Air conditioner pipeline mildew spot area identification method fusing GMM and adaptive threshold segmentation

Abstract

The invention discloses an air conditioner pipeline mildew spot area identification method fusing GMM and adaptive threshold segmentation, and relates to the technical field of intelligent image detection and air conditioner pipeline maintenance. Mildew spot features are accurately extracted and the area is calculated through an image analysis technology, so that the detection precision is remarkably improved; dependence on various sensors is reduced, image acquisition equipment and an algorithm are used for replacement, and hardware and maintenance cost is reduced; image data are processed in real time, the detection period is shortened, and quick response is achieved; the combination of Gaussian GMM clustering and adaptive threshold segmentation can effectively deal with the condition that the wall surface of the air conditioner pipeline is complex, and the GMM can distinguish interference factors from mildew spots through modeling of pixel distribution; and adaptive threshold segmentation can remove interference according to local features, so that the recognition result is more stable and reliable.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent image detection and air-conditioning duct maintenance, and specifically to a method for identifying mold spots on air-conditioning ducts by integrating GMM and adaptive threshold segmentation. Background Art

[0002] Currently, in the field of mildew detection on the walls of air-conditioning ducts, existing technologies mainly use sensor-based detection methods and manual detection methods.

[0003] Sensor-based detection solutions deploy multiple sensors within pipelines, such as those for humidity, temperature, and gas, to collect real-time environmental parameters. When these data trigger preset thresholds, the system issues a warning about mold risk. However, this method only provides indirect assessments based on environmental parameters and cannot precisely quantify the extent of mold.

[0004] Manual inspection relies on regular inspections by professionals, who use visual observation or simple tools to identify mold spots. This method is not only inefficient and costly, but the test results are also significantly affected by subjective factors.

[0005] In addition, there are many problems that need to be solved:

[0006] First, detection accuracy is limited. Sensor-based detection can only indirectly determine the presence of mold through environmental parameters, but cannot accurately identify the area of ​​mold. Manual detection is difficult to provide accurate detection results due to the limitations of visual observation and the lack of quantitative standards.

[0007] Secondly, the high cost of installing multiple sensors and performing regular maintenance, coupled with the labor cost of manual inspections, results in high overall inspection costs.

[0008] Furthermore, data processing is delayed. The data collected by the sensor needs to be transmitted to the control end for analysis, which is a time-consuming process and makes it difficult to detect mold problems in a timely manner.

[0009] Furthermore, the level of intelligence is low. Existing systems mostly rely on threshold-triggered warnings, lacking in-depth data analysis and autonomous learning capabilities. Finally, environmental adaptability is poor. Due to the significant differences in operating environments for different air conditioners, existing technologies are unable to dynamically adjust detection strategies based on the actual environment, significantly reducing the reliability of detection results in complex environments.

[0010] Therefore, a new solution to the above problems needs to be proposed. Summary of the Invention

[0011] The purpose of the present invention is to provide a method for identifying mold spots on air-conditioning ducts by integrating GMM and adaptive threshold segmentation, so as to solve the technical problems raised in the background technology.

[0012] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying mold spots on air conditioning ducts by integrating GMM and adaptive threshold segmentation, comprising at least the following steps:

[0013] S1: Use a 4K high-resolution camera or an autofocus HD camera, illuminated by auxiliary lighting equipment, to capture high-quality images of mold spots on the walls of air conditioning ducts;

[0014] S2: Preprocessing the collected high-quality images of mildew on the walls of air-conditioning ducts to construct a database of typical characteristics of mildew areas on the walls of air-conditioning ducts, wherein the preprocessing includes at least grayscale conversion, filtering and denoising, and histogram equalization;

[0015] S3: performing feature extraction and edge detection on the surface mold spot image, wherein the feature extraction and edge detection at least include color feature extraction, texture feature extraction, and shape feature extraction;

[0016] S4: Integrate Gaussian GMM clustering and adaptive threshold segmentation technology, integrate the deep learning U-Net algorithm to segment the mold area, and combine it with a multi-dimensional mold feature extraction method to build an automatic mold area recognition model for the air conditioning duct wall. The multi-dimensional features include at least color, texture, and shape.

[0017] S5: Optimize and train the automatic recognition model for the area of ​​mildew on the walls of air-conditioning ducts. Use evaluation indicators such as accuracy, recall rate, and F1 value to verify the reliability of the optimized automatic recognition model. Based on the verification results, improve the automatic recognition model for the area of ​​mildew on the walls of air-conditioning ducts. Add the improved model to a database of typical features of mildew on duct walls to form an automatic recognition system for the area of ​​mildew on duct walls. Improving the automatic recognition model for the area of ​​mildew on the walls of air-conditioning ducts includes at least model fine-tuning and parameter optimization.

[0018] S6: Obtain an image of the mildew on the wall of the air-conditioning duct to be processed, input the image of the area of ​​the mildew on the wall of the duct to be processed into an established automatic recognition system for the area of ​​mildew on the wall of the duct, and obtain a recognition result of the input image, wherein the recognition result is specific pollution information, and the pollution information includes at least the area and pollution degree of the mildew on the wall of the duct in the input image.

[0019] Furthermore, the optimization training of the automatic recognition model for the area of ​​mildew on the wall of the air-conditioning duct includes at least the following steps:

[0020] Collect and integrate image information of pipeline wall mold area to form a pipeline wall mold area training dataset;

[0021] An image feature analysis algorithm is used to classify the color, texture, and shape features of the pipe wall mold area in the training images, and multiple groups of pipe wall mold area training images with different feature combinations are obtained.

[0022] Based on the morphological characteristics of mildew on the pipeline wall in the training image, an optimization algorithm is used to conduct targeted training on the basic image recognition model to form an automatic recognition system for the area of ​​mildew on the pipeline wall. The optimization algorithm includes at least gradient descent and back propagation.

[0023] Furthermore, the color feature extraction is to convert the image from RGB color space to HSV color space, calculate statistics, analyze the characteristic distribution of the mold spots in different color spaces, describe the color characteristics of the mold spots, and distinguish the mold spots from the background. The statistics include at least color mean, variance and histogram;

[0024] The texture feature extraction is to use the gray level co-occurrence matrix (GLCM) to calculate parameters, and adopt the local binary pattern (LBP) statistical coding histogram to capture the texture information such as the roughness and granularity of the mold surface. The parameters include at least energy, entropy, contrast and correlation.

[0025] The shape feature extraction is to first extract the mold spot contour through the Canny edge detection algorithm and the contour tracking algorithm, and then calculate the shape description to characterize the appearance characteristics of the mold spot. The shape description at least includes the perimeter, area, circularity and rectangularity.

[0026] Furthermore, the construction of the automatic identification model for the area of ​​mildew on the wall of the air-conditioning duct comprises at least the following steps:

[0027] Convert color images into grayscale images based on the RGB to HSV color space conversion algorithm;

[0028] Adopting automatic threshold selection algorithms based on image grayscale distribution, such as the OTSU method, to automatically determine the segmentation threshold that adapts to different image grayscale distributions;

[0029] Use a rectangular kernel structure element of size 3×3 to perform morphological processing of dilation and erosion on the image;

[0030] Canny edge detection, binarization and pollutant red marking operations are performed in sequence to make the marking of pollutant areas more complete and accurate;

[0031] When introducing a deep learning algorithm, pre-configure a basic model. The configuration of the basic model includes selecting U-Net as the preset model, setting the test data type to image, saving the prediction results in a path, setting the input image size to 640×640 pixels, setting the confidence threshold to 0.80 by default, and setting the result display settings to improve the training speed of the basic model and the recognition speed of the automatic recognition model for mold area on pipeline walls;

[0032] Based on the computer hardware parameters, the input image resolution is set to 640×640, and the confidence threshold is set to 0.80 by default;

[0033] The number of training rounds is set to 200, and the pre-set basic model is trained to ensure that the training results reach a sufficient confidence level, thereby forming a reliable automatic recognition model for mold spots on pipeline walls.

[0034] Furthermore, the segmentation of the mold spot area includes at least the following steps:

[0035] Adaptive threshold method is used to automatically determine the segmentation threshold according to the local features of the image using algorithms such as Otsu method, which can adapt to the segmentation of mold spots under different lighting and background conditions. Compared with the fixed threshold method, it is more flexible and accurate.

[0036] The Gaussian mixture model (GMM) is used. It is assumed that the image pixels are composed of a mixture of multiple Gaussian distributions. The distribution parameters are estimated through the expectation maximization (EM) algorithm, and the pixels are clustered to separate the mold spots from the background. It is suitable for the segmentation of mold spots in complex backgrounds.

[0037] The deep learning U-Net is used, and the pre-trained U-Net convolutional neural network model is used. Through training with a large number of labeled images, the characteristics and boundaries of mold spots are automatically learned, achieving high-precision mold spot area segmentation, which is especially suitable for complex and changeable mold spot images.

[0038] Furthermore, obtaining the area of ​​the mold spots on the pipe wall includes at least the following steps:

[0039] For a single large area of ​​mildew, the segmented mildew image was converted into a binary image using the binary image method, and the number of pixels with a value of 1 (representing mildew) was counted to obtain the pixel area of ​​the mildew.

[0040] For multiple unconnected mildew areas, the connected region marking method is used. A connected region marking algorithm such as the seed filling algorithm is used to mark each area. The number of pixels is counted and accumulated to accurately calculate the total mildew pixel area. Combined with the image scale, the actual area is converted to the actual area according to the formula "actual surface = (number of marked pixels / total number of pixels) × actual area" in cm. 2 , and output the result.

[0041] Furthermore, the pollution degree is determined based on a preset three-level threshold value:

[0042] Mild warning: the pollutant coverage area accounts for 10%-30% of the air outlet area;

[0043] Moderate warning: pollutant coverage area accounts for 31%-60% of the air outlet area;

[0044] Severe warning: the pollutant coverage area accounts for more than 60% of the air outlet area.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] The present invention uses image analysis technology to accurately extract mildew features and calculate area, significantly improving detection accuracy; reduces dependence on multiple sensors, replacing them with image acquisition equipment and algorithms, reducing hardware and maintenance costs; processes image data in real time, shortens the detection cycle, and achieves rapid response; the combination of Gaussian GMM clustering and adaptive threshold segmentation can effectively cope with the complex situation of the air-conditioning duct wall. GMM can distinguish interference factors from mildew by modeling pixel distribution; adaptive threshold segmentation can remove interference based on local features, making the recognition results more stable and reliable.

[0047] Combined with the deep learning U-net algorithm, the system is given the ability to learn independently and improve its intelligence level. At the same time, it can optimize the detection model according to image data in different environments, enhance environmental adaptability, and support an early warning mechanism with custom thresholds, which can flexibly respond to mold problems of varying degrees. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0049] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0050] Figure 2 This is a schematic diagram of the process for identifying the area of ​​mildew on the wall of an air-conditioning duct according to the present invention;

[0051] Figure 3 is a schematic diagram of the pollution degree of the present invention;

[0052] Figure 4 This is a flow chart of image feature extraction and edge detection in the present invention. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0054] Example 1:

[0055] See also Figure 1-Figure 4 The method for identifying mold spots on air-conditioning ducts by integrating GMM and adaptive threshold segmentation includes at least the following steps:

[0056] S1: Use a 4K high-resolution camera or an autofocus HD camera, illuminated by auxiliary lighting equipment, to capture high-quality images of mold spots on the walls of air-conditioning ducts;

[0057] S2: Preprocessing the collected high-quality images of mildew on the walls of air-conditioning ducts to construct a database of typical characteristics of mildew areas on the walls of air-conditioning ducts, wherein the preprocessing includes at least grayscale conversion, filtering and denoising, and histogram equalization;

[0058] S3: performing feature extraction and edge detection on the surface mold spot image, wherein the feature extraction and edge detection at least include color feature extraction, texture feature extraction, and shape feature extraction;

[0059] S4: Integrate Gaussian GMM clustering and adaptive threshold segmentation technology, and integrate the deep learning U-Net algorithm to segment the mold area. Combined with the multi-dimensional mold feature extraction method, an automatic recognition model for the mold area on the wall of the air conditioning duct is constructed. The multi-dimensionality includes at least color, texture and shape;

[0060] S5: Optimize and train the automatic recognition model for the area of ​​mildew on the walls of air-conditioning ducts. Use evaluation indicators such as accuracy, recall rate, and F1 value to verify the reliability of the optimized automatic recognition model. Based on the verification results, improve the automatic recognition model for the area of ​​mildew on the walls of air-conditioning ducts. Add the improved model to the database of typical characteristics of mildew on duct walls to form an automatic recognition system for the area of ​​mildew on duct walls. Improving the automatic recognition model for the area of ​​mildew on the walls of air-conditioning ducts includes at least model fine-tuning and parameter optimization.

[0061] S6: Obtain an image of the mildew on the wall of the air-conditioning duct to be processed, input the image of the area of ​​the mildew on the wall of the duct to be processed into an established automatic recognition system for the area of ​​mildew on the wall of the duct, and obtain a recognition result of the input image, where the recognition result is specific pollution information, and the pollution information includes at least the area and pollution degree of the mildew on the wall of the duct in the input image.

[0062] The optimization training of the automatic recognition model for the mold area on the wall of the air-conditioning duct includes at least the following steps:

[0063] Collect and integrate image information of pipeline wall mold area to form a pipeline wall mold area training dataset;

[0064] An image feature analysis algorithm is used to classify the color, texture, and shape features of the pipe wall mold area in the training images, and multiple groups of pipe wall mold area training images with different feature combinations are obtained.

[0065] Based on the morphological characteristics of mildew on the pipe wall in the training image, an optimization algorithm is used to conduct targeted training on the basic image recognition model to form an automatic recognition system for the area of ​​mildew on the pipe wall. The optimization algorithm includes at least gradient descent and back propagation.

[0066] Color feature extraction involves converting the image from RGB color space to HSV color space, calculating statistics, analyzing the characteristic distribution of mold spots in different color spaces, and describing the color characteristics of mold spots to distinguish them from the background. The statistics include at least color mean, variance, and histogram.

[0067] Texture feature extraction uses gray-level co-occurrence matrix (GLCM) to calculate parameters and adopts local binary pattern (LBP) statistical coding histogram to capture texture information such as roughness and granularity of the mold surface. The parameters include at least energy, entropy, contrast and correlation.

[0068] Shape feature extraction is to first extract the mold spot contour through the Canny edge detection algorithm and contour tracking algorithm, and then calculate the shape description to characterize the appearance characteristics of the mold spot. The shape description at least includes perimeter, area, circularity and rectangularity.

[0069] Building an automatic identification model for the area of ​​mold spots on the wall of air-conditioning ducts includes at least the following steps:

[0070] Convert color images into grayscale images based on the RGB to HSV color space conversion algorithm;

[0071] Adopting automatic threshold selection algorithms based on image grayscale distribution, such as the OTSU method, to automatically determine the segmentation threshold that adapts to different image grayscale distributions;

[0072] Use a rectangular kernel structure element of size 3×3 to perform morphological processing of dilation and erosion on the image;

[0073] Canny edge detection, binarization and pollutant red marking operations are performed in sequence to make the marking of pollutant areas more complete and accurate;

[0074] When introducing a deep learning algorithm, pre-configure the basic model. This includes selecting U-Net as the default model, setting the test data type to image, saving the prediction results to a path, setting the input image size to 640×640 pixels, setting the confidence threshold to 0.80 by default, and setting the result display settings. This speeds up basic model training and the recognition speed of the automatic identification model for mold stains on pipe walls.

[0075] Based on the computer hardware parameters, the input image resolution is set to 640×640, and the confidence threshold is set to 0.80 by default;

[0076] The number of training rounds is set to 200, and the pre-set basic model is trained to ensure that the training results reach a sufficient confidence level, thereby forming a reliable automatic recognition model for mold spots on pipeline walls.

[0077] The segmentation of the mold spot area includes at least the following steps:

[0078] Adaptive threshold method is used to automatically determine the segmentation threshold according to the local features of the image using algorithms such as Otsu method, which can adapt to the segmentation of mold spots under different lighting and background conditions. Compared with the fixed threshold method, it is more flexible and accurate.

[0079] The Gaussian mixture model (GMM) is used. It is assumed that the image pixels are composed of a mixture of multiple Gaussian distributions. The distribution parameters are estimated through the expectation maximization (EM) algorithm, and the pixels are clustered to separate the mold spots from the background. It is suitable for the segmentation of mold spots in complex backgrounds.

[0080] The deep learning U-Net is used, and the pre-trained U-Net convolutional neural network model is used. Through training with a large number of labeled images, the characteristics and boundaries of mold spots are automatically learned, achieving high-precision mold spot area segmentation, which is especially suitable for complex and changeable mold spot images.

[0081] Obtaining the area of ​​mold spots on the pipe wall includes at least the following steps:

[0082] For a single large area of ​​mildew, the segmented mildew image was converted into a binary image using the binary image method, and the number of pixels with a value of 1 (representing mildew) was counted to obtain the pixel area of ​​the mildew.

[0083] For multiple unconnected mildew areas, the connected region marking method is used. A connected region marking algorithm such as the seed filling algorithm is used to mark each area. The number of pixels is counted and accumulated to accurately calculate the total mildew pixel area. Combined with the image scale, the actual area is converted to the actual area according to the formula "actual area = (number of marked pixels / total number of pixels) × actual area" in cm. 2 , and output the result.

[0084] The degree of contamination is determined based on a three-level preset threshold:

[0085] Mild warning: the pollutant coverage area accounts for 10%-30% of the air outlet area;

[0086] Moderate warning: pollutant coverage area accounts for 31%-60% of the air outlet area;

[0087] Severe warning: the pollutant coverage area accounts for more than 60% of the air outlet area.

[0088] In Example 1, by further combining another deep learning U-Net algorithm with the recognition process of the automatic recognition system for mold spot area on the pipeline wall, the deep learning U-Net algorithm is used to model the variation characteristics of component dust pollution, thereby forming a component dust pollution risk prediction model;

[0089] In embodiment 1, an intelligent replacement warning system can be introduced and developed in the automatic identification system for the area of ​​mildew on the pipe wall. The dust pollution risk of components can be monitored and analyzed in real time through a prediction model. When the dust pollution condition of the air-conditioning system components reaches a certain threshold, an early warning is issued through an indicator light (green light, yellow light, red light) or a text message notification to remind the user to take measures to clean the air-conditioning pipes.

[0090] Based on the above embodiment 1, a specific application is proposed, which constitutes embodiment 2:

[0091] First, image data from various air conditioning duct surfaces was collected on-site. Using 4K high-resolution webcams, mounted at key locations on the duct walls, the images were captured from multiple angles and under various environmental conditions, encompassing varying degrees of mildew (mild, moderate, and severe). After acquisition, pre-processing techniques such as image enhancement and noise reduction were applied to enhance image quality. The system integrates image processing algorithms implemented in programming languages ​​like Python, encompassing a variety of functions including grayscale processing, binarization, morphological operations, edge detection, and area calculation.

[0092] Subsequently, a combination of manual and semi-automated annotation was used to mark the mold spots in the image and extract typical features such as color, texture, and shape. The processed image information and corresponding feature parameters were integrated to establish a typical feature database for mold spots on the walls of air conditioning ducts. The most representative features from the extracted color and texture features were selected for subsequent classification and segmentation.

[0093] Secondly, image preprocessing uses the weighted averaging method (Gray = 0.299R + 0.587G + 0.114B) to convert the color image into a grayscale image, remove color interference, simplify the data dimension, and improve the efficiency of subsequent processing. Then, adaptive binarization based on the OTSU algorithm is used to automatically calculate the optimal threshold and convert the grayscale image into a binary image. By traversing all possible thresholds and calculating the inter-class variance, the threshold with the largest inter-class variance is selected as the segmentation point. Pixels below the threshold are set to white (background), and pixels above the threshold are set to black (potential mold area). This allows the algorithm to adapt to different lighting conditions and automatically distinguish mold from the background. Morphological operations are then performed on the binary image using a 3×3 rectangular structuring element. Each pixel is dilated twice, and if there is at least one white pixel within the area covered by the structuring element, the target pixel is set to white. This fills small holes within the mold spot, connects adjacent scattered patches, and enhances regional integrity. The dilated image is then eroded once using the same structuring element. The target pixel remains white only when the area covered by the structuring element is entirely white; otherwise, it is set to black. This removes edge burrs and unnecessary protrusions generated by the dilation and restores the true boundary details of the mold spot. Edge detection and region fusion are performed by first applying a 5×5 Gaussian filter to the grayscale image to remove noise. The gradient magnitude and direction are then calculated. Edges are refined using non-maximum suppression, while retaining local maxima in the gradient direction. Edges are filtered using a dual threshold (low threshold 50, high threshold 150). Pixels above the low threshold and connected to the high threshold edge are connected to obtain the complete mold spot edge contour. The binary image is then fused with the Canny edge image using a bitwise OR operation, integrating the complete area marked by the morphological operation with the detailed contours supplemented by edge detection to form a precise mold spot marking area. After extracting the mold features, a region-segmented Gaussian mixture model (GMM) is used to classify the pixels in the image using a probabilistic model. When identifying mold on the walls of air conditioning ducts, since the color, texture, and other characteristics of the mold may vary, the GMM can accurately separate mold and non-mold areas by learning the distribution of these variations, thereby improving recognition accuracy. Adaptive threshold segmentation automatically adjusts the threshold based on the local features of the image, adapting to varying lighting conditions and changes in the duct wall background. Inside air conditioning ducts, lighting can be uneven, and the wall color and reflectivity may vary at different locations. Adaptive threshold segmentation calculates an appropriate threshold for each local region, allowing the mold area to be more clearly separated from the background, further improving recognition accuracy.

[0094] Based on the basic model, new images of the air conditioning duct wall are continuously collected as images to be processed, input into the model for recognition, and the recognition results are obtained. A certain proportion of samples are randomly extracted from the recognition results, and verification analysis is carried out through manual review or comparison with the test results of high-precision detection equipment. For recognition errors or deviations, the causes are deeply analyzed, the model parameters are reversely optimized, and the image feature extraction algorithm and edge detection strategy are adjusted to eliminate the influence of uncertain factors such as lighting changes and differences in the pipe wall material. Through multiple iterations, the accuracy of the mold spot automatic recognition results is improved, the reliability of the model is verified, and an optimized mold spot automatic recognition model is formed. Based on the optimized automatic recognition model, combined with the powerful feature extraction capabilities of the convolutional neural network U-Net, the changing characteristics of the mold spot on the air conditioning duct wall are modeled. Through multiple convolutional layers and pooling layers, the deep features of the mold spot at different growth stages are extracted.

[0095] Finally, the area of ​​the mold spot is calculated based on the camera calibration parameters. For example, if it is known that the pre-processed 420×420 pixel image corresponds to an actual area of ​​1.7cm×1.7cm, the number of pixels in the marked area is converted into the actual area through a conversion formula. The mold spot pollution classification is based on a preset three-level warning threshold: mild warning (pollutant coverage area accounts for 10%-30% of the air outlet area), moderate warning (pollutant coverage area accounts for 31%-60% of the air outlet area), and high warning (pollutant coverage area accounts for >60% of the air outlet area), and supports user-defined adjustments; when the detection area exceeds the corresponding threshold, a real-time alarm is issued through an indicator light (green / yellow / red), SMS notification, or APP push notification, reminding the user that the pipe wall needs to be cleaned and maintained.

[0096] In summary:

[0097] This invention addresses existing issues such as low recognition efficiency due to reliance on manual debugging, the difficulty of fixed-threshold algorithms adapting to complex lighting and background variations, poor recognition accuracy due to incomplete capture of mold features by a single algorithm, high hardware deployment costs and a lack of versatility, and a lack of intelligent analysis and graded warning mechanisms. By combining the modeling capabilities of Gaussian GMM clustering for complex feature distributions with the dynamic adjustment advantages of adaptive threshold segmentation, this invention achieves high-precision mold spot recognition while also offering strong environmental adaptability, full-process automation, and multi-dimensional risk assessment, significantly enhancing the intelligent level and economic benefits of air conditioning duct maintenance.

[0098] The present invention adopts a fusion method of adaptive threshold Otsu method (OTSU) and Gaussian GMM clustering. Otsu method binarizes the image by automatically calculating the optimal threshold value, and can dynamically adjust the segmentation threshold according to the grayscale distribution of the image, thereby overcoming the defect that the traditional fixed threshold method is difficult to adapt to different lighting and changes in mold characteristics; Canny edge detection further accurately extracts the edge contour of the mold spot, and the combination of the two provides accurate basic data for subsequent Gaussian GMM clustering. Gaussian GMM clustering can capture the complex color, texture and other characteristic distributions of mold spots by probabilistic modeling of pixel distribution, and distinguish the mold spots from the background more accurately. In practical applications, the traditional fixed threshold segmentation method often suffers from over-segmentation or under-segmentation in scenes with uneven lighting and various colors of mold on the walls of air-conditioning ducts, resulting in errors in edge contour extraction, which in turn affects the accuracy of area calculation. The combination of the adaptive threshold Otsu method and Canny edge detection in the present invention improves the accuracy of edge contour extraction by more than 30% for pipeline wall images with different lighting and materials; on this basis, Gaussian GMM clustering further refines the classification, and ultimately achieves an area calculation error of less than 5%, significantly improving the accuracy of mold spot recognition.

[0099] First, the color image is converted into a grayscale image through grayscale processing to remove the interference caused by color information, so that the algorithm focuses on grayscale features for processing; morphological operations (dilation and erosion) are used to optimize the binarized image, fill the small holes inside the mildew spots, connect the broken edges, and suppress the influence of noise and other stains on the pipe wall. These preprocessing steps provide clean and accurate image data for subsequent Gaussian GMM clustering and adaptive threshold segmentation, so that the algorithm can run stably in complex environments. The internal environment of air-conditioning ducts is complex, and there are problems such as metal reflection, dust adhesion, and pipe texture interference. The existing technology is prone to false detection and missed detection in such an environment. The present invention effectively filters out irrelevant interference factors through grayscale and morphological operations, providing stable input for the core algorithm. After actual testing, in a complex pipeline environment, the missed detection rate and false detection rate of the present invention are reduced to less than 1%. Compared with the existing technology, the adaptability and reliability of the algorithm in complex environments are greatly improved.

[0100] The present invention has constructed a complete set of automated processing procedures, starting from the automatic acquisition of air-conditioning duct wall images by high-definition cameras, and sequentially undergoing image preprocessing (grayscale, denoising, morphological operations), mildew feature extraction (adaptive threshold Otsu method, Gaussian GMM clustering), area calculation to the final graded warning output, and the entire process does not require human intervention. The system adopts efficient algorithms and optimized program architecture, which can quickly process image data and achieve high-frequency real-time monitoring. The existing technology often relies on manual operation or semi-automatic processing, which is not only inefficient, but also difficult to meet the maintenance requirements of 24-hour uninterrupted operation of air-conditioning systems. After the present invention realizes full-process automation, it supports all-weather real-time monitoring, and the detection frequency can reach 1 time / minute. It can timely detect the generation and development of mildew, provide strong support for the intelligent maintenance of air-conditioning systems, and greatly improve maintenance efficiency and timeliness.

[0101] The present invention can not only identify the presence of mildew, but also classify the degree of mildew contamination (mild, moderate, severe) through precise area calculation and combined with a pre-set area threshold. Based on different pollution levels, the system gives corresponding early warning information and maintenance recommendations, providing a quantitative basis for the formulation of rapid and accurate maintenance strategies for air-conditioning ducts. The existing technology can usually only open the duct and simply judge whether mildew exists with the naked eye. It cannot provide a scientific basis for maintenance decisions and can easily lead to excessive maintenance (causing waste of resources) or insufficient maintenance (affecting the performance of the air-conditioning system). The present invention uses machine learning and image recognition to intelligently evaluate the risk of mildew in air-conditioning ducts in multiple dimensions, helping operation and maintenance personnel to formulate reasonable cleaning and maintenance plans for air-conditioning ducts and avoid unnecessary maintenance costs. According to actual case calculations, the operation and maintenance costs can be reduced by more than 20%, while ensuring the healthy and stable operation of the air-conditioning system.

[0102] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A method for identifying mold spots on air conditioning ducts that integrates GMM and adaptive threshold segmentation, characterized by: At least the following steps are included: S1: Use a 4K high-resolution camera or an autofocus HD camera, illuminated by auxiliary lighting equipment, to capture high-quality images of mold spots on the walls of air conditioning ducts; S2: Preprocessing the collected high-quality images of mildew on the walls of air-conditioning ducts to construct a database of typical characteristics of mildew areas on the walls of air-conditioning ducts, wherein the preprocessing includes at least grayscale conversion, filtering and denoising, and histogram equalization; S3: performing feature extraction and edge detection on the mold spot image, wherein the feature extraction and edge detection at least include color feature extraction, texture feature extraction, and shape feature extraction; S4: Integrate Gaussian GMM clustering and adaptive threshold segmentation technology, integrate the deep learning U-Net algorithm to segment the mold area, and combine it with a multi-dimensional mold feature extraction method to build an automatic mold area recognition model for the air conditioning duct wall. The multi-dimensional features include at least color, texture, and shape. S5: Optimizing and training the automatic recognition model for the area of ​​mildew on the walls of air-conditioning ducts, using evaluation indicators to verify the reliability of the optimized automatic recognition model, and based on the verification results, improving the automatic recognition model for the area of ​​mildew on the walls of air-conditioning ducts, adding the improved model to a database of typical features of mildew on duct walls to form an automatic recognition system for the area of ​​mildew on duct walls, wherein improving the automatic recognition model for the area of ​​mildew on the walls of air-conditioning ducts includes at least model fine-tuning and parameter optimization. S6: Obtain an image of the mildew on the wall of the air-conditioning duct to be processed, input the image of the area of ​​the mildew on the wall of the duct to be processed into an established automatic recognition system for the area of ​​mildew on the wall of the duct, and obtain a recognition result of the input image, wherein the recognition result is specific pollution information, and the pollution information includes at least the area and pollution degree of the mildew on the wall of the duct in the input image.

2. The method for identifying mold spots on air conditioning ducts by integrating GMM and adaptive threshold segmentation according to claim 1 is characterized by: The optimization training of the automatic recognition model for the area of ​​mildew on the wall of the air-conditioning duct comprises at least the following steps: Collect and integrate image information of pipeline wall mold area to form a pipeline wall mold area training dataset; An image feature analysis algorithm is used to classify the color, texture, and shape features of the pipe wall mold area in the training images, and multiple groups of pipe wall mold area training images with different feature combinations are obtained. Based on the morphological characteristics of mildew on the pipeline wall in the training image, an optimization algorithm is used to conduct targeted training on the basic image recognition model to form an automatic recognition system for the area of ​​mildew on the pipeline wall. The optimization algorithm includes at least gradient descent and back propagation.

3. The method for identifying mold spots on air conditioning ducts by integrating GMM and adaptive threshold segmentation according to claim 1 is characterized by: The color feature extraction is to convert the image from RGB color space to HSV color space, calculate statistics, analyze the characteristic distribution of mold spots in different color spaces, describe the color characteristics of mold spots, and distinguish between mold spots and background. The statistics include at least color mean, variance and histogram. The texture feature extraction is to use the gray level co-occurrence matrix to calculate parameters, and at the same time adopt the local binary pattern statistical coding histogram to capture the texture information such as the roughness and granularity of the mold surface. The parameters include at least energy, entropy, contrast and correlation. The shape feature extraction is to first extract the mold spot contour through the Canny edge detection algorithm and the contour tracking algorithm, and then calculate the shape description to characterize the appearance characteristics of the mold spot. The shape description at least includes the perimeter, area, circularity and rectangularity.

4. The method for identifying mold spots on air conditioning ducts by integrating GMM and adaptive threshold segmentation according to claim 1 is characterized by: The construction of the automatic identification model for the area of ​​mildew on the wall of the air-conditioning duct comprises at least the following steps: Convert color images into grayscale images based on the RGB to HSV color space conversion algorithm; Adopting the automatic threshold selection algorithm based on the grayscale distribution of the image to automatically determine the segmentation threshold that adapts to different image grayscale distribution conditions; Use a rectangular kernel structure element of size 3×3 to perform morphological processing of dilation and erosion on the image; Canny edge detection, binarization and pollutant red marking operations are performed in sequence to make the marking of pollutant areas more complete and accurate; When introducing a deep learning algorithm, pre-configure the basic model. This includes selecting U-Net as the default model, setting the test data type to image, the prediction result save path, setting the input image size to 640×640 pixels, setting the confidence threshold to 0.80 by default, and setting the result display. Based on the computer hardware parameters, the input image resolution is set to 640×640, and the confidence threshold is set to 0.80 by default; The number of training rounds is set to 200, and the pre-set basic model is trained to ensure that the training results reach a sufficient confidence level, thereby forming a reliable automatic recognition model for mold spots on pipeline walls.

5. The method for identifying mold spots on air conditioning ducts by integrating GMM and adaptive threshold segmentation according to claim 4 is characterized by: The mold spot area segmentation comprises at least the following steps: Adopting the adaptive threshold method, the algorithm automatically determines the segmentation threshold according to the local features of the image, and adapts to the mold spot segmentation under different lighting and background conditions; The Gaussian mixture model (GMM) is used. It is assumed that the image pixels are a mixture of multiple Gaussian distributions. The distribution parameters are estimated through the expectation maximization algorithm, and the pixels are clustered to separate the mold spots from the background. Using deep learning U-Net and pre-trained U-Net convolutional neural network model, the system automatically learns the characteristics and boundaries of mold spots through training with a large number of labeled images, achieving high-precision mold spot area segmentation.

6. The method for identifying mold spots on air conditioning ducts by integrating GMM and adaptive threshold segmentation according to claim 1 is characterized by: Obtaining the area of ​​mold spots on the pipe wall includes at least the following steps: For a single large area of ​​mildew, the segmented mildew image is converted into a binary image using the binary image method. The number of pixels with a value of 1 is counted, where 1 represents mildew, and the pixel area of ​​the mildew is obtained. For multiple unconnected mildew areas, the connected region marking method is used. Each area is marked using the connected region marking algorithm. The number of pixels is counted and accumulated to accurately calculate the total mildew pixel area. Combined with the image scale, the actual area is converted to the actual area according to the formula "actual surface = (number of marked pixels / total number of pixels) × actual area" in cm. 2 , and output the result.

7. The method for identifying mold spots on air conditioning ducts by integrating GMM and adaptive threshold segmentation according to claim 6 is characterized by: The degree of contamination is determined based on a three-level preset threshold for the degree of mold contamination: Mild warning: the pollutant coverage area accounts for 10%-30% of the air outlet area; Moderate warning: pollutant coverage area accounts for 31%-60% of the air outlet area; Severe warning: the pollutant coverage area accounts for more than 60% of the air outlet area.

Images