Flame detection method and flame detector suitable for dangerous environment
Patent Information
- Application Number
- CN202511666113.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-14
- Publication Date
- 2026-01-20
AI Technical Summary
Existing flame detectors have difficulty distinguishing between thermal radiation from equipment and flame signals in environments where high-temperature industrial equipment and fire sources coexist, leading to misjudgment or underreporting of real fires, especially when high-temperature equipment and specific organic compounds are burning, resulting in severe overlap of spectral features.
The system uses infrared and visible light imaging modules to collect data synchronously. Through timestamp matching, light source interference analysis, heat wave trajectory analysis, and multimodal feature fusion, combined with a random forest classifier and diffusion model, it accurately determines the combustion status and generates evacuation plans.
It effectively distinguishes between real combustion and external heat source interference, reduces the false alarm rate, improves the accuracy of judgment, and generates precise flame detection and evacuation plans.
Smart Images
Figure CN121366468A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of flame detection technology, and particularly relates to a flame detection method and a flame detector suitable for a dangerous environment. BACKGROUND
[0002] At present, fire detectors mainly include three categories of temperature sensing, smoke sensing and light sensing, which respectively identify fires according to temperature changes, smoke characteristics and light characteristics generated when flames burn. The temperature sensing fire detector and the smoke sensing fire detector are relatively simple in production process, but are prone to false alarms of temperature and smoke generated by non-fire, and have slow response speed. At the same time, although the light sensing fire detector has relatively fast response capability, it is easily disturbed by sunlight and artificial light sources, resulting in false alarm.
[0003] A Chinese patent with the application number CN202010907770.X discloses a comprehensive flame detection method based on ultraviolet, infrared and vision, including that a data output end of an ultraviolet detector is connected with a first data input end of a controller, a data output end of an infrared detector is connected with a second data input end of the controller, a data output end of a vision detector is connected with a third data input end of the controller, a trigger data end of the vision detector is connected with a trigger data end of the controller, and a communication end of the controller is connected with a communication end of a communication module; image data collected by the vision detector is analyzed and processed at a server end to determine a fire condition. The image type fire detection technology is combined with ultraviolet and infrared detection, which more greatly meets the detection needs of people for high sensitivity, high reliability and visible to the human eye.
[0004] Although the foregoing patent document combines the image type fire detection technology with ultraviolet and infrared detection, reduces the occurrence of false alarms of the light sensing fire detector, however, in actual engineering application scenarios, the technical solution still faces a core technical bottleneck that needs to be broken through: the medium and far infrared energy spectrum band continuously radiated by high-temperature industrial equipment significantly overlaps with the flame radiation spectrum generated when a specific organic compound (a flammable liquid represented by methanol) burns in the 3.0-5.0 μm band range. This spectral coupling effect causes the existing flame detector to be difficult to accurately extract the flame signal through a single characteristic parameter, especially in industrial scenarios where high-temperature equipment and fire sources coexist, the detector is prone to misjudging the equipment thermal radiation as a flame signal, or missing the real fire due to the flame spectrum being submerged by the thermal radiation background, which seriously restricts the engineering application performance of the multi-modal fire detection technology. SUMMARY
[0005] The present application aims to provide a flame detection method and a flame detector suitable for a dangerous environment to solve the technical problems proposed in the background.
[0006] To achieve the above object, the present application provides the following technical solutions: A flame detection method suitable for dangerous environments, the flame detection method comprising: S1, synchronously collecting data of a set region through an infrared light imaging module and a visible light imaging module, and acquiring infrared light imaging data and visible light imaging data; S2, time stamp matching the infrared light imaging data and the visible light imaging data on a time axis to form a synchronous data pair; S3, detecting temperature abnormalities of the set region based on the infrared light imaging data, if the temperature abnormalities do not exceed a set threshold, continuing data collection, and if the temperature abnormalities exceed the set threshold, starting light source interference analysis; S4, judging whether there is an interference light source by analyzing light source features in the visible light image, if there is an interference light source, performing image repair and feature enhancement on the interference region; S5, analyzing dynamic features of heat wave trajectories in combination with the repaired visible light data and infrared data, and determining a combustion state according to symmetry or randomness of the trajectories; S6, performing multi-modal feature fusion and physical law cross-validation on the center-symmetric trajectories through an incomplete combustion judgment model, and outputting a diagnosis report through a random forest classifier; S7, if the output result of the diagnosis report is determined as no-color flame combustion or incomplete combustion, establishing a diffusion model and generating an evacuation plan.
[0007] Preferably, the light source interference analysis comprises: S401, performing multi-spectral parameter extraction on the synchronous data, including infrared temperature value, temperature rise rate, and visible light brightness distribution and color space features; S402, analyzing frequency domain features of the visible light sequence through Fourier transform, and excluding light sources without flickering signals; S403, positioning the interference region in combination with a U-Net image segmentation algorithm based on flame edge irregularity and centroid movement features; S404, reconstructing a background of the interference region through Poisson image editing, and generating a flame probability thermodynamic map through multi-spectral data fusion.
[0008] Preferably, the heat wave trajectory analysis comprises: S501, applying a ByteTrack algorithm to the temperature abnormal region to perform continuous frame bounding box matching, and generating a trajectory line; S502, extracting motion direction entropy, curvature fluctuation rate, and diffusion rate of the trajectory as feature parameters; S503, when the trajectory is in a dynamic random change state, determining as no-color flame combustion; S504, when the trajectory is in a center-symmetric state, whether it is affected by an external heat source is verified in combination with an incomplete combustion judgment model.
[0009] Preferably, the incomplete combustion judgment model comprises: S601, a heat wave diffusion equation is constructed, a heat conduction process of an external heat source is modeled as a diffusion equation driven by Fourier's law, and a thermal radiation distribution stability threshold is set; S602, wavelet transform analysis is performed on an incomplete combustion heat source term Q, high-frequency components of a temperature signal are extracted, and when a high-frequency energy proportion exceeds a set value, intermittent pulse characteristics are determined; S603, a combustion efficiency is calculated in combination with a thermal radiation intensity and an oxygen consumption rate, and when the combustion efficiency is less than a set value and the oxygen consumption rate deviates from a theoretical value, incomplete combustion is confirmed.
[0010] Preferably, the multi-modal feature fusion comprises: Infrared modal: extracting thermal radiation distribution uniformity, heat wave edge complexity, and centroid motion stability; Visible light modal: detecting smoke color characteristics, texture contrast, and light flow field direction consistency; Physical constraint: establishing a nonlinear mapping relationship between thermal radiation intensity and combustion products to verify data rationality.
[0011] Preferably, the random forest classifier is implemented as: Infrared thermal radiation standard deviation, visible light smoke texture contrast, light flow field main direction variance, heat zone centroid drift speed, and combustion efficiency are input as input features into the random forest classifier; When the input features meet a first preset value, the random forest classifier outputs “incomplete combustion”; when the input features meet a second preset value, the random forest classifier outputs “external heat source”; For some feature combinations that partially meet the conditions, a physical constraint model is started for secondary verification.
[0012] Preferably, the diffusion model generates an evacuation plan through the following steps: S701, based on a diffusion rate and a direction entropy of a heat wave trajectory, a propagation path of a colorless flame is simulated; S702, in combination with environmental sensor data, a three-dimensional space thermal field distribution map is constructed; S703, according to a personnel density and a safety distance threshold, a dynamic evacuation route is planned and output to a display terminal.
[0013] A flame detector, comprising a data acquisition unit, a data processing unit, and a decision unit; The data acquisition unit comprises an infrared imaging module and a visible light imaging module. The data processing unit comprises: a timestamp alignment module which aligns the infrared light data and the visible light data by adopting a PTP (Precision Time Protocol); a light source interference analysis module which identifies and processes the light source interference in the set region by integrating an FFT (Fast Fourier Transform) frequency domain analysis, edge shape identification and multispectral correlation verification algorithm; a heat wave trajectory generation module which analyzes and identifies the heat wave trajectory in the set region by a ByteTrack algorithm; an incomplete combustion judgment module which further judges the identification of the heat wave trajectory by multi-modal feature fusion; The decision unit comprises: a diffusion model generation module which outputs an evacuation scheme according to the diffusion rate and direction entropy of the heat wave trajectory.
[0014] Preferably, the data acquisition unit further comprises a parameter configuration interface for adjusting the parameters of the infrared imaging module and the visible light imaging module according to actual scene requirements.
[0015] Preferably, the data processing unit further comprises a data storage module for storing the acquired original data, intermediate data in the processing process and final result data.
[0016] Technical effects and advantages of the present application: 1. The present application effectively distinguishes real combustion from external heat source interference by analyzing the symmetry or randomness of the heat wave trajectory; when the trajectory changes dynamically and randomly, it is determined to be real combustion; when the trajectory is center-symmetric, it is further verified whether it is an external heat source, thereby improving the judgment accuracy.
[0017] 2. The present application extracts infrared and visible light modal features and combines physical constraints by multi-modal feature fusion, comprehensively analyzes the data, effectively distinguishes external heat sources and incomplete combustion; at the same time, it uses a random forest classifier, combines dynamic threshold rules and a physical constraint model for secondary verification, accurately determines whether the temperature anomaly is an external heat source or incomplete combustion, and reduces the misjudgment rate. BRIEF DESCRIPTION OF DRAWINGS
[0018] Figure 1 is a schematic diagram of the main process of the flame detection method of the present application; Figure 2 is a schematic diagram of the steps of the flame detection method of the present application; Figure 3 is a schematic diagram of the steps of the light source interference analysis of the present application; Figure 4 is a schematic diagram of the steps of the heat wave trajectory analysis of the present application; Figure 5A schematic diagram of steps of the incomplete combustion judgment model of the present application; Figure 6 A schematic diagram of steps of the diffusion model of the present application; Figure 7 A schematic diagram of the structure of the flame detector of the present application. DETAILED DESCRIPTION
[0019] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.
[0020] Embodiment one Referring to Figures 1 to 7 The present application provides a flame detection method suitable for a dangerous environment, which comprises the following steps: S1, synchronously collecting data of a set region by an infrared light imaging module and a visible light imaging module to obtain infrared light imaging data and visible light imaging data; An infrared thermal imaging module and a visible light imaging module are arranged in the set region, and the infrared imaging data and the visible light imaging data of the set region are collected by the two modules.
[0021] After the data collection is completed, the infrared imaging data and the visible light data are processed by a data processing unit, which specifically includes the following contents: The preprocessing of the infrared imaging data includes the following aspects: The CLAHE algorithm is used to improve the contrast of thermal radiation details to suppress local overexposure. Specifically, the original infrared imaging data is converted into a 14-bit grayscale image to retain more dynamic range; then the image is divided into several non-overlapping small blocks by identifying high-temperature / low-temperature regions, and the histogram equalization is performed on each small block independently, the contrast is adjusted according to the local pixel distribution, and the brightness difference of different regions is adapted. In order to avoid noise amplification caused by local over-enhancement, a clipping threshold is set for the histogram of each small block. If the histogram height exceeds the threshold, the exceeding part is evenly distributed to other gray levels. In this way, overexposure is suppressed and low-temperature zone details are improved.
[0022] It should be noted that when the infrared imaging data is processed, the equalization results of adjacent small blocks are smoothly spliced by using bilinear interpolation to eliminate block effects and avoid boundary artifacts.
[0023] The preprocessing of the visible light imaging data includes the following aspects: Firstly, through the way of bilateral filtering, the image noise is smoothed while avoiding the edge area from being blurred due to excessive smoothing; after completing the bilateral filtering denoising, the whole image may be in a dark state, at this time, through the light intensity meter deployed in the set area, the ambient light intensity is collected in real time, and a nonlinear mapping model of light intensity and brightness adjustment coefficient is established. According to the real-time light data, the global brightness of the image is adjusted to avoid the whole image being dark after bilateral filtering, which cannot meet the requirements of computer vision tasks.
[0024] S2, time stamp matching of the infrared light imaging data and the visible light imaging data on the time axis to form a synchronous data pair; After the infrared imaging data and the visible light imaging data are preprocessed, the two are corresponded in the time dimension to avoid errors in subsequent analysis due to time deviation. Specifically, the following steps are included: First, each frame of infrared or visible light image is assigned an accurate time label to record the acquisition time of the frame.
[0025] Then, by comparing the time stamps of the infrared frame and the visible light frame, the closest frame pair in time is found to associate the data at the same time.
[0026] S3, detecting temperature anomaly of a set area based on the infrared light imaging data, if the set threshold is not exceeded, continue data collection, if the set threshold is exceeded, start light source interference analysis; The temperature of the set area is monitored by the infrared light imaging module; when the temperature of the set area does not exceed the temperature threshold set by the infrared light imaging module, the infrared light imaging module and the visible light imaging module continue to collect data for the set area; when the temperature of the set area exceeds the temperature threshold set by the infrared light imaging module, the system transmits this situation to the light source interference analysis module of the data processing unit, and the light source interference analysis module analyzes and judges the light source interference of the set area.
[0027] S4, by analyzing the light source features in the visible light image, it is judged whether there is an interference light source, if there is, the image repair and feature enhancement of the interference area are carried out; Specifically, the light source interference analysis includes the following steps: S401, multi-spectral parameter extraction of synchronous data, including infrared temperature value, temperature rise rate, visible light brightness distribution and color space feature; After the infrared image data and the visible light image data are time stamped, a time data set is cut back from the current time point as the starting point, which includes all the infrared image data and visible light image data within the cut time. The time cut length is preferably 0-3min data set cut back from the current time point as the starting point.
[0028] By extracting the corresponding infrared data and visible light data in the data set, the infrared image data includes temperature values and temperature rise rates within the interception time, and the temperature rise rate of the temperature reflects whether a flame burning phenomenon occurs, specifically, when the temperature rises at a rate of 3°C / s per second, it is a typical characteristic of a flame burning phenomenon; and the visible light image data includes the brightness distribution and color space of the set region within the interception time. The irregular highlight area is detected by the brightness distribution to exclude uniform light sources; and the orange-red color unique to the flame is identified by the color space feature to exclude white light and other interference light sources.
[0029] S402, analyze the frequency domain features of the visible light sequence by Fourier transform to exclude light sources without flicker signals; First, the highlight area is extracted from the intercepted visible light image data sequence as the region of interest (ROI), and the noise is eliminated by brightness threshold screening and morphological optimization to form a connected domain, and the continuity of the ROI across frames is dynamically tracked. Then, the average value of the pixel brightness in each ROI is taken to generate a time domain signal, which is filtered by sliding average filtering to remove noise and used as the input for frequency domain analysis.
[0030] The time domain brightness signal is converted into a frequency domain amplitude spectrum by fast Fourier transform, focusing on the flicker frequency of the flame to exclude direct current components and high frequency noise; wherein the flicker frequency of the flame is preferably 1-12 Hz.
[0031] Significant frequency components are screened by setting an amplitude threshold, and if the peak frequency falls within the interval of 1-12 Hz and the amplitude meets the standard, it is determined to be a potential flame; otherwise, it is determined to be a stable light source.
[0032] Temporal and spatial correlation verification is performed in combination with the infrared temperature rise data: it is confirmed that the visible light ROI and the infrared high temperature region are spatially coincident, and the infrared temperature rise time point and the visible light brightness flicker time point are synchronized.
[0033] S403, based on the irregularity of the flame edge and the centroid movement feature, the interference area is located by combining the U-Net image segmentation algorithm; The edge of the highlight area in the visible light image is extracted by the Canny algorithm, the number of sharp corners of the flame edge of each frame and the dynamic change of the shape are counted, and the irregularity feature is quantified; at the same time, the weighted centroid of the highlight area is calculated, the moving range and speed stability are analyzed, and a flame dynamic feature library is constructed.
[0034] A lightweight U-Net model is trained to realize pixel-level segmentation under the input of 256x256 pixels, and the edge accuracy is optimized by the Dice loss function; further, the edge irregularity score, the centroid stability score and the U-Net segmentation probability are weighted and fused to generate a comprehensive flame probability map, and the interference area is located by setting a threshold, when the flame probability value is less than the set value, the area is defined as the interference area.
[0035] Finally, through infrared thermal imaging data cross-validation, it is confirmed that the interference area has no significant temperature rise in the infrared image, further excluding false judgments.
[0036] S404, the background of the interference area is reconstructed by using Poisson image editing, and a flame probability thermal map is generated by fusing multispectral data.
[0037] Through the analysis of the irregularity of the flame edge in the early stage and the U-Net segmentation result, the pixel-level boundary of the interference light source is determined, and a binary mask is generated; The color gradient of the neighborhood around the interference area is extracted, and the background image is reconstructed using the Poisson equation to ensure that the reconstructed area is continuous with the surrounding gradient and the boundary is naturally transitioned.
[0038] The reconstructed background and the original image are weighted and fused to focus on retaining the texture of the non-interference area, and Gaussian blur is used to eliminate the splicing marks on the edge to generate a clean background without interference.
[0039] By inputting the infrared temperature matrix and visible light texture features into a convolutional neural network (CNN), a flame probability thermal map is output, thereby suppressing false positives of a single sensor.
[0040] Embodiment Two Although the above embodiment can remove light source interference factors through light source interference analysis, in actual use, the colors produced by compounds such as methanol and ethanol during combustion are difficult to distinguish by the visible light imaging module, and if there is a heat source near the combustion area, the infrared imaging module is likely to make a false judgment when detecting the area. In view of this, technical improvements are made to the technology of Embodiment One, and the improved technical solution is as follows: Based on Embodiment One, after completing the removal of the light source interference term, if the temperature of the set area still exceeds the temperature threshold set by the infrared light imaging module, the specific state of the set area can be determined through heat wave trajectory analysis: Referring to Figures 1 to 7 The present application provides a kind of flame detection method suitable for dangerous environment, the flame detection method includes: S5, combine the visible light data and infrared data after repair, analyze the dynamic characteristics of heat wave trajectory, determine the combustion state according to the symmetry or randomness of trajectory; The heat wave trajectory analysis includes: S501, the ByteTrack algorithm is applied to the temperature anomaly area to match the boundary box of consecutive frames, and a trajectory line is generated; First, the temperature anomaly area of each frame is binarized and segmented to generate a minimum circumscribed rectangle as a boundary box, the center point coordinates, size and temperature peak of the box are recorded, and an unconfirmed trajectory is created.
[0041] Secondly, a bidirectional soft matching strategy is adopted to implement the matching of the boundary boxes of consecutive frames. The intersection over union (IOU) of the detection box of the current frame and the non-disappeared tracking box of the previous frame is calculated to realize the IOU matching of the detection box and the tracking box. If the IOU exceeds a threshold, it is considered to belong to the same target, and the position and temperature information of the tracking box are updated. For the unmatched detection box, a secondary similarity matching is performed with the tracking box of the previous frame based on the appearance features such as the temperature change rate and the size difference of the boundary box, so as to avoid the track breakage caused by occlusion or temperature fluctuation.
[0042] Then, the center points of the boundary boxes of each frame that are successfully matched are connected in time sequence to form an initial trajectory line, and the jitter is eliminated through moving average filtering, and the breakpoints are completed by linear interpolation to ensure the continuity of the trajectory.
[0043] Finally, the parameters are recorded and the results are output. Each trajectory needs to save the start / stop frame number, center point coordinate sequence, temperature peak value sequence and other information, and finally output the complete trajectory line containing the spatial path (coordinate sequence) and the time range (frame number) for subsequent heat wave trajectory analysis.
[0044] S502, extract the motion direction entropy, curvature fluctuation rate and diffusion rate of the trajectory as feature parameters; By dividing the trajectory direction into several intervals, the probability distribution of the direction change in each interval is counted, and the direction entropy is calculated through the Shannon entropy formula. The higher the direction entropy value is, the more disordered the direction change is.
[0045] By segmenting the trajectory line, the curvature of each segment is calculated, and the variance of all segment curvatures is counted to obtain the curvature fluctuation rate. The higher the value is, the more tortuous the path is.
[0046] By recording the area of the heat wave affected area of each frame, the ratio of the area increment of adjacent frames to the time interval is calculated to obtain the diffusion rate. The higher the value is, the faster the heat wave expands.
[0047] It should be noted that when the motion direction entropy is high, the curvature fluctuation rate is large, and the diffusion rate is fast, the trajectory is in a random state. When the trajectory is in a dynamic random change state, it is biased towards real combustion; otherwise, it is in a symmetric state. When the trajectory is in a center-symmetric state.
[0048] S503, when the trajectory is in a dynamic random change state, it is determined as non-color flame combustion; S504, when the trajectory is in a center-symmetric state, whether it is affected by an external heat source is verified in combination with an incomplete combustion judgment model.
[0049] Example Three Although the above embodiment can determine whether the temperature anomaly of the set region is affected by the external heat source or the real combustion phenomenon through the heat wave trajectory analysis, in actual use, when the combustion of the compound presents the colorless incomplete combustion, the heat wave trajectory generated by the compound is basically consistent with the heat wave trajectory generated by the external heat source, and it is difficult to determine the specific reason for the temperature anomaly of the set region only through the heat wave trajectory analysis. In view of this, the technical scheme is improved on the basis of embodiment two, and the improved technical scheme is as follows: Based on embodiment two, after completing the analysis of the heat wave trajectory analysis, if the heat wave trajectory is in a central symmetric state, whether it is affected by the external heat source is verified by combining the incomplete combustion judgment model.
[0050] Referring to Figures 1 to 7 The flame detection method provided by the application is suitable for a dangerous environment, and the flame detection method comprises the following steps: S6, the central symmetric trajectory is subjected to multi-modal feature fusion and physical law cross verification through an incomplete combustion judgment model, and a diagnosis report is output through a random forest classifier; The incomplete combustion judgment model comprises: S601, a heat wave diffusion equation is constructed, a heat conduction process of an external heat source is modeled as a diffusion equation driven by Fourier's law, and a heat radiation distribution stability threshold is set; S602, wavelet transform analysis is performed on an incomplete combustion heat source term Q, high-frequency components of a temperature signal are extracted, and when the proportion of high-frequency energy exceeds a set value, intermittent pulse characteristics are determined; S603, combustion efficiency is calculated in combination with heat radiation intensity and oxygen consumption rate, and when the combustion efficiency is less than a set value and the oxygen consumption rate deviates from a theoretical value, incomplete combustion is confirmed.
[0051] The multi-modal feature fusion comprises infrared modal features and visible light modal features; Among them, the infrared modal features mainly extract heat radiation distribution uniformity, heat wave edge complexity and centroid motion stability; The visible light modal features mainly detect smoke color characteristics, texture contrast and light flow field direction consistency; By establishing a nonlinear mapping relationship between heat radiation intensity and combustion products, the data rationality is verified.
[0052] The implementation mode of the random forest classifier is as follows: The infrared heat radiation standard deviation, the visible light smoke texture contrast, the light flow field main direction variance, the heat area centroid drift speed and the combustion efficiency are input into the random forest classifier as input features; When the input features meet the first preset value, the random forest classifier outputs "incomplete combustion"; when the input features meet the second preset value, the random forest classifier outputs "external heat source"; For part of the feature combination that meets the condition, the physical constraint model is started for secondary verification.
[0053] In the judgment by the incomplete combustion judgment model: first, the infrared data and visible light data are extracted, including the infrared modal features of heat radiation distribution entropy, heat wave edge irregularity, and heat area centroid drift speed, and the visible light modal features of smoke color features, texture contrast, and optical flow field direction.
[0054] The heat radiation distribution entropy is obtained by calculating the standard deviation of the heat radiation intensity time series, which quantifies the fluctuation degree of the target heat radiation signal in the time dimension to distinguish between external heat sources and incomplete combustion. The main reason is that the heat radiation intensity of the external heat source changes little over time, showing a stable state. While the incomplete combustion has a dramatic change in heat radiation intensity due to the fluctuation of combustion efficiency, resulting in significant fluctuations.
[0055] By calculating the standard deviation of the heat radiation intensity time series, the degree of deviation of the data from the mean can be directly quantified, which can effectively distinguish between the two states; the smaller the value of the standard deviation, the more stable the heat radiation intensity; the larger the value, the more intense the heat radiation fluctuation.
[0056] The heat area edge complexity is quantified by the fractal dimension algorithm to quantify the complexity of the heat area boundary, thereby distinguishing between external heat sources and incomplete combustion. If the heat area edge presents a regular state, it is biased towards external heat sources; if the heat area edge presents an irregular fuzzy state, it is biased towards incomplete combustion. Specifically, the heat area image is divided into different size grids using the box counting method, and the number of grids covering the heat area edge is counted. By changing the grid size, the logarithmic relationship between the number of grids and the grid size is obtained, and then the fractal dimension is calculated.
[0057] The heat area centroid drift speed directly reflects the stability of the heat source by quantifying the position change of the heat source in the time dimension. External heat sources are easily affected by environmental wind speed, and the centroid drift speed is positively correlated with the wind speed. That is, the larger the wind speed, the stronger the driving force on the heat area, resulting in faster movement of the heat area centroid; on the contrary, the smaller the wind speed, the slower the heat area centroid drift speed.
[0058] The texture contrast of the smoke is extracted by a gray level co-occurrence matrix. Specifically, first, the visible light image is converted into a gray image, and the gray value combinations of all adjacent pixel pairs are counted to extract the smoke-covered area. Then, the number of occurrences of the gray value combinations is arranged into a matrix, and each element in the matrix represents the distribution frequency of a specific gray combination in the image. Finally, the roughness or delicacy of the texture is evaluated by the distribution difference of the gray combinations in the matrix; the rougher the texture, the greater the gray difference between adjacent pixels, the higher the contrast, and the greater the possibility of burning; on the contrary, the more delicate the texture, the smaller the gray difference between adjacent pixels, the lower the contrast, and the smaller the possibility of burning.
[0059] The optical flow vector of each pixel in the consecutive frame image is calculated by the Lucas-Kanade algorithm, the main direction of all pixel optical flow vectors is counted, and the variance of the main direction is calculated. The higher the variance, the more dispersed the direction of the pixels in the image over time, and the smaller the possibility of burning; the lower the variance, the more concentrated the direction of the pixels in the image over time, and the greater the possibility of burning.
[0060] The visible light image is converted from the RGB color space to the HSV color space; the hue parameter in the HSV is extracted, and the incomplete combustion and external heat source are judged by setting the hue value. The main reason is that during incomplete combustion, the characteristic yellow light emitted by the incompletely oxidized hydrocarbons can be identified by color space.
[0061] Cross-validation is performed on the multi-modal feature extraction results based on the basic principles of thermodynamics and combustion science to ensure that the classification results conform to the physical laws and solve the misjudgment problem that may be caused by simply relying on data features. The verification methods mainly include heat wave diffusion equation verification and combustion efficiency verification.
[0062] The heat wave diffusion equation mainly verifies the dynamic coupling process of the heat conduction of the external heat source and the incomplete combustion heat source, and its core mechanism can be divided into the following two parts: The heat conduction of the external heat source follows Fourier's law, and the temperature distribution changes stably over time. Assuming that the heat source is an external independent heat source, the heat conduction equation is established: ∂T / ∂t=α∇²T (α is the thermal diffusion coefficient); the theoretical temperature distribution is generated by numerical simulation, and the actual infrared data is compared with the theoretical distribution. If the deviation is less than the set threshold, it is determined to be an external heat source.
[0063] The heat release of incomplete combustion presents an intermittent pulse feature, and the temperature fluctuation contains high-frequency components. By extracting the infrared temperature time series, wavelet transform analysis is performed; the signal is decomposed into different frequency components, and the high-frequency energy ratio is calculated; if the high-frequency energy ratio is less than the set value, it is determined to be incomplete combustion.
[0064] The combustion efficiency is mainly based on the heat radiation intensity and the oxygen consumption rate, so as to verify the external heat source and incomplete combustion. That is, if the heat radiation intensity in the environment is high, but the oxygen consumption rate is low or has no obvious change, it means that the heat comes from an external independent heat source, not from flame combustion. If the oxygen consumption rate is lower than the theoretical value, and the heat radiation intensity may be low, it can be determined as incomplete combustion.
[0065] Through the random forest classifier, the multi-modal features extracted in the previous stage are combined with the dynamically adjusted threshold rule to finally determine whether the temperature anomaly in the set area is caused by an external heat source or incomplete combustion.
[0066] Specifically, the infrared thermal radiation standard deviation, the visible light smoke texture contrast, the heat region centroid drift speed, and the combustion efficiency are input into the random forest classifier as input features. The random forest classifier determines whether it is an external heat source or incomplete combustion according to the pre-input threshold of the input features.
[0067] When the input features meet the second preset value, the random forest classifier outputs a diagnostic report that the infrared data exceeds the value due to an external heat source. When the input features meet the first preset value, the random forest classifier outputs a diagnostic report that the infrared data exceeds the value due to incomplete combustion. For some feature combinations that partially meet the conditions, a physical constraint model is started for secondary verification.
[0068] After determining that it is caused by an external heat source, the spatial range is determined by heat wave trajectory analysis and incomplete combustion determination model, the occurrence period is recorded and compared with the typical characteristics of flame combustion, the affected infrared and visible light data frames are marked as interference frames, and are temporarily excluded in subsequent analysis.
[0069] For the interference frames, the infrared background is reconstructed using the Poisson image editing to eliminate the fixed heat source signal; combined with the light intensity meter data, the visible light image brightness is normalized. The repaired data is re-fused to generate a "clean" data set.
[0070] The standard deviation of the heat radiation intensity time series in the interference area is checked, which is lower than the threshold value and the heat wave trajectory center is symmetrical without high-frequency fluctuations, and then the incomplete combustion model is verified (such as oxygen consumption and heat radiation matching) to ensure that there is no residual interference.
[0071] Embodiment Four Although the above embodiments can determine whether the infrared data exceeds the value due to an external heat source or incomplete combustion through the incomplete combustion determination model, in actual use, the existing infrared flame detector cannot generate a corresponding flame combustion diffusion model according to the combustion, and generate a corresponding evacuation plan. In view of this, the technology is improved based on Embodiment Three, and the improved technical solution is as follows: Based on example two and example three, after determining that the set area has combustion or incomplete combustion phenomenon, a diffusion model is established, and a corresponding evacuation plan is generated.
[0072] Referring to Figures 1 to 7 The present application provides a flame detection method suitable for a dangerous environment, the flame detection method comprising: S7, if the output result of the diagnosis report is determined to be no-color flame combustion or incomplete combustion, a diffusion model is established and an evacuation plan is generated.
[0073] The diffusion model generates an evacuation plan by the following steps: S701, based on the diffusion rate and direction entropy of the heat wave trajectory, the propagation path of the no-color flame is simulated; The heat wave trajectory diffusion rate is calculated by the heat wave affected area change rate, the flame lateral expansion speed is determined; the disorder of the flame spread direction is quantified by using the Shannon entropy value of the heat wave trajectory direction distribution, the higher the entropy value, the more random the path; combined with the terrain characteristics of the combustion area, the fluid dynamics model is used to simulate the flame propagation path. At the same time, combined with the environmental sensor data, the deflection direction of the flame affected by the wind force is adjusted in real time, a plurality of potential propagation paths are generated and the high-risk area is marked.
[0074] S702, combined with the environmental sensor data, a three-dimensional space heat field distribution map is constructed; By mapping the infrared thermal radiation intensity to the three-dimensional space coordinate system, the obstacle contour in the visible light image is superimposed, a real-time updated three-dimensional heat field model is constructed; at the same time, according to the distance from the combustion center point, the conduction and convection attenuation process of heat in the air is simulated; taking the area with thermal radiation intensity exceeding the safety threshold as the core, combined with the diffusion range of combustion products, the high-risk, warning, and observation three-level warning areas are demarcated.
[0075] S703, according to the personnel density and safety distance threshold, a dynamic evacuation route is planned and output to the display terminal.
[0076] According to the factory's own personnel positioning system data and environmental heat field distribution, a personalized evacuation plan is generated, the evacuation plan takes the shortest safety distance as the primary goal, avoids the heat field high-risk area and obstacles, and preferentially selects the channel with good ventilation and far away from the combustion source.
[0077] At the same time of evacuation, the personnel density in each area is detected by camera or radar, the standby route is enabled for crowded sections to avoid the risk of trampling.
[0078] By refreshing the heat field data and personnel position every second, the route is dynamically adjusted and pushed to the terminal device, and the safe exit and temporary shelter position is marked.
[0079] The final evacuation scheme is presented in the form of a three-dimensional visual map on the monitoring terminal, triggers a sound and light alarm device synchronously, and pushes the route information to the on-site personnel equipment through the Internet of Things. The system continuously monitors the flame spread and evacuation progress, and if it detects that the path is blocked or the thermal field is mutated, it immediately starts the backup scheme and updates the guidance information.
[0080] Finally, it should be noted that: the above only for the preferred embodiments of the present application, and not for the purpose of limiting the present application, although the foregoing embodiments of the present application have been described in detail, for those skilled in the art, it still can be modified, or part of the technical features of the equivalent replacement, within the spirit and principles of the present application, any modification, equivalent replacement, improvement, etc., should be included within the scope of the present application.
[0081] Although embodiments of the present application have been shown and described, those of ordinary skill in the art can make various changes, modifications, replacements and deformations to these embodiments without departing from the principles and spirits of the present application, the scope of the present application is defined by the appended claims and their equivalents.
Claims
1. A method of flame detection suitable for use in a hazardous environment, characterised in that, The flame detection method comprises: S1, data collection: synchronous data acquisition of the set region is performed by the infrared light imaging module and the visible light imaging module, and infrared light imaging data and visible light imaging data are obtained; S2, data matching: the infrared light imaging data and the visible light imaging data are time-stamped matched on a time axis to form a synchronous data pair; S3, infrared judgment: temperature abnormalities of the set region are detected based on the infrared light imaging data, if the temperature abnormalities do not exceed a set threshold, data collection is continued, if the temperature abnormalities exceed the set threshold, light source interference analysis is started; S4, light source interference analysis: whether there is an interference light source is judged by analyzing light source features in the visible light image, if there is an interference light source, image restoration and feature enhancement of the interference region are performed; S5, heat wave trajectory analysis: dynamic characteristics of a heat wave trajectory are analyzed in combination with the restored visible light data and the infrared data, and a combustion state is judged according to symmetry or randomness of the trajectory; S6, incomplete combustion judgment: a multi-modal feature fusion and a physical law cross verification are performed on a center-symmetrical trajectory by an incomplete combustion judgment model, and a diagnosis report is output by a random forest classifier; S7, diffusion model establishment: if an output result of the diagnosis report judges that there is no color flame combustion or incomplete combustion, a diffusion model is established and an evacuation plan is generated.
2. The flame detection method suitable for hazardous environments according to claim 1, characterized in that, The light source interference analysis comprises: S401, multi-spectral parameter extraction is performed on the synchronous data, including infrared temperature value, temperature rise rate, visible light brightness distribution and color space feature; S402, frequency domain features of the visible light sequence are analyzed by Fourier transform, and light sources without flicker signals are excluded; S403, based on flame edge irregularity and centroid movement feature, an interference region is located in combination with a U-Net image segmentation algorithm; S404, background reconstruction of the interference region is performed by Poisson image editing, and a flame probability thermodynamic map is generated by multi-spectral data fusion.
3. The flame detection method suitable for hazardous environments according to claim 1, characterized in that, The heat wave trajectory analysis comprises: S501, ByteTrack algorithm is applied to the temperature abnormal region to perform continuous frame bounding box matching, and a trajectory line is generated; S502, motion direction entropy, curvature fluctuation rate and diffusion rate of the trajectory are extracted as feature parameters; S503, when the trajectory is in a dynamic random change state, it is judged as no color flame combustion; S504, when the trajectory is in a center-symmetrical state, whether it is influenced by an external heat source is verified in combination with an incomplete combustion judgment model.
4. The flame detection method suitable for hazardous environments according to claim 3, characterized in that, The incomplete combustion judgment model comprises: S601, a heat wave diffusion equation is constructed, a heat conduction process of an external heat source is modeled as a diffusion equation driven by Fourier's law, and a thermal radiation distribution stability threshold is set; S602, wavelet transform analysis is performed on an incomplete combustion heat source term Q, high frequency components of a temperature signal are extracted, and when a high frequency energy proportion exceeds a set value, it is judged as an intermittent pulse feature; S603, combustion efficiency is calculated in combination with thermal radiation intensity and oxygen consumption rate, when the combustion efficiency is less than a set value and the oxygen consumption rate deviates from a theoretical value, incomplete combustion is confirmed.
5. The flame detection method suitable for hazardous environments according to claim 3, characterized in that, The multi-modal feature fusion comprises: infrared mode: thermal radiation distribution uniformity, heat wave edge complexity and centroid motion stability are extracted; Visible light modalities: detecting smoke color features, texture contrast, and light flow field direction consistency; Physical constraints: establishing a nonlinear mapping relationship between thermal radiation intensity and combustion products, and verifying data rationality.
6. The flame detection method suitable for hazardous environments according to claim 5, characterized in that, The implementation of the random forest classifier is: The infrared thermal radiation standard deviation, visible light smoke texture contrast, light flow field main direction variance, thermal zone centroid drift speed, and combustion efficiency are input into the random forest classifier as input features; When the input features meet the first preset value, the random forest classifier outputs "incomplete combustion"; when the input features meet the second preset value, the random forest classifier outputs "external heat source"; For some feature combinations that partially meet the conditions, a secondary verification of the physical constraint model is started.
7. The flame detection method suitable for hazardous environments according to claim 3, characterized in that, The diffusion model generates an evacuation plan by the following steps: S701, based on the diffusion rate and direction entropy of the heat wave trajectory, simulating the propagation path of the colorless flame; S702, combining environmental sensor data, constructing a three-dimensional space thermal field distribution map; S703, according to the personnel density and safety distance threshold, planning dynamic evacuation routes and outputting to the display terminal.
8. A flame detector for carrying out the method of any one of claims 1 to 7, characterized in that The flame detector includes a data acquisition unit, a processing unit, and a decision unit; The data acquisition unit includes an infrared imaging module and a visible light imaging module; The data processing unit includes: A timestamp alignment module that aligns infrared light data and visible light data using the PTP precise clock protocol; A light source interference analysis module that integrates FFT frequency domain analysis, edge shape recognition, and multispectral correlation verification algorithms to identify and process light source interference in the specified area; A heat wave trajectory generation module that analyzes and identifies the heat wave trajectory in the specified area using the ByteTrack algorithm; An incomplete combustion judgment module that further determines the identification of the heat wave trajectory through multi-modal feature fusion; The decision unit includes: A diffusion model generation module that outputs an evacuation plan based on the diffusion rate and direction entropy of the heat wave trajectory.
9. The flame detector of claim 8, wherein, The data acquisition unit also includes a parameter configuration interface for adjusting the parameters of the infrared imaging module and the visible light imaging module according to actual scene requirements.
10. The flame detector of claim 8, wherein, The data processing unit also includes a data storage module for storing collected raw data, intermediate data during processing, and final result data.