of temperature, humidity, air flow velocity and air composition change data in the monitoring area;

[0080] The alarm trigger module is used to receive fire identification data and determine whether the preset fire alarm conditions are met; when the fire identification data meets the preset fire alarm conditions, an alarm signal is output.

[0081] In an embodiment of the present invention, the flame spectrum acquisition module acquires multi-spectral image data of the monitored area, including infrared spectrum, ultraviolet spectrum and visible spectrum data, so that the detector can obtain comprehensive and multi-dimensional image information. The infrared spectrum is used to sense the thermal radiation of the flame, the ultraviolet spectrum is used to capture the short-wave flame characteristics, and the visible spectrum can provide color and brightness information. This combined application of multiple spectra not only enhances the accuracy of flame detection, but also effectively reduces false alarms under different light sources. Traditional single-spectrum or dual-spectrum fire detection methods are usually unable to accurately distinguish between flames and noise sources when dealing with complex light sources or changing environments. The present device greatly improves the accuracy of detection by combining three spectra.

[0082] After receiving the multispectral image data from the flame spectrum acquisition module, the image preprocessing module first performs denoising, edge enhancement and light intensity smoothing to generate more characteristic multispectral data. In traditional fire detection technology, noise interference often leads to false alarms, while the image preprocessing module in this device effectively eliminates the interference of environmental noise through a denoising algorithm. Edge enhancement processing can highlight the boundary characteristics of flames and smoke, ensuring the prominence of flames or smoke in the image. Adaptive light intensity smoothing can smoothly process image data according to the intensity changes of background light, improving the adaptability of the device in complex lighting environments, thereby ensuring that fire features are still clearly visible under changing lighting conditions.

[0083] The feature extraction module further analyzes the pre-processed multispectral data, and generates partition feature data by dividing the image into several analysis areas and independently extracting the features of each area. The module extracts features including the morphology, edges, colors, and textures of flames and smoke, and integrates the partition feature data to obtain the final multidimensional feature data. In traditional fire detectors, feature extraction is usually limited to a single or limited feature dimension, and cannot effectively cope with the changes in complex flame and smoke features. The multidimensional feature extraction method of this device can obtain fire feature information more comprehensively, providing a more accurate data basis for subsequent classification.

[0084] The environmental condition acquisition module provides key environmental reference data for fire detection by real-time acquisition of temperature, humidity, air flow velocity and air composition changes in the monitoring area. Traditional detectors often ignore the impact of environmental data, which affects the detection performance when the environment changes drastically. This device can adjust the sensitivity of fire detection in time according to environmental changes through the collection and analysis of environmental data, thereby improving the detection efficiency of the early stage of fire.

[0085] After receiving the multi-dimensional feature data, the AI ​​classification module uses a pre-trained multi-layer neural network model to analyze and classify the fire features layer by layer, identifying multiple feature combinations of flames or smoke. Unlike traditional simple threshold or rule judgments, the AI ​​classification module of this device has adaptive learning capabilities and can optimize the recognition effect based on different fire feature combinations. The module ultimately generates preliminary fire recognition results that significantly improve the reliability of fire detection.

[0086] The background adaptation module receives the preliminary fire identification results and environmental condition data, removes background noise and optimizes features on the preliminary results to generate fire identification data. The application of this module effectively solves the problem of false alarms caused by background interference, ensures the accuracy and effectiveness of fire identification data, and improves the stability and applicability of detection.

[0087] The alarm trigger module determines whether the fire identification data meets the preset alarm conditions, and outputs an alarm signal after confirming that the fire identification data meets the alarm conditions. The application of this module ensures that the system can respond immediately when a fire occurs, provides reliable protection for fire alarms, and meets the needs of early fire detection and rapid alarms.

[0088] In a preferred embodiment of the present invention, the flame spectrum acquisition module further includes:

[0089] A spectral image acquisition unit, used to acquire infrared spectrum, ultraviolet spectrum and visible spectrum data of the monitored area;

[0090] A light source identification unit is used to identify the type of light source in the monitoring area by detecting the wavelength and spectral characteristics of the light source in the environment, and generate light source classification data;

[0091] A background analysis unit is used to detect background spectrum changes of the multi-spectral image data at different time intervals, generate spectrum change trend data, and generate a spectrum adjustment signal according to the spectrum change trend data;

[0092] The spectrum adjustment unit is used to adjust the acquisition parameters of infrared spectrum, ultraviolet spectrum and visible spectrum data in real time according to the light source classification data, spectrum change trend data and spectrum adjustment signal to obtain multi-spectral image data. The adjustment formula of the multi-spectral image data is as follows:

[0093] + + ,

[0094] in, is the adjusted multispectral image data, , , To represent infrared, ultraviolet and visible light image data respectively, is the wavelength shift of the infrared spectrum, is the weight parameter, is the wavelength shift of the ultraviolet spectrum, To adjust the constant of the UV spectrum sensitivity to wavelength shift, is the wavelength shift of the visible spectrum, is the maximum wavelength of the visible light spectrum, is the circumference constant, is the light source brightness, , , , , , is a constant parameter.

[0095] In the embodiment of the present invention, a spectral image acquisition unit, a light source identification unit, a background analysis unit and a spectrum adjustment unit are provided. The spectral image acquisition unit can accurately acquire infrared, ultraviolet and visible spectrum data, thereby ensuring that the collected spectrum information is comprehensive and accurate. The light source identification unit identifies the light source type by detecting the wavelength and spectral characteristics of the light source, generates light source classification data, and performs a more accurate analysis of fire characteristics under different light source conditions. The background analysis unit detects the background spectrum changes of the multi-spectral image at different time intervals, generates spectrum change trend data, thereby improving the accuracy of flame and smoke recognition.

[0096] The spectrum adjustment unit adjusts the acquisition parameters of infrared, ultraviolet and visible spectra in real time according to the light source classification data, spectrum change trend data and spectrum adjustment signal, ensuring the acquisition of the best multi-spectral image data in various environments. For example, when the background light changes greatly, the spectrum adjustment unit will automatically adjust the acquisition parameters to avoid the influence of light source changes on fire feature recognition. This dynamic adjustment greatly improves the applicability and accuracy of fire detection.

[0097] The use of adjustment formulas further optimizes multispectral data. In the formula, parameters and functions are used to dynamically adjust the proportion of infrared, ultraviolet and visible light data to ensure that the weight and response of each spectrum are optimal under different light source intensities and wavelength offsets, thereby improving the clarity of multispectral images and the robustness of fire identification.

[0098] In a preferred embodiment of the present invention, the image preprocessing module further includes:

[0099] The multidimensional noise filtering unit is used to suppress the dynamic noise of the multispectral image data by using the spatiotemporal filtering algorithm and combining the time series characteristics of the image to generate time series image data; the denoising formula is as follows:

[0100] ,

[0101] in, is the pixel value after denoising, For the time series The pixel value of the frame, is the Gaussian spatiotemporal filter weight function, Controls the width of the filter, is the total number of filtering frames;

[0102] The edge saliency enhancement unit is used to identify and enhance the edge saliency features of flames and smoke in different spectra according to the time series image data, and generate saliency image data; the edge saliency feature enhancement formula is as follows:

[0103] ,

[0104] in, is the image after edge enhancement, is the time series image data, is the Laplace operator;

[0105] The adaptive light intensity smoothing unit performs adaptive smoothing adjustment on the light intensity change of the saliency image according to the contrast difference between the background light intensity of the monitoring area and the target spectral data to generate multi-spectral data; the smoothing formula is as follows:

[0106] ,

[0107] in, is the pixel value after light intensity smoothing, For location and location The spatial distance For the location The pixel value of the image after edge enhancement at and are the control coefficients of spatial distance and pixel difference respectively.

[0108] In the embodiment of the present invention, by adding a multidimensional noise filtering unit, an edge significance enhancement unit and an adaptive light intensity smoothing unit, the fire feature extraction is made more accurate. The multidimensional noise filtering unit uses a spatiotemporal filtering algorithm to suppress dynamic noise according to the time series characteristics of the image and generate time series image data. Compared with traditional spatial filtering, this spatiotemporal filtering method can more effectively retain the time series characteristics of flames and smoke, thereby accurately identifying fire features in dynamic scenes. The denoising formula uses Gaussian weights and combines multi-frame information to remove background noise to ensure image quality in dynamic fire scenes.

[0109] The edge saliency enhancement unit performs Laplace operator processing on the edges of the time series image data to highlight the edge contours of the flame and smoke. This processing method can significantly enhance the clarity of the edges, effectively distinguish the boundaries between flames and smoke and the background, and reduce false alarms caused by blurred boundaries. In addition, the adaptive light intensity smoothing unit uses a bilateral filtering algorithm to adaptively smooth the light intensity of the image to reduce the impact of different lighting conditions on fire identification. For example, in an environment with alternating strong light and shadows, adaptive smoothing can eliminate the uneven intensity in the image, making the performance of fire features in multispectral images more stable and clear.

[0110] The introduction of formulas makes the functions of each processing module more precise. The denoising formula achieves multi-frame noise suppression through time series filtering; the edge saliency enhancement formula strengthens the boundary features through the Laplace operator; and the light intensity smoothing formula achieves adaptive light intensity regulation through bilateral filtering. These formulas not only enhance the stability of fire characteristics, but also enable the detector to have higher recognition accuracy in complex lighting and noise environments.

[0111] Specifically, the Laplacian operator is a mathematical operator commonly used in image processing to detect edges in images. It is based on the concept of second-order derivatives and calculates the degree of change in pixel brightness in an image, thereby highlighting areas of the image that have undergone significant changes (i.e., edges). In applications such as fire detectors, the Laplacian operator can help accurately extract boundary features of flames and smoke. The Laplacian operator can quickly identify areas in an image where there are drastic changes in brightness, such as object boundaries. This is particularly important for fire detectors, as the edges of flames and smoke are key identification features.

[0112] In a preferred embodiment of the present invention, the feature extraction module includes:

[0113] A partition mapping unit is used to perform spatial mapping partition processing on multispectral data, divide the image into multiple spatial analysis areas, and generate corresponding partition mapping data for each area;

[0114] The morphology distribution extraction unit is used to analyze the flame and smoke morphology in each analysis area, calculate its edge curvature, shape contour and structural complexity, generate regional morphological feature data, and integrate them to obtain global morphological distribution data; the edge curvature, shape contour and structural complexity are calculated by the following formulas:

[0115] ,in, is the edge curvature, which is used to measure the morphological changes of the edge. is the symbol of partial derivative;

[0116] ,in, is the fractal dimension, which indicates the complexity of the shape outline. The minimum size required to cover the edge is The number of units, The scale for edge measurement;

[0117] ,in, is the structural complexity, which indicates the complexity inside the shape. For the The probability of occurrence of a structural feature, is the total number of structural features;

[0118] A feature comparison and analysis unit is used to compare and analyze the color difference, edge clarity and texture characteristics of flames and smoke in each analysis area according to the partition mapping data, and generate a regional feature matrix;

[0119] The feature integration unit is used to integrate the morphological distribution, color difference, edge clarity and texture characteristics of flames and smoke according to the global morphological distribution data and the regional feature matrix to generate multi-dimensional feature data.

[0120] In an embodiment of the present invention, the feature extraction module includes a partition mapping unit, a morphological distribution extraction unit, a feature comparison analysis unit and a feature integration unit, and the accurate extraction of fire features is achieved by partitioning the multispectral data. The partition mapping unit divides the multispectral image data into several spatial regions, which can separate the local features of the flame and smoke and generate partition mapping data. The morphological distribution extraction unit performs a more detailed feature analysis on each partition, calculates the edge curvature, shape contour and structural complexity of the flame and smoke. The edge curvature formula calculates the degree of edge change, which helps to identify the boundary of complex flames; the shape contour complexity formula quantifies the self-similarity of the flame boundary based on the fractal dimension, and the structural complexity formula can describe the multi-level details inside the flame based on information entropy, so that the detector can more accurately identify the fire characteristics.

[0121] The feature comparison and analysis unit compares and analyzes the color, edge clarity and texture features in the partition mapping data to generate a regional feature matrix. By integrating the features of each region, this matrix can effectively identify the differences between flames and smoke under different spectra. Finally, the feature integration unit comprehensively integrates the multidimensional feature data of flames and smoke based on the global morphological distribution data and the regional feature matrix, thereby improving the comprehensiveness and accuracy of fire identification. Compared with traditional technologies, this feature extraction module can build a more sophisticated fire model based on multidimensional feature data, achieving efficient identification of complex fire scenes.

[0122] The application of the formula further improves the accuracy of feature extraction. The edge curvature, shape contour complexity and structural complexity formulas of the morphological distribution extraction unit can fully describe the edge characteristics and internal structure of the flame, effectively improving the accuracy and robustness of fire identification.

[0123] Specifically, the main function of the feature comparison and analysis unit is to compare and analyze the multi-dimensional features of flames and smoke based on the partition mapping data. For the feature data of each partition, the system will perform similarity analysis between features according to the defined comparison criteria, such as using color similarity to calculate the color difference of the flame or smoke area, and calculating the granular characteristics of the smoke through texture analysis. The result of the comparison analysis will generate a regional feature matrix, and each unit in the matrix represents the comparison analysis results of different features in each area. The matrix can show the degree of difference between flame or smoke features in different partitions, providing a data basis for the subsequent feature integration unit.

[0124] Specifically, after receiving the regional feature matrix from the feature comparison and analysis unit, the feature integration unit integrates the morphological distribution, color difference, edge clarity and texture features of flames and smoke in each area to generate the final multi-dimensional feature data. The feature integration unit integrates the global feature data into a multi-dimensional feature vector to describe the overall performance of the fire characteristics. The multi-dimensional feature vector contains data in multiple dimensions such as color, morphology, edge clarity and texture, providing an accurate and comprehensive description of fire characteristics for the AI ​​classification module.

[0125] In a preferred embodiment of the present invention, the AI ​​classification module further includes:

[0126] A feature data hierarchical processing unit, used for receiving multi-dimensional feature data and inputting the multi-dimensional feature data layer by layer into a pre-trained multi-layer neural network model for feature analysis;

[0127] An adaptive feature enhancement unit, used to dynamically adjust the weight of the intermediate classification data generated by the feature data hierarchical processing unit according to the environmental condition data, so as to obtain enhanced feature data;

[0128] The model iteration tuning unit is used to adjust the learning rate and bias value of each layer of the neural network according to the enhanced feature data.

[0129] In the embodiment of the present invention, a deep fire feature analysis of a multi-layer neural network is realized through a feature data hierarchical processing unit, an adaptive feature enhancement unit and a model iteration tuning unit. The feature data hierarchical processing unit receives the multi-dimensional feature data generated by the feature extraction module and inputs these data layer by layer into the pre-trained multi-layer neural network model. Each layer of the neural network analyzes different features, thereby extracting multi-level flame and smoke features. The layer-by-layer feature analysis helps to fully understand the fire features.

[0130] The adaptive feature enhancement unit adjusts the feature weights of the intermediate classification data based on the environmental condition data. By analyzing data such as ambient temperature, humidity, and light intensity, the adaptive feature enhancement unit can dynamically adjust the feature weights for different environmental conditions, so that the classification model can still maintain a high recognition accuracy in complex environments. Finally, the model iteration tuning unit further adjusts the learning rate and bias value of the enhanced feature data, and optimizes the recognition performance of the model through multiple iterations, making the fire detection system more adaptable in different scenarios.

[0131] This multi-level feature classification method significantly improves the accuracy and stability of fire detection, especially in complex and changing environments. The AI ​​classification module in this embodiment uses dynamic adjustment of feature weights and iterative tuning to enable the model to be optimized according to actual environmental changes, effectively solving the problem of high false alarm rate of traditional detection technology in complex environments.

[0132] Specifically, the model iteration tuning unit is responsible for adaptively adjusting the parameters of the AI ​​classification model to improve the accuracy of fire identification. The implementation steps are as follows:

[0133] Learning rate initialization: When the system starts the fire recognition task, it will assign initial values ​​to the learning rates of each layer of the neural network model. The size of the learning rate determines the step size of the model parameter update, so the system will set an appropriate initial learning rate based on the requirements of the recognition task and the complexity of the environmental characteristics.

[0134] Feature data input and confidence calculation: The model iteration tuning unit receives the enhanced feature data output by the adaptive feature enhancement unit and calculates the confidence of the current recognition result after each classification task. Confidence, as the basis for model adaptive tuning, will determine the direction and magnitude of the learning rate and bias adjustment.

[0135] Adaptive adjustment of learning rate and bias: Based on the confidence level, the system will dynamically adjust the learning rate and bias value of the model. When the confidence level is high, the system will appropriately reduce the learning rate to avoid over-adjustment of the model; when the confidence level is low, the system will increase the learning rate so that the model can learn new features faster. In addition, the bias value will be fine-tuned according to the feature sensitivity of different levels to ensure the accuracy of the model in specific feature dimensions.

[0136] Parameter update and iteration: After each round of recognition tasks, the system will apply the new learning rate and bias value to the next round of feature data processing, forming an iterative adjustment mechanism. The model iteration tuning unit will further optimize the learning rate and bias according to the changes in the recognition results after each round of recognition until the model converges or reaches the preset recognition accuracy standard.

[0137] Dynamic feedback and real-time tuning: The model iteration tuning unit monitors the recognition results in real time and adjusts the parameters when necessary to further improve the recognition ability of the model in complex environments. The system can automatically adjust the model parameters according to environmental changes to ensure the accuracy and adaptability of fire detection.

[0138] In a preferred embodiment of the present invention, the feature data hierarchical processing unit includes:

[0139] The multi-layer feature input unit is used to receive multi-dimensional feature data and input it layer by layer into the first layer of the multi-layer neural network model; the multi-layer feature input unit performs preliminary normalization processing on the input data;

[0140] The layer-by-layer feature extraction unit is used to extract the morphological distribution, color difference, edge definition and texture features of flames and smoke in each layer of the multi-layer neural network model; the layer-by-layer feature extraction unit generates intermediate feature data according to the input features of the current layer, and identifies the key feature combination related to flames and smoke in each layer to generate inter-layer correlation data;

[0141] The inter-layer feature association unit is used to transfer and integrate inter-layer association data between layers of the multi-layer neural network; the inter-layer feature association unit aggregates the association feature data according to the performance differences of flame and smoke features at different levels;

[0142] The hierarchical feature output unit is used to integrate the finally generated multi-layer associated feature data into comprehensive intermediate classification data and output it to the adaptive feature enhancement unit.

[0143] In the embodiment of the present invention, the accuracy of fire identification is improved through multi-level feature processing by further subdividing into a multi-layer feature input unit, a layer-by-layer feature extraction unit, an inter-layer feature association unit, and a layered feature output unit. The multi-layer feature input unit first normalizes the multi-dimensional feature data so that the feature data has a consistent scale at different levels, thereby avoiding the influence of feature value differences on the model. The layer-by-layer feature extraction unit extracts the morphological distribution, color difference, edge clarity, and texture features of flames and smoke in each layer of the neural network, ensuring that key information of fire features can be captured at each layer.

[0144] The inter-layer feature association unit is responsible for transmitting and integrating fire features between different levels. By analyzing the correlation between layers, a data set with multi-level features is generated, so that the classification model can understand the fire features as a whole. Finally, the hierarchical feature output unit outputs the inter-layer integrated feature data to provide optimized feature input for subsequent classification and recognition. Compared with traditional feature extraction methods, the multi-level correlation feature extraction method of this module can effectively improve the deep learning ability of the model, making the classification results of fire detection more accurate and stable.

[0145] Specifically, the layer-by-layer feature extraction unit refines and extracts the multi-dimensional features of flames and smoke in each layer of the neural network model, ensuring that each layer can gradually capture the key information of fire characteristics. The implementation steps are as follows:

[0146] Feature data input and normalization processing: The layer-by-layer feature extraction unit first receives and normalizes the input multi-dimensional feature data to ensure that the data has a consistent scale between each layer to avoid affecting the accuracy of subsequent feature extraction due to numerical differences.

[0147] Layer-by-layer feature analysis: The system analyzes and extracts the normalized feature data in each layer of the neural network. Each layer processes different features, such as shape, color, edge definition and texture, independently. By extracting feature combinations at different levels, the fire features are refined and strengthened in the layer-by-layer processing.

[0148] Intermediate feature data generation: After each layer of feature analysis is completed, the system will generate corresponding intermediate feature data, which represents the current layer’s understanding and classification effect of fire features. The intermediate feature data will be used as input for the next layer, so as to gradually capture the global features of fire in the multi-layer structure.

[0149] Feature transfer: The intermediate feature data of each layer is automatically transferred to the next layer, allowing the multi-layer neural network to accumulate fire feature information layer by layer. This layer-by-layer transfer ensures that the model's understanding of fire features gradually deepens, improving the accuracy and robustness of the classification model.

[0150] Specifically, the inter-layer feature association unit is used to establish feature association relationships between different layers in order to integrate and transmit fire feature information in multiple layers. The implementation steps are as follows:

[0151] Intermediate feature reception: This unit receives the intermediate feature data generated by the layer-by-layer feature extraction unit, collects fire feature information at each level, and provides basic data for inter-layer correlation analysis.

[0152] Correlation calculation: The inter-layer feature correlation unit calculates the correlation between feature data of different layers. The system uses similarity metrics, such as cosine similarity or Euclidean distance, to calculate the similarity between feature data of each layer and generate a correlation matrix. This matrix can show the degree of change and correlation of fire features between layers.

[0153] Feature aggregation: After completing the correlation calculation, the system aggregates the features with higher similarity, so that the fire features are strengthened and optimized when they are transmitted between layers. Through feature aggregation, the system can eliminate redundant feature data and ensure that the transmitted feature information is more concise and accurate.

[0154] Correlation feature output: The inter-layer feature correlation unit passes the aggregated correlation feature data as output to the next module, providing optimized feature data for subsequent classification and analysis.

[0155] In a preferred embodiment of the present invention, the adaptive feature enhancement unit comprises:

[0156] An environmental condition analysis unit, used to receive environmental condition data, analyze the changing trends of temperature, humidity, air flow velocity and air composition, and generate environmental impact parameters;

[0157] A feature weight adjustment unit, used to adjust the weight of the multi-dimensional feature data of flame and smoke according to the environmental impact parameters;

[0158] ,in, After adjustment feature weights, For the original feature weights, is the temperature, For humidity, , , , , , is the weight coefficient;

[0159] A dynamic weight updating unit, used to monitor the output of the feature weight adjustment unit in real time and dynamically update the feature weight;

[0160] The weight fusion unit is used to apply the dynamically updated weights to the multi-dimensional feature data, distribute the adjusted feature weights to the feature layers such as morphology, color, edge and texture, and generate enhanced feature data.

[0161] In the embodiment of the present invention, the environmental condition analysis unit, the feature weight adjustment unit, the dynamic weight update unit and the weight fusion unit jointly realize the optimization of fire identification based on environmental adaptation. The environmental condition analysis unit receives and analyzes the changing trends of temperature, humidity, air flow velocity and air composition in real time, generates environmental impact parameters, and provides a basis for feature weight adjustment. The feature weight adjustment unit dynamically adjusts the weights of the multi-dimensional feature data of flames and smoke according to the environmental impact parameters, and through weight adjustment, the fire features can be effectively expressed in different environments.

[0162] The dynamic weight update unit monitors the output data of the feature weight adjustment unit and updates the feature weight in real time according to the changes in the environment. The weight fusion unit applies the adjusted feature weights to multi-dimensional features such as morphology, color, edge and texture to generate enhanced feature data that is adaptive to the environment, ensuring that fire features can still be accurately extracted under conditions of drastic environmental changes. Compared with traditional methods, this module can respond to environmental changes in real time and perform adaptive feature adjustments on fire features, significantly improving the accuracy and adaptability of fire detection.

[0163] The application of the feature weight adjustment formula further optimizes the adaptability of the features. In the formula, the dynamic changes of environmental parameters such as temperature, humidity and brightness affect the weight distribution of features, which greatly improves the accuracy of fire detection in different environments.

[0164] Specifically, the weight fusion unit is responsible for applying the feature weights adjusted in the adaptive feature enhancement unit to each feature layer to generate enhanced feature data that is adaptive to the environment. The implementation steps are as follows:

[0165] Weight data reception: The weight fusion unit first receives the weight data output by the dynamic weight update unit, including the adjustment weights of the morphology, color, edge and texture features.

[0166] Weight allocation and application: The system applies the received weight data to each feature layer, adjusting the weight ratio of each feature layer one by one. For example, the weight of the morphological feature can be applied to the contour feature of the edge of the flame, and the weight of the color feature is applicable to the color distribution of the flame or smoke.

[0167] Multi-dimensional feature fusion: After the weights are applied, the system will perform multi-dimensional fusion of the adjusted morphological, color, edge and texture feature data to ensure that the data of different feature layers are integrated according to the set weight ratio to generate comprehensive and enhanced multi-dimensional feature data.

[0168] Enhanced feature output: The weight fusion unit outputs the fused enhanced feature data, providing complete feature information for the classification module, making fire identification more adaptable and accurate under different environmental conditions.

[0169] In a preferred embodiment of the present invention, the background adaptation module includes:

[0170] A multi-dimensional background elimination unit is used to eliminate background noise data in the preliminary fire identification result according to the environmental condition data set to generate fire characteristic data;

[0171] The characteristic dynamic balancing unit is used to adjust the weight of fire characteristic data under different environmental conditions and generate fire identification data.

[0172] In the embodiment of the present invention, the multi-dimensional background elimination unit and the feature dynamic balance unit jointly realize the effective suppression of environmental interference, so that the accuracy of fire identification under different environmental conditions is significantly improved. These environmental data can reflect the real-time changes in the monitored area, especially sudden environmental fluctuations or feature changes under special conditions.

[0173] The multi-dimensional background rejection unit performs multi-dimensional background noise rejection processing on the preliminary fire identification results based on the environmental condition data set. By using data such as temperature, humidity and airflow to detect and eliminate false alarm signals introduced by non-fire factors in the background, it ensures that the final fire identification data only contains flame and smoke features. For example, in an industrial environment, since the operation of equipment may generate high temperature or smoke, the background rejection unit can eliminate these false alarm sources based on real-time environmental data, thereby ensuring the reliability of the detection results.

[0174] When generating fire identification data, the feature dynamic balance unit adaptively adjusts the weight of the fire feature data according to different environmental conditions, so that the flame and smoke features remain accurate in complex environments. The dynamic balance mechanism makes the system more flexible in dealing with different environmental challenges. For example, in an environment with high humidity or strong airflow, the system can balance the weights of different features in real time to improve the accuracy of recognition. Compared with traditional single feature pattern recognition, this module can dynamically optimize fire features under a variety of environmental factors, significantly reducing false alarms and missed alarms, and improving the robustness of the system.

[0175] Specifically, the multi-dimensional background elimination unit performs multi-dimensional elimination of background noise on the preliminary fire identification results based on the environmental condition data. The implementation steps are as follows:

[0176] Environmental data reception: The multi-dimensional background elimination unit receives the environmental condition data generated by the real-time environment acquisition unit, including information such as temperature, humidity, air flow velocity and air composition changes, providing a basis for eliminating background noise.

[0177] Background model establishment: The system establishes a background noise model based on environmental data to identify normal noise characteristics in the current environment. For example, in a high temperature environment, the background model will treat the high temperature area as normal background rather than fire characteristics.

[0178] Noise Removal: The system uses the background noise model to remove non-fire features from the initial fire identification results, eliminating false alarm signals caused by temperature, humidity or air flow fluctuations. This step ensures the clarity and effectiveness of fire features.

[0179] Background filtering data output: After removing the background noise, the multi-dimensional background elimination unit outputs the filtered fire feature data to provide accurate feature data for further fire identification and classification.

[0180] Specifically, when generating fire identification data, the feature dynamic balance unit adaptively adjusts the weight of the fire feature data according to different environmental conditions. The implementation steps are as follows:

[0181] Feature weight reception: The feature dynamic balance unit receives the fire feature data output by the multi-dimensional background rejection unit, as well as the changing parameters in the environmental condition data, including temperature, humidity, and airflow.

[0182] Dynamic weight adjustment: The system adjusts the weight of fire feature data in real time according to changes in environmental conditions. For example, when humidity is high, the system will reduce the weight of smoke features to avoid false alarms caused by humidity; under high temperature conditions, the weight of temperature-related flame features will be increased.

[0183] Balanced data generation: The dynamic balancing unit integrates the adjusted fire feature data to generate balanced fire identification data. This data can maintain a stable feature distribution under various environmental conditions, effectively improving the accuracy of fire detection.

[0184] Balanced data output: The balanced fire identification data is transmitted to the alarm trigger module to ensure accurate identification of fire characteristics in complex environments and provide reliable data support for subsequent alarms.

[0185] In a preferred embodiment of the present invention, the alarm triggering module includes:

[0186] A credibility evaluation unit is used to calculate the confidence of the fire identification data and use the confidence as the basis for determining the trigger priority;

[0187] A multi-level trigger condition determination unit is used to set trigger conditions of different priorities. When the fire identification data meets the priority conditions, an alarm signal of the corresponding level is generated;

[0188] The multi-channel alarm output unit is used to output alarm signals through sound, light, SMS, and network channels according to the alarm level when the trigger conditions are met.

[0189] In the embodiment of the present invention, a hierarchical response mechanism for fire alarm is implemented through a credibility evaluation unit, a multi-level trigger condition determination unit and a multi-channel alarm output unit. The credibility evaluation unit performs credibility calculation on the fire identification data and determines the trigger priority of the alarm by analyzing the confidence of the identification result. This credibility evaluation greatly reduces the possibility of false alarms, so that the detector can avoid false alarms when the fire characteristics are unclear or the confidence is low.

[0190] The multi-level trigger condition determination unit makes trigger determinations of different priorities based on the credibility of the fire identification data. According to the confidence of the fire identification data, the alarm is divided into three priorities: low, medium, and high. In the case of low priority, the system can trigger an early warning signal to remind the staff to check; in the case of medium priority, the alarm signal will be activated to prompt personnel to respond quickly; and in the case of high priority, the system will immediately trigger the highest level of global alarm to ensure a rapid response in an emergency. The multi-level trigger conditions enable the system to flexibly adjust the alarm response according to the severity of the fire and the reliability of the identification results, avoiding frequent interference with the normal production or living environment.

[0191] The multi-channel alarm output unit outputs the alarm signal through sound, light, SMS, network and other methods to ensure that the fire alarm in different environments and scenarios can be received in time. This multi-channel alarm output design enables the system to adapt to the needs of various application scenarios, whether it is a factory, office building or residential area, to ensure that the alarm signal can be conveyed to relevant personnel in the first place, thereby reducing the harm caused by the fire.

[0192] Specifically, the credibility evaluation unit is used to calculate the confidence of the fire identification data in order to provide a basis for judgment in the alarm response. Its implementation steps are as follows:

[0193] Fire identification data reception: The credibility assessment unit receives the fire identification data from the feature dynamic balance unit to calculate the confidence of the current identification result.

[0194] Confidence calculation: The system calculates the confidence of the recognition based on factors such as the clarity, continuity and consistency of the fire recognition data. Recognition results with high clarity, continuous data and strong feature consistency will obtain higher confidence.

[0195] Priority setting: The system automatically assigns recognition results to different alarm priorities based on the confidence level. Recognition results with high confidence levels will be assigned high priority to ensure that the system issues an alarm quickly; while recognition results with low confidence levels will be assigned low priority to avoid false alarms.

[0196] Confidence data output: The credibility assessment unit transmits the confidence data to the multi-level trigger condition determination unit to provide accurate priority determination basis for the multi-level alarm mechanism.

[0197] In a preferred embodiment of the present invention, the multi-level trigger condition determination unit further includes:

[0198] A low priority determination unit, used to determine whether to generate a low priority alarm signal according to the frequency and duration of changes in flame or smoke characteristics when the confidence level of fire identification data is lower than a preset threshold range;

[0199] A medium priority determination unit, used to generate a medium priority alarm signal according to the composite characteristics of flame and smoke features under multiple spectra when the confidence level of the fire identification data is within a preset threshold range;

[0200] The high priority determination unit is used to generate a high priority alarm signal and trigger a global alarm device when the confidence level of the fire identification data is higher than a preset threshold range.

[0201] In the embodiment of the present invention, by further subdividing into low priority determination units, medium priority determination units and high priority determination units, a hierarchical alarm mechanism provides more precise control for fire response. When the confidence of fire identification data is low, the low priority determination unit determines whether it is necessary to generate a low priority alarm signal by analyzing the frequency and duration of changes in flame and smoke characteristics. At this time, the system may only issue a slight warning to indicate that there is an abnormality in the monitored area, thereby guiding the staff to conduct further inspections. This low priority alarm is suitable for the early stage of a fire or the stage where the characteristics are not obvious, which helps to prevent the fire in time at the early stage of the incident.

[0202] When the confidence level of the fire identification data is in the medium range, the medium priority determination unit generates a medium priority alarm signal based on the composite characteristics of flames and smoke under multi-spectral conditions. At this level, the possibility of a fire is high, but it has not yet reached an emergency state. The system will alert relevant personnel through sound and light alarms or SMS notifications, prompting them to take preventive measures immediately. This type of alarm is suitable for helping relevant personnel respond quickly when the fire characteristics are gradually becoming obvious but not yet out of control.

[0203] The high priority determination unit is used to immediately trigger an emergency alarm signal and activate the global alarm device when the confidence level of the fire identification data reaches the preset high priority threshold. This level of alarm is suitable for scenarios where a fire has clearly occurred and may spread rapidly. Through the global alarm, the system can notify all relevant parties to conduct emergency evacuation and rescue to ensure the maximum safety of personnel and property.

[0204] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. AI-based three-light image fire detector, characterized in that: include: The flame spectrum acquisition module is used to obtain multi-spectral image data of the monitoring area, wherein the multi-spectral image data includes infrared spectrum, ultraviolet spectrum and visible spectrum data; An image preprocessing module is used to perform denoising, edge enhancement and light intensity smoothing on multispectral image data to generate multispectral data; A feature extraction module is used to receive multispectral data, divide it into several analysis areas, and independently extract feature information in each analysis area to generate partition feature data, integrate the partition feature data to obtain multidimensional feature data, and the multidimensional feature data includes the shape, edge, color and texture features of flame and smoke; Environmental condition acquisition module, used to collect environmental condition data of the monitored area, including temperature, humidity, air flow velocity and air composition; AI classification module, which uses a pre-trained multi-layer neural network model to perform layer-by-layer analysis and classification of multi-dimensional feature data, identify multiple feature combinations of flames or smoke, and generate preliminary fire identification results; The background adaptation module is used to receive the preliminary fire identification results and environmental condition data, and perform background noise elimination and feature optimization processing on the preliminary fire identification results to obtain fire identification data; the environmental condition data includes real-time acquisition of temperature, humidity, air flow velocity and air composition change data in the monitoring area; An alarm trigger module is used to receive fire identification data and determine whether the preset fire alarm conditions are met; When the fire identification data meets the preset fire alarm conditions, an alarm signal is output; The flame spectrum acquisition module also includes: A spectral image acquisition unit, used to acquire infrared spectrum, ultraviolet spectrum and visible spectrum data of the monitored area; A light source identification unit is used to identify the type of light source in the monitoring area by detecting the wavelength and spectral characteristics of the light source in the environment, and generate light source classification data; A background analysis unit is used to detect background spectrum changes of the multi-spectral image data at different time intervals, generate spectrum change trend data, and generate a spectrum adjustment signal according to the spectrum change trend data; The spectrum adjustment unit is used to adjust the acquisition parameters of infrared spectrum, ultraviolet spectrum and visible spectrum data in real time according to the light source classification data, the spectrum change trend data and the spectrum adjustment signal to obtain multi-spectral image data. The adjustment formula of the multi-spectral image data is as follows: + + ， in, is the adjusted multispectral image data, , , To represent infrared, ultraviolet and visible light image data respectively, is the wavelength shift of the infrared spectrum, is the weight parameter, is the wavelength shift of the ultraviolet spectrum, To adjust the constant of the UV spectrum sensitivity to wavelength shift, is the wavelength shift of the visible spectrum, is the maximum wavelength of the visible light spectrum, is the circumference constant, is the light source brightness, , , , , , is a constant parameter.

2. The AI-based three-light image-type fire detector according to claim 1, characterized in that: The image preprocessing module also includes: The multidimensional noise filtering unit is used to suppress the dynamic noise of the multispectral image data by using the spatiotemporal filtering algorithm and combining the time series characteristics of the image to generate time series image data; the denoising formula is as follows: ， in, is the pixel value after denoising, For the time series The pixel value of the frame, is the Gaussian spatiotemporal filter weight function, Controls the width of the filter, is the total number of filtering frames; The edge saliency enhancement unit is used to identify and enhance the edge saliency features of flames and smoke in different spectra according to the time series image data, and generate saliency image data; the edge saliency feature enhancement formula is as follows: ， in, is the image after edge enhancement, is the time series image data, is the Laplace operator; The adaptive light intensity smoothing unit performs adaptive smoothing adjustment on the light intensity change of the saliency image according to the contrast difference between the background light intensity of the monitoring area and the target spectral data to generate multi-spectral data; the smoothing formula is as follows: ， in, is the pixel value after light intensity smoothing, For location and location The spatial distance For the location The pixel value of the image after edge enhancement at and are the control coefficients of spatial distance and pixel difference respectively.

3. The AI-based three-light image-type fire detector according to claim 2 is characterized in that: The feature extraction module comprises: A partition mapping unit is used to perform spatial mapping partition processing on multispectral data, divide the image into multiple spatial analysis areas, and generate corresponding partition mapping data for each area; The morphology distribution extraction unit is used to analyze the flame and smoke morphology in each analysis area, calculate its edge curvature, shape contour and structural complexity, generate regional morphological feature data, and integrate them to obtain global morphological distribution data; the edge curvature, shape contour and structural complexity are calculated by the following formulas: ,in, is the edge curvature, which is used to measure the morphological changes of the edge. is the symbol of partial derivative; ,in, is the fractal dimension, which indicates the complexity of the shape outline. The minimum size required to cover the edge is The number of units, The scale when measuring the edge; ,in, is the structural complexity, which indicates the complexity inside the shape. For the The probability of occurrence of a structural feature, is the total number of structural features; A feature comparison and analysis unit is used to compare and analyze the color difference, edge clarity and texture characteristics of flames and smoke in each analysis area according to the partition mapping data, and generate a regional feature matrix; The feature integration unit is used to integrate the morphological distribution, color difference, edge clarity and texture characteristics of flames and smoke according to the global morphological distribution data and the regional feature matrix to generate multi-dimensional feature data.

4. The AI-based three-light image-type fire detector according to claim 3 is characterized in that: The AI ​​classification module also includes: A feature data hierarchical processing unit, used for receiving multi-dimensional feature data and inputting the multi-dimensional feature data layer by layer into a pre-trained multi-layer neural network model for feature analysis; An adaptive feature enhancement unit, used to dynamically adjust the weight of the intermediate classification data generated by the feature data hierarchical processing unit according to the environmental condition data, so as to obtain enhanced feature data; The model iteration tuning unit is used to adjust the learning rate and bias value of each layer of the neural network according to the enhanced feature data.

5. The AI-based three-light image-type fire detector according to claim 4, characterized in that: The adaptive feature enhancement unit comprises: The multi-layer feature input unit is used to receive multi-dimensional feature data and input it layer by layer into the first layer of the multi-layer neural network model; the multi-layer feature input unit performs preliminary normalization processing on the input data; The layer-by-layer feature extraction unit is used to extract the morphological distribution, color difference, edge definition and texture features of flames and smoke in each layer of the multi-layer neural network model; the layer-by-layer feature extraction unit generates intermediate feature data according to the input features of the current layer, and identifies the key feature combination related to flames and smoke in each layer to generate inter-layer correlation data; The inter-layer feature association unit is used to transfer and integrate inter-layer association data between layers of the multi-layer neural network; the inter-layer feature association unit aggregates the association feature data according to the performance differences of flame and smoke features at different levels; The hierarchical feature output unit is used to integrate the finally generated multi-layer associated feature data into comprehensive intermediate classification data and output it to the adaptive feature enhancement unit.

6. The AI-based three-light image-type fire detector according to claim 5, characterized in that: The adaptive feature enhancement unit comprises: An environmental condition analysis unit, used to receive environmental condition data, analyze the changing trends of temperature, humidity, air flow velocity and air composition, and generate environmental impact parameters; A feature weight adjustment unit, used to adjust the weight of the multi-dimensional feature data of flame and smoke according to the environmental impact parameters; ,in, After adjustment feature weights, For the original feature weights, is the temperature, For humidity, , , , , , is the weight coefficient; A dynamic weight updating unit, used to monitor the output of the feature weight adjustment unit in real time and dynamically update the feature weight; The weight fusion unit is used to apply the dynamically updated weights to the multi-dimensional feature data, distribute the adjusted feature weights to the feature layers such as morphology, color, edge and texture, and generate enhanced feature data.

7. The AI-based three-light image-type fire detector according to claim 6, characterized in that: The background adaptation module comprises: A multi-dimensional background elimination unit is used to eliminate background noise data in the preliminary fire identification result according to the environmental condition data set to generate fire characteristic data; The characteristic dynamic balancing unit is used to adjust the weight of fire characteristic data under different environmental conditions and generate fire identification data.

8. The AI-based three-light image-type fire detector according to claim 7, characterized in that: The alarm triggering module comprises: A credibility evaluation unit is used to calculate the confidence of the fire identification data and use the confidence as the basis for determining the trigger priority; A multi-level trigger condition determination unit is used to set trigger conditions of different priorities. When the fire identification data meets the priority conditions, an alarm signal of the corresponding level is generated; The multi-channel alarm output unit is used to output alarm signals through sound, light, SMS, and network channels according to the alarm level when the trigger conditions are met.

9. The AI-based three-light image-type fire detector according to claim 8, characterized in that: The multi-stage trigger condition determination unit also includes: A low priority determination unit, used to determine whether to generate a low priority alarm signal according to the frequency and duration of changes in flame or smoke characteristics when the confidence level of fire identification data is lower than a preset threshold range; A medium priority determination unit, used to generate a medium priority alarm signal according to the composite characteristics of flame and smoke features under multiple spectra when the confidence level of the fire identification data is within a preset threshold range; The high priority determination unit is used to generate a high priority alarm signal and trigger a global alarm device when the confidence level of the fire identification data is higher than a preset threshold range.

Images