Window frost inhibition and micro-heating synergic control method based on image segmentation

Abstract

The present application relates to a window frost inhibition and micro-heating collaborative control method based on image segmentation, belonging to the technical field of image segmentation, comprising the following steps: step 1, synchronously collecting visible light images and thermal images by a visible light camera and a long-wave infrared thermal imager and registering to a window physical coordinate system, generating a direction prior channel based on a Gabor filter bank; step 2, inputting a frost dendrite segmentation input tensor into a frost dendrite UNet, outputting a current control period condensation / frost segmentation mask from a semantic segmentation branch, outputting a next control period predicted barcode length map under a reference control condition from a topological feedforward branch, and inverting a dominant cold bridge position; step 3, determining three types of trigger partitions according to the dominant cold bridge position, the condensation / frost segmentation mask and the next control period predicted barcode length map, solving a transparent heating partition heating power sequence and a frame vent opening degree sequence by quadratic programming, and taking the first control cycle solution to execute. The present application realizes active positioning and early feedforward intervention of the window cold bridge position, makes the heating energy supply have spatial specificity, and realizes multi-objective collaborative optimization among transparency loss, energy consumption and thermal comfort.

Description

Technical Field

[0001] This invention belongs to the field of image segmentation technology, specifically relating to a method for coordinated control of window frosting suppression and micro-heating based on image segmentation. Background Technology

[0002] Condensation and frost formation on windows are common physical phenomena in building envelopes and various transparent observation windows under low temperature and high humidity conditions. In scenarios such as exterior windows of civil buildings, windshields of vehicles, observation windows of cold storage, light-transmitting covers of greenhouses, and observation windows of medical equipment, condensation and frost formation can lead to a sharp decrease in transparency, obstruction of vision, and secondary problems such as accelerated aging of seals, increased energy consumption, and mold growth.

[0003] To address this issue, existing technologies can be broadly categorized into three types. The first type is an electric heating defrosting and defogging solution based on temperature and humidity thresholds. This method calculates the dew point temperature using indoor dry-bulb temperature and relative humidity, and initiates uniform heating across the entire window when the window surface temperature falls below the dew point. While structurally simple, this solution suffers from a significant drawback: the heating energy is not directed to truly high-risk areas, resulting in wasted energy. Furthermore, it exhibits a delayed response, often intervening only after condensation is visually apparent. The second type is a fog area recognition solution based on visible light images. This method statistically analyzes image clarity, contrast, or texture entropy to determine the degree of fogging on the window surface and triggers heating. However, the sensitivity of visible light features alone to early micro-liquid films and sparse frost dendrites is insufficient, and it lacks the physical identification capability to determine the cause of condensation, failing to distinguish between conditions driven by external humidity and those driven by indoor moisture sources. The third category is the partitioned electrochromic or partitioned transparent heating film solution, which divides the window surface into multiple independent addressable areas. However, its partitioning drive mostly adopts on / off control based on fixed timing or simple thresholds. It neither forms a closed loop with image perception nor establishes a predictive model between partitioned surface temperature and control variables, making it difficult to make a constrained multi-objective trade-off between transparency loss, energy consumption, and thermal comfort.

[0004] In summary, existing technologies generally suffer from the following technical problems: First, they lack the ability to actively identify the location of cold bridges on the window surface, and the heating energy delivery is not spatially targeted; second, they lack feedforward prediction of the early evolution trend of condensation and frost, resulting in a delayed intervention time; third, the two types of actuators, heating and ventilation, lack a unified rolling time-domain collaborative control framework, making it difficult to balance suppression effect, energy consumption, and thermal comfort. Summary of the Invention

[0005] The main objective of this invention is to provide a method for coordinated control of window frosting suppression and micro-heating based on image segmentation. This invention achieves active localization and early feedforward intervention of the cold bridge position on the window, enabling spatial targeting of heating energy delivery and achieving multi-objective coordinated optimization between transparency loss, energy consumption and thermal comfort.

[0006] To solve the above problems, the technical solution of the present invention is implemented as follows:

[0007] A method for coordinated control of window frosting suppression and micro-heating based on image segmentation includes the following steps:

[0008] Step 1: Simultaneously acquire visible light images and thermal images using a visible light camera and a long-wave infrared thermal imager, and register them to the window physical coordinate system; generate a directional prior channel for the visible light image based on the Gabor filter bank, including a directional energy map, a principal axis direction cosine map, and a principal axis direction sine map; stitch the visible light image, thermal image, and directional prior channel along the channel dimension to obtain the frost dendrite segmentation input tensor;

[0009] Step 2: The frost dendrite segmentation input tensor is fed into the pre-trained frost dendrite UNet. The frost dendrite UNet includes a semantic segmentation branch and a topology feedforward branch. The semantic segmentation branch outputs the condensation / frost segmentation mask for the current control cycle, and the topology feedforward branch outputs the predicted barcode length map for the next control cycle under the baseline control conditions. The topology feedforward branch is pre-trained using the topology supervision channel obtained from the thermal image through persistent cohomology calculation as the supervision ground value. The dominant cold bridge position is inverted based on the condensation / frost segmentation mask, the orientation prior channel, and the temperature gradient of the thermal image.

[0010] Step 3: Divide the window surface into multiple independently addressable transparent heating zones, and set multiple adjustable ventilation openings on the window frame. The collaborative controller determines the cold bridge trigger zone, follow-up trigger zone, and feedforward trigger zone based on the dominant cold bridge location, condensation / frost separation mask, and the predicted barcode length map for the next control cycle. Based on the zone surface temperature state prediction model and dew point margin, the heating power sequence of all transparent heating zones and the ventilation opening sequence of all frame ventilation openings are solved in the prediction time domain through quadratic programming. The solution of the first control cycle is then sent out for execution.

[0011] Furthermore, in step 1, the visible light camera and the long-wave infrared thermal imager synchronously acquire visible light images and original thermal images using the same hardware trigger signal. The original thermal image is then subjected to response uniformity correction and emissivity correction to obtain the thermal image. Through checkerboard calibration and homography matrix transformation, the visible light image and the thermal image are registered to a window physical coordinate system established with the lower left corner of the window as the origin, the width direction of the window as the X-axis, and the height direction of the window as the Y-axis. This ensures that pixels at the same physical location in the visible light image and the thermal image correspond one-to-one, resulting in a multimode image pair.

[0012] Furthermore, the Gabor filter bank comprises eight directions: 0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135°, and 157.5°. Each filter has the same center wavelength and the same Gaussian envelope width. The Gabor filters in each of the eight directions are convolved with the visible light image in two dimensions. The square of the complex modulus value of the response at the same pixel position in the resulting eight directional response maps is taken to obtain the energy values ​​of the eight directions at the corresponding pixel positions. The direction energy value with the largest value at each pixel position is selected as the principal axis direction energy value of the corresponding pixel, and the direction corresponding to the largest direction energy value is taken as the principal axis direction of the corresponding pixel. The directional energy map is composed of the principal axis direction energy values ​​of all pixels. The cosine and sine values ​​of the principal axis direction of each pixel after taking a double angle are used to form the principal axis direction cosine map and principal axis direction sine map, respectively.

[0013] Furthermore, the FrostDendric UNet includes an encoder, a decoder, and two parallel output branches, with skip connections between the encoder and decoder. The semantic segmentation branch consists of two convolutional layers and one sigmoid activation function connected in series, outputting a condensation / frost confidence map for the current control cycle. The condensation / frost confidence map is binarized according to a preset binarization threshold to obtain the condensation / frost segmentation mask for the current control cycle. The topology feedforward branch consists of two convolutional layers and one ReLU activation function connected in series, outputting a predicted barcode length map for the next control cycle under the baseline control conditions. The baseline control conditions refer to the evolutionary conditions under which the heating power of each transparent heating zone and the ventilation opening of each frame vent maintain the values ​​of the current cycle.

[0014] Furthermore, the topological supervision channel is obtained as follows: For the thermal images acquired in the training samples during the next control cycle under baseline control conditions, multiple level set thresholds are set from low to high within their temperature range using a preset fixed temperature step size; for each level set threshold, pixels with temperatures less than or equal to the level set threshold are set as foreground, and pixels with temperatures greater than the level set threshold are set as background, resulting in corresponding level set binary slices; the level set binary slices are traversed one by one in order of increasing level set thresholds, and foreground connected components are extracted according to the eight-connectivity rule; when a foreground connected component first appears, the current level set threshold is recorded as the birth threshold of the corresponding connected component; when two or more foreground connected components merge into a new foreground connected component under the same level set threshold, the component with the lower birth threshold is retained as the surviving component, and the extinction threshold of the remaining foreground connected components is recorded as the current level set threshold; the thermal images are pre-defined... Figure 4An image boundary protection band is set. Foreground connected components generated within the image boundary protection band are still tracked as candidate connected components for cold bridges when they meet the image boundary, until they are merged with other foreground connected components or traversed to the highest level set threshold. The fusion threshold or the highest level set threshold is used as the extinction threshold of the corresponding connected component. The extinction threshold of each foreground connected component is subtracted from the generation threshold to obtain the barcode length of the corresponding connected component. The barcode length is backfilled to all pixel positions occupied by the corresponding connected component in the binary slice of the level set before the extinction threshold. When the same pixel is backfilled by multiple connected components, the maximum value of the barcode length is taken. During the training of Frost Dendrite UNet, the binary mask of the condensation / frost region of the current control cycle is used as the supervision ground value of the semantic segmentation branch, and the binary cross-entropy loss is used. The topological supervision channel is used as the supervision ground value of the topological feedforward branch, and the smooth L1 loss is used. The sum of the binary cross-entropy loss and the smooth L1 loss is used as the total loss. The Adam optimizer is used to iteratively update the parameters of Frost Dendrite UNet until the total loss converges.

[0015] Furthermore, the method for inverting the dominant cold bridge location is as follows: The temperature gradient magnitude and direction of each pixel are calculated using the Sobel operator on the thermal image. The low-temperature pointing direction of the corresponding pixel is obtained along the temperature decrease direction. Each foreground pixel is traversed in the condensation / frost segmentation mask. The principal axis direction is obtained by inversely solving the cosine and sine values ​​of the principal axis direction of the corresponding foreground pixel. The principal axis direction extends along the original direction as a forward ray and extends in the reverse direction as a reverse ray. The side with the smaller angle to the low-temperature pointing direction among the two rays is selected as the directed backtracking direction of the corresponding foreground pixel. Foreground pixels that simultaneously meet the following three screening conditions are included in the set of valid voting pixels: the angle between the principal axis direction of the foreground pixel and the low-temperature pointing direction is less than or equal to... The following criteria are met: First, the orientation consistency threshold, the principal axis energy value is greater than or equal to a preset energy threshold, and the temperature gradient amplitude is greater than or equal to a preset gradient threshold. For each foreground pixel in the set of valid voting pixels, a ray is drawn from the corresponding position along the corresponding directional backtracking direction. The first intersection point of the ray within the window's physical coordinate system border area is taken as the border hit point of the corresponding foreground pixel. The window border is divided into multiple border segments with a preset fixed arc length step. All border hit points are counted according to their respective border segments. The border segment with the highest hit point count is identified as the dominant cold bridge position. The window condensation state description result is composed of the condensation / frost segmentation mask, the barcode length map predicted for the next control cycle, and the dominant cold bridge position.

[0016] Furthermore, the collaborative controller determines the cold bridge triggering partition, follow-up triggering partition, and feedforward triggering partition in the following ways: Cold bridge directional triggering: the transparent heating partition adjacent to the dominant cold bridge position in the window physical coordinate system is set as the cold bridge triggering partition, and the frame vent closest to the dominant cold bridge position along the arc length of the window frame is set as the cold bridge triggering vent; Condensation area follow-up triggering: the window block covered by each transparent heating partition is traversed in the condensation / frost segmentation mask, the number of foreground pixels in each window block is counted, and the transparent heating partition with the number of foreground pixels exceeding the preset follow-up threshold is set as the follow-up triggering partition; Topology feedforward triggering: the window block covered by each transparent heating partition is traversed in the predicted barcode length map in the next control cycle, the average predicted barcode length of all pixels in each window block is calculated, and the transparent heating partition with the average predicted barcode length exceeding the preset feedforward threshold is set as the feedforward triggering partition.

[0017] Furthermore, in step 3, the window surface is pre-divided into M rows and N columns of independently addressable transparent heating zones, where M and N are both positive integers; P frame vents are uniformly arranged along the circumference of the window frame area, where P is a positive integer, and each frame vent is connected to the indoor dehumidification duct to deliver dehumidified airflow to the window edge area to reduce the partial pressure of water vapor near the wall layer of the window surface; one surface temperature sensor is arranged on the window surface of each transparent heating zone, one dry bulb temperature sensor and one relative humidity sensor are arranged indoors, and one outdoor temperature sensor is arranged outdoors; the zone surface temperature state prediction model and dew point margin are constructed as follows: read the current measurement values ​​of each surface temperature sensor, dry bulb temperature sensor, relative humidity sensor and outdoor temperature sensor; predict the surface temperature of each transparent heating zone in the next cycle, which is composed of the current cycle surface temperature of the corresponding zone, the current cycle heating power of the corresponding zone, the current cycle surface temperature of the adjacent zone of the corresponding zone and the corresponding... The summation of the differences in surface temperatures of the current period for each zone, the current period ventilation opening of the nearest frame vent to the corresponding zone, the current indoor dry-bulb temperature, and the current outdoor temperature are all calculated by multiplying each of these six factors by a zone superposition coefficient. The predicted indoor dry-bulb temperature for the next period is calculated by multiplying each of the current indoor dry-bulb temperature, the current outdoor temperature, and the current period ventilation opening of all frame vents by an indoor superposition coefficient. Both the zone superposition coefficient and the indoor superposition coefficient were obtained from the step response identification experiment before the collaborative controller was put into operation and are stored in the collaborative controller's memory. The current dew point temperature is calculated using the Magnus method based on the current indoor dry-bulb temperature and relative humidity. The difference between the predicted surface temperature of the next period for each zone and the current dew point temperature is used as the dew point margin for the corresponding zone in the next period. If the predicted surface temperature of the next period for each zone is below 0 degrees Celsius, the frost point temperature, corrected from the current dew point temperature, is used instead of the current dew point temperature in the calculation of the dew point margin for the corresponding zone.

[0018] Furthermore, the objective function used in the quadratic programming solution is the sum of the following five terms: The first term is the sum of squares of the negative dew point margins of all transparent heating zones within the prediction time domain. The negative dew point margins are represented by auxiliary variables, which are greater than or equal to 0 and greater than or equal to the difference between the current dew point temperature and the predicted surface temperature for the next cycle; The second term is the sum of squares of the instantaneous heating power of all transparent heating zones within the prediction time domain; The third term is the sum of squares of the instantaneous ventilation openings of all frame vents within the prediction time domain; The fourth term is the sum of squares of the difference in heating power between adjacent cycles of the same transparent heating zone and the difference in ventilation opening between adjacent cycles of the same frame vent within the prediction time domain; The fifth term is the sum of squares of the difference between the predicted indoor dry-bulb temperature and the preset comfort temperature for the next cycle within the prediction time domain.

[0019] Furthermore, the constraints used in the quadratic programming solution include: upper and lower limits of heating power for each transparent heating zone, upper and lower limits of ventilation opening for each frame vent, upper limit of surface temperature for each transparent heating zone, upper limits of heating power change rate and ventilation opening change rate during adjacent periods, state transition equation constraints given by the zone surface temperature state prediction model, and indoor dry-bulb temperature prediction equation constraints; a mandatory constraint greater than or equal to a preset safety margin lower limit is applied to the dew point margin at the end of the prediction time domain for the cold bridge trigger zone, follow-up trigger zone, and feedforward trigger zone, with the safety margin lower limit being a constant greater than 0; a mandatory constraint greater than or equal to 0 is applied to the dew point margin at the end of the prediction time domain for the remaining transparent heating zones; and the cold bridge trigger zone... The following steps are used as the initial values ​​for warm-start optimization variables: 1) The heating power of the following trigger zones and feedforward trigger zones is set to a preset proportion of the upper limit of the heating power in the current cycle; the heating power of the remaining transparent heating zones is set to another lower preset proportion of the upper limit of the heating power; and the ventilation opening of the cold bridge trigger vent is set to a preset proportion of the upper limit of the ventilation opening in the current cycle. 2) A quadratic programming solver is used to simultaneously solve all heating power sequences and ventilation opening sequences in the prediction time domain at the beginning of each control cycle. 3) The solution of the heating power sequence and ventilation opening sequence in the prediction time domain for the first control cycle is taken as the actual command to be issued in the current control cycle and issued to the corresponding transparent heating zones and frame vents. 4) Steps 1 to 3 are repeated at the beginning of the next control cycle.

[0020] This invention offers the following advantages: By deeply integrating multimodal sensing, persistent coherence-supervised image segmentation, and rolling temporal collaborative control, it achieves several beneficial effects in suppressing condensation and frost formation on window surfaces. Firstly, through synchronous acquisition by a visible light camera and a long-wave infrared thermal imager, and registration with the physical coordinate system of the window surface, geometrically strictly corresponding multimodal image pairs are obtained. This compensates for the insufficient sensitivity of single-modal sensing to early microliquid films and sparse frost dendrites, making early intervention possible. Secondly, by generating directional prior channels containing principal axis direction cosine and principal axis direction sine diagrams through a Gabor filter bank, the inherent 180-degree axial ambiguity of the Gabor response is eliminated using a double-angle encoding method, allowing the downstream convolutional neural network to learn a smooth directional manifold. Third, by performing zero-sustainability long-term homology calculation on the low-temperature level set of the thermal image and using the barcode length as the supervisory ground truth of the topological feedforward branch, the network can distinguish transient disturbances from stable cold sources at the topological scale. Furthermore, by using the predicted barcode length map for the next control cycle under baseline control conditions as the feedforward signal, control shifts from a delayed response to proactive intervention. Fourth, by combining condensation / frost segmentation masks, directional prior channels, and thermal image temperature gradient directions, the undirected texture axis is uniquely upgraded to a directed backtracking ray, enabling proactive localization of the dominant cold bridge location and giving the heating energy delivery a clear spatial targeting. Fifth, through the construction of a zoned surface temperature state prediction model and dew point margin, quadratic programming is used to simultaneously solve the heating power sequence of the transparent heating zone and the ventilation opening sequence of the frame vents in the prediction time domain. This achieves constrained multi-objective collaborative optimization between transparency loss, energy consumption, and thermal comfort, effectively maintaining transparency while reducing overall window energy consumption. Attached Figure Description

[0021] Figure 1 A schematic diagram illustrating the geometric arrangement principle of dual-mode image synchronous acquisition and window physical coordinate system registration provided in an embodiment of the present invention;

[0022] Figure 2 This is a schematic diagram illustrating how a pixel position is encoded to a unit circle at a double angle along the main axis, according to an embodiment of the present invention.

[0023] Figure 3 A schematic diagram of the geometric relationship between the directed backtracking ray and the window frame during the dominant cold bridge location inversion provided in an embodiment of the present invention;

[0024] Figure 4 This is a schematic diagram illustrating the time sequence of commands issued by the collaborative controller to three types of transparent heating zones and cold bridge-triggered ventilation openings within a 60-second simulation interval, as provided in an embodiment of the present invention. Detailed Implementation

[0025] A method for coordinated control of window frosting suppression and micro-heating based on image segmentation includes the following steps:

[0026] Step 1: Simultaneously acquire visible light images and thermal images using a visible light camera and a long-wave infrared thermal imager, and register them to the window physical coordinate system; generate a directional prior channel for the visible light image based on the Gabor filter bank, including a directional energy map, a principal axis direction cosine map, and a principal axis direction sine map; stitch the visible light image, thermal image, and directional prior channel along the channel dimension to obtain the frost dendrite segmentation input tensor;

[0027] Step 2: The frost dendrite segmentation input tensor is fed into the pre-trained frost dendrite UNet. The frost dendrite UNet includes a semantic segmentation branch and a topology feedforward branch. The semantic segmentation branch outputs the condensation / frost segmentation mask for the current control cycle, and the topology feedforward branch outputs the predicted barcode length map for the next control cycle under the baseline control conditions. The topology feedforward branch is pre-trained using the topology supervision channel obtained from the thermal image through persistent cohomology calculation as the supervision ground value. The dominant cold bridge position is inverted based on the condensation / frost segmentation mask, the orientation prior channel, and the temperature gradient of the thermal image.

[0028] Step 3: Divide the window surface into multiple independently addressable transparent heating zones, and set multiple adjustable ventilation openings on the window frame. The collaborative controller determines the cold bridge trigger zone, follow-up trigger zone, and feedforward trigger zone based on the dominant cold bridge location, condensation / frost separation mask, and the predicted barcode length map for the next control cycle. Based on the zone surface temperature state prediction model and dew point margin, the heating power sequence of all transparent heating zones and the ventilation opening sequence of all frame ventilation openings are solved in the prediction time domain through quadratic programming. The solution of the first control cycle is then sent out for execution.

[0029] The visible light camera is an industrial-grade complementary metal-oxide-semiconductor (CMOS) area array camera with an effective resolution of 1280×720 pixels, a lens focal length of 6 mm, and a maximum frame rate of 60 frames per second. The long-wave infrared thermal imager uses an uncooled vanadium oxide microbolometer array with an effective resolution of 640×480 pixels, a spectral response range of 8 to 14 micrometers, a noise equivalent temperature difference better than 0.05 Kelvin, and a maximum frame rate of 50 frames per second. Both are mounted parallel to each other on a rigid aluminum alloy bracket at a distance of 0.8 to 1.5 meters directly in front of the window, with the optical axis pointing towards the center of the window and the baseline distance limited to within 50 mm. This mounting size is designed to reduce parallax residuals to below one pixel, allowing subsequent dual-mode registration to only compensate for scale and rotation without introducing complex parallax correction.

[0030] To ensure strict temporal alignment between the two image streams, a microcontroller generates a square wave synchronization trigger signal with a period equal to the control period. In a typical implementation scenario, the control period is 1 second, meaning the acquisition frame rate is 1 Hz. The falling edge of the synchronization trigger pulse is simultaneously input to the external trigger pins of both cameras. Both cameras initiate exposure after responding to the same falling edge, with a measured exposure time deviation of less than 100 microseconds. This hardware-level synchronization is based on the fact that while the evolution of condensation and frost is a slow process, it is still within seconds. Indoor footsteps, opening and closing of doors and windows, and sunlight projection can significantly alter the visible light image content. If there is a time difference of tens of milliseconds or more between the acquisition of the two images, the subsequent Gabor principal axis alignment and thermal imaging low-temperature gradient alignment will produce false positives or false negatives, directly undermining the reliability of cold bridge inversion. Optionally, in networked smart building bus scenarios, precise time protocol synchronization or satellite synchronization can be used instead of hardware trigger lines, suitable for situations where the control period is extended to 5 or 10 seconds and the nanosecond-level synchronization requirements are relaxed.

[0031] The raw thermal image obtained is denoted as ,in These are pixel coordinates. Due to the limitations of microbolometer pixel manufacturing processes, there are differences in pixel bias and gain, resulting in fixed-mode noise between the direct readout value and the true radiative flux. Response uniformity correction is performed pixel-by-pixel using the following formula: ;in This is the corrected radiative flux response. The pixel offset was obtained by averaging 100 frames at a low temperature using a uniform blackbody source before shipment. The pixel gain is obtained by normalizing the difference between the time average of 100 frames at the high-temperature end and the low-temperature frame for a uniform blackbody source. In some implementations, a single-point offset update can be triggered at the moment the shutter calibration wheel built into the thermal imager closes to compensate for the gradual offset change caused by ambient temperature drift.

[0032] Emissivity correction further eliminates the contamination of temperature measurement by the non-blackbody characteristics of the glass surface and environmental reflections. Let the emissivity of the glass surface be... The empirical value ranges from 0.85 to 0.92, and in one embodiment, it is taken as 0.88; let the equivalent temperature of environmental reflection be... The temperature was measured by an infrared radiation thermometer placed indoors, and approximated by the indoor dry-bulb temperature in a simplified scenario. The corrected true temperature of the glass surface. satisfy: ;in Let be the Stefan-Boltzmann constant, with a value of . 4 Kelvin per square meter For thermal imagers The apparent temperature is obtained by looking up the built-in calibration table. Solve. After that, with Replace the original apparent temperature, backfill pixel by pixel, and obtain the corrected thermal image, denoted as . The consideration in this correction step is that glass does not radiate as a perfect blackbody in the 8-14 micrometer wavelength range. Infrared radiation from indoor lighting fixtures, people, and walls will be reflected by the glass surface and enter the thermal imager, causing the uncorrected apparent temperature to be 2 to 4 degrees Celsius higher than the actual glass temperature. If ignored, the subsequent dew point margin calculation will have a systematic positive bias, causing the co-controller to misjudge the area as a safe zone and thus miss the control.

[0033] Dual-mode registration utilizes a dual-mode calibration board with a black and white grid array and embedded thin-film resistance wire heating units within the grid. The grid has a physical side length of 30 mm and a grid configuration of 9 rows and 6 columns of internal corner points. After power-on, the temperature difference between the white and black grid areas stabilizes at approximately 8 degrees Celsius, allowing the visible light camera to see a normal light-dark checkerboard pattern, and the long-wave infrared thermal imager to see a clear hot-cold checkerboard pattern. The calibration board is placed in front of a window, its pose is varied, and 15 to 20 sets of synchronized images are acquired. For each set of images, the sub-pixel coordinates of the internal corner points are detected in both the visible light and thermal images. The homography matrix from the thermal image plane to the visible light image plane is solved using the one-to-one correspondence of the corner points. , It has 3 rows and 3 columns and 8 degrees of freedom.

[0034] refer to Figure 1 , Figure 1 The geometric arrangement for simultaneous dual-mode image acquisition and registration with the physical coordinate system of the window is shown. The visible light camera and the long-wave infrared thermal imager are mounted in front of the window with their optical axes approximately parallel. They are arranged at intervals in the vertical direction, with the baseline length limited to within 50 mm. The field of view cones of the two cameras together cover the entire window area. Figure 1 The left side is set with a synchronous triggering component, the middle is drawn with a dual-mode calibration board, and the right side is drawn with the window surface and multiple independently addressable transparent heating zones and four aluminum cross reference points. A two-dimensional physical coordinate system of the window surface is drawn with the lower left corner of the window surface as the origin, the width direction of the window surface as the X-axis, and the height direction of the window surface as the Y-axis. During operation, any pixel in the long-wave infrared image is first mapped to the visible light image pixel coordinates through the homography matrix H, and then mapped to the window surface physical coordinates through the homography matrix Hwp.

[0035] Visible light images are converted to grayscale before being fed into Gabor convolution. One implementation uses the weighted average of the International Telecommunication Union (ITU) BT.709 standard. ,in , , These represent the values ​​for the red, green, and blue channels, respectively. To output grayscale values. The consideration for this weighting method is that the scattering spectrum of frost dendrites is close to the full visible light, and the contrast in the green channel is the most sufficient. BT.709 just raises the weight of the green channel to 0.7152. In low-cost scenarios, the ITU BT.601 weighting can also be used instead, with a difference of about 2%.

[0036] A Gabor filter bank with eight directions is constructed, with the eight directions taking values ​​of 0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135°, and 157.5°, uniformly covering the range from 0° to 180°. This corresponds to the... The Gabor cores in each direction are: ; ;in These are the relative pixel coordinates with the kernel center as the origin. For the first Angles in each direction ( Take 0 to 7). The wavelength is a sinusoidal carrier (in one embodiment, it is 8 pixels, which is equivalent to the spacing of a typical frost dendrite trunk on a 512×512 grid). The standard deviation of the Gaussian envelope is 0.56. (approximately 4.5 pixels). The aspect ratio of the ellipse with the Gaussian envelope is set to 0.5 to make the kernel narrower along the carrier direction and improve directional selectivity. The unit is the imaginary number. The kernel size is set to 25×25 pixels, and the amplitude at the kernel boundary is attenuated to less than 1% of the center value, so that the error introduced by zero extension can be ignored. Optionally, Log-Gabor can be used instead of standard Gabor. Log-Gabor has a log-Gaussian distribution in the frequency domain, is not sensitive to low-frequency DC bias, and is suitable for scenarios with strong background temperature drift.

[0037] The complex response is obtained by performing two-dimensional convolutions of the Gabor kernels in eight directions with the visible light grayscale image. The square of the complex modulus of the complex response is taken as the pixel position at the 1st position. Directional energy value in the direction: ;in and These are the real and imaginary parts of the complex response, respectively. The reason for using the square of the complex modulus value instead of any single-sided response is that the center of the bright fringe of the same frost dendrite will fall on the pixel or between adjacent pixels due to the different alignment phases of the pixel sampling grid between different frames, causing the sign of the real response to jump repeatedly between frames; the square of the complex modulus value folds away this phase difference, making the directional energy value stable between frames, and also improving the robustness to subtle texture displacements.

[0038] For each pixel, the maximum value among the eight energy values ​​is taken as the principal axis energy value for that pixel, denoted as . Simultaneously, record the direction index where the maximum value was obtained. The main axis direction of this pixel Just put Using Gabor responses as input to subsequent convolutional neural networks presents challenges—Gabor responses have 180-degree equivalence, meaning 0 degrees and 157.5 degrees are physically almost on the same axis, yet numerically differ by 157.5 degrees, making it difficult for the network to learn a smooth directional manifold. To eliminate this discontinuity, the principal axis directions are encoded onto the unit circle in a double-angle format: ;in and These represent the values ​​of the cosine and sine axes along the principal axis at this pixel, respectively, and both have a value range of [missing value]. This transformation maps both 0 degrees and 180 degrees to... This, along with a single point, completely eliminates axial ambiguity; adjacent directions (such as 0 degrees and 22.5 degrees) correspond to 0 degrees and 45 degrees at double angles, forming a 45-degree angle on the unit circle, resulting in a smooth transition. (This is achieved from all pixels.) The directional energy diagram is denoted as ; composed of all pixels , The cosine and sine graphs along the principal axis are respectively constructed and denoted as follows: , The three together constitute a directional prior channel.

[0039] refer to Figure 2 , Figure 2 Eight sampling directions are marked at equal intervals on the outer circumference: 0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135°, and 157.5°. These eight directions cover the semi-circular interval from 0° to 157.5° with a uniform step size of 22.5°. Since the Gabor filter has 180° equivalence for texture directions (i.e., 0° and 180° correspond to the same physical axis), the range from 0° to 157.5° completely covers all poses of the undirected axis. Three concentric circular reference lines are drawn at equal intervals along the polar radius direction of the pole figure, corresponding to directional energy values ​​of 0.3, 0.6, and 0.9, respectively.

[0040] The eight directional energy values ​​are represented on the pole figure by radial line segments extending outward from the pole center. The length of each line segment strictly corresponds to the directional energy value in that direction. The eight radial line segments are connected by a broken line to form a closed directional energy profile. Considering the 180-degree equivalence of the Gabor response, eight radial line segments located on the opposite side (180° to 337.5°) are also plotted on the pole figure, giving the entire pole figure a centrally symmetrical shape about the pole center. This symmetrical display clearly characterizes the undirected nature of the principal axis directions.

[0041] exist Figure 2 In the shown operating conditions, the energy values ​​for the eight directions are 0.22, 0.30, 0.55, 0.85, 0.92, 0.70, 0.40, and 0.25, respectively. The highest energy value for one direction occurs at 90 degrees, with a value of 0.92. This highest energy value is taken as the principal axis energy value of that pixel, denoted as... The direction corresponding to the maximum directional energy value is taken as the principal axis direction of that pixel, denoted as . In the pole figure, the radial line segment corresponding to 90 degrees is marked in bold along with the solid dot at its endpoint, serving as a visual marker of the principal axis direction. The energy values ​​of the other seven directions represent the texture response intensity along other directions at this pixel location, and their values ​​are much lower than the energy values ​​of the principal axis direction, indicating that the texture of this pixel has significant directional selectivity, and the determination of the principal axis direction has sufficient discriminative power.

[0042] A directional energy map is composed of the principal axis energy values ​​of all pixels. Subsequent steps use this directional energy map for two purposes: first, as one channel of the input tensor for frost dendrite segmentation, enabling the segmentation network to perceive the spatial distribution of texture intensity; second, as the energy threshold for selecting valid voting pixels in the dominant cold bridge location inversion stage. Foreground pixels with an energy level greater than or equal to a preset energy threshold are allowed to participate in the border hit point voting, thereby eliminating weak texture pixels with unclear orientation from the voting.

[0043] The visible light image, the corrected thermal image, and the orientation prior channel are sequentially concatenated along the channel dimension to obtain the frost dendrite segmentation input tensor. The spatial resolution is 512×512, and the channel arrangement is as follows: visible light red component, visible light green component, visible light blue component, corrected thermal image, orientation energy map, principal axis orientation cosine map, and principal axis orientation sine map, for a total of 7 channels (if the visible light is taken as a single-channel grayscale image weighted by BT.709, then there are 5 channels). Before being fed into the subsequent frost dendrite UNet, each channel is normalized to zero mean and unit variance using the following formula: ;in For a certain channel to be normalized, and These are the global mean and standard deviation of each channel, pre-calculated on the training set, and are embedded in the input preprocessing parameters of the Frost Dendrite UNet. This processing ensures that the channels are comparable on a numerical scale; for example, the amplitude of thermal imaging channels is typically... Celsius In Celsius, the direct magnitude will dominate the early gradient of the network, while after normalization, the variance of each channel is pulled down to about 1, allowing the network to pay equal attention to each mode.

[0044] Optionally, in embedded neural network accelerator deployment scenarios, the normalized tensor can be further linearly mapped to 8-bit integers from 0 to 255 to adapt to the input format of the fixed-point inference engine; in real-time control scenarios that pursue the lowest latency, the principal axis direction cosine graph and principal axis direction sine graph can be directly obtained by summing the real and imaginary parts of the Gabor complex response and normalizing them, eliminating the need for explicit direction maximum search. The cost is a slight decrease in direction selectivity, which is suitable for operating modes where the frost dendrite texture is already quite significant and the controller mainly undertakes steady-state maintenance.

[0045] The FrostDevice UNet generally adopts an encoder-decoder symmetrical topology, with two task branches connected in parallel at the end of the decoder. The encoder consists of four downsampling stages starting from the input. Each stage comprises a double convolutional block consisting of two 3×3 convolutions, batch normalization, and ReLU activation, followed by a 2×2 max pooling operation with a stride of 2. The number of channels increases with depth, starting at 64, 128, 256, and 512 from the input, with the deepest bottleneck layer having 1024 channels. The decoder mirrors this architecture. Each stage first doubles the spatial resolution using transposed convolutions or nearest-neighbor upsampling, then concatenates encoder features of the same resolution along the channel dimension using skip connections, followed by a double convolutional block. The number of decoder output channels decreases sequentially to 512, 256, 128, and 64. The input tensor space resolution is 512×512. After 4 downsampling steps, the bottleneck layer resolution is 32×32. The consideration for this depth selection is that the feature span of a typical frost dendrite trunk on the original image is 8 to 32 pixels. The receptive field of the bottleneck layer covers more than 8 times this scale, which can accommodate the overall morphology of the dendrite without over-pooling and causing the loss of small-scale condensation details.

[0046] The 64-channel feature map at the end of the decoder is denoted as ,in . Simultaneously, two task branches are entered. The semantic segmentation branch consists of two concatenated 3×3 convolutional layers, with a ReLU activation in the middle. The channel dimension is compressed from 64 to 32, and a Sigmoid activation is appended at the end to obtain the condensation / frost confidence map for the current control cycle. ,in These are pixel coordinates. According to the preset binarization threshold Binarization is performed to obtain the condensation / frost segmentation mask for the current control cycle. : ;in In one implementation, a value of 0.5 is used, but it can also be fine-tuned to the range of 0.4 to 0.6 before deployment, based on the intersection of the precision-recall curves on the validation set. The topological feedforward branch structure is isomorphic to the semantic segmentation branch, but the terminal activation is changed to ReLU. This is because the barcode length to be predicted is itself a non-negative continuous physical quantity (temperature difference, in degrees Celsius), and the Sigmoid function will suppress the value. The ReLU algorithm retains positive values ​​for loss magnitude and naturally suppresses weak responses near zero. This branch outputs a graph predicting the barcode length for the next control cycle under baseline control conditions. The numerical unit is degrees Celsius. The baseline control condition refers to the evolutionary condition under which the heating power of each transparent heating zone and the ventilation opening of each frame vent maintain the current period value. This limitation is designed to decouple "image evolution" from "control intervention" at the supervised learning level, so that the network learns the inherent law of "how the window will deteriorate without new intervention", and then the collaborative controller adds intervention when running online.

[0047] The training process of the Frost Dendrite UNet is divided into two parts: offline supervised ground truth preparation and supervised parameter update. Offline supervised ground truth preparation includes manually labeled masks for the semantic segmentation branch and topological supervision channels for the topological feedforward branch. The former is obtained by trained annotators who delineate condensation and frost areas on the visible light image to obtain binary masks. The typical annotation rule is: connected regions where water film reflections or crystal textures are visible are all considered foreground.

[0048] The true value of the topology feedforward branch's supervisory function—the topology supervisory channel—is calculated from the corrected thermal image acquired under the reference control conditions in the next control cycle using zero-duration long-term homology. This thermal image is denoted as... Let the lower bound of temperature sampling be set. With the upper realm In one implementation, take Celsius Celsius, temperature step Celsius, level set threshold sequence satisfy ,common Zhang's horizontal set binary slice image. For the first... For each slice image, position it as follows: That is, pixels with temperature values ​​less than or equal to the horizontal set threshold are set as foreground, and pixels with temperature values ​​greater than the horizontal set threshold are set as background. The slice images are traversed one by one from low to high. For each slice image, the foreground connected component is extracted according to the 8-connectivity rule. The 8-connectivity rule means that if any pixel's 8 adjacent pixels (top, bottom, left, right, and 4 diagonals) are all foreground, they are classified into the same connected component. This connectivity definition is more compact than 4-connectivity under the condition that frost crystals tend to spread along the diagonal, and it is less likely to break a single oblique dendrite into multiple components.

[0049] To track the birth and death of connected components in the evolution of level sets, a disjoint-set data structure is introduced. Each connected component first appears at level 1. When creating a slice of the image, a new node is created in the union-find set, and its birth threshold is recorded as . And use a set of foreground pixel coordinates as the representative element set for this component. When When increasing the threshold causes two or more foreground connected components to become adjacent for the first time at the same level set and then merge into a new foreground connected component, the Elder rule is executed: among all connected components involved in the fusion, the one with the lowest birth threshold is found and retained as a surviving component; the death threshold of the remaining connected components is uniformly recorded as the current threshold. The Elder rule is based on the idea that the lowest birth threshold means that the component grows from a colder core and is physically closer to the center of the real cold bridge. The warmer components that appear later are often substructures that split off after the core spreads outward. Merging them and causing them to disappear is more in line with the frost formation mechanism of "diffusion from cold to the outside".

[0050] In advance The four sides shrink inwards The ring band of a pixel is defined as the image boundary protection band. In one implementation, 25 pixels are used, approximately 5% of the 512-pixel side length. For foreground connected components generated within the image boundary protection band, even if they are in a certain... The first time it encounters an image boundary, it is not immediately assigned a disappearance threshold. Instead, it is marked as a candidate connected component for a cold bridge in the union-find set and continues to be processed. Advance tracking; for candidate connected components of a cold bridge in the border, their extinction threshold is only applied when they are merged with other connected components or traversed to the highest level set threshold. The time square is assigned a value, and the assignment rule is to take the threshold of the level set corresponding to the triggering event. This retention strategy is aimed at the physical characteristics of the cold bridge of the border - the low temperature area near the border is often attached to the edge of the image as soon as it is created. If the usual practice is to make it disappear when it is connected to the boundary, a barcode length close to 0 will be obtained, and the border cold bridge that should be identified will be removed from the topologically significant area.

[0051] The barcode length of each foreground connected component is obtained by subtracting its birth threshold from its extinction threshold, denoted as . subscript Identifier The barcode length physically represents the temperature span over which the low-temperature cluster can persist on the horizontal set temperature scale—the higher the temperature rises and the more the cluster maintains its low-temperature island status, the more robust its cold source; conversely, clusters that are submerged by surrounding warm areas with a slight temperature increase are usually noise or transient disturbances. The backfilling rule is: take the binary slice of the horizontal set preceding the disappearance threshold of the connected component (i.e., the first...). All pixel positions occupied in the slice image) ,Bundle These are used as topological supervision values ​​for these pixels; when backfilling conflicts occur at the same pixel location among different connected components, the maximum value of the barcode length is taken: ;in That is, the topology supervision channel at the pixel The value is taken in degrees Celsius. Using the maximum value as the decision-making method tends to preserve significant topological events, allowing downstream networks to focus more on stable cold sources rather than short-lived disturbances. Alternatively, the sum of the barcode lengths can be used as the conflict decision, at the cost of the values ​​being superimposed multiple times and losing their clear physical dimensions. This method is typically only used in early warning scenarios where high sensitivity is required.

[0052] In the supervised parameter update phase, the semantic segmentation branch uses manually labeled masks. To supervise the truth value, a binary cross-entropy loss is used: ;in , This represents the number of pixels in the height and width directions of the image. It is the natural logarithm. The topological feedforward branch uses a topological supervision channel. To supervise the truth value, a smoothed L1 loss is used: ;in For Huber functions, To predict the residuals, a smoothed L1 loss model is chosen instead of a pure L2 loss model because thermal imaging measurements are affected by non-uniform residuals and emissivity drift, resulting in a thicker tail in the residual distribution. A pure L2 loss model would be dominated by extreme samples at the tail. A smoothed L1 loss model degenerates into linear growth when the residuals are large, significantly suppressing the pull of extreme samples on the gradient. The total loss is the direct sum of the two loss models: The Adam optimizer is used to iteratively update the parameters of the frost dendrite UNet. A typical hyperparameter configuration is as follows: initial learning rate... The decay rates of the first and second moments are set to 0.9 and 0.999, respectively. The mini-batch size is 16, and the training is performed for 80 epochs. The learning rate is decayed to 0.1 and 0.01 of the initial value in epochs 40 and 60, respectively. Convergence is defined as the total loss changes by less than 1% over 5 consecutive epochs on the validation set. Optionally, learnable homoscedasticity uncertainty weights can be added to the two loss terms to automatically balance the gradient magnitudes of the two branches in the early stages of training. In embedded deployment scenarios, 8-bit fixed-point quantization and channel pruning can be applied to the converged parameters, reducing the model size to one-quarter of the original size, and the accuracy loss can be controlled within 2%.

[0053] The inversion of the dominant cold bridge location is initiated after the frost dendrite UNet completes forward inference. This is first performed on the corrected thermal image. The temperature gradient of each pixel is calculated using the Sobel operator. The horizontal and vertical kernels of the Sobel operator are both 3×3 difference kernels, and the horizontal gradient component is obtained after the operation. With vertical gradient components Both are measured in units of the same dimension as temperature (degrees Celsius), representing the differential rate of change of temperature along the X and Y directions. Temperature gradient amplitude. with respect to the direction of the temperature gradient It is given by the following formula: ; ;in It is a two-parameter arctangent function, with a range of values. This returns the directed argument of the gradient vector. The opposite direction of the temperature gradient, i.e., along the direction of decreasing temperature, is defined as the direction of low temperature. To merge the scope to This definition gives each pixel a clear arrow pointing to the "cooler side," which is the "compass" needed to upgrade the undirected Gabor axis to a directed backtracking ray.

[0054] refer to Figure 3 , Figure 3 The window surface is represented by a rectangular outline, with the lower left corner of the rectangle serving as the origin of the physical coordinate system. The small rectangles that are deepened in the middle of the upper edge of the rectangle represent the location of the dominant cold bridge identified under this working condition. The irregular polygons in the upper left area inside the window surface represent the foreground connected region of the condensation / frost segmentation mask in the current control cycle. Several sample foreground pixels are marked with small solid dots. The solid lines with arrows drawn from each sample foreground pixel along its directional backtracking direction present a geometric shape that converges in segments towards the frame where the dominant cold bridge is located, intuitively reflecting the physical idea of ​​tracing back from the condensation area along the dendrite axis to the cold source.

[0055] Traversing within the condensation / frost separation mask For each foreground pixel, read the cosine value of that pixel from the principal axis cosine map and the principal axis sine map. With sine value Solve the following equation in reverse for a double angle and return it to the principal axis direction of a single angle: Main axis direction The value falls within , representing one standard representative angle of the undirected axis within this range. The direction of the ray extending along the original direction is... The direction of the ray extending in the opposite direction from the original direction is The two rays are physically two segments of the same straight line. The segment with the smaller angle to the low-temperature pointing direction is selected as the directed tracing direction for the foreground pixel. Specifically, the determination is as follows: Calculation and The absolute value of the included angle and folded back , recorded as ;calculate and The absolute value of the included angle and folded back , recorded as ;like Then take Otherwise take The rationale behind this choice is that frost dendrites physically grow along the direction of decreasing temperature (i.e., from a relatively warm area to a relatively cold area). By defining the section along the axis pointing to the colder side as the retrograde direction, the undirected texture axis can be uniquely upgraded into a directed retrograde ray.

[0056] To eliminate the contamination of voting results by weak information pixels, a three-stage screening process is introduced: 1) It must be less than or equal to the preset directional consistency threshold. 2 is the energy value in the principal axis direction. It must be greater than or equal to the preset energy threshold. 3 is the amplitude of the temperature gradient. It must be greater than or equal to the preset gradient threshold. Foreground pixels that simultaneously satisfy all three conditions are included in the set of valid voting pixels. In one implementation method, take That is, 30 degrees Celsius; Take all foreground pixels within the control period The median is 0.5 times the median, to accommodate different illuminance levels; A 0.05°C gradient per pixel is used to mask spurious gradients caused by background noise in thermal images. This triple screening process removes pixels with weak directional and low-temperature gradients from the voting process, retaining only pixels with both clear directionality and a clear cold source orientation for cold bridge inversion.

[0057] For each pixel in the set of valid voting pixels, start from its position and trace back along its directed direction. Draw a ray; the first intersection point of the ray within the window's physical coordinate system border region is taken as the border hit point for that pixel. The intersection of the ray and the border region is calculated analytically: the window border is abstracted into four line segments (bottom edge, top edge, left edge, and right edge), and the intersection points of the ray's parametric equation and the equations of the four border line segments are calculated respectively. The intersection point with the smallest parameter and greater than 0 is selected as the hit point. The ray is then drawn along the window border with a preset fixed arc length step. Divided into multiple border segments, In one implementation, one-twentieth of the window width is taken, typically 50 mm; the arc length is zero at the lower left edge, and increases clockwise. All border hit points are counted according to the border segments to which the arc length belongs, and the hit point count of each border segment is obtained; the border segment with the maximum hit point count is identified as the dominant cold bridge location, and the window surface physical coordinates of the midpoint of this border segment are used as the representative point of the dominant cold bridge location.

[0058] Under certain operating conditions, the hit point count will show two close maxima on the frame, reflecting the simultaneous presence of two cold bridges on the window surface (common in double-glazed windows with metal spacers along both the top and bottom edges). Optionally, in an implementation that aims to identify dual cold bridges, the hit point count sequence can be smoothed by a 1D convolution with a Gaussian kernel width of 3 frame segments, and then all local maxima can be extracted using non-maximum suppression. Local maxima that are higher than 0.7 times the global maximum and are spaced more than 5% of the total length of the window frame are identified as dominant cold bridge locations. The subsequent collaborative controller will simultaneously initiate cold bridge directional triggering for each dominant cold bridge location.

[0059] At this point, the condensation / frost separation mask is complete. Next control cycle predicted barcode length diagram Together with the location of the dominant cold bridge (characterized by one or more representative points in the physical coordinate system of the window), the three constitute the window condensation state description result, which serves as the state input for all subsequent collaborative control links.

[0060] The window surface is physically divided into OK Transparent heating zones that are independently addressable from each other. and In a typical implementation, 6 and 8 are selected respectively, resulting in a total of 48 transparent heating zones in the entire window. The conductive medium of each transparent heating zone is an indium tin oxide thin film or a silver nanowire composite film, with a sheet resistance of 10 ohms to 30 ohms per square, a visible light transmittance higher than 85%, and a haze lower than 2%. These are laid out in a regular rectangular grid in the physical coordinate system of the window surface. Each transparent heating zone is connected at both ends to a power adjustment unit located within the window frame via a thin copper strip running along the frame. Power adjustment is achieved by pulse width modulation with a switching frequency of 20 kHz. The duty cycle can be independently set for each control cycle, thereby independently setting the instantaneous heating power. The upper limit of the instantaneous heating power for a single zone is... Set 15 watts per zone, 0 watts as the lower limit, and 0.1% duty cycle resolution to ensure sufficiently fine power adjustment.

[0061] The frame ventilation openings total They are evenly distributed around the perimeter of the window frame. In one implementation, the value is 8. Each frame vent is composed of a thin-blade louver mechanism driven by a stepper motor. The ventilation opening is linearly adjustable from fully closed to fully open, with the mechanical full stroke corresponding to 1024 steps of the stepper motor. Therefore, the digital resolution of the ventilation opening is approximately one-thousandth. Each frame vent is connected to the indoor dehumidification duct via a flexible duct. The dehumidification duct is maintained by an independently configured rotary dehumidifier or membrane dehumidifier, and the outlet dew point temperature is more than 3 degrees Celsius lower than the indoor dew point temperature. The reason for choosing dehumidified airflow instead of directly blowing indoor air is that once the water vapor partial pressure near the window surface exceeds the saturated water vapor pressure of the window surface, condensation is triggered. The dew point of the incoming airflow must be significantly lower than the window surface temperature to dilute the water vapor partial pressure near the wall surface. Directly blowing indoor air may carry an equal amount of water vapor into the near wall surface under humid and cold conditions, which would have the opposite effect.

[0062] In each transparent heating zone, a thin-film negative temperature coefficient thermistor is attached to the inner side of the glass interior as a surface temperature sensor, with a sampling accuracy of 0.1 degrees Celsius. Inside the zone, a dry-bulb temperature sensor and a capacitive relative humidity sensor are positioned 1.5 meters from the window surface, with sampling accuracies of 0.1 degrees Celsius and 1% relative humidity, respectively. Outside the zone, near the eaves on the outer side of the window, an outdoor temperature sensor is positioned with a sampling accuracy of 0.5 degrees Celsius. All sensors sample at a frequency of 1 Hz, the same as the control cycle, and the data is sent to the co-controller after low-pass digital filtering.

[0063] The collaborative controller operates in sync with the generation rhythm of the window condensation state description results. At the beginning of each control cycle, the dominant cold bridge location and condensation / frost separation mask are read from the window condensation state description results. Predicted barcode length chart for the next control cycle Based on this, three types of trigger partitions are identified.

[0064] The determination of cold bridge directional triggering is based on the representative point of the dominant cold bridge location in the window surface physical coordinate system. The geometric center coordinates and coverage rectangle of each transparent heating zone are pre-fixed in the memory of the co-controller. The transparent heating zone corresponding to the coverage rectangle where the representative point of the dominant cold bridge location is located, as well as transparent heating zones that share edges with this zone in the window surface physical coordinate system, are collectively identified as cold bridge triggering zones. The shared edge relationship includes adjacent zones in four directions: top, bottom, left, and right. This extension is because the thermal effect of the cold bridge has a lateral diffusion effect within the glass surface, with a typical diffusion distance of 50 mm to 100 mm on insulated glass, close to the side length of a single zone. Therefore, including adjacent zones in the cold bridge triggering zone ensures that the compensation heat precisely covers the lateral influence domain of the cold bridge. Among the frame vents, the frame vent closest to the representative point of the dominant cold bridge location along the arc length of the window frame is identified as the cold bridge triggering vent. When there are two or more representative points at the dominant cold bridge location, the above identification is performed on each representative point, and the union of the cold bridge trigger zone and the cold bridge trigger vent is taken as the final result.

[0065] The condensation area is determined by the triggering criteria, targeting the real-time area where condensation or frost has already occurred. For each transparent heating zone, statistics are compiled within its coverage rectangle. The number of pixels, denoted as ( , (For partition row and column indexes); if The transparent heating zone is identified as the follow-triggered zone. In one implementation, the threshold is set to 5% of the total number of pixels covered by the partition. That is, the system is triggered as long as more than 5% of the pixels in the partition are identified as foreground. The consideration for this threshold is that the signal-to-noise ratio of the micro liquid film in the initial stage is relatively weak in visible light and thermal images. The 5% coverage rate corresponds to the early stage when the naked eye has not yet perceived the blur, but the instrument can stably identify it. Intervening at this time, rather than waiting until it is visible to the naked eye, can suppress the continued development of the liquid film with low energy consumption.

[0066] The topology feedforward triggering decision is geared towards high-risk areas that have not yet occurred but are about to. For each transparent heating zone, within its coverage rectangle, [the following is applied]... Take the arithmetic mean, denoted as ;like The transparent heating zone was identified as a feedforward triggered zone. The unit is degrees Celsius, and in one implementation, it is taken as 1.5 degrees Celsius. This value is chosen because the topological response barcode length caused by typical sporadic disturbances (such as a person's instantaneous exhalation or a passing sunspot) usually falls between 0.5 and 1.0 degrees Celsius. Only low-temperature clumps supported by a stable cold source can sustain above 1.5 degrees Celsius. This boundary effectively eliminates transient noise and selects the stable cold sources that truly require feedforward intervention. The three types of trigger zones are allowed to overlap. One transparent heating zone can be simultaneously identified as a cold bridge trigger zone, a follow-up trigger zone, and a feedforward trigger zone. In the subsequent quadratic programming solution, all three identities are recorded in the end-point dew margin mandatory constraint.

[0067] The core of the collaborative controller is the rolling time-domain collaborative optimization. The predicted time-domain length is taken as... One control cycle, In one implementation, the value is 6, corresponding to a 6-second look-ahead; the optimization variable is the heating power sequence of all transparent heating zones in the prediction time domain. The sequence of ventilation openings along the entire frame , To predict the relative periodic index in the time domain, values ​​are taken from 0 to... , For the edge vent index, take 1 to... .

[0068] The zoned surface temperature state prediction model characterizes the impact of heating power and ventilation opening on future surface temperature, using the following linear difference form: ;in For the first Transparent heating partition in the prediction time domain Surface temperature per control cycle, in degrees Celsius; For this partition in the th Heating power per control cycle, in watts; This is the set of indices of adjacent partitions that share an edge in the physical coordinate system of the window; The nearest frame vent to this partition is in the first The ventilation opening degree per control cycle, dimensionless. For the index of the nearest border vent; and The first Indoor dry-bulb temperature and outdoor temperature for each control cycle, in degrees Celsius; to These are the coefficients for partitioned superposition, dimensionless (some coefficients inherently include physical dimension conversion factors). In terms of physical meaning... Reflects the inherent thermal inertia of the glass, typically ranging from 0.6 to 0.9; This reflects the conversion ratio from heating power to temperature rise, typically ranging from 0.05 to 0.2. Reflects the intensity of lateral thermal diffusion, typically ranging from 0.02 to 0.08; This reflects the coupling of ventilation opening to the surface temperature of that zone, and is typically taken as a small negative value, with a typical range of... to This is because the temperature of the dehumidifying airflow is usually lower than the temperature of the glass surface. and These values ​​reflect the convective heat transfer weights of indoor and outdoor air temperatures on the glass surface, respectively, with the sum typically ranging from 0.05 to 0.15.

[0069] The predicted indoor dry-bulb temperature for the next cycle will be updated using the following formula: ;in , , This is the indoor superposition coefficient, which is dimensionless. Reflects indoor thermal inertia, typically ranging from 0.95 to 0.99; Reflects the overall heat transfer from the outside to the inside through walls and windows, typically ranging from 0.005 to 0.02. Reflecting the overall disturbance of indoor air temperature caused by the frame vents, it is usually taken as a smaller negative value, typically within a certain range. to .

[0070] The aforementioned zone superposition coefficients and indoor superposition coefficients are obtained using step response identification. Specifically, before the collaborative controller is put into operation, the window surface is placed in a steady state, and several typical operating conditions (indoor-outdoor temperature difference of 5°C, 10°C, and 20°C) are selected. A square wave step with an amplitude of 50% of the rated heating power upper limit is applied sequentially to each transparent heating zone for 300 seconds, and the time response curves of the surface temperature of each transparent heating zone and the indoor dry-bulb temperature are recorded. Then, a fully open step is applied to each frame vent in the same way. The resulting response curves are regressed to the aforementioned linear difference form using the least squares method to obtain the zone superposition coefficients and indoor superposition coefficients, which are then stored in the collaborative controller's memory. Optionally, during the initial two weeks after new window installation, the collaborative controller can identify coefficients online using recursive least squares to allow the model to adapt to the gradual parameter changes caused by the stress release during new window assembly and the initial softening of the seals. After two weeks, the coefficients are switched back to their original state.

[0071] Dew point temperature is calculated from the current indoor dry-bulb temperature and relative humidity using the Magnus method: ;in This is the current dew point temperature, in degrees Celsius. This refers to the current indoor dry-bulb temperature, in degrees Celsius. The current relative humidity, expressed as a percentage; , , Magnus constant, Take 17.625, Take 243.04 degrees Celsius. Use 243.04 degrees Celsius. Dew point margin for the next cycle in each transparent heating zone. Defined as: ; among which reference temperature exist Take Celsius ,exist The temperature is taken from Celsius. Corrected frost point temperature The frost point temperature correction is given according to the following empirical relationship: This correction reflects the physical fact that the saturated vapor pressure on the ice surface is lower than that on the supercooled water surface, making the criteria more stringent in the low-temperature range and prompting heating to intervene earlier. This prevents the controller from still interpreting the dew point margin as safe and thus missing control when thin frost forms before thin dew.

[0072] The quadratic programming objective function is the sum of five terms. The first term is the sum of squares of the negative dew point margins for all transparent heating zones within the prediction time domain. (Negative dew point margin) Represented by a non-negative auxiliary variable, this auxiliary variable simultaneously satisfies and The consideration behind constructing this auxiliary variable is to directly address... Taking the square introduces piecewise nonlinearity, making it unsuitable for direct input into a quadratic programming solver. By leveraging these two inequality constraints, driven by the minimization objective, The optimal solution must be the larger of these two lower bounds, that is, exactly equal to... This effectively equates piecewise nonlinearity to linear constraints. The second term is the sum of squares of the instantaneous heating power of all transparent heating zones within the prediction time domain, representing an energy consumption penalty. The third term is the sum of squares of the instantaneous ventilation openings of all frame vents within the prediction time domain, representing a ventilation disturbance penalty. The fourth term is the sum of the squares of the difference in heating power between adjacent cycles of the same transparent heating zone and the squares of the difference in ventilation openings between adjacent cycles of the same frame vent, representing a control smoothness penalty. This avoids drastic jumps in power and opening between adjacent cycles, leading to frequent actuator movements and accelerated wear. The fifth term is the sum of squares of the difference between the predicted indoor dry-bulb temperature and the preset comfort temperature for the next cycle within the prediction time domain. In one implementation, the preset comfort temperature is 22 degrees Celsius. The terms are weighted and summed to form a total objective, with the weighting coefficients taking typical values ​​depending on the operating conditions. Under humid and cold conditions, the weight of the first term is amplified to strictly control condensation, while in energy-saving mode, the weight of the second term is amplified to suppress energy consumption.

[0073] The secondary planning constraints include: the maximum heating power of each transparent heating zone. The upper limit is 0 watts, and the lower limit is 0 watts; the upper limit of ventilation opening for each frame vent is 1 watt, and the lower limit is 0 watts; the upper limit of surface temperature for each transparent heating zone is 40 degrees Celsius to prevent excessive heating from causing thermal stress cracking of the glass; the upper limit of heating power change rate between adjacent weeks is 5 watts per cycle, and the upper limit of ventilation opening change rate between adjacent weeks is 0.2 watts per cycle; the state transition equations given by the zone surface temperature state prediction model and the indoor dry bulb temperature prediction model are respectively embedded as equation constraints.

[0074] To reflect the differentiated protection of the three types of trigger partitions, at the end of the prediction time domain... A mandatory constraint on the end-point dew point margin is introduced at all times. For cold bridge triggered partitions, follow-triggered partitions, and feedforward triggered partitions, the following requirements are specified. , To preset a lower safety margin, 2 degrees Celsius is used in one embodiment; the remaining transparent heating zones require... This end-of-phase constraint mechanism replaces the practice of directly incorporating the current segmentation mask into the constant term of the objective function, and forcibly projects "where the risk is high" into the optimization solution space through constraints. Partitions with three types of trigger identities are forced to have a safety margin of 2 degrees Celsius at the end of the prediction time domain. In order to achieve this margin, the optimizer will actively allocate sufficient heating power and ventilation opening to them. Non-trigger partitions only need to maintain a non-negative dew point margin. Based on this, the optimizer will allocate excess resources to high-risk partitions.

[0075] To accelerate the iterative convergence of the quadratic programming solver at the start of each control cycle, a warm-start strategy is adopted to provide initial values ​​for the optimization variables. The initial heating power for the cold bridge triggering partition, follow-up triggering partition, and feedforward triggering partition is taken as the upper limit of the heating power for the current cycle. The preset ratio value is typically 0.6 times, or 9 watts; the other transparent heating zones take another lower preset ratio value, typically 0.2 times, or 3 watts; the initial value of the ventilation opening of the cold bridge triggered vents in the current cycle takes the preset ratio value of the upper limit of the ventilation opening, typically 0.5 times; the other frame vents take 0. This initial value for warm start is physically close to the directional characteristics of the optimal solution, and the quadratic programming solver can converge after 3 to 5 iterations starting from this initial value, which is far less than the 15 to 20 iterations required for cold start.

[0076] At the start of each control cycle, the co-controller invokes a quadratic programming solver to simultaneously solve for all heating power sequences and ventilation opening sequences within the prediction time domain. The solver utilizes mature implementations based on the interior-point method or the effective set method, and in one implementation, it is deployed on an industrial-grade embedded platform with a 128-bit floating-point coprocessor, with a single solution taking less than 100 milliseconds. Only the solution from the first control cycle in the prediction time domain of the obtained sequence is used as the actual command issued for the current control cycle. and As the current cycle value, the transparent heating zone is driven by pulse width modulation, and the louvers of the frame ventilation openings are driven by stepper motors. The remaining values ​​in the prediction time domain... The solution for each cycle is discarded. This "multiple solutions, one solution" rolling strategy ensures that the actual instructions issued in each cycle are always re-optimized based on the latest observations, effectively suppressing the accumulation of modeling errors and external disturbances.

[0077] At the start of the next control cycle, new visible light images and thermal images are simultaneously acquired, registered, and directional prior channels are generated and sent to the frost dendrite UNet to obtain updated window condensation state description results. Based on this, three types of trigger zones are re-identified, and the initial values ​​of the zone surface temperature state prediction model and the indoor dry-bulb temperature prediction model are refreshed with updated sensor readings. The quadratic programming is solved again. The above steps are executed cyclically in each control cycle to achieve closed-loop operation of window frost image segmentation and recognition, dominant cold bridge inversion, and coordinated control of transparent heating zones and frame ventilation openings.

[0078] Optionally, in a smart building scenario with multi-window joint control, the collaborative controllers of each window can be connected to a single upper-level scheduling coordinator to apply a global upper limit to the total energy consumption and total noise. The weighting coefficients of the energy consumption term and ventilation disturbance term in the local objective function of each window are allocated online by the coordinator using the Lagrange relaxation method, so that the global energy consumption meets the upper limit while preserving the condensation suppression effect of each window as much as possible. Under extreme outdoor conditions (such as outdoor temperatures below -30 degrees Celsius for more than 2 hours), the safety margin lower limit of the cold bridge triggering zone can be temporarily increased from 2 degrees Celsius to 3 degrees Celsius to provide a thicker anti-condensation buffer.

[0079] refer to Figure 4 The horizontal axis represents the control cycle in seconds, from 0 to 59 seconds, with main scale intervals of 10 seconds. The left vertical axis represents the instantaneous heating power of a single zone in watts, with scale intervals of 3 watts, ranging from 0 watts to 15 watts and then extending upwards to 17 watts to allow space for upper limit power markings. The right vertical axis represents the ventilation opening, dimensionless, with a scale range from 0 to 1.05, and main scale values ​​at 0, 0.25, 0.50, 0.75, and 1.0. (The entire sheet...) Figure 4 Using time as the common horizontal axis, the two heterogeneous control variables, heating power and ventilation opening, are superimposed on the same time series diagram to intuitively compare their coordinated response within the same time window.

[0080] Figure 4Four curves are plotted. The first curve, represented by the thickest solid line, shows the heating power timing of the cold bridge trigger zone. It maintains a moderate level of approximately 9 watts for the first 10 seconds, then rises and stabilizes at around 12.5 watts between the 10th and 25th seconds to address the preset high-risk phase. After the 25th second, it drops back to a steady-state level of approximately 8 watts. The second curve, represented by a medium-thickness dashed line, shows the heating power timing of the follow-trigger zone. It is approximately 5.5 watts for the first 10 seconds, rises to approximately 9.5 watts during the high-risk phase, and then drops back to approximately 5 watts after the 25th second. Its overall level remains lower than that of the cold bridge trigger zone, demonstrating the higher priority protection status of the cold bridge trigger zone in the controller compared to the follow-trigger zone. The third curve, represented by the thinnest dotted line, shows the heating power timing of the ordinary zone, which does not belong to any trigger zone. It remains at a low level of approximately 2.5 watts throughout the entire time window, accompanied by slight sinusoidal fluctuations, reflecting the ordinary zone's operation mode of only undertaking basic insulation tasks and not participating in high-risk suppression.

[0081] The fourth curve represents the timing of the ventilation opening of the vents triggered by the cold bridge. It is plotted on the right-hand vertical axis and presented as a solid line connecting circular data markers with white fill. Initially, it maintains a moderate opening level of approximately 0.45 for the first 10 seconds, increasing to approximately 0.75 during the high-risk phase, and then dropping back to approximately 0.40 after 25 seconds. The trend of this curve is synchronized in time phase with the heating power curve of the cold bridge-triggered zone, indicating that the co-controller increases both the heating power near the cold bridge and the opening of the vents near the cold bridge, thereby simultaneously engaging two actuators to jointly intervene in the moisture partial pressure near the window wall and the glass surface temperature.

[0082] Figure 4 A dashed line is drawn at 15 watts on the vertical axis of heating power, representing the preset upper limit of heating power Qmax for a single zone; all three heating power curves are below this upper limit throughout the entire period, indicating that the solution output by the quadratic programming solver strictly satisfies the upper and lower limit constraints of heating power.

[0083] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for coordinated control of window frosting suppression and micro-heating based on image segmentation, characterized in that, Includes the following steps: Step 1: Simultaneously acquire visible light images and thermal images using a visible light camera and a long-wave infrared thermal imager, and register them to the window physical coordinate system; generate a directional prior channel for the visible light image based on the Gabor filter bank, including a directional energy map, a principal axis direction cosine map, and a principal axis direction sine map; stitch the visible light image, thermal image, and directional prior channel along the channel dimension to obtain the frost dendrite segmentation input tensor; Step 2: Input the frost dendrite segmentation input tensor into the pre-trained frost dendrite UNet. The frost dendrite UNet includes a semantic segmentation branch and a topology feedforward branch. The semantic segmentation branch outputs the condensation / frost segmentation mask for the current control cycle. The topology feedforward branch outputs the predicted barcode length map for the next control cycle under the baseline control condition. The topology feedforward branch is pre-trained using the topology supervision channel obtained from the thermal image through persistent cohomology calculation as the supervision ground value. The location of the dominant cold bridge is determined by inverting the condensation / frost segmentation mask, the orientation prior channel, and the temperature gradient of the thermal image. Step 3: Divide the window into multiple independently addressable transparent heating zones, and set multiple adjustable ventilation openings on the window frame. The collaborative controller determines the cold bridge trigger zone, follow-up trigger zone, and feedforward trigger zone based on the dominant cold bridge location, condensation / frost segmentation mask, and the predicted barcode length map for the next control cycle. Based on the zoned surface temperature state prediction model and dew point margin, the heating power sequence of all transparent heating zones and the ventilation opening sequence of all frame vents are solved in the prediction time domain through quadratic programming. The solution of the first control cycle is then sent out for execution.

2. The method as described in claim 1, characterized in that, In step 1, the visible light camera and the long-wave infrared thermal imager synchronously acquire visible light images and raw thermal images using the same hardware trigger signal. The raw thermal image is then subjected to response uniformity correction and emissivity correction to obtain the thermal image. Through checkerboard calibration and homography matrix transformation, the visible light image and the thermal image are registered to a window physical coordinate system established with the lower left corner of the window as the origin, the width direction of the window as the X-axis, and the height direction of the window as the Y-axis. This ensures that pixels at the same physical location in the visible light image and the thermal image correspond one-to-one, resulting in a multimode image pair.

3. The method as described in claim 1, characterized in that, The Gabor filter bank comprises eight directions: 0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135°, and 157.5°. Each filter has the same center wavelength and the same Gaussian envelope width. The Gabor filters in each of the eight directions are convolved with the visible light image in two dimensions. The square of the complex modulus value of the response at the same pixel position in the resulting eight directional response maps is taken to obtain the energy values ​​of the eight directions at the corresponding pixel positions. The direction energy value with the largest value at each pixel position is selected as the principal axis direction energy value of the corresponding pixel, and the direction corresponding to the largest direction energy value is taken as the principal axis direction of the corresponding pixel. A directional energy map is composed of the principal axis direction energy values ​​of all pixels. The cosine and sine values ​​of the principal axis direction of each pixel, after taking a double angle, are used to form the principal axis direction cosine and principal axis direction sine maps, respectively.

4. The method as described in claim 1, characterized in that, The FrostDendric UNet consists of an encoder, a decoder, and two parallel output branches, with skip connections between the encoder and decoder. The semantic segmentation branch is composed of two convolutional layers and one sigmoid activation function, outputting the condensation / frost confidence map for the current control cycle. The condensation / frost confidence map is binarized according to a preset binarization threshold to obtain the condensation / frost segmentation mask for the current control cycle. The topology feedforward branch is composed of two convolutional layers and one ReLU activation function, outputting the predicted barcode length map for the next control cycle under the baseline control conditions. The baseline control conditions refer to the evolutionary conditions under which the heating power of each transparent heating zone and the ventilation opening of each frame vent maintain the values ​​of the current cycle.

5. The method as described in claim 4, characterized in that, The topology supervision channel is obtained as follows: for the thermal images acquired in the training samples during the next control cycle under the baseline control conditions, multiple level set thresholds are set from low to high within its temperature range using a preset fixed temperature step size. For each level set threshold, pixels with temperature values ​​less than or equal to the level set threshold are set as foreground, and pixels with temperature values ​​greater than the level set threshold are set as background, thus obtaining the corresponding level set binary slice image. The binary slices of the level set are traversed one by one in order of increasing level set threshold, and the foreground connected components are extracted according to the 8-connectivity rule. When a foreground connected component first appears, the current level set threshold is recorded as the birth threshold of the corresponding connected component. When two or more foreground connected components merge into a new foreground connected component under the same level set threshold, the component with the lower birth threshold is retained as a surviving component, and the extinction threshold of the remaining foreground connected components is recorded as the current level set threshold. Image boundary protection bands are pre-set on the four sides of the thermal image. Foreground connected components born within the image boundary protection band that connect to the image boundary are still tracked as candidate connected components for cold bridges in the border until they merge with other foreground connected components or are traversed to the highest level set threshold. The fusion threshold or the highest level set threshold is used as the extinction threshold of the corresponding connected component. Each foreground connected component... The barcode length of the corresponding connected component is obtained by subtracting the birth threshold from the extinction threshold. The barcode length is then backfilled to all pixel positions occupied by the corresponding connected component in the binary slice of the level set before the extinction threshold. When the same pixel is backfilled by multiple connected components, the maximum barcode length is taken. During the training of the Frost Dendrite UNet, the manually labeled binary mask of the condensation / frost region in the current control cycle is used as the supervision ground value of the semantic segmentation branch, and the binary cross-entropy loss is adopted. The topological supervision channel is used as the supervision ground value of the topological feedforward branch, and the smooth L1 loss is adopted. The sum of the binary cross-entropy loss and the smooth L1 loss is used as the total loss. The Adam optimizer is used to iteratively update the parameters of the Frost Dendrite UNet until the total loss converges.

6. The method as described in claim 1, characterized in that, The method for retrieving the dominant cold bridge location is as follows: calculate the temperature gradient magnitude and temperature gradient direction of each pixel in the thermal image using the Sobel operator, and obtain the low temperature pointing direction of the corresponding pixel along the temperature decrease direction; In the condensation / frost segmentation mask, each foreground pixel is traversed. The principal axis direction is obtained by inversely solving the cosine and sine values ​​of the principal axis direction of the corresponding foreground pixel. The principal axis direction extends along the original direction as a forward ray and extends in the reverse direction as a backward ray. The side with the smaller angle with the low temperature pointing direction is selected as the directed backtracking direction of the corresponding foreground pixel. Foreground pixels that simultaneously meet the following three screening conditions are included in the set of valid voting pixels: the angle between the principal axis direction of the foreground pixel and the low temperature pointing direction is less than or equal to a preset direction consistency threshold, the energy value of the principal axis direction is greater than or equal to a preset energy threshold, and the temperature gradient amplitude is greater than or equal to a preset energy threshold. The gradient threshold is greater than or equal to the preset gradient threshold. For each foreground pixel in the set of valid voting pixels, a ray is drawn from the corresponding position along the corresponding directed backtracking direction. The first intersection point of the ray within the window's physical coordinate system border area is taken as the border hit point of the corresponding foreground pixel. The window border is divided into multiple border segments with a preset fixed arc length step. All border hit points are counted according to the border segment to which they belong. The border segment with the maximum hit point count is identified as the dominant cold bridge position. The condensation / frost segmentation mask, the barcode length map predicted for the next control cycle, and the dominant cold bridge position together constitute the window condensation state description result.

7. The method as described in claim 6, characterized in that, The collaborative controller determines the cold bridge trigger zone, the follow trigger zone and the feedforward trigger zone in the following ways: cold bridge directional triggering, the transparent heating zone that is adjacent to the position of the dominant cold bridge in the window physical coordinate system is set as the cold bridge trigger zone, and the frame vent that is closest to the position of the dominant cold bridge along the arc length of the window frame is set as the cold bridge trigger vent. Condensation area follow-triggered: In the condensation / frost segmentation mask, traverse the window blocks covered by each transparent heating zone, count the number of foreground pixels in each window block, and set the transparent heating zone with the number of foreground pixels exceeding the preset follow-triggered threshold as the follow-triggered zone; Topology feedforward triggered: In the next control cycle, traverse the window blocks covered by each transparent heating zone in the predicted barcode length map, calculate the average predicted barcode length of all pixels in each window block, and set the transparent heating zone with the average predicted barcode length exceeding the preset feedforward threshold as the feedforward triggered zone.

8. The method as described in claim 7, characterized in that, In step 3, the window surface is pre-divided into M rows and N columns of independently addressable transparent heating zones, where M and N are both positive integers. P frame vents, where P is a positive integer, are evenly distributed along the circumference of the window frame. Each frame vent is connected to an indoor dehumidification duct to deliver dehumidified airflow to the window edge area, reducing the partial pressure of water vapor near the wall layer. One surface temperature sensor is placed on the window surface of each transparent heating zone, one dry-bulb temperature sensor and one relative humidity sensor are placed indoors, and one outdoor temperature sensor is placed outdoors. The zone surface temperature state prediction model and dew point margin are constructed as follows: reading data from each surface temperature sensor, dry-bulb temperature sensor, ... The current measurements from the relative humidity sensor and the outdoor temperature sensor; the predicted surface temperature for each transparent heating zone in the next cycle is obtained by multiplying the current cycle surface temperature of the corresponding zone, the current cycle heating power of the corresponding zone, the sum of the differences between the current cycle surface temperatures of adjacent zones and the current cycle surface temperatures of the corresponding zone, the current cycle ventilation opening of the nearest frame vent to the corresponding zone, the current indoor dry-bulb temperature, and the current outdoor temperature by a zone superposition coefficient and then summing them; the predicted indoor dry-bulb temperature for the next cycle is obtained by multiplying the current indoor dry-bulb temperature, the current outdoor temperature, and the current cycle ventilation opening of all frame vents by an indoor superposition coefficient and then summing them. Both the zonal superposition coefficient and the indoor superposition coefficient were obtained from the step response identification experiment before the collaborative controller was put into operation and were stored in the memory of the collaborative controller; the current dew point temperature was calculated by converting the current indoor dry bulb temperature and relative humidity according to the Magnus method. The difference between the predicted surface temperature for the next cycle of a zone and the current dew point temperature is used as the dew point margin for the next cycle of the corresponding zone. If the predicted surface temperature for the next cycle of a partition is below 0 degrees Celsius, the frost point temperature, which is corrected from the current dew point temperature, will be used instead of the current dew point temperature in the difference calculation of the dew point margin for the corresponding partition.

9. The method as described in claim 8, characterized in that, The objective function used in the quadratic programming solution is the sum of the following five terms: The first term is the sum of squares of the negative dew point margins of all transparent heating zones within the prediction time domain. The negative dew point margins are represented by auxiliary variables, which are greater than or equal to 0 and greater than or equal to the difference between the current dew point temperature and the predicted surface temperature for the next cycle; The second term is the sum of squares of the instantaneous heating power of all transparent heating zones within the prediction time domain; The third term is the sum of squares of the instantaneous ventilation openings of all frame vents within the prediction time domain; The fourth term is the sum of the squares of the difference in heating power between adjacent cycles of the same transparent heating zone and the squares of the difference in ventilation openings between adjacent cycles of the same frame vent within the prediction time domain; The fifth term is the sum of squares of the difference between the predicted indoor dry-bulb temperature and the preset comfort temperature for the next cycle within the prediction time domain.

10. The method as described in claim 9, characterized in that, The constraints used in the quadratic programming solution include: upper and lower limits of heating power for each transparent heating zone, upper and lower limits of ventilation opening for each frame vent, upper limit of surface temperature for each transparent heating zone, upper limits of heating power change rate and ventilation opening change rate during adjacent periods, state transition equation constraints given by the zone surface temperature state prediction model, and indoor dry-bulb temperature prediction equation constraints; a mandatory constraint greater than or equal to a preset safety margin lower limit is applied to the dew point margin at the end of the prediction time domain for cold bridge triggering zones, follow-up triggering zones, and feedforward triggering zones, with the safety margin lower limit being a constant greater than 0; a mandatory constraint greater than or equal to 0 is applied to the dew point margin at the end of the prediction time domain for the remaining transparent heating zones; using cold bridge triggering zones, follow-up triggering zones, and feedforward triggering zones as examples... The heating power of the trigger zone and the feedforward trigger zone in the current cycle is taken as a preset proportion of the upper limit of the heating power, the other transparent heating zones are taken as another lower preset proportion of the upper limit of the heating power, and the ventilation opening of the cold bridge trigger vent in the current cycle is taken as a preset proportion of the upper limit of the ventilation opening, which are used as the initial values ​​of the warm start of the optimization variables. A quadratic programming solver is used to solve all heating power sequences and ventilation opening sequences in the prediction time domain at the beginning of each control cycle. The solution of the heating power sequence and the ventilation opening sequence in the prediction time domain of the first control cycle is taken as the actual command issued in the current control cycle and issued to the corresponding transparent heating zone and the frame vent. Steps 1 to 3 are repeated at the beginning of the next control cycle.

Images