A method for predicting heat release rate based on flame fusion characterization

Abstract

The present application relates to the field of fire safety monitoring and computer vision, in particular to a heat release rate prediction method based on flame fusion representation, obtains fire video and heat release rate and matches evolution trend to realize synchronization, extracts regional color, brightness and saturation distribution to construct multi-dimensional fusion feature, divides enhancement, active and decay stage along color migration path, uses continuous time unit to process non-equidistant sampling and corrects smoke shielding state, combines hidden state to associate combustion inertia, and generates heat release rate prediction result. According to the evolution continuous relationship, the dynamic time correction is constructed, the video and the heat release record time sequence are consistent, the cross-section correlation deviation interference is reduced, the color brightness representation is formed by fusing the boundary and the core area, the combustion state distinguishing ability is enhanced, the state constraint is formed by the migration path and the continuous trajectory, the shielding before and after is connected stably, and the stage recognition accuracy and the continuous prediction reliability are improved.

Description

Technical Field

[0001] This invention relates to the fields of fire safety monitoring and computer vision technology, and in particular to a method for predicting heat release rate based on flame fusion characterization. Background Technology

[0002] The field of fire safety monitoring and computer vision technology includes flame image acquisition, video acquisition, flame region segmentation, boundary extraction, visual feature analysis, and combustion process characterization. Its core lies in depicting the spatial morphology and temporal evolution of flames based on continuous video data, and establishing the correlation between flame visual information and heat release rate. Heat release rate is used to describe fire intensity, combustion scale, and degree of danger. Traditional acquisition methods usually rely on contact or semi-contact means such as oxygen consumption measurement, mass loss measurement, or calorimetry devices, which require specialized equipment and controlled environmental conditions, making them difficult to apply to real-time monitoring in complex real-world scenarios.

[0003] With the development of image processing, computer vision, and deep learning technologies, video-based non-contact heat release rate prediction methods have gradually attracted attention. Typically, a correspondence between video information and heat release rate is established by extracting visual features of the flame region. In existing solutions, some methods rely only on morphological parameters such as flame area, height, or width, which are insufficient to fully characterize complex visual information such as flame texture distribution, brightness changes, local combustion state, and smoke obstruction. Some methods use convolutional neural networks to directly extract deep features, but the physical meaning of the features is not intuitive enough and lacks a clear correspondence with changes in flame morphology, contour fluctuations, and scale changes. Existing discrete-time modeling methods are insufficient to fully describe the cross-frame evolution process in continuous fire videos. Furthermore, in cases of short-term flame flickering, smoke obstruction, or no-flame frames, the lack of constraint processing can easily affect the stability and reliability of the prediction results.

[0004] Among them, the heat release rate prediction method based on flame fusion characterization refers to processing continuous frame images in flame videos, extracting information on flame area changes, boundary contours, color components, grayscale intensity, and texture distribution, and combining these multiple visual features to form a unified flame fusion feature sequence. This sequence describes the change relationship between adjacent frames in the time dimension, establishes the correspondence between the flame feature sequence and the heat release rate, and covers flame area extraction, color calculation, brightness statistics, contour acquisition, and feature sequence construction. It also incorporates constraints based on flame flickering, smoke obstruction, and frames without flames, thereby describing the flame evolution process and predicting the heat release rate based on continuous frame data.

[0005] When existing technologies use fixed correspondences to connect video information and heat release data, correlation shifts are prone to occur across time segments. It is difficult to maintain a stable correspondence between the continuous combustion process and the heat release change process. There is a lack of coordinated representation of color changes, brightness differences and spatial evolution relationships between local areas of the flame. Local combustion state recognition is prone to confusion. When encountering scenarios such as smoke obstruction, flame interruption or changes in sampling interval in continuous video processing, there is a lack of effective connection basis for the combustion states before and after, which can easily cause stage transition distortion, abrupt changes in heat release, and insufficient continuity of prediction results. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this invention provides a method for predicting the heat release rate based on flame fusion characterization. The technical solution is as follows: A method for predicting heat release rate based on flame fusion characterization includes the following steps: S1: Acquire the flame video frame sequence, heat release rate record and acquisition time, establish an initial time mapping, identify the evolution trend of the video flame and the change trend of heat release, divide into consistent and inconsistent segments, dynamically adjust the mapping for inconsistent segments, reconstruct the correlation relationship, and generate a time-synchronized flame sequence. S2: Based on the time-synchronized flame sequence, determine the synchronized flame video frame, extract the flame boundary region and the flame core region, read the RGB information of the pixels in the two types of regions and classify the colors, organize the brightness median feature and saturation distribution, and combine the spatial position relationship to construct the flame color brightness fusion feature. S3: Based on the flame color brightness fusion characteristics, divide the local flame region in the synchronization frame, match adjacent regions according to spatial continuity, color consistency or color migration order, connect them to form a color migration path, track the color brightness saturation changes and mark the stages, and form a combustion stage change sequence. S4: Based on the combustion stage change sequence, construct a closed continuous time state modeling unit, update the local combustion state according to the actual time interval, generate time constant and update coefficient from the fusion features and historical hidden states, locate the disappearing frame, verify the occlusion re-connection and reconnection mark, and generate the continuous combustion stage trajectory.

[0007] As a further aspect of the present invention, the time-synchronized flame sequence includes a synchronization timestamp, a corresponding flame frame, a heat release rate value, and a time-related segment; the flame color brightness fusion feature includes a boundary color ratio, a core color ratio, regional brightness features, and a saturation distribution state; the combustion stage change sequence specifically includes an enhancement stage marker, an active stage marker, a decay stage marker, and a stage transition sequence; and the continuous combustion stage trajectory includes a stage continuation state, an occlusion connection state, a regional recurrence state, and a continuous evolution state.

[0008] As a further aspect of the present invention, the step of obtaining S1 is as follows: S11: Obtain the flame video frame sequence, heat release rate record and corresponding acquisition time in the building fire monitoring scene, match the video frame acquisition time and heat release rate acquisition time in sequence, organize the correspondence between video frame number, heat release rate sequence number and acquisition time, and establish the initial time mapping relationship. S12: Based on the initial time mapping relationship, identify the trends of flame outline expansion, outline contraction, brightness enhancement and color migration in adjacent video frames, identify the rising, stable and falling trends in the heat release rate record, judge the flame evolution trend and the heat release rate change trend segment by segment, and generate trend matching segments. S13: Call the trend matching segment, retain the time correspondence between video frames and heat release rate records in the trend consistent segment, and for the trend inconsistent segment, adjust the time mapping according to the adjacent trend consistent segment, the flame evolution continuity relationship and the heat release rate change continuity relationship, and establish a time-synchronized flame sequence.

[0009] As a further aspect of the present invention, the step of obtaining S2 is as follows: S21: Based on the association between the reconstructed flame video frames and heat release rate records in the time-synchronized flame sequence, read the frame items and heat release rate record items in the association according to the video frame number order, filter out the frame items with acquisition time links, mark the reading position of the frame item area, separate the flame boundary area and the flame core area from the synchronized flame video frames, and generate a set of area coordinates. S22: Call the region coordinate set, read the pixel RGB information of the flame boundary region and the flame core region, determine the pixel color belonging according to the proportion of red channel, green channel and blue channel, classify the pixel into red, yellow, bright white and dark gray distribution items, organize the color classification position and region belonging, and generate color distribution quantity. S23: Based on the color distribution, organize the brightness distribution of the flame boundary region and the flame core region, determine the median brightness feature corresponding to the middle position of the brightness ranking, statistically analyze the saturation distribution of the two types of regions, and combine the spatial positional relationship between the flame boundary region and the flame core region to fuse the color distribution, median brightness feature and saturation distribution to establish the flame color brightness fusion feature.

[0010] As a further aspect of the present invention, the step of obtaining S3 is as follows: S31: Based on the spatial positional relationship, color distribution, median brightness feature and saturation distribution in the flame color and brightness fusion features, the flame boundary region and flame core region in the synchronized flame video frame are divided according to the adjacent pixel connection relationship, the region boundary closure state and the core adjacency state are determined, and a local region sequence is generated. S32: Call the local region sequence, read the position order and color distribution type of the local flame region between adjacent time frames, determine the spatial position continuity, color distribution type consistency, and preset color migration order correspondence, connect the matched local flame regions, and obtain the color migration path; S33: Based on the color migration path, track the changes in the median characteristics and saturation distribution of red, yellow, bright white, dark gray, and brightness, determine that the red-yellow change corresponds to the enhancement stage, the yellow-white change corresponds to the active stage, and the white-gray change corresponds to the decay stage, connect the stage markers in chronological order, and generate a combustion stage change sequence.

[0011] As a further aspect of the present invention, the step of obtaining S4 is as follows: S41: Based on the local flame region stage markers, color migration paths, regional spatial positional relationships, and flame color brightness fusion features in the combustion stage change sequence, a closed continuous time state modeling unit is constructed. According to the time interval between adjacent synchronization frames, the current fusion features and historical hidden states are called to generate a dynamic time constant and obtain the hidden state update relationship. S42: Call the hidden state update relationship, merge the historical burning state and the current burning state according to the update ratio, and locate the flame area disappearance frame based on the hidden state continuation of adjacent synchronization frames, combined with the changes in flame burning stage, color migration speed and flame spatial shape changes, and generate a flame disappearance frame identifier. S43: Based on the flame disappearance frame identifier, read the visible state, color migration path continuity, and hidden state continuity of the local flame area before and after the disappearance frame, determine that the position where the area changes from visible to invisible is the flame-obscured smoke area, determine that the subsequent visible position is the flame reappearance area, and generate the obscured reappearance interval. S44: For the obscuration and reproduction interval, read the position sequence, color migration direction and hidden state change of the flame obscuration smoke area and the flame reproduction area. When the position is continuous, the direction is consistent and the state is continuous, the stage mark before disappearance is used. When the position jumps, the direction changes or the state changes abruptly, the stage mark is updated to establish a continuous combustion stage trajectory.

[0012] As a further aspect of the present invention, the closed-loop continuous-time state modeling unit updates the hidden state used to characterize the continuous combustion state of the flame according to the flame color brightness fusion characteristics, combustion stage markers, color migration paths, and occlusion connection states in the following manner: ;in: State retention coefficient satisfy: ; Variable time constant satisfy: ; Candidate hidden state satisfy: ; in, This represents the actual time interval between the t-th time step and the (t-1)-th time step; Let represent the variable time constant at time t, and >0, and Use the same time unit or the same normalized time scale; This represents the state update coefficient, used to control the proportion of candidate combustion hidden states in the current hidden state update. The value range is from 0 to 1; 1- This indicates the percentage of historical burning hidden states that are retained in the current hidden state update; This represents the hidden state of continuous flame burning at time t. This represents the hidden state of continuous flame combustion at time t-1; This represents the candidate combustion hidden state generated by the current flame color brightness fusion feature and the flame continuous combustion hidden state of the previous moment; represents the flame color brightness fusion feature vector at time t, which includes the boundary color ratio, core color ratio, regional brightness features, and saturation distribution; ⊙ represents element-wise multiplication; , , , This represents a trainable weight matrix; , Indicates the bias term; This represents the element-wise positiveization function used to generate positive time constants; This represents the element-wise nonlinear transformation function used to generate candidate combustion hidden states.

[0013] As a further aspect of the present invention, the closed continuous-time state modeling unit is executed by the processor of the building fire monitoring equipment. It is used to generate a continuous combustion state vector based on the synchronization timestamp, flame color and brightness fusion characteristics, combustion stage change sequence, and historical combustion hidden state. The continuous combustion state vector is then output to the heat release rate prediction unit to dynamically constrain the flame continuous combustion inertia, stage transition continuity, and state continuity relationship after short-term obstruction, thereby generating a heat release rising segment, a heat release peak segment, a heat release falling segment, and a continuous heat release change curve.

[0014] As a further aspect of the present invention, the step of obtaining S5 is as follows: S51: Based on the local flame region stage markers reconnected in time sequence in the continuous combustion stage trajectory, read the flame morphology change records corresponding to the enhancement stage, active stage, and decay stage, combine them with continuous time dynamic state modeling to obtain hidden state information, determine the correlation between stage markers and flame continuous evolution characteristics, and generate stage morphology records. S52: Call the stage morphology record, read the flame area expansion and red distribution migration to yellow distribution process in the enhancement stage, read the bright white distribution persistence and saturation convergence process in the active stage, read the dark gray distribution expansion and flame area contraction process in the decay stage, determine the heat release segment assignment, and obtain the heat release segment identifier. S53: Based on the heat release segment identifier, the heat release rising segment, heat release peak segment and heat release falling segment are connected in the time sequence of the continuous combustion stage trajectory. Combined with the hidden state, the continuous combustion inertia of the flame, the continuity of the stage transition and the state continuation relationship after short-term obstruction are dynamically constrained to generate the heat release rate prediction result.

[0015] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this invention, a dynamic time correction mechanism is constructed based on the continuous relationship between flame evolution and heat release change. This mechanism maintains the temporal consistency between the video frame sequence and the heat release record, reduces the interference caused by cross-segment correlation offset to heat release analysis, and combines the flame boundary region and core region to form a color brightness fusion characterization, thereby enhancing the ability to distinguish local combustion states. Furthermore, it combines the color migration path and the stage continuous trajectory to form a continuous state constraint, maintaining the stability of the connection between combustion stages before and after flame obstruction, and improving the smoothness of the heat release change process, the accuracy of stage identification, and the reliability of continuous prediction under complex combustion environments. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart illustrating the acquisition process of S1 in this invention; Figure 3 This is a flowchart illustrating the acquisition process of S2 in this invention; Figure 4 This is a flowchart illustrating the acquisition process of S3 in this invention; Figure 5 This is a flowchart illustrating the acquisition process of S4 in this invention; Figure 6 This is a flowchart of the acquisition process for S5 of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0019] Please see Figure 1 This invention provides a technical solution: a method for predicting the heat release rate based on flame fusion characterization, comprising the following steps: S1: Acquire the flame video frame sequence, heat release rate record, and corresponding acquisition time in the building fire monitoring scene. Based on the acquisition time, establish an initial time mapping relationship between the flame video frame sequence and the heat release rate record. Identify the flame outline expansion trend, flame outline contraction trend, flame brightness enhancement trend, and flame color migration trend in adjacent video frames. Identify the upward trend, stable trend, and downward trend in the heat release rate record. Match the flame evolution trend with the heat release rate change trend according to the initial time mapping relationship. Divide the time segments with consistent trends into time segments with inconsistent trends. For time segments with consistent trends, retain the original time mapping relationship between the video frame and the heat release rate record as the synchronization benchmark for adjacent time segments. For time segments with inconsistent trends, dynamically adjust the time mapping between the video frame and the heat release rate record based on the time continuity relationship, flame evolution continuity relationship, and heat release rate change continuity relationship in adjacent time segments with consistent trends. Reconstruct the association relationship between the flame video frame sequence and the heat release rate record based on the adjusted time mapping relationship to generate a time-synchronized flame sequence. S2: Based on the correlation between the reconstructed flame video frames and heat release rate records in the time-synchronized flame sequence, determine each synchronized flame video frame, extract the flame boundary region and flame core region from the synchronized flame video frames, read the RGB information of the pixels in the two types of regions, classify the pixels in the flame boundary region and flame core region according to the color channel ratio, forming red, yellow, bright white and dark gray distributions. After the color distribution is formed, organize the brightness distribution in the flame boundary region and flame core region respectively, obtain the median brightness feature of the flame boundary region and the median brightness feature of the flame core region, and count the saturation distribution in the two types of regions to form a concentrated color distribution and a discrete color distribution. Combine the spatial position relationship between the flame boundary region and the flame core region, fuse the color distribution, median brightness feature and saturation distribution to construct the flame color brightness fusion feature. S3: Based on the spatial positional relationship, color distribution, median brightness feature, and saturation distribution in the flame color and brightness fusion features, the flame boundary region and flame core region in the synchronized flame video frame are divided into several local flame regions. Between adjacent time frames, local flame regions that meet the requirements of spatial positional continuity, consistent color distribution type, or conform to the preset color migration order are matched. The matched local flame regions are connected along the time sequence to form a color migration path. The dynamic changes of red, yellow, bright white, dark gray, median brightness feature, and saturation distribution are tracked along the color migration path. When the red distribution weakens and the yellow distribution strengthens and is accompanied by an increase in brightness, it is marked as the enhancement stage. When the yellow distribution weakens and the bright white distribution expands and is accompanied by a convergence in saturation, it is marked as the active stage. When the bright white distribution shrinks and the dark gray distribution expands and is accompanied by a decrease in brightness, it is marked as the decay stage. The enhancement stage, active stage, and decay stage are connected in the time sequence to form a combustion stage change sequence. S4: Based on the local flame region stage markers, color migration paths, regional spatial relationships, and flame color-brightness fusion features in the combustion stage change sequence, a closed continuous-time state modeling unit is constructed. According to the actual time interval between adjacent synchronized video frames, the continuous combustion state of the local flame region is dynamically updated. A dynamic time constant is generated using the flame color-brightness fusion features at the current time step and the historical hidden state. A state update coefficient is generated based on the dynamic time constant and the actual time interval to control the proportion of historical and current combustion states in the hidden state update. This allows the hidden state to adaptively update according to the flame combustion stage change rate, color migration rate, and flame spatial morphology change rate, thus addressing the continuous evolution of the flame, the non-equidistant sampling process, and the continuous change of the combustion state. The continuous-time dynamic modeling of the combustion process is used to locate the flame region disappearance frame based on the continuous-time dynamic state modeling results. According to the visible state changes, color migration path continuity, and hidden state continuity of the local flame region before and after the flame region disappears, the flame-obscured smoke region and the flame reappearance region are identified. The connection relationship of the combustion stages is verified based on the positional continuity, color migration direction, and hidden state change continuity between the flame-obscured smoke region and the flame reappearance region. When the flame-obscured smoke region and the flame reappearance region are positionally continuous, have the same color migration direction, and have continuous hidden state changes, the stage marker before disappearance is used. When the position jumps, the color migration direction changes, or the hidden state changes abruptly, the stage marker after reappearance is updated. The stage markers of each local flame region are reconnected in chronological order to generate a continuous combustion stage trajectory. S5: Based on the stage markers of local flame regions reconnected in chronological order in the continuous combustion stage trajectory, flame morphology change records corresponding to the enhancement stage, active stage, and decay stage are extracted respectively. Combined with the hidden state information obtained from continuous time dynamic state modeling, dynamic correlation analysis is performed on the continuous evolution characteristics of flames under different combustion stages. The process of flame region expansion and red distribution migrating to yellow distribution in the enhancement stage is identified as the heat release rising segment. The process of bright white distribution persisting in the core area of ​​the flame in the active stage and accompanied by saturation convergence is identified as the heat release peak segment. The process of dark gray distribution expansion and flame region contraction in the decay stage is identified as the heat release falling segment. The heat release rising segment, heat release peak segment, and heat release falling segment are connected according to the chronological order of the continuous combustion stage trajectory. Combined with the hidden state, the continuous combustion inertia of flames, the continuity of stage transitions, and the state continuity relationship after short-term flame obstruction are dynamically constrained to generate heat release rate prediction results.

[0020] The time-synchronized flame sequence includes a synchronization timestamp, corresponding flame frame, heat release rate value, and time-related segments. The flame color and brightness fusion features include boundary color ratio, core color ratio, regional brightness features, and saturation distribution. The combustion stage change sequence specifically includes enhancement stage markers, active stage markers, decay stage markers, and stage transition order. The continuous combustion stage trajectory includes stage continuation state, occlusion connection state, regional recurrence state, and continuous evolution state. The heat release rate prediction results include the heat release rise segment, heat release peak segment, heat release fall segment, and continuous heat release change curve.

[0021] Please see Figure 2 The steps to obtain S1 are as follows: S11: Obtain the flame video frame sequence, heat release rate record and corresponding acquisition time in the building fire monitoring scene, match the video frame acquisition time and heat release rate acquisition time in sequence, organize the correspondence between video frame number, heat release rate sequence number and acquisition time, and establish the initial time mapping relationship. To acquire flame video frame sequences, heat release rate records, and corresponding acquisition times in building fire monitoring scenarios, high-definition surveillance cameras deployed on-site continuously acquire video streams. Heat release rate sensors or calorimeters based on the oxygen consumption principle are used to collect power data. An NTP time synchronization server is used to unify the system clocks of all acquisition devices to obtain the corresponding acquisition times. The acquired initial video files are then processed according to frame rate parameters. The frames are broken down every second, and the millisecond-level timestamps corresponding to each frame are extracted. And assign a unique integer number Synchronously retrieve floating-point power data uploaded by the heat release rate sensor. Record the corresponding sensor sampling time point. And label the data sequence number. Set time alignment window threshold for Seconds, timestamp each group of video frames. Sampling time with heat release rate Perform subtraction to obtain the absolute deviation value When the deviation value is within the range Within seconds, determine the video frame number. With heat release rate number It has the property of physical time coincidence, for example, in the 1st Captured in seconds with the number The frame image, while the sensor is in Recorded in seconds The heat release rate of kilowatts, with a deviation of If the second is less than the threshold, then... and Address binding is performed, and when the deviation value is greater than... If the time interval is determined to be out of sync, it indicates a delay in data acquisition and transmission. In this case, a linear interpolation method is used to calculate the estimated power at the corresponding moment based on adjacent heat release rate sampling points, and the results are replaced according to the increasing order of the video frames. With heat release rate data sequence Perform many-to-one or one-to-one linear mapping storage to establish the initial time mapping relationship.

[0022] S12: Based on the initial time mapping relationship, identify the trends of flame outline expansion, outline contraction, brightness enhancement and color migration in adjacent video frames, identify the rising, stable and falling trends in the heat release rate record, judge the flame evolution trend and the heat release rate change trend segment by segment, and generate trend matching segments. Based on the initial time mapping relationship, the computer vision image reading algorithm is used to call the number. and Extract pixel coordinates from two adjacent frames of images. grayscale value at Calculate the number of pixels within the outline region Change ,when Greater than the set area growth benchmark value When there are 100 pixels, it is determined to be a contour expansion trend. In the range The system is determined to show a contour contraction trend, and the average brightness of the region is calculated simultaneously. The difference ,like In the positive range This indicates a brightness enhancement trend, and the RGB color components are extracted. Channel mean and Ratio of channel means migration amount ,like If an offset occurs, record the color migration trend and retrieve the corresponding sequence number at the same time. and Heat release rate value and Perform subtraction to obtain ,when Greater than the power fluctuation benchmark value Kilowatt-hours are marked as having an upward trend, when In The kilowatt range is marked as a stable trend, when Less than The kilowatt-hours are marked as a downward trend. The expansion or enhancement features identified in the video frame are logically ANDed with the upward features in the heat release rate. If the two types of trend features exist simultaneously in the same time period, the data segment is defined as a trend matching segment. If the continuity of the evolution trend is determined to be inconsistent, that is, the video shows expansion while the heat release rate shows a decrease or stabilization, it means that the fire source may be blocked or the sensor response is delayed. This segment is defined as a trend inconsistency segment for subsequent alignment correction.

[0023] S13: Call the trend matching segment, retain the time correspondence between video frames and heat release rate records in the trend consistent segment, and for the trend inconsistent segment, adjust the time mapping according to the adjacent trend consistent segment, the flame evolution continuity relationship and the heat release rate change continuity relationship, and establish a time-synchronized flame sequence. Call the trend matching segment to extract the starting value of the video frame number within the segment. To the termination value and its associated heat release rate sequence number arrive For segments with consistent trends, the original linear correspondence table index is directly retained. For abnormal segments with inconsistent trends, the last time point of the forward nearest neighbor segment with consistent trends is selected. The starting time point of a trend-consistent segment with the backward nearest neighbor. Calculate the video frame time span between two reference points. Time span of heat release rate records Calculate the slope of flame evolution continuation. (i.e., the average rate of change of the contour area over time) and the continuous slope of the heat release rate change (i.e., the average rate of change of power with respect to data), set a dimensionless time adjustment factor. Used to quantify the scaling ratio of the time axis. For example, when the video shows that the flame has begun to shrink but the sensor value is still at a high plateau, the video frame time axis is scaled according to the continuity logic of adjacent segments. Perform translation and scaling operations Using the starting frame of flame contour contraction as the synchronization alignment anchor point for the beginning of the heat release rate decline trend, the time offset in the mapping table is adjusted by addition and subtraction to establish a time-synchronized flame sequence.

[0024] Please see Figure 3 The steps to obtain S2 are as follows: S21: Based on the correlation between the reconstructed flame video frames and heat release rate records in the time-synchronized flame sequence, read the frame items and heat release rate record items in the correlation according to the video frame number order, filter the frame items with acquisition time links, mark the reading position of the frame item area, separate the flame boundary area and the flame core area from the synchronized flame video frames, and generate a set of area coordinates. Based on the correlation between the reconstructed flame video frames and heat release rate records in the time-synchronized flame sequence, the corrected linear correspondence table index from the previous step S13 is retrieved to extract the video frame number at each synchronization moment. and the corresponding heat release rate values Iterate through each row of data in the mapping table, remove invalid entries with missing video frames or heat release rate power values ​​of zero, filter out link frame entries with completely overlapping timestamps and complete pixel matrices, and set a flame detection brightness threshold. for Traverse all coordinate points within the image frame. pixel values ,Will The set of pixels is determined to be the flame connected component. If no connected component that meets the threshold is detected, the frame is recorded as an empty frame and subsequent region division is skipped. The geometric center coordinates of the connected component are obtained. Calculate the Euclidean distance from each pixel in the connected component to the center coordinates. Set boundary decision coefficients for Calculate the maximum distance This will satisfy the distance interval The pixels marked as the location of the flame core region will satisfy the distance range The pixels are marked as the locations of the flame boundary region, for example, for a certain location with a heat release rate of... Image of a kilowatt-level fire source, with its flame area spanning... The pixel range is calculated to determine the number of pixels in its core region. Each, the boundary region contains pixels. Each of the two types of regions is recorded with its own horizontal and vertical coordinates to generate a region coordinate set.

[0025] S22: Call the region coordinate set, read the pixel RGB information of the flame boundary region and the flame core region, determine the pixel color belonging based on the proportion of red channel, green channel and blue channel, classify the pixel into red, yellow, bright white and dark gray distribution items, organize the color classification position and region belonging, and generate color distribution quantity. Call the regional coordinate set to retrieve the red coordinates corresponding to each coordinate point within the core and boundary regions. ,green ,blue To prevent calculation errors caused by a zero denominator in the three-channel component values, a minimum value is introduced. Perform normalization to obtain the percentage of each channel, i.e. , , Set the color classification logic judgment criteria, when and When the pixel is assigned to the red distribution term, , and When the pixel is assigned to the yellow distribution term, All three values ​​are greater than And the absolute values ​​of their differences are all less than When the pixel is classified as a white brightness distribution term, All three values ​​are within the range. Pixels are categorized as dark gray in the distribution term, while pixels that do not meet any of the above conditions are categorized as background interference. The total number of pixels in the core region for each distribution term is then counted. Total number of pixels in the boundary region The proportion of pixels detected, for example, bright white pixels detected in the core area. Individual, accounting for Red pixels were detected in the boundary area. Individual, accounting for Record the mapping relationship between these color category labels and coordinate points, and calculate the pixel frequency of each category in the two regions. Generate color distribution.

[0026] S23: Based on the color distribution, organize the brightness distribution of the flame boundary region and the flame core region, determine the median brightness feature corresponding to the middle position of the brightness ranking, statistically analyze the saturation distribution of the two types of regions, and combine the spatial position relationship between the flame boundary region and the flame core region to fuse the color distribution, median brightness feature and saturation distribution to establish flame color brightness fusion features. Based on the color distribution, extract the grayscale response values ​​of all pixels within the flame boundary region and flame core region, and sort them in ascending order from low to high. Locate the pixel in the sorted sequence that is at the [number]th position. The position value is used as the median feature of brightness. Retrieve saturation calculation parameters for the two types of regions: ,statistics Value at The probability distribution within the range is used to calculate the mean saturation value of the core region and the boundary region respectively. and and saturation standard deviation ,like If the color is concentrated, then... The color distribution is determined to be discrete; to standardize the units, the median luminance value is divided by... Mapped to Normalized brightness is obtained from the interval The proportion of bright white pixels in the core area, the proportion of red pixels in the boundary area, and the median normalized brightness of the two areas are calculated. and And the saturation distribution type is used to perform matrix concatenation, introducing spatial location weight coefficients. Its value is determined by the ratio of the area of ​​the core area to the area of ​​the boundary area. For example, the current core area is pixels, with a boundary area of pixels, then Set as Perform weighted fusion operation: Establish flame color and brightness fusion features.

[0027] Please see Figure 4 The steps to obtain S3 are as follows: S31: Based on the spatial positional relationship, color distribution, median brightness feature and saturation distribution in the flame color and brightness fusion features, the flame boundary region and flame core region in the synchronized flame video frame are divided according to the connection relationship of adjacent pixels. The region boundary closure state and the core adjacency state are determined to generate a local region sequence. Based on the spatial relationship, color distribution, median brightness feature, and saturation distribution in the flame color-brightness fusion features, the feature matrix generated in the previous step S23 is retrieved, and the core pixel proportion is extracted. The proportion of red borders and median brightness for The feature vectors are used to traverse the pixel coordinates in the synchronized flame video frame using the eight-neighbor pixel connectivity rule. When the Euclidean distance between adjacent pixels is... for If all pixels belong to the flame boundary region or the flame core region, logical merging is performed. For the boundary region, closed path search is performed, and the area of ​​the polygon formed by the boundary pixels is calculated. Total number of actual pixels at the boundary The difference, when the absolute value of the difference is in the interval When the difference is within a pixel, it is determined to be a closed region boundary. If the absolute value of the difference is greater than 1, it is considered a closed region boundary. If a pixel is determined to be unclosed, the convex hull algorithm is used to forcibly extract the outermost contour points to close the edge. For the pixel set in the core region, the shortest distance from each pixel to the nearest boundary pixel is calculated. ,like All greater than If the condition is determined to be a core adjacency state, the connected pixel clusters that satisfy the above geometric topological constraints are marked as independent local flame units, for example, for a preceding heat release rate of... A kilowatt-level heat source is divided into sections numbered as follows: , The core local unit and numbered The boundary local units generate a local region sequence.

[0028] S32: Call the local region sequence, read the position order and color distribution type of the local flame region between adjacent time frames, determine the spatial position continuity, color distribution type consistency, and preset color migration order correspondence, connect the matched local flame regions, and obtain the color migration path; Call the local region sequence to extract local area at any time center coordinates With color distribution retrieval Time corresponding area center coordinates With color distribution Calculate the displacement vector between two time points. Set a spatial continuous distance threshold for Pixels, if In the range Then it is determined to be spatially continuous, if Greater than If a pixel is determined to be broken or not part of the same local region, the current region link is terminated and a new index is assigned for comparison. and The color proportions in the two are considered identical if their dominant color components are the same, and the preset color migration order is set to "red". yellow Bai Liang "Dark gray", when The dominant color at all times is red and When the dominant color changes to yellow, it is determined to conform to the preset migration order. A logical OR operation is then performed to process the consistency determination result, establishing unique index links for local regions across frames. For example, the first frame... Frame coordinates The red boundary area at the location and the first Frame coordinates In areas where the yellow proportion increases, two-way pointer binding is performed, and the matched area coordinates and feature vectors are accumulated frame by frame according to the time step to obtain the color migration path.

[0029] S33: Based on the color migration path, track the changes in the median characteristics and saturation distribution of red, yellow, bright white, dark gray, and brightness, determine that the red-yellow changes correspond to the enhancement stage, the yellow-white changes correspond to the active stage, and the white-gray changes correspond to the decay stage, connect the stage markers in chronological order, and generate a combustion stage change sequence. Based on the color migration path, retrieve the percentage of red corresponding to each node on the path. Yellow percentage White and bright proportion and the proportion of dark gray The change in brightness was used to simultaneously extract the median brightness feature. Increment with saturation standard deviation The changes, the execution phase logical judgment operation, when the condition is met , and At that time, mark the path segment as a numerical value. The corresponding enhancement phase, when the following conditions are met , and The change is in When dealing with intervals, mark the path segment as a numerical value. The corresponding active phase, when the conditions are met , and At that time, mark the path segment as a numerical value. For the corresponding decay phase, set the phase transition weight coefficient. Its value is determined with reference to the slope of the heat release rate change. ,when The time-enhanced phase label weight is set to ,when The time decay phase label weight is set to If the above stage characteristics are not met or the process is in a transitional state, the previous time marker is used. The above numerical stage markers are arranged in ascending order according to the millisecond timestamps corresponding to the video frames, and combined into a time-series vector containing different local evolution states of the fire source, generating a combustion stage change sequence.

[0030] Please see Figure 5 The steps to obtain S4 are as follows: S41: Based on the local flame region stage markers, color migration paths, regional spatial positional relationships, and flame color brightness fusion features in the combustion stage change sequence, a closed continuous time state modeling unit is constructed. According to the time interval between adjacent synchronization frames, the current fusion features and historical hidden states are called to generate a dynamic time constant and obtain the hidden state update relationship. Based on the local flame region stage markers, color migration paths, regional spatial relationships, and flame color brightness fusion features in the combustion stage change sequence, a hidden state dimension is constructed. (For example, set to) The closed continuous-time state modeling unit (3D) retrieves the enhancement stage marker with a value of 1 from the preceding step S33 and the corresponding color migration vector based on the flame color brightness fusion characteristics, combustion stage markers, color migration paths, and occlusion connection states. It also reads the time difference between adjacent synchronized video frames as the first... The time step and the first The actual time interval between time steps For example, set it to Seconds, retrieve the current number Flame color brightness fusion feature vector at any moment The flame color brightness fusion feature vector includes the boundary color ratio, core color ratio, regional brightness features, and saturation distribution state, and simultaneously extracts the data retained from the previous time step. The flame burns continuously in a hidden state. Perform matrix dot product and addition operations to obtain the candidate combustion hidden state. ,in , This represents a trainable weight matrix. Indicates the bias term. This represents the element-wise nonlinear transformation function used to generate candidate combustion hidden states, and the variable time constant is calculated. ,in , This represents a trainable weight matrix. Indicates the bias term. This represents the element-wise positiveization function used to generate positive time constants, and , and Using the same time unit or the same normalized time scale, for example, when the fire source is in an active phase and the proportion of the core bright white distribution is detected as... Upgraded to At that time, the calculated Value seconds, and then calculate the state preservation coefficient. , This represents the state update coefficient, used to control the hidden state of candidate combustion. The percentage of the current hidden state update, its value range is: to , This indicates the percentage of historical burning hidden states retained in the current hidden state update, according to the formula. Update is used to characterize the first The hidden state of the flame burning continuously at all times ,in This indicates element-wise multiplication, which yields the hidden state update relationship.

[0031] S42: Call the hidden state update relationship, merge the historical burning state and the current burning state according to the update ratio, and locate the flame area disappearance frame based on the hidden state continuation of adjacent synchronization frames, combined with the changes in flame burning stage, color migration speed and flame spatial shape changes, and generate flame disappearance frame identifier. The hidden state update relation is invoked, and based on the extracted hidden state feature vectors, the Euclidean distance between the hidden states of adjacent synchronization frames is retrieved in real time. Retrieve the color migration speed calculated by computer vision. (i.e., the percentage change in the proportion of red migrating to yellow per unit time) and the rate of change in flame area. Set the disappearance judgment benchmark value for When the flame area is from Pixel count drops sharply pixels and below In the range When the range of the mutation is within a certain range, determine that the current frame is the trigger point for flame occlusion. If the area shrinks gradually and In the range This indicates that the flame has entered a natural decay state rather than being obstructed. In this case, the obstruction determination is skipped, and the video frame number at the moment of obstruction triggering is recorded. For example, when the heat release rate is continuously increasing, but the video is suddenly unable to identify the condition due to the influence of black smoke, it may not meet the requirements. When the connected component is connected, the logical bit corresponding to the frame is set to true, and a flame disappearance frame identifier is generated.

[0032] S43: Based on the flame disappearance frame identifier, read the visible state, color migration path continuity, and hidden state continuity of the local flame area before and after the disappearance frame, determine that the position where the area changes from visible to invisible is the flame-occupied smoke area, determine the subsequent visible position as the flame reappearance area, and generate the occlusion reappearance interval. Based on the flame disappearance frame identifier, trace back to the time before the disappearance frame. Frame to after Visibility status bit of local flame region in the frame Extract the coordinates of the end point of the color migration path in the last frame before it disappears. With hidden state vector Extract the starting coordinates of the first frame to reproduce With hidden state Perform spatial distance subtraction to obtain ,like In the range Within a pixel and during the disappearance period, the sensor acquires the heat release rate. The fluctuation value is less than kilowatts, then coordinates The pixel region where it is located is determined to be a smoke area obscured by flames, and the coordinates are... The region where the re-detected connected component is identified as the flame re-emergence region, if Greater than Pixels represent newly generated independent fire sources in that area, and are not associated with recurrence. The time span between the occlusion start frame and the recurrence start frame is calculated. ,like In Instant If the interval is a second, then the index sequence within that time period is recorded to generate the occlusion reproduction interval.

[0033] S44: For the occlusion and reproduction interval, read the position sequence, color migration direction and hidden state change of the flame occlusion and smoke area and the flame reproduction area. When the position is continuous, the direction is consistent and the state is continuous, the stage mark before disappearance is used. When the position jumps, the direction changes or the state changes abruptly, the stage mark is updated to establish a continuous combustion stage trajectory. For the obstruction reproduction range, read the coordinate displacement vector direction between the center of the flame-obstructed smoke area and the center of the reproduction area. Read the tangent direction of the color migration path before it disappears. ; Perform angle deviation calculation ,like In the range And the hidden state after reappearance Hidden state before disappearing The cosine similarity is greater than When the positions are determined to be continuous and the states are consecutive, the markers for the enhancement phase or active phase are directly filled into each frame within the occluded region. Greater than or similarity is If the low-order range indicates a sudden change in state or interference from disruptive airflow, the previous state is no longer used. Instead, the color distribution of the reproduced frame is retrieved, and the S33 stage determination is re-executed. For example, if the proportion of dark gray changes after reproduction... Sudden increase to The stage markers are updated to decay stages, and the corrected stage attributes of each local area are connected in series according to the time axis to establish a continuous combustion stage trajectory.

[0034] Please see Figure 6 The steps to obtain S5 are as follows: S51: Based on the stage markers of local flame regions reconnected in time sequence in the continuous combustion stage trajectory, read the flame morphology change records corresponding to the enhancement stage, active stage, and decay stage, combine them with continuous time dynamic state modeling to obtain hidden state information, determine the correlation between stage markers and flame continuous evolution characteristics, and generate stage morphology records. Based on the local flame region stage markers reconnected in time sequence in the continuous combustion stage trajectory, the time sequence marker vector containing the enhancement stage value 1, the active stage value 2, and the decay stage value 3 established in the previous step S44 is retrieved, and the corresponding flame area change rate obtained from image processing is retrieved. Change in perimeter of the profile and centroid offset distance Synchronously retrieve the data generated in the preceding step S41 3D hidden state vector Using the magnitude value and the stage marker as an index, the pixel growth gradient when the stage is marked as enhancement stage is extracted. Extract the duration of the bright white distribution in the core region during the active phase. Extract the edge grayscale diffusion coefficient marked as the decay stage. Calculate the flame area in each stage. Compared with the median value of brightness The product term sets the threshold for the correlation of evolutionary features. for Pearson correlation coefficient calculation between execution phase markers and evolutionary features ,like In the range Then it is determined that the evolutionary feature has a strong logical correlation with the corresponding combustion stage. Less than If the correlation is weak, a preset empirical threshold for heat release is used as a fallback, for example, when the heat source is in the enhancement phase and the heat release rate is high. Showing kilowatts increased to When recording the evolution path of kilowatts, the area is recorded from... Pixel linear expansion to The shape of the pixels is recorded, and the relationship is mapped and stored with the inertial components in the hidden state to generate the stage shape record.

[0035] S52: Call the stage morphology record, read the flame area expansion and red distribution migration to yellow distribution process in the enhancement stage, read the bright white distribution persistence and saturation convergence process in the active stage, read the dark gray distribution expansion and flame area contraction process in the decay stage, determine the heat release segment assignment, and obtain the heat release segment identifier. The process involves retrieving the phase morphology record and reading the total number of pixels in the flame region obtained using the finite difference method during the enhancement phase. The second derivative Simultaneously, the proportion of red components in the color migration path was retrieved. Yellow component percentage Conversion rate ,like Greater than and In the range Determine the entry point of the heat release rising phase during that time segment, and read the proportion of bright white pixels in the core region coordinate set during the active phase. ,like In the range High fluctuation range and saturation standard deviation In the range The extremely narrow convergence range is used to determine the peak heat release segment of that time period, and the proportion of dark gray distribution during the decay phase is read. Cumulative increment over time steps ,like Greater than And the total area shrinkage rate In the range The system identifies the heat release and subsequent decline phase during this time period. If none of the above conditions are met, it is marked as a transitional fluctuation zone, and the data recorded in the preceding S11 is retrieved. Using kilowatts as the peak reference point, numerical values ​​are set for each segment identifier, i.e., the rising segment is set to... Peak segment set The fallback segment is set as Establish segment identifiers and address indexes of video frame sequences to obtain heat release segment identifiers.

[0036] S53: Based on the heat release segment identifier, the heat release rising segment, heat release peak segment and heat release falling segment are connected in the time sequence of the continuous combustion stage trajectory. Combined with the hidden state, the flame continuous combustion inertia, stage transition continuity and state continuation relationship after short-term shielding are dynamically constrained to generate heat release rate prediction results. According to the heat release segmentation markers, and following the trajectory of the continuous combustion stage... Based on the timestamp order, perform concatenation and alignment operations on the rising segment, peak segment, and falling segment, and retrieve the temporal evolution weights from the hidden state vector. Set the combustion inertia coefficient for Referring to the predicted heat release rate value at the previous moment Execute prediction function calculation ,in This is a linear weighted empirical function that combines the fusion characteristics and area proportions at the current stage. Its specific form is expressed as ( These are the preset fitting weight constants corresponding to each stage. (Assuming the flame area ratio), this calculation can effectively suppress abnormal power spikes caused by interference. For the occlusion reproduction interval identified in S43, the hidden state of the previous frame is adjusted and canceled. Perform linear extrapolation and set a short-term occlusion duration threshold. for seconds, if the occlusion duration is less than Seconds then command Maintaining the smooth evolution of the slope trend before disappearance, when the segmentation marker is detected, the slope trend changes smoothly. jump to At that time, the second-order difference compensation amount before and after the transition point in the calculation stage is added to... In order to ensure the tangent continuity of the power curve, the predicted power values ​​contributed by each local region are ultimately weighted according to spatial location. Perform weighted accumulation, for example, the core unit generated The kilowatts generated by the boundary unit The kilowatts are summed to generate the predicted heat release rate.

[0037] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting heat release rate based on flame fusion characterization, characterized in that, Includes the following steps: S1: Acquire the flame video frame sequence, heat release rate record and acquisition time, establish an initial time mapping, identify the evolution trend of the video flame and the change trend of heat release, divide into consistent and inconsistent segments, dynamically adjust the mapping for inconsistent segments, reconstruct the correlation relationship, and generate a time-synchronized flame sequence. S2: Based on the time-synchronized flame sequence, determine the synchronized flame video frame, extract the flame boundary region and the flame core region, read the RGB information of the pixels in the two types of regions and classify the colors, organize the brightness median feature and saturation distribution, and combine the spatial position relationship to construct the flame color brightness fusion feature. S3: Based on the flame color brightness fusion characteristics, divide the local flame region in the synchronization frame, match adjacent regions according to spatial continuity, color consistency or color migration order, connect them to form a color migration path, track the color brightness saturation changes and mark the stages, and form a combustion stage change sequence. S4: Based on the combustion stage change sequence, construct a closed continuous time state modeling unit, update the local combustion state according to the actual time interval, generate time constant and update coefficient from the fusion features and historical hidden states, locate the disappearing frame, verify the occlusion re-connection and reconnection mark, and generate the continuous combustion stage trajectory. The steps for obtaining S4 are as follows: S41: Based on the local flame region stage markers, color migration paths, regional spatial positional relationships, and flame color brightness fusion features in the combustion stage change sequence, a closed continuous time state modeling unit is constructed. According to the time interval between adjacent synchronization frames, the current fusion features and historical hidden states are called to generate a dynamic time constant and obtain the hidden state update relationship. S42: Call the hidden state update relationship, merge the historical burning state and the current burning state according to the update ratio, and locate the flame area disappearance frame based on the hidden state continuation of adjacent synchronization frames, combined with the changes in flame burning stage, color migration speed and flame spatial shape changes, and generate a flame disappearance frame identifier. S43: Based on the flame disappearance frame identifier, read the visible state, color migration path continuity, and hidden state continuity of the local flame area before and after the disappearance frame, determine that the position where the area changes from visible to invisible is the flame-obscured smoke area, determine that the subsequent visible position is the flame reappearance area, and generate the obscured reappearance interval. S44: For the obstruction and reproduction interval, read the position sequence, color migration direction and hidden state change of the flame obstruction smoke area and the flame reproduction area. When the position is continuous, the direction is consistent and the state is continuous, the stage mark before disappearance is used. When the position jumps, the direction changes or the state changes abruptly, the stage mark is updated to establish a continuous combustion stage trajectory. The continuous combustion stage trajectory includes stage continuation state, shielding connection state, regional recurrence state, and continuous evolution state.

2. The method for predicting heat release rate based on flame fusion characterization according to claim 1, characterized in that: The time-synchronized flame sequence includes a synchronization timestamp, a corresponding flame frame, a heat release rate value, and a time-related segment. The flame color brightness fusion features include boundary color ratio, core color ratio, regional brightness features, and saturation distribution state. The combustion stage change sequence specifically includes an enhancement stage marker, an active stage marker, a decay stage marker, and a stage transition sequence.

3. The method for predicting heat release rate based on flame fusion characterization according to claim 1, characterized in that: The steps for obtaining S1 are as follows: S11: Obtain the flame video frame sequence, heat release rate record and corresponding acquisition time in the building fire monitoring scene, match the video frame acquisition time and heat release rate acquisition time in sequence, organize the correspondence between video frame number, heat release rate sequence number and acquisition time, and establish the initial time mapping relationship. S12: Based on the initial time mapping relationship, identify the trends of flame outline expansion, outline contraction, brightness enhancement and color migration in adjacent video frames, identify the rising, stable and falling trends in the heat release rate record, judge the flame evolution trend and the heat release rate change trend segment by segment, and generate trend matching segments. S13: Call the trend matching segment, retain the time correspondence between video frames and heat release rate records in the trend consistent segment, and for the trend inconsistent segment, adjust the time mapping according to the adjacent trend consistent segment, the flame evolution continuity relationship and the heat release rate change continuity relationship, and establish a time-synchronized flame sequence.

4. The method for predicting heat release rate based on flame fusion characterization according to claim 1, characterized in that: The steps for obtaining S2 are as follows: S21: Based on the association between the reconstructed flame video frames and heat release rate records in the time-synchronized flame sequence, read the frame items and heat release rate record items in the association according to the video frame number order, filter out the frame items with acquisition time links, mark the reading position of the frame item area, separate the flame boundary area and the flame core area from the synchronized flame video frames, and generate a set of area coordinates. S22: Call the region coordinate set, read the pixel RGB information of the flame boundary region and the flame core region, determine the pixel color belonging according to the proportion of red channel, green channel and blue channel, classify the pixel into red, yellow, bright white and dark gray distribution items, organize the color classification position and region belonging, and generate color distribution quantity. S23: Based on the color distribution, organize the brightness distribution of the flame boundary region and the flame core region, determine the median brightness feature corresponding to the middle position of the brightness ranking, statistically analyze the saturation distribution of the two types of regions, and combine the spatial positional relationship between the flame boundary region and the flame core region to fuse the color distribution, median brightness feature and saturation distribution to establish the flame color brightness fusion feature.

5. The method for predicting heat release rate based on flame fusion characterization according to claim 1, characterized in that: The steps for obtaining S3 are as follows: S31: Based on the spatial positional relationship, color distribution, median brightness feature and saturation distribution in the flame color and brightness fusion features, the flame boundary region and flame core region in the synchronized flame video frame are divided according to the adjacent pixel connection relationship, the region boundary closure state and the core adjacency state are determined, and a local region sequence is generated. S32: Call the local region sequence, read the position order and color distribution type of the local flame region between adjacent time frames, determine the spatial position continuity, color distribution type consistency, and preset color migration order correspondence, connect the matched local flame regions, and obtain the color migration path; S33: Based on the color migration path, track the changes in the median characteristics and saturation distribution of red, yellow, bright white, dark gray, and brightness, determine that the red-yellow change corresponds to the enhancement stage, the yellow-white change corresponds to the active stage, and the white-gray change corresponds to the decay stage, connect the stage markers in chronological order, and generate a combustion stage change sequence.

6. The method for predicting heat release rate based on flame fusion characterization according to claim 1, characterized in that: The closed-loop continuous-time state modeling unit updates the hidden state representing the continuous combustion state of the flame according to the flame color brightness fusion characteristics, combustion stage markers, color migration paths, and occlusion connection states in the following manner: ;in: State retention coefficient satisfy: ; Variable time constant satisfy: ; Candidate hidden state satisfy: ; in, Indicates the time step t and the time step t. The actual time interval between each time step; Let represent the variable time constant at time t, and >0, and Use the same time unit or the same normalized time scale; This represents the state update coefficient, used to control the proportion of candidate combustion hidden states in the current hidden state update. The value range is from 0 to 1; 1- This indicates the percentage of historical burning hidden states that are retained in the current hidden state update; This represents the hidden state of continuous flame burning at time t. This represents the hidden state of continuous flame combustion at time t-1; This represents the candidate combustion hidden state generated by the current flame color brightness fusion feature and the flame continuous combustion hidden state of the previous moment; represents the flame color brightness fusion feature vector at time t, which includes the boundary color ratio, core color ratio, regional brightness features, and saturation distribution; ⊙ represents element-wise multiplication; , , , This represents a trainable weight matrix; , Indicates the bias term; This represents the element-wise positiveization function used to generate positive time constants; This represents the element-wise nonlinear transformation function used to generate candidate combustion hidden states.

7. The method for predicting heat release rate based on flame fusion characterization according to claim 1, characterized in that: The closed-loop continuous-time state modeling unit is executed by the processor of the building fire monitoring equipment. It is used to generate a continuous combustion state vector based on the synchronization timestamp, flame color and brightness fusion characteristics, combustion stage change sequence, and historical combustion hidden state. The continuous combustion state vector is then output to the heat release rate prediction unit to dynamically constrain the flame continuous combustion inertia, stage transition continuity, and state continuity relationship after short-term obstruction, and to generate heat release rising segment, heat release peak segment, heat release falling segment, and continuous heat release change curve.

8. The method for predicting heat release rate based on flame fusion characterization according to claim 1, characterized in that: The method further includes: S5: Based on the continuous combustion stage trajectory, extract the flame morphology change records of the enhancement, activity and decay stages, combine them with the hidden state to carry out dynamic correlation analysis, determine the heat release rising segment, peak segment and falling segment, constrain combustion inertia, stage transition and shielding continuation, and generate heat release rate prediction results. The predicted heat release rate includes the heat release rising segment, the heat release peak segment, the heat release falling segment, and the continuous heat release change curve.

9. The method for predicting heat release rate based on flame fusion characterization according to claim 8, characterized in that: The steps for obtaining S5 are as follows: S51: Based on the local flame region stage markers reconnected in time sequence in the continuous combustion stage trajectory, read the flame morphology change records corresponding to the enhancement stage, active stage, and decay stage, combine them with continuous time dynamic state modeling to obtain hidden state information, determine the correlation between stage markers and flame continuous evolution characteristics, and generate stage morphology records. S52: Call the stage morphology record, read the flame area expansion and red distribution migration to yellow distribution process in the enhancement stage, read the bright white distribution persistence and saturation convergence process in the active stage, read the dark gray distribution expansion and flame area contraction process in the decay stage, determine the heat release segment assignment, and obtain the heat release segment identifier. S53: Based on the heat release segment identifier, the heat release rising segment, heat release peak segment and heat release falling segment are connected in the time sequence of the continuous combustion stage trajectory. Combined with the hidden state, the continuous combustion inertia of the flame, the continuity of the stage transition and the state continuation relationship after short-term obstruction are dynamically constrained to generate the heat release rate prediction result.

Images