Automatic tracking and monitoring system for target in high-temperature furnace based on multispectral infrared
The multispectral infrared automatic tracking and monitoring system solves the problem of judgment errors caused by reliance on human experience in high-temperature forging, achieves high-precision temperature control and improves the quality of finished products, and reduces equipment risks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- MAANSHAN KETAI ELECTRICAL ENG CO LTD
- Filing Date
- 2026-01-14
- Publication Date
- 2026-05-12
AI Technical Summary
In the current high-temperature forging process, the judgment of the workpiece forging stage mainly relies on the operator's experience, which is highly subjective and has a large judgment error. This leads to inaccurate temperature control, affecting the workpiece forming quality and the defect rate of finished products.
An automatic target tracking and monitoring system based on multispectral infrared is adopted in a high-temperature furnace. The YOLO target detection algorithm is used to identify and track targets inside the furnace. The surface temperature distribution field of the target is inverted through a multispectral radiative transfer model. The correction factor is generated by combining deformation rate characteristics and apparent morphology, the acquisition frequency is dynamically adjusted, and differentiated control commands are executed.
It enables accurate judgment during the workpiece forging stage, reduces false alarm and false alarm rates, improves the finished product qualification rate, ensures process quality, reduces the risk of equipment damage, and improves system operating efficiency and economy.
Smart Images
Figure CN122023767A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of target tracking technology, specifically to an automatic target tracking and monitoring system for high-temperature furnaces based on multispectral infrared. Background Technology
[0002] High-temperature forging is a core process in metallurgy, heavy machinery manufacturing, and aerospace component processing. Its core requirement is to precisely control the temperature and deformation of the workpiece at different forging stages to ensure its mechanical properties and forming accuracy. Automatic target tracking and monitoring within the high-temperature furnace is the core support for achieving precise process control, equipment safety protection, and intelligent production upgrades in high-temperature forging. In a forging furnace environment, automatic tracking and monitoring can continuously lock onto the workpiece, accurately determine the initial forging, intermediate forming, and final cooling stages, and achieve phased closed-loop temperature control. Without tracking and monitoring, the workpiece is prone to internal structural defects due to stage determination errors and excessive temperature fluctuations, leading to an increased defect rate.
[0003] In current forging furnace production processes, the determination of the workpiece forging stage mainly relies on the operator's personal experience, which is highly subjective and inconsistent, resulting in low stage identification accuracy and large fluctuations in results. Due to the lack of clear and consistent criteria for stage determination, subsequent temperature control is difficult to achieve accurately and stably, thereby affecting the heating uniformity of the workpiece, the metal microstructure and properties, and the final forming quality, and causing defects such as oxidation, overheating, and deformation. Summary of the Invention
[0004] The purpose of this invention is to provide an automatic target tracking and monitoring system for high-temperature furnaces based on multispectral infrared, in order to solve the problem that the current system relies heavily on the experience of operators to judge the forging stage of the workpiece, which has the problems of strong subjectivity and large judgment error, resulting in no clear stage basis for subsequent temperature control, and thus affecting the forming quality of the workpiece.
[0005] The objective of this invention can be achieved through the following technical solution: an automatic tracking and monitoring system for targets inside a high-temperature furnace based on multispectral infrared, comprising: a tracking and acquisition module: identifying and tracking targets inside the furnace based on the YOLO target detection algorithm, presetting an initial acquisition frequency and synchronously acquiring radiation image data of the target in different infrared bands and acquiring environmental parameters inside the furnace, and further comprising;
[0006] Identification and Judgment Module: Based on the acquisition of multi-band target radiation image data, the surface temperature distribution field of the target is inverted through the multispectral radiative transfer model. The target appearance morphology and deformation rate features are extracted simultaneously. The deformation rate features are matched with the preset deformation rate feature vector library to obtain preliminary labels. The surface temperature distribution field and appearance morphology are then fused to generate correction factors. The preliminary labels are corrected using the correction factors to generate the final labels.
[0007] Adjustment and comparison module: dynamically adjusts the initial acquisition frequency based on the importance of the current process status, and compares the target surface temperature distribution field with the preset process temperature reference field point by point to calculate the average temperature deviation value;
[0008] Execution module: Classifies abnormalities based on average temperature deviation, generates differentiated control commands, and executes them.
[0009] Preferably, the specific steps for identifying and tracking targets inside the furnace based on the YOLO target detection algorithm are as follows:
[0010] First, multispectral infrared images containing targets were collected under different forging stages and different levels of interference. Boundary boxes were marked on the targets in the images to form a furnace-specific target sample library. The YOLOv8 model was then trained based on the sample library.
[0011] Secondly, the current image of the target inside the furnace is acquired from the multispectral infrared camera at a preset or dynamically adjusted frequency. The image is preprocessed and then input into the trained YOLOv8 model. The model outputs the bounding box coordinates, confidence level, and target category of the detected target in the image.
[0012] Then, the information of the target in the current image output by the YOLOv8 model is input into the BoT-SORT tracker. For each existing target trajectory, the Kalman filter is used to predict its bounding box position in the current image based on its position in the previous image. BoT-SORT performs optimal matching between the detection box in the current image and the existing predicted trajectory by calculating motion similarity and appearance similarity.
[0013] Finally, the information of the matched detection boxes is updated to the corresponding trajectory, and the unique ID of the trajectory remains unchanged, thus completing the trajectory continuation. Unmatched detection boxes are initialized as new trajectories, and a new unique ID is assigned by the tracker. They are then re-matched in subsequent consecutive images. If the match is successful, the trajectory continuation is confirmed; otherwise, the trajectory is terminated.
[0014] Preferably, the specific steps for inverting the target surface temperature distribution field using a multispectral radiative transfer model are as follows:
[0015] First, subpixel-level phase correlation is used to align the pixels of the acquired multi-band target radiation images. Then, the factory calibration coefficients and on-site calibration parameters of the multispectral infrared camera are used to adjust the image pixel grayscale values. Converted to actual radiance value of the corresponding band The specific conversion formula is as follows: ,in For different infrared bands, Image pixel coordinates, For band The radiation gain coefficient, For band radiative offset coefficient;
[0016] Secondly, establish actual radiance values. relative to the target true temperature Relationship: ,in, For flue gas band transmittance, For smoke and dust on the band transmittance, To obtain the target in the band through laboratory calibration emission rate, For band Lower temperature The corresponding blackbody radiation brightness, The ambient incident radiation brightness, The radiation emitted by the flue gas itself;
[0017] Finally, for each pixel In this case, we can solve the problem by combining the relationships of all bands and then using the least squares method iteratively.
[0018] Preferably, the specific steps for extracting the target's apparent morphology and deformation rate features are as follows:
[0019] Target appearance morphology: First, select the band with the highest contrast between the target and the background inside the furnace from the multi-band image as the feature image I(x,y,t). Filter the feature image to suppress noise and stretch the contrast to highlight the target edge. Based on the area in the target surface temperature distribution field that is higher than the preset temperature threshold of the background as the initial position, use the edge detection algorithm on the feature image to determine the outline of the target and obtain the binarized image B(x,y,t). Finally, calculate its area, perimeter and density feature parameters based on the binarized image B(x,y,t).
[0020] Deformation rate feature: Retrieve contour features from two adjacent images with the same target ID, perform contour matching using an iterative nearest-neighbor algorithm, and calculate the rate of change of the width of the minimum bounding rectangle between the two images based on the matching results. Altitude change and area change and in combination with time intervals We obtain the deformation rates of width, height, and area, which are collectively referred to as deformation rate characteristics.
[0021] Preferably, the specific steps for obtaining the initial tag are as follows:
[0022] First, a vector library is constructed by summarizing the standard deformation rate characteristics of typical process stages in m from historical data, and the process stage name corresponding to each deformation rate characteristic is identified.
[0023] Then, the Euclidean geometric distance between the current deformation rate feature and the standard deformation rate feature of each process stage in the vector library is calculated using the Euclidean geometric distance formula. This distance is used to measure the similarity between the current deformation rate feature and the standard deformation rate feature of each process stage in the vector library.
[0024] Finally, the standard deformation rate feature with the smallest Euclidean geometric distance from the current deformation rate feature is selected, and the process stage name corresponding to the standard deformation rate feature is output as the initial label.
[0025] Preferably, the specific steps for obtaining the correction factor are as follows:
[0026] First, define an ideal characteristic range and an acceptable characteristic range for each process state;
[0027] Secondly, the average temperature and temperature uniformity parameters in the temperature distribution field, as well as the area, perimeter, and density parameters in the apparent morphology features, are extracted.
[0028] Then, the extracted parameters are compared one by one with the preset ideal feature range and the acceptable feature range. If the extracted parameter is within the ideal feature range, the score is 1; if the extracted parameter is outside the ideal range but within the acceptable feature range, the score is 0.5; otherwise, the score is 0. The score for each parameter is obtained.
[0029] Finally, weights are assigned to each parameter, and a correction factor is obtained through weighted fusion.
[0030] Preferably, the specific steps for revising the initial labels using a correction factor to generate the final labels are as follows:
[0031] Preset high threshold Y1 and low threshold Y2, and And denote the correction factor as Y;
[0032] when At this point, the final label is confirmed as the preliminary label; when At this point, the final label is determined to be a process abnormality; when At this point, the final label retains the initial label, but with an additional low-confidence flag.
[0033] Preferably, the specific steps for obtaining the average temperature deviation value are as follows:
[0034] Based on the current process status, retrieve the corresponding preset process temperature reference field, and align the coordinates of each measurement point in the target surface temperature distribution field with the coordinates in the process temperature reference field;
[0035] The absolute difference between the measured temperature and the reference temperature at all measurement points on the target surface is calculated sequentially. Finally, the absolute differences of all measurement points are summed and divided by the total number of measurement points to obtain the average temperature deviation value.
[0036] The beneficial effects of this invention are:
[0037] 1. This invention obtains preliminary labels by matching deformation rate features, and then generates correction factors by combining temperature distribution field and apparent morphology to complete secondary calibration. This achieves cross-verification of multi-source information, greatly reducing false alarm rate and false alarm rate, providing clear and reliable stage basis for subsequent temperature control, reducing internal structural defects in workpieces caused by stage judgment errors, and improving the finished product qualification rate.
[0038] 2. This invention dynamically adjusts the acquisition frequency of multispectral images based on the final determined process state. During periods of rapid change and high risk, the sampling rate is automatically increased to capture critical transient details, while the frequency is appropriately reduced during stable phases. This effectively reduces real-time data processing while ensuring no loss of critical process information, thus improving the overall system's operational efficiency and economy.
[0039] 3. This invention compares the inverted high-precision temperature field point-by-point with the ideal process curve, quantifies the average temperature deviation, and classifies it into distinct anomaly levels. For different levels, the system automatically generates and executes differentiated control commands. This enables proactive adjustment of process parameters such as heating power to maintain process quality, while effectively preventing equipment damage and major production accidents. Attached Figure Description
[0040] The present invention will now be further described with reference to the accompanying drawings.
[0041] Figure 1 This is a flowchart of the system provided in an embodiment of the present invention. Detailed Implementation
[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0043] Please see Figure 1 As shown, the present invention is an automatic target tracking and monitoring system for high-temperature furnaces based on multispectral infrared, comprising:
[0044] Tracking and acquisition module S100: Based on the YOLO target detection algorithm, it identifies and tracks targets inside the furnace, presets the initial acquisition frequency, and synchronously acquires radiation image data of the target in different infrared bands as well as acquires environmental parameters inside the furnace.
[0045] In this embodiment of the invention, the specific steps for identifying and tracking targets inside the furnace based on the YOLO target detection algorithm are as follows: First, multispectral infrared images containing targets are acquired under different forging stages and different levels of interference. Boundary boxes are labeled on the targets in the images to form a dedicated sample library of targets inside the furnace. The labeling information includes the target category, bounding box coordinates, and the forging stage to which it belongs. Based on the sample library, the YOLOv8 model is specifically trained to optimize its internal parameters, enabling it to accurately identify target objects from background interference such as flames and smoke, and ensuring that the confidence level of the judgment is greater than a preset feasibility threshold. Second, current images of targets inside the furnace are acquired from the multispectral infrared camera at a preset or dynamically adjusted frequency. These images are preprocessed, and the preprocessed current images are input into the trained YOLOv8 model. The model outputs the bounding box coordinates, confidence level, and target category of the detected targets in the image. Then, the target information in the current image output by the YOLOv8 model is input into the BoT-SORT tracker. For each existing target trajectory, a Kalman filter is used to predict its bounding box position in the current image based on its position in the previous image. BoT-SORT performs optimal matching between the detection boxes in the current image and the existing predicted trajectories by calculating motion similarity and appearance similarity. Finally, the information of the matched detection boxes (position, appearance features) is updated to the corresponding trajectory, and the unique ID of the trajectory remains unchanged, completing trajectory continuation. Unmatched detection boxes are initialized as new trajectories, assigned a new unique ID by the tracker, and re-matched in subsequent consecutive images. If a match is successful, trajectory continuation is confirmed; otherwise, the trajectory is terminated. For each image, the system outputs normalized information for all active trajectories, including: target ID, precise bounding box in the current image, and confidence level.
[0046] The specific steps for setting an initial acquisition frequency and simultaneously acquiring radiation image data of the target in different infrared bands include: first, setting an initial acquisition frequency based on the basic working conditions of the furnace forging process and the imaging performance of the multispectral infrared camera; selecting an infrared band suitable for the furnace scene; and setting camera parameters to ensure that the acquisition field of view can accurately cover the initial area of the target in the furnace. Specific camera parameters include exposure time and gain coefficient. Finally, after the YOLOv8 model initializes the target, the multispectral infrared camera is automatically triggered to acquire data according to the preset initial acquisition frequency, simultaneously acquiring infrared radiation images of the target in different bands.
[0047] The furnace internal environment parameters are collected, specifically including gas concentration, flue gas temperature, radiation path length, dust concentration, typical dust particle size, and furnace wall temperature. Specifically, gas concentration and flue gas temperature can be directly measured using a flue gas analyzer, dust concentration can be directly measured using a dust concentration meter, furnace wall temperature can be directly obtained using an infrared thermometer, the typical dust particle size can be set as a statistical value based on fuel type and combustion experience, and the radiation path length can be calculated from the equipment drawings.
[0048] Furthermore, the high-temperature furnace target automatic tracking and monitoring system based on multispectral infrared also includes:
[0049] The identification and judgment module S200: Based on the acquired multi-band target radiation image data, it inverts the target surface temperature distribution field through a multispectral radiative transfer model, simultaneously extracts the target appearance morphology and deformation rate features, matches the deformation rate features with a preset deformation rate feature vector library to obtain preliminary labels, then fuses the surface temperature distribution field and appearance morphology to generate correction factors, and uses the correction factors to correct the preliminary labels to generate final labels.
[0050] In this embodiment of the invention, the specific steps for inverting the target surface temperature distribution field using a multispectral radiative transfer model are as follows: First, the acquired multi-band target radiation images are pixel-aligned using a sub-pixel-level phase correlation method. Specifically, the two-dimensional Fourier transform phase difference of the multi-band target radiation images is calculated, and the translation, rotation, and scaling parameters between the bands are solved to control the registration error within 0.1 pixels, ensuring that the spatial positions of the same target pixel are completely consistent in the multi-band images. Then, the image pixel grayscale values are adjusted using the factory calibration coefficients and on-site calibration parameters of the multispectral infrared camera. Converted to actual radiance value of the corresponding band The specific conversion formula is as follows: ,in For different infrared bands, Image pixel coordinates, For band The radiation gain coefficient, For band The radiation shift coefficient, and All values were obtained through laboratory blackbody calibration. Secondly, actual radiance values were established. relative to the target true temperature Relationship: ,in, For flue gas to band The transmittance is calculated using the following formula: ,in, For other absorption coefficients, C is the gas volume concentration, and L is the radiation path length, determined based on the geometric model of the furnace and the multispectral infrared camera installation. For smoke and dust on the band The transmittance is calculated using the following formula: ,in, For the smoke and dust scattering cross section, ,in, This is a coefficient representing the strength of an individual dust particle's ability to block infrared light, obtained through Mie scattering theory. This represents the typical size of smoke and dust particles. This refers to the concentration of smoke and dust. To obtain the target in the band through laboratory calibration Emission rate. For band Lower temperature The corresponding blackbody radiance is calculated using the following formula: , where h is Planck's constant, c is the speed of light, k is Boltzmann's constant, and T is the current temperature of the target surface. Environmental incident radiance refers to the radiation emitted by environmental sources such as furnace walls and flames onto a target. ,in, This refers to the effective temperature of the furnace wall. The radiation emitted by the flue gas itself. Where T2 is the flue gas temperature. Finally, for each pixel... In other words, the relationships between all bands are solved simultaneously, and then the least squares method is used iteratively. Specifically, the temperature values are initialized. Substituting the values into the simultaneous equations, the difference E between the theoretical and actual radiance is calculated using the following formula: ,in, The temperature value is calculated theoretically and continuously corrected using Newton's iteration method until the difference E is less than a preset difference threshold, and the temperature deviation between two adjacent iterations is less than 0.5℃. This temperature is then considered the pixel value. The final temperature at the target surface is calculated by performing the above temperature calculation process on all target pixels in the image, mapping the temperature values to a two-dimensional matrix according to the pixel coordinates, and generating a global temperature distribution field on the target surface.
[0051] The specific steps for extracting the target's apparent morphology and deformation rate features are as follows: Target Apparent Morphology: First, select the band with the highest contrast between the target and the furnace background from the multi-band image as the feature image I(x,y,t). Filter the feature image to suppress noise and perform contrast stretching to highlight the target edges. Using the region in the target surface temperature distribution field that is higher than the preset temperature threshold of the background as the initial position, use an edge detection algorithm on the feature image to determine the target's contour, obtaining a binarized image B(x,y,t), where the target pixel is 1 and the furnace background pixel is 0. Finally, calculate the feature parameters describing its geometric attributes based on the binarized image B(x,y,t), including area, perimeter, and density. Specifically, the area S is obtained by counting all pixels belonging to the target. Where W and H are the width and height of the image. A boundary tracing algorithm is used to find all the contour pixels of the target, resulting in an ordered sequence of contour points. Where n is the number of contour pixels, and then the Euclidean geometric distance between contour pixels is calculated. The sum of these gives the perimeter L. The density D is obtained by calculating how close the target shape is to a circle, and its calculation formula is as follows: ,in, Pi (π) represents the mathematical constant pi. The closer the value is to 1, the closer the shape is to a circle, indicating a tighter shape; the smaller the value, the more irregular the shape. Deformation rate feature: Contour features are retrieved from two adjacent images of the same target ID. An iterative nearest-neighbor algorithm is used to perform contour matching. Based on the matching results, the rate of change of the width of the minimum bounding rectangle of the two images is calculated. Altitude change and area change and in combination with time intervals The deformation rate characteristics are obtained, namely the deformation rate characteristics of width, height, and area. Their calculation methods are all the same, with the width deformation rate as the starting point. For example, the calculation formula is as follows: ,in, This represents the target width at time t.
[0052] Preliminary labeling refers to a preliminary judgment about the current process stage based solely on the target deformation motion characteristics and comparison with historical data. Specifically, firstly, a vector library is constructed by summarizing standard deformation rate characteristics of m typical process stages from historical data. Each deformation rate characteristic corresponds to a different deformation rate in width, height, and area, and is labeled with the name of the corresponding process stage, such as initial heating, mid-forging, late-forging, and process anomaly. Then, the similarity between the current deformation rate characteristic and the standard deformation rate characteristics of each process stage in the vector library is calculated using Euclidean geometric distance; a smaller Euclidean geometric distance indicates higher similarity. Finally, the standard deformation rate characteristic with the smallest Euclidean geometric distance to the current deformation rate characteristic is selected, and the name of the corresponding process stage is output as the preliminary label.
[0053] The correction factor is an index derived through rule fusion based on temperature distribution field and apparent morphology features, used for verifying and calibrating initial labels. Specifically, the steps for obtaining the correction factor are as follows: First, define an ideal feature range and an acceptable feature range for each process state (initial heating, mid-forging, late-forging, process anomaly). The acceptable feature range includes the ideal feature range but is larger. Second, extract the average temperature and temperature uniformity parameters from the temperature distribution field, and the area, perimeter, and density parameters from the apparent morphology features. The average temperature is obtained by summing the temperatures of all effective pixels in the temperature distribution field and then dividing by the total number of pixels. Temperature uniformity is obtained by calculating the standard deviation of the temperature. Then, compare each extracted parameter with the preset ideal and acceptable feature ranges. If the extracted parameter is within the ideal feature range, the score is 1; if the extracted parameter is outside the ideal range but within the acceptable feature range, the score is 0.5; otherwise, the score is 0. Finally, assign weights to each parameter, with the sum of the weights being 1. Obtain the correction factor through weighted fusion. The value range of the correction factor is... The closer the value is to 1, the more supportive the temperature distribution field and apparent morphology features are for the initial label. The above feature ranges, weights, and scoring rules can all be dynamically adjusted using process samples to adapt to different materials such as carbon steel and aluminum alloys.
[0054] The specific steps for refining the initial labels using correction factors to generate the final labels are as follows: Preset a high threshold Y1 and a low threshold Y2, and... Its value can be dynamically adjusted according to the actual situation, and the correction factor is denoted as Y. When At this point, the final label is confirmed as the preliminary label; when At this point, the final label is determined to be a process abnormality; when At this point, the final label retains the initial label but is supplemented with a low-confidence flag. Cross-validation of multi-source information is achieved through a correction factor, significantly reducing both false positive and false negative rates.
[0055] Furthermore, the high-temperature furnace target automatic tracking and monitoring system based on multispectral infrared also includes:
[0056] Adjustment and comparison module S300: Dynamically adjusts the initial acquisition frequency based on the importance of the current process status, and compares the target surface temperature distribution field with the preset process temperature reference field point by point to calculate the average temperature deviation value.
[0057] In this embodiment of the invention, the specific steps for dynamically adjusting the initial acquisition frequency based on the importance determined by the current process state are as follows: First, a basic acquisition frequency F is preset. Then, based on the physical characteristics and monitoring requirements of each process stage, a frequency adjustment coefficient is set for each process stage. For example, the frequency adjustment coefficient for initial heating is K1, the frequency adjustment coefficient for the middle stage of forging is K2, the frequency adjustment coefficient for the later stage of forging is K3, and the frequency adjustment coefficient for process abnormalities is K4. Finally, when forging begins, the acquisition frequency for the initial stage is... When the final label output by the identification and judgment module changes, the system switches to the corresponding acquisition frequency. This frequency is the base acquisition frequency multiplied by a frequency adjustment coefficient corresponding to the process state. Adjusting the acquisition frequency according to different process states helps adapt to the high-density data requirements of critical stages without causing data redundancy and reducing data processing pressure.
[0058] The average temperature deviation value refers to the difference between the actual temperature of each measurement point on the target surface and the ideal temperature at the corresponding position and time in the preset process temperature curve. Specifically, based on the current process state, the corresponding preset process temperature reference field is retrieved, and the coordinates of each measurement point in the target surface temperature distribution field are aligned with the coordinates in the process temperature reference field to ensure that the same coordinate represents the same physical position on the workpiece surface. The absolute difference between the measured temperature and the reference temperature at all measurement points on the target surface is calculated sequentially. Finally, the absolute differences of all measurement points are summed and divided by the total number of measurement points to obtain the average temperature deviation value.
[0059] Furthermore, the high-temperature furnace target automatic tracking and monitoring system based on multispectral infrared also includes:
[0060] Execution module S400: Classifies abnormal levels based on average temperature deviation values, generates differentiated control commands, and executes them.
[0061] In this embodiment of the invention, the specific steps for classifying the abnormality level are as follows: First, based on process knowledge and historical data, a warning threshold A1 and a danger threshold A2 are preset, and the average temperature deviation value is denoted as A. When When the temperature meets expectations, it is set to level 1; when At this point, it indicates that the temperature has slightly deviated from the expected value and the process parameters need to be adjusted for correction; this is set to level 2. At this point, it indicates a severe temperature anomaly, posing a risk of equipment damage, and is set to Level 3.
[0062] The differentiated control commands specifically include: At Level 1, normal operation is maintained at the preset dynamic acquisition frequency, and temperature deviation values and process stage data are recorded. At Level 2, the measured temperature at each measurement point is compared with the preset temperature to determine whether the overall deviation is too high or too low, and the difference is calculated to obtain the corresponding adjustment amount, which is then sent to the heating control room for adjustment. At Level 3, an emergency shutdown command is triggered, data is saved, and relevant personnel are notified for maintenance.
[0063] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.
Claims
1. An automatic target tracking and monitoring system for high-temperature furnaces based on multispectral infrared, comprising: Tracking and acquisition module: Based on the YOLO target detection algorithm, it identifies and tracks targets inside the furnace, presets an initial acquisition frequency, and synchronously acquires radiation image data of the target in different infrared bands as well as acquires environmental parameters inside the furnace. Its characteristic is that it also includes: Identification and Judgment Module: Based on the acquisition of multi-band target radiation image data, the surface temperature distribution field of the target is inverted through the multispectral radiative transfer model. The target appearance morphology and deformation rate features are extracted simultaneously. The deformation rate features are matched with the preset deformation rate feature vector library to obtain preliminary labels. The surface temperature distribution field and appearance morphology are then fused to generate correction factors. The preliminary labels are corrected using the correction factors to generate the final labels. Adjustment and comparison module: dynamically adjusts the initial acquisition frequency based on the importance of the current process status, and compares the target surface temperature distribution field with the preset process temperature reference field point by point to calculate the average temperature deviation value; Execution module: Classifies abnormality levels based on average temperature deviation values, generates differentiated control commands, and executes them.
2. The automatic target tracking and monitoring system for high-temperature furnaces based on multispectral infrared as described in claim 1, characterized in that, The specific steps for identifying and tracking targets inside the furnace based on the YOLO target detection algorithm are as follows: Multispectral infrared images containing targets were collected under different forging stages and different levels of interference. Boundary boxes were marked on the targets in the images to form a dedicated sample library of targets inside the furnace. The YOLOv8 model was trained based on the sample library. The current image of the target inside the furnace is acquired from the multispectral infrared camera at a preset or dynamically adjusted frequency. The current image is preprocessed and then input into the trained YOLOv8 model. The YOLOv8 model outputs the target information detected in the current image, where the template information includes bounding box coordinates, confidence level and target category. The target information of the current image is input into the BoT-SORT tracker. For each existing target trajectory, the Kalman filter is used to predict its bounding box position in the current image based on its position in the previous image. The BoT-SORT tracker calculates motion similarity and appearance similarity to perform optimal matching between the detection box in the current image and the existing predicted trajectory. The information of the matched detection boxes is updated to the corresponding trajectory, and the unique ID of the trajectory remains unchanged, thus completing the trajectory continuation. Unmatched detection boxes are initialized as new trajectories, and a new unique ID is assigned by the BoT-SORT tracker. The tracks are then re-matched in subsequent consecutive images. If a match is successful, the trajectory continuation is confirmed; otherwise, the trajectory is terminated.
3. The automatic target tracking and monitoring system for high-temperature furnaces based on multispectral infrared as described in claim 2, characterized in that, The specific steps for inverting the target surface temperature distribution field using a multispectral radiative transfer model are as follows: A subpixel-level phase correlation method was used to align the pixels of the acquired multi-band target radiation images. Then, the image pixel grayscale values were adjusted using the factory calibration coefficients of the multispectral infrared camera and the field calibration parameters. Converted to actual radiance value of the corresponding band The specific conversion formula is as follows: ,in For different infrared bands, Image pixel coordinates, For band The radiation gain coefficient, For band radiative offset coefficient; Establish actual radiance values relative to the target true temperature Relationship: ,in, For flue gas band transmittance, For smoke and dust on the band transmittance, To obtain the target in the band through laboratory calibration emission rate, For band Lower temperature The corresponding blackbody radiation brightness, The ambient incident radiation brightness, The radiation emitted by the flue gas itself; For each pixel The relationships between all bands are combined and then solved iteratively using the least squares method.
4. The automatic target tracking and monitoring system for high-temperature furnaces based on multispectral infrared as described in claim 3, characterized in that, The specific steps for extracting the target's apparent morphology and deformation rate features are as follows: First, the band with the highest contrast between the target and the background inside the furnace is selected from the multi-band image as the feature image I(x,y,t). The feature image is filtered to suppress noise and contrast is stretched to highlight the target edge. Based on the area in the target surface temperature distribution field that is higher than the preset temperature threshold of the background as the initial position, the edge detection algorithm is used on the feature image to determine the outline of the target and obtain the binarized image B(x,y,t). Finally, the area, perimeter and density feature parameters are calculated based on the binarized image B(x,y,t). Contour features of two adjacent images with the same target ID are retrieved, and contour matching is performed using an iterative nearest-neighbor algorithm. Based on the matching results, the rate of change of the width of the minimum bounding rectangle of the two images is calculated. Altitude change and area change and in combination with time intervals We obtain the deformation rates of width, height, and area, which are collectively referred to as deformation rate characteristics.
5. The automatic target tracking and monitoring system for high-temperature furnaces based on multispectral infrared as described in claim 4, characterized in that, The specific steps for obtaining the initial tag are as follows: We summarize the standard deformation rate characteristics of m typical process stages from historical data to construct a vector library, and identify the process stage name corresponding to each deformation rate characteristic; The Euclidean geometric distance is calculated using the Euclidean geometric distance formula between the current deformation rate feature and the standard deformation rate feature of each process stage in the vector library. This distance is used to measure the similarity between the current deformation rate feature and the standard deformation rate feature of each process stage in the vector library. Select the standard deformation rate feature with the smallest Euclidean geometric distance from the current deformation rate feature, and output the process stage name corresponding to the standard deformation rate feature as the initial label.
6. The automatic target tracking and monitoring system for high-temperature furnaces based on multispectral infrared as described in claim 5, characterized in that, The specific steps for obtaining the correction factor are as follows: Define an ideal characteristic range and an acceptable characteristic range for each process state; Extract the average temperature and temperature uniformity parameters from the temperature distribution field, as well as the area, perimeter, and density parameters from the apparent morphology features; The extracted parameters are compared one by one with the preset ideal feature range and the acceptable feature range. If the extracted parameter is within the ideal feature range, the score is 1. If the extracted parameter is outside the ideal range but within the acceptable feature range, the score is 0.
5. Otherwise, the score is 0. The score for each parameter is obtained. Each parameter is assigned a weight, and then a correction factor is obtained through weighted fusion.
7. The automatic target tracking and monitoring system for high-temperature furnaces based on multispectral infrared as described in claim 6, characterized in that, The specific steps for revising the initial labels using correction factors to generate the final labels are as follows: Preset high threshold Y1 and low threshold Y2, and And denote the correction factor as Y; when At this point, the final label is confirmed as the preliminary label; when At this point, the final label is determined to be a process abnormality; when At this point, the final label retains the initial label, but with an additional low-confidence flag.
8. The automatic target tracking and monitoring system for high-temperature furnaces based on multispectral infrared as described in claim 7, characterized in that, The specific steps for obtaining the average temperature deviation value are as follows: Based on the current process status, retrieve the corresponding preset process temperature reference field, and align the coordinates of each measurement point in the target surface temperature distribution field with the coordinates in the process temperature reference field; The absolute difference between the measured temperature and the reference temperature at all measurement points on the target surface is calculated sequentially. Finally, the absolute differences of all measurement points are summed and divided by the total number of measurement points to obtain the average temperature deviation value.