A visual detection method and system for early warning of icing environment

Abstract

The application discloses a visual detection method for early warning of icing environment, and relates to the technical field of aviation icing detection, comprising the following steps: constructing a two-dimensional joint feature residual space based on real-time contour vectors and real-time gray field vectors, generating a candidate region set, constructing a three-dimensional feature vector for maximum points in the candidate region set, clustering by using an affinity propagation clustering algorithm, and obtaining structured ice phase units; combining the affinity propagation clustering algorithm with an effective rank quantization method based on singular value entropy of a trajectory matrix to construct an ice layer analysis mechanism with multiple singular structure decoupling capabilities and continuous morphological phase state quantization capabilities; and constructing a spatial distribution sequence of internal deformation variables of each ice phase unit into a trajectory matrix, introducing singular value decomposition and Shannon entropy calculation to obtain an effective rank continuous scalar, and improving the perception granularity of the icing visual detection system on the spatial non-uniformity of the ice layer.

Description

Technical Field

[0001] This invention relates to the field of aviation icing detection technology, and in particular to a visual detection method and system for early warning of icing environments. Background Technology

[0002] With the increasing demand for aircraft flights under extreme weather conditions, icing environments pose an increasingly serious threat to flight safety. In recent years, visual perception-based icing detection methods have been gradually applied. These methods utilize image processing technology to acquire real-time images from the outside and extract icing features through edge detection, texture analysis, and other means. They have high sensitivity and spatial resolution, enabling refined analysis of icing morphology.

[0003] However, existing technologies have limitations: existing machine vision-based icing detection methods are limited by the holistic analysis paradigm and cannot effectively decouple and independently quantify overlapping and coexisting multiphase ice structures. The features extracted by existing technologies, such as area, mean thickness, and skewness, are all global aggregated descriptions. However, in the actual mixed phase or initial icing stage, multiple ice structures with completely different morphological features often coexist at the leading edge of the component. Large-area disc-shaped frost ice constitutes a stable low-frequency component, while discretely grown needle-shaped ice ridges exhibit sparse high-frequency peaks. When the two overlap, the global aggregation operation causes the intrinsic signal of the needle-shaped ice ridge to be submerged by the disc-shaped ice layer, resulting in statistical smoothing annihilation, which makes it impossible to independently identify and classify high-risk ice tumors. Furthermore, existing methods for quantifying ice-shaped spectra mostly employ binary criteria, lacking a continuous scalar framework to finely characterize the phase transition from disk-shaped uniformity to needle-shaped concentration, making it difficult to establish a quantitative mapping between detection output and aerodynamic risk. Summary of the Invention

[0004] In a first aspect, the present invention provides a visual detection method for early warning of icing environments, comprising: Real-time images of the pilot icing sensing component are acquired and preprocessed to extract real-time contour vectors and real-time grayscale field vectors. Based on real-time contour vectors and real-time grayscale field vectors, a two-dimensional joint feature residual space is constructed to generate a candidate region set. A three-dimensional feature vector is constructed for the maxima points in the candidate region set, and the affinity propagation clustering algorithm is used to cluster them to obtain structured ice phase units. Based on structured ice phase units, a trajectory matrix is ​​generated. The effective rank is obtained through singular value decomposition, Shannon entropy calculation, and exponential function. The structured ice phase units are then morphologically classified. A four-dimensional physical property vector is constructed for each structural ice phase unit, and the fusion ice uniformity index is calculated; Based on morphological classification results and the fusion of icing uniformity index, a comprehensive judgment is made on the icing environment, and a graded early warning is issued based on the judgment results. Based on the early warning level, physical action control is executed.

[0005] As a preferred embodiment of the visual detection method for early warning of icing environment described in this invention, the step of acquiring and preprocessing real-time images of the pilot icing sensing component to extract real-time contour vectors and real-time grayscale field vectors includes: Real-time images of the pilot icing sensing component are acquired and denoised using Gaussian filtering to obtain the pre-processed image; Subpixel-level Canny edge detection is performed on the preprocessed image to extract the contour of the leading edge of the icing sensing component and obtain the real-time contour. Using a template matching algorithm based on normalized cross-correlation, the real-time contour is rigidly registered with the reference contour, and the contour is parameterized in polar coordinates to generate a real-time contour vector. In the coordinate system after rigid registration, the ROI region is defined as the reference contour, which is extended by M pixels in both the inner and outer directions along the normal direction. The ROI region of the preprocessed image is sampled by gray level at each angle to generate a real-time gray field vector.

[0006] As a preferred embodiment of the visual detection method for early warning of icing environment described in this invention, the method involves: constructing a two-dimensional joint feature residual space based on real-time contour vectors and real-time grayscale field vectors, generating a candidate region set, constructing a three-dimensional feature vector for the maxima points in the candidate region set, and using an affinity propagation clustering algorithm to cluster the regions to obtain structured ice phase units, including: Based on the angular coordinate axis of the leading edge of the icing sensing component, the contour deformation residual field and the gray-scale symmetry residual field are calculated and weighted and stitched together to construct a two-dimensional joint feature residual space. The Suzuki85 contour edge tracking algorithm is applied to the two-dimensional joint feature residual space to generate a candidate region set. From the candidate region set, maxima are selected in ascending order of angular position, and a three-dimensional feature vector is constructed for each maxima point; The three-dimensional feature vector has three dimensions: first, contour shape variable; second, local texture entropy; and third, local gray-level skewness. Among them, the contour deformation variable refers to the contour deformation residual field at the maximum point, the local texture entropy is the information entropy at the maximum point, and the local gray level skewness is the skewness at the maximum point. The squared negative Euclidean distance between the three-dimensional feature vectors of two maxima points is used as a similarity metric to construct a similarity matrix; Initialize two matrices with the same dimensions as the similarity matrix, namely the suitable matrix and the usable matrix; In this context, the points in the suitability matrix represent the suitability of a maximum point w as a cluster center for a maximum point j. The points in the available matrix represent the degree to which a maximum point j can serve as a cluster center for a maximum point w. The two matrices are transmitted in opposite directions; Using the affinity propagation clustering algorithm, the appropriate and available matrices are iteratively updated until the maximum number of iterations is reached, decoupling all the maxima into structured glacial units.

[0007] As a preferred embodiment of the visual detection method for early warning of icing environments described in this invention, the step of generating a trajectory matrix based on structured ice phase units, obtaining the effective rank through singular value decomposition, Shannon entropy calculation, and exponential functions, and performing morphological classification of the structured ice phase units includes: The maximum points in the structured ice phase unit are sorted according to their angular positions, and the corresponding contour deformations are generated into column vectors. Perform an outer product operation between the column vector and a spatial location encoding vector to obtain a column matrix, and then perform column expansion to convert it into a trajectory matrix; The trajectory matrix is ​​decomposed by singular value decomposition to obtain a set of singular values. Each singular value is normalized, and the Shannon entropy of the normalized singular values ​​is calculated. Based on Shannon entropy, the effective rank is calculated using an exponential function, and then truncated and scaled to obtain the scaled rank. Map the scaling rank to a fixed interval; Based on prior data from fluid dynamics simulations and wind tunnel experiments, two fixed classification thresholds, including an upper threshold and a lower threshold, are set to determine the morphological classification of structural ice phase units.

[0008] As a preferred embodiment of the visual detection method for early warning of icing environment described in this invention, the step of constructing a four-dimensional physical property vector for each structural ice phase unit and calculating the fused icing uniformity index includes: A four-dimensional physical property vector is constructed for each structural ice phase unit, including the center angle coordinates, angle span, volume estimate, and effective rank; The angular span is calculated from the angle difference between the leftmost maximum point and the rightmost maximum point within the unit. The volume estimate is obtained by multiplying the mean of the profile deformation in each structural ice phase unit by its angular span. Based on the four-dimensional physical property vector, the risk factor of each structural ice phase unit is calculated; The risk factor is composed of a weighted average of morphological risk components, volumetric risk components, and location risk components. Among them, the morphological risk component is calculated from the effective rank using the Sigmoid function; The volume risk component is obtained by normalizing the volume estimate of the current structural ice phase unit relative to the global maximum volume in the same frame; The location risk component was calculated using a linear weighted method based on normalized distance. The risk factors of all structural ice phase units in each candidate region are summed to obtain the comprehensive risk factor of each candidate region. The comprehensive risk factors of all candidate regions are then summed to obtain the total global structural risk value. The fusion icing uniformity index is calculated based on the volume estimate, risk factors, and total global structural risk.

[0009] As a preferred embodiment of the visual detection method for early warning of icing environment described in this invention, the step of constructing a four-dimensional physical property vector for each structural ice phase unit and calculating the fused icing uniformity index includes: A four-dimensional physical property vector is constructed for each structural ice phase unit, including the center angle coordinates, angle span, volume estimate, and effective rank; The angular span is calculated from the angle difference between the leftmost maximum point and the rightmost maximum point within the unit. The volume estimate is obtained by multiplying the mean of the profile deformation in each structural ice phase unit by its angular span. Based on the four-dimensional physical property vector, the risk factor of each structural ice phase unit is calculated; The risk factor is composed of a weighted average of morphological risk components, volumetric risk components, and location risk components. Among them, the morphological risk component is calculated from the effective rank using the Sigmoid function; The volume risk component is obtained by normalizing the volume estimate of the current structural ice phase unit relative to the global maximum volume in the same frame; The location risk component was calculated using a linear weighted method based on normalized distance. The risk factors of all structural ice phase units in each candidate region are summed to obtain the comprehensive risk factor of each candidate region. The comprehensive risk factors of all candidate regions are then summed to obtain the total global structural risk value. The fusion icing uniformity index is calculated based on the volume estimate, risk factors, and total global structural risk.

[0010] As a preferred embodiment of the visual detection method for early warning of icing environments according to the present invention, the step of executing physical motion control based on the warning level includes: When a routine warning is issued, the standard de-icing mode will be executed. When an enhanced warning is issued, a zone-specific de-icing mode will be implemented. When an emergency warning is issued, the powerful de-icing mode will be activated. The standard de-icing mode refers to uniform heating throughout the entire system. The partitioned differentiated de-icing mode sends a global uniform heating command to the electrothermal anti-icing and de-icing system, with all heating elements operating at 100% of their rated power, while simultaneously sending an icing presence status marker to the flight control system and imposing no aerodynamic restrictions. The partitioned differential de-icing mode refers to increasing the power of the heating element in the corresponding angle range to 150% of the rated value according to the angular coordinate list of the transitional structure ice phase unit, while maintaining the standard heating power in the remaining uniform area, and sending the non-uniform icing state mark and the angle range of the non-uniform area to the flight control system at the same time. The powerful de-icing mode directs all heating elements to be heated at maximum power for emergency heating, while simultaneously transmitting the four-dimensional physical property vector of the needle-shaped ice phase unit and risk factors to the flight control system in real time for it to perform aerodynamic protection actions. During the de-icing process, the risk factors in the heated area are continuously monitored; If the risk factor in a certain region continues to decline, the current power level will be maintained. If the risk factor in a certain area increases instead of decreasing, or if the needle-like ice phase unit continues to exist at maximum power for more than the set time limit, an instantaneous overpower pulse will be triggered in that area. After all the needle-like units disappear, the heating power will be gradually restored to the standard mode. When the signs of icing disappear, the warning is lifted, a stop heating command is sent to the anti-icing system and the power outage is confirmed, a return to normal mode command is sent to the flight control system, all aerodynamic restrictions are canceled, the internal state of the system is reset, and it returns to the normal monitoring state.

[0011] Secondly, this invention provides a visual detection system for early warning of icing environments, comprising, The acquisition and extraction module is used to acquire and preprocess real-time images of the pilot icing sensing component, and extract real-time contour vectors and real-time grayscale field vectors. A clustering module is constructed to build a two-dimensional joint feature residual space based on real-time contour vectors and real-time grayscale field vectors, generate a candidate region set, construct a three-dimensional feature vector for the maxima points in the candidate region set, and use the affinity propagation clustering algorithm to cluster them to obtain structured ice phase units. The decomposition and classification module is used to generate trajectory matrices based on structured ice phase units, obtain effective rank through singular value decomposition, Shannon entropy calculation and exponential function, and perform morphological classification of structured ice phase units. The computational fusion module is used to construct a four-dimensional physical property vector for each structural ice phase unit and calculate the fusion ice uniformity index. The judgment execution module is used to make a comprehensive judgment on the icing environment based on morphological classification results and fused icing uniformity index, and to issue graded warnings based on the judgment results. Based on the warning level, physical action control is executed.

[0012] The beneficial effects of this invention are: By combining affinity propagation clustering algorithm with efficient rank quantization method based on singular value entropy of trajectory matrix, an ice layer analysis mechanism with decoupling capability for multiple singular structures and quantization capability for continuous morphological phase states is constructed. Specifically: This invention takes the joint feature residual space of the contour deformation field and the gray-scale symmetry field as input, and utilizes the adaptive classification characteristics of affinity propagation clustering under the condition of no prior number of categories. It decouples the ice layer region, which is traditionally regarded as a single whole for statistical description, into several physically independent structured ice phase units. This allows local high-risk ice cores such as needle-shaped ice ridges, which were originally submerged by the stable component of the disk-shaped uniform ice layer in the calculation of global mean or higher-order moments, to be independently identified. Based on this, by constructing the spatial distribution sequence of deformation variables within each ice phase unit into a trajectory matrix, and by introducing singular value decomposition and Shannon entropy calculation to obtain the effective rank as a continuous scalar, a detailed characterization of the complete phase spectrum of ice layers from disk-shaped uniform distribution to needle-shaped concentrated distribution is achieved. This overcomes the shortcomings of existing technologies that rely on binary criteria for coarse classification, and significantly improves the perception granularity of ice layer spatial non-uniformity of the icing visual inspection system as well as the quantitative mapping accuracy between detection output and aerodynamic risk assessment. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart of the visual detection method for early warning of icing environment in Example 1; Figure 2 This is a structural diagram of the visual detection system for early warning of icing environment in Example 1. Detailed Implementation

[0015] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0016] Example 1, referring to Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a visual detection method for early warning of icing environments, including the following steps: S1. Acquire real-time images of the pilot icing sensing component and preprocess them to extract real-time contour vectors and real-time grayscale field vectors. Specifically, real-time images of the pilot icing sensing component are acquired and denoised using Gaussian filtering to obtain the pre-processed image; Subpixel-level Canny edge detection is performed on the preprocessed image to extract the contour of the leading edge of the icing sensing component and obtain the real-time contour. A template matching algorithm based on normalized cross-correlation is used to rigidly register the real-time profile with the reference profile to compensate for the translational and rotational offsets introduced by airborne vibration. The profile is then parameterized in polar coordinates to generate a real-time profile vector. The template matching algorithm based on normalized cross-correlation described above rigidly registers the real-time contour with the reference contour. If the peak value of the cross-correlation matching coefficient is lower than the matching threshold, it indicates that the contour is severely deformed due to icing, causing conventional registration to fail. The root region of the geometrically stable component of the first icing sensing component is selected, and the feature corner points in this region are extracted using the FAST algorithm. The homography matrix is ​​solved by feature matching, and the image is corrected by perspective transformation. The reference contour and geometric center are mapped to the corrected image coordinate system to eliminate the perspective distortion caused by airborne vibration and ensure the reliability of ROI positioning. The matching threshold is set based on the 3σ principle. The aforementioned baseline profile is obtained by generating an ice-free baseline image from an image acquired when there is no ice or contaminants on the surface of the pilot icing sensing component, through a multi-frame averaging noise reduction method, and then performing sub-pixel level Canny edge detection. The corresponding baseline grayscale field is also generated by sampling the grayscale mean value of the ice-free baseline image at each angle in the same polar coordinate system. The above polar coordinate parameterization refers to discretizing the contour into 360 equal-angle sampling points, with the geometric center of the leading edge surface of the component as the origin, and each angle corresponding to a radius value, to generate a real-time contour vector; In the coordinate system after rigid registration, the ROI region is defined as the reference contour, which is extended by M pixels in both the inner and outer directions along the normal direction. The ROI region of the preprocessed image is sampled by gray level from angle to angle to generate a real-time gray field vector, where M is set based on the sum of the maximum vibration residual and the detection error. The aforementioned angle-by-angle grayscale sampling refers to calculating the average grayscale value by taking a strip-shaped area along the normal direction.

[0017] S2. Based on real-time contour vectors and real-time grayscale field vectors, a two-dimensional joint feature residual space is constructed to generate a candidate region set. A three-dimensional feature vector is constructed for the maximum points in the candidate region set, and the affinity propagation clustering algorithm is used to cluster them to obtain structured ice phase units. Specifically, based on the angular coordinate axis of the leading edge of the icing sensing component, the contour deformation residual field and the gray-scale symmetry residual field are calculated and weighted and stitched together to construct a two-dimensional joint feature residual space; The aforementioned angular coordinate axes of the leading edge refer to the continuous arc surface from one end of the leading edge to the other end, with values ​​ranging from 0° to 180°. The above-mentioned contour deformation residual field, gray-level symmetry residual field, and two-dimensional joint feature residual space are represented by the following formula: ; ; ; in, and At angles The residual field of contour deformation and the residual field of gray-level symmetry at the location, and At angles The coordinates of the leading edge contour point and the reference contour point of the component are as follows: and At angles The average grayscale value within a 5-pixel-wide strip area on either side of the component's center line. Use small positive numbers to prevent division by zero. Let be a two-dimensional joint feature residual space, where Let be the coordinates of the two-dimensional joint feature residual space. In this space, angles are used as rows, and feature difference levels are used as columns, where the feature difference level values ​​are... The value, and These represent the weights of the profile deformation residual field and the gray-scale symmetry residual field, respectively. Because profile deformation is the most direct geometric representation of icing, it has a slightly higher weight. Gray-scale symmetry, being an auxiliary optical representation, has a slightly lower weight. Based on engineering experience, such as , , The feature difference level threshold is automatically calculated using the Otsu method. The Suzuki85 contour edge tracking algorithm is applied to the two-dimensional joint feature residual space to generate a candidate region set. The above-mentioned application of the Suzuki85 contour edge tracing algorithm to the two-dimensional joint feature residual space to generate a candidate region set means that the Suzuki85 contour edge tracing algorithm starts from the first position in the upper left corner of the two-dimensional joint feature residual space and scans each position row by row. When it encounters a position with a value of 1 and that position has not yet been marked as visited, the algorithm enters the tracing mode. After the tracing mode ends, it continues to scan the next unvisited pixel with a value of 1 until all pixels with a value of 1 have been visited. The above tracking mode refers to starting from this position and walking along the boundary positions with a value of 1 in its 8 neighborhoods. At each step, the coordinates of the current position are recorded. When returning to the starting position, a closed contour is formed. The Suzuki85 contour edge tracking algorithm marks all positions inside the closed contour as a candidate region. From the candidate region set, maxima are selected in ascending order of angular position, and a three-dimensional feature vector is constructed for each maxima point; The aforementioned maxima represent the centroid location of a potential, geometrically independent ice layer deposition structure within the candidate region: When ice forms on the surface of a component, it does not spread evenly, but tends to thicken first at multiple nucleation points, forming multiple local ice nuclei or ice ridges. The maximum point is the candidate location marker for these local ice nuclei. The three-dimensional feature vector has three dimensions: first, contour shape variable; second, local texture entropy; and third, local gray-level skewness. Among them, the contour deformation variable refers to the contour deformation residual field at the maximum point, the local texture entropy is the information entropy at the maximum point (the information entropy of the local binary pattern features within a circular neighborhood with a radius of 5 pixels centered on the maximum point), and the local grayscale skewness is the skewness at the maximum point (the skewness of all pixel grayscale values ​​within a span of 5° on the left and right sides of the maximum point). The squared negative Euclidean distance between the three-dimensional feature vectors of two maxima points is used as a similarity metric to construct a similarity matrix; It should be noted that, in the absence of prior knowledge, the diagonal elements in the similarity matrix are set to have the same value; Initialize two matrices with the same dimensions as the similarity matrix, namely the suitable matrix and the usable matrix; In this context, the points in the suitability matrix represent the suitability of a maximum point w as a cluster center for a maximum point j. The points in the available matrix represent the degree to which a maximum point j can serve as a cluster center for a maximum point w. The two matrices are transmitted in opposite directions; Using the affinity propagation (AP) clustering algorithm, the appropriate and available matrices are iteratively updated until the maximum number of iterations is reached. All maxima are decoupled into structured glacial units, where each structured glacial unit contains a cluster center maxima and the maxima of all its members. Specifically, suppose a candidate region originally contains a relatively large uniform thin ice layer (with similar characteristics of candidate points) and an isolated ice nodule located on the thin ice layer (with thick and coarse candidate points). Traditional methods would treat the entire region as a whole for analysis, while this scheme uses AP clustering to automatically classify the candidate points (maximum points) of the uniform thin ice layer into one category and the candidate points of the isolated ice nodule into another category. This decoupling capability stems from the design of the three-dimensional feature vectors—candidate points within the same ice structure exhibit significant consistency in thickness, texture, and grayscale symmetry, thus naturally clustering in the feature space.

[0018] S3. Based on the structured ice phase unit, a trajectory matrix is ​​generated. The effective rank is obtained through singular value decomposition, Shannon entropy calculation and exponential function. The structured ice phase unit is then morphologically classified. Specifically, the maximum points in the structured ice phase units are sorted according to their angular positions, and the corresponding contour deformations are generated into column vectors. Perform an outer product operation between the column vector and a spatial location encoding vector to obtain a column matrix, and then perform column expansion to convert it into a trajectory matrix; The above column expansion refers to the use of delayed embedding operations; The spatial location encoding vector is a vector of length 1 consisting entirely of 1s; If the ice phase unit is a disk-shaped uniform ice layer, it means that the thickness of the ice layer is approximately equal at multiple adjacent angular positions. In this case, the values ​​of each element in the column vector will be very close. In this case, the singular value energy of the matrix will be distributed across multiple singular values. If the ice phase unit is a needle-shaped ice ridge, it means that the ice thickness is highly concentrated in a certain or extremely narrow angular range. In this case, the column vector will contain a peak value that is significantly higher than other elements, while the other elements are close to 0. In this case, the singular value energy of the matrix will be highly concentrated on the first singular value. The trajectory matrix is ​​decomposed by singular value decomposition to obtain a set of singular values. Each singular value is normalized, and the Shannon entropy of the normalized singular values ​​is calculated. Shannon entropy directly quantifies the degree of disorder or dispersion of deformation energy distribution within an ice phase unit. For a disk-shaped homogeneous ice layer, energy is evenly distributed in multiple directions, the probability distribution is flat, and the entropy value is relatively large. For needle-shaped ice ridges, energy is highly concentrated in the first direction, the probability distribution is sharp, and the entropy value is small. Based on Shannon entropy, the effective rank is calculated using an exponential function, and then truncated and scaled to obtain the scaled rank. ; The above effective rank It is a continuous scalar that satisfies: D, where The corresponding energy is completely concentrated in a single direction, while D corresponds to energy that is uniformly distributed in all D directions, where D is the number of rows in the trajectory matrix; Effective rank can be understood as—equivalent to the number of effective directions, for example, =1.3 means that the energy distribution of the structural ice phase unit is equivalent to 1.3 non-zero singular values—that is, most of the energy is concentrated in one direction, but a small amount of energy is dispersed in other directions; The above truncation process, formula: ; Mapping the scaling rank to a fixed interval, such as [1,2], yields: This indicates that the deformation energy is evenly distributed at multiple angular positions, and the ice layer appears as a large area with uniform thickness (disc-shaped). The aerodynamic effect is similar to an increase in surface roughness, and the risk is relatively low. Here, aerodynamic is an abbreviation for aerodynamic characteristics. This indicates that the deformation energy is distributed in two main directions, the ice layer presents a transitional state with certain textures or wave-like morphology, the aerodynamic effects begin to show local separation, and the risk increases; This indicates that the deformation energy is almost entirely concentrated in a very narrow angular range, and the ice layer appears as sharp, towering ice ridges or ice nodules (needle-like). The aerodynamic effect is equivalent to local bulging, which can induce strong turbulence and separation, posing an extremely high risk. The system is based on prior data from fluid dynamics simulation and wind tunnel experiments. It sets two fixed classification thresholds, including an upper threshold and a lower threshold, and determines the morphological classification of structural ice phase units based on the classification thresholds. If the upper threshold is 1.7 and the lower threshold is 1.2, then the morphological classification of the structural ice phase unit is as follows: ; in, It is a disc-shaped, uniform ice layer. This is a transition state. C represents needle-shaped ice ridges / ice tumors.

[0019] S4. Construct a four-dimensional physical property vector for each structural ice phase unit and calculate the fusion ice uniformity index; Specifically, a four-dimensional physical property vector is constructed for each structural ice phase unit, including the center angular coordinates, angular span, volume estimate, and effective rank; The angular span is calculated from the angle difference between the leftmost maximum point and the rightmost maximum point within the unit. The volume estimate is obtained by multiplying the mean of the profile deformation in each structural ice phase unit by its angular span. Based on the four-dimensional physical property vector, the risk factor of each structural ice phase unit is calculated; The risk factor is composed of a weighted average of morphological risk components, volumetric risk components, and location risk components. The sum of the weights of the three components mentioned above is 1, and they decrease in the order of morphological risk component, volumetric risk component, and positional risk component. For example, the weight of the morphological risk component is 0.5, the weight of the volumetric risk component is 0.3, and the weight of the positional risk component is 0.2. This order depends on the fact that morphological influence is the most critical aerodynamic disturbance factor, followed by volumetric, and finally position. It can be adjusted based on actual usage. Among them, the morphological risk component is calculated from the effective rank using the Sigmoid function; The volume risk component is obtained by normalizing the volume estimate of the current structural ice phase unit relative to the global maximum volume in the same frame; The location risk component was calculated using a linear weighted method based on normalized distance. Morphological risk component calculation formula: ; in, For the i-th candidate region, the first... Morphological risk component values ​​of individual structural ice phase units K is the kurtosis coefficient of the sigmoid function, which controls the sensitivity of the morphological risk function from low to high risk. The larger the value of k, the narrower the transition band. near The risk value changes drastically; the smaller the k value, the wider the transition band, and the smoother the change in risk value with effective rank. The value is based on empirical settings, such as 5. For the i-th candidate region, the first... Effective value of each structural ice phase unit The middle threshold is the average of the upper and lower thresholds. Volumetric risk component calculation formula: ; in, For the i-th candidate region, the first... Volumetric risk component value of each structural ice phase unit For the i-th candidate region, the first... Volume estimates for each structural ice phase unit. This represents the global maximum value among the volume estimates of all glacial units within all candidate regions. A global index for candidate regions. This is a global index for structural ice phase units; Location risk component calculation formula: ; in, For the i-th candidate region, the first... Location risk component values ​​of each structural ice phase element For the i-th candidate region, the first... The center angle coordinates of each structural ice phase unit The stagnation angle at the leading edge of the component. For angular span; The leading edge stagnation point angle of the above-mentioned component is 90° when the actual component is a symmetrical airfoil. This value is based on the assumption that the leading icing sensing component is a symmetrical airfoil or an approximately symmetrical structure, and its leading edge stagnation point is located at the midpoint of the arc surface (angle position 90°). If the actual component is an asymmetric airfoil, this value can be obtained through aerodynamic data calibration; The risk factors of all structural ice phase units in each candidate region are summed to obtain the comprehensive risk factor of each candidate region. The comprehensive risk factors of all candidate regions are then summed to obtain the total global structural risk value. Based on the volume estimate, risk factors, and total global structural risk, the fusion icing uniformity index is calculated. ,formula: ; in, This represents the total value of global structural risk. and These are the estimated total global ice volume and the global maximum risk factor, respectively. The calculation method is similar to the calculation process of the total global structural risk value. The global maximum risk factor is the maximum risk factor in all candidate regions.

[0020] S5. Based on the morphological classification results and the fusion of the icing uniformity index, a comprehensive judgment is made on the icing environment, and a graded early warning is issued based on the judgment results. Based on the early warning level, physical action control is executed. Specifically, the mean values ​​of the real-time contour vector and the real-time grayscale field vector are calculated respectively. If the mean values ​​of the real-time contour vector and the real-time grayscale field vector are both greater than the preset contour threshold and grayscale threshold, it is determined that there are signs of icing in this frame and enters the continuous frame stability judgment. Otherwise, it is determined to be normal and the detection of the next frame continues. The aforementioned contour threshold and grayscale threshold are set using the 3σ principle based on images acquired when there is no ice or contaminants on the surface of the leading icing sensing component. For example, the contour threshold is 2.5 pixels and the grayscale threshold is 8 pixels. If N consecutive frames are determined to show signs of icing, the icing environment is officially confirmed and a graded warning judgment is entered. If any frame does not meet the conditions during this period, the counter is reset to zero, the count is restarted, and it is determined that no icing is detected. The above values ​​for N are set based on the specific engineering requirements; The aforementioned graded early warning judgment refers to classifying the fusion icing uniformity index calculated in real time through preset regular thresholds and emergency thresholds, including regular early warning, enhanced early warning and emergency early warning; In ice wind tunnel experiments, a sample set was generated encompassing various ice shapes, including disc-shaped uniform ice, transitional textured ice, and needle-shaped ice nodules. Each sample was manually labeled with its actual hazard level, and the fusion ice uniformity index was calculated for each sample. Regular and emergency thresholds were set, such as... and ; The grading rules for the fusion icing uniformity index are as follows: ; in, This is a routine warning, indicating the presence of icing, but the distribution is uniform (mainly in disc-shaped patterns), with minimal aerodynamic impact. To strengthen early warning, this indicates that icing exhibits a certain degree of non-uniformity, requiring attention to specific areas. This is an emergency warning, indicating the presence of severe asymmetric icing or high-risk icing tumors, requiring immediate intervention.

[0021] Furthermore, when a routine warning is issued, the standard de-icing mode will be implemented; When an enhanced warning is issued, a zone-specific de-icing mode will be implemented. When an emergency warning is issued, the powerful de-icing mode will be activated. The standard de-icing mode refers to uniform heating throughout the entire system. The partitioned differentiated de-icing mode sends a global uniform heating command to the electrothermal anti-icing and de-icing system, with all heating elements operating at 100% of their rated power, while simultaneously sending an "icing presence" status flag to the flight control system and imposing no aerodynamic restrictions. The aforementioned electrothermal anti-icing and de-icing system is an airborne system that is already installed on the aircraft. It uses resistance heating elements to heat key surfaces such as the leading edge of the wing to prevent the formation of ice or to melt existing ice. The flight control system is the avionics system on an aircraft responsible for stabilizing and controlling its flight attitude and responding to pilot commands or autopilot inputs. It can adjust control surfaces, engine thrust, etc. The partitioned differential de-icing mode refers to increasing the power of the heating element in the corresponding angle range to 150% of the rated value according to the angular coordinate list of the transitional structure ice phase unit, while maintaining the standard heating power in the remaining uniform area (disc-shaped uniform ice layer), and sending a "non-uniform icing" status mark and the angle range of the non-uniform area to the flight control system. The powerful de-icing mode directs all heating elements to be heated urgently at maximum power (200% of the rated value). At the same time, it transmits the four-dimensional physical attribute vector of the needle-shaped ice phase unit (Ph=C) and the risk factor to the flight control system in real time, so that it can perform aerodynamic protection actions, including limiting the deflection rate and amplitude of the control surface in high-risk areas, lowering the angle of attack / sideslip angle protection threshold, and activating the stability augmentation control law. During the de-icing process, the risk factors in the heated area are continuously monitored; if the risk factor in a certain area continues to decrease, the current power is maintained. If the risk factor in a certain area increases instead of decreasing, or if the needle-like ice phase unit continues to exist at maximum power for more than the set time limit, an instantaneous overpower pulse will be triggered in that area. After all the needle-like units disappear, the heating power will be gradually restored to the standard mode. When the signs of icing disappear, the warning is lifted, a stop heating command is sent to the anti-icing and de-icing system and the power is confirmed to be cut off, a "ice environment removed, return to normal mode" command is sent to the flight control system, all aerodynamic restrictions are canceled, the internal state of the system is reset, and it returns to the normal monitoring state.

[0022] This embodiment also provides a visual detection system for early warning of icing environments, including: The acquisition and extraction module is used to acquire and preprocess real-time images of the pilot icing sensing component, and extract real-time contour vectors and real-time grayscale field vectors. A clustering module is constructed to build a two-dimensional joint feature residual space based on real-time contour vectors and real-time grayscale field vectors, generate a candidate region set, construct a three-dimensional feature vector for the maxima points in the candidate region set, and use the affinity propagation clustering algorithm to cluster them to obtain structured ice phase units. The decomposition and classification module is used to generate trajectory matrices based on structured ice phase units, obtain effective rank through singular value decomposition, Shannon entropy calculation and exponential function, and perform morphological classification of structured ice phase units. The computational fusion module is used to construct a four-dimensional physical property vector for each structural ice phase unit and calculate the fusion ice uniformity index. The judgment execution module is used to make a comprehensive judgment on the icing environment based on morphological classification results and fused icing uniformity index, and to issue graded warnings based on the judgment results. Based on the warning level, physical action control is executed.

[0023] It should be noted that 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A visual detection method for early warning of icing environments, characterized in that, include: Real-time images of the pilot icing sensing component are acquired and preprocessed to extract real-time contour vectors and real-time grayscale field vectors. Based on real-time contour vectors and real-time grayscale field vectors, a two-dimensional joint feature residual space is constructed to generate a candidate region set. A three-dimensional feature vector is constructed for the maxima points in the candidate region set, and the affinity propagation clustering algorithm is used to cluster them to obtain structured ice phase units. Based on structured ice phase units, a trajectory matrix is ​​generated. The effective rank is obtained through singular value decomposition, Shannon entropy calculation, and exponential function. The structured ice phase units are then morphologically classified. A four-dimensional physical property vector is constructed for each structural ice phase unit, and the fusion ice uniformity index is calculated; Based on morphological classification results and the fusion of icing uniformity index, a comprehensive judgment is made on the icing environment, and a graded early warning is issued based on the judgment results. Based on the early warning level, physical action control is executed.

2. The visual detection method for early warning of icing environments as described in claim 1, characterized in that, The process of acquiring and preprocessing real-time images of the pilot icing sensing component to extract real-time contour vectors and real-time grayscale field vectors includes: Real-time images of the pilot icing sensing component are acquired and denoised using Gaussian filtering to obtain the pre-processed image; Subpixel-level Canny edge detection is performed on the preprocessed image to extract the contour of the leading edge of the icing sensing component and obtain the real-time contour. Using a template matching algorithm based on normalized cross-correlation, the real-time contour is rigidly registered with the reference contour, and the contour is parameterized in polar coordinates to generate a real-time contour vector. In the coordinate system after rigid registration, the ROI region is defined as the reference contour, which is extended by M pixels in both the inner and outer directions along the normal direction. The ROI region of the preprocessed image is sampled by gray level at each angle to generate a real-time gray field vector.

3. The visual detection method for early warning of icing environments as described in claim 2, characterized in that, The process involves constructing a two-dimensional joint feature residual space based on real-time contour vectors and real-time grayscale field vectors, generating a candidate region set, constructing three-dimensional feature vectors for the maxima in the candidate region set, and using an affinity propagation clustering algorithm to cluster them, resulting in structured glacial units, including: Based on the angular coordinate axis of the leading edge of the icing sensing component, the contour deformation residual field and the gray-scale symmetry residual field are calculated and weighted and stitched together to construct a two-dimensional joint feature residual space. The Suzuki85 contour edge tracking algorithm is applied to the two-dimensional joint feature residual space to generate a candidate region set. From the candidate region set, maxima are selected in ascending order of angular position, and a three-dimensional feature vector is constructed for each maxima point; The three-dimensional feature vector has three dimensions: first, contour shape variable; second, local texture entropy; and third, local gray-level skewness. Among them, the contour deformation variable refers to the contour deformation residual field at the maximum point, the local texture entropy is the information entropy at the maximum point, and the local gray level skewness is the skewness at the maximum point. The squared negative Euclidean distance between the three-dimensional feature vectors of two maxima points is used as a similarity metric to construct a similarity matrix; Initialize two matrices with the same dimensions as the similarity matrix, namely the suitable matrix and the usable matrix; In this context, the points in the suitability matrix represent the suitability of a maximum point w as a cluster center for a maximum point j. The points in the available matrix represent the degree to which a maximum point j can serve as a cluster center for a maximum point w. The two matrices are transmitted in opposite directions; Using the affinity propagation clustering algorithm, the appropriate and available matrices are iteratively updated until the maximum number of iterations is reached, decoupling all the maxima into structured glacial units.

4. The visual detection method for early warning of icing environments as described in claim 3, characterized in that, The method involves generating a trajectory matrix based on structured ice phase units, obtaining the effective rank through singular value decomposition, Shannon entropy calculation, and exponential functions, and performing morphological classification of the structured ice phase units, including: The maximum points in the structured ice phase unit are sorted according to their angular positions, and the corresponding contour deformations are generated into column vectors. Perform an outer product operation between the column vector and a spatial location encoding vector to obtain a column matrix, and then perform column expansion to convert it into a trajectory matrix; The trajectory matrix is ​​decomposed by singular value decomposition to obtain a set of singular values. Each singular value is normalized, and the Shannon entropy of the normalized singular values ​​is calculated. Based on Shannon entropy, the effective rank is calculated using an exponential function, and then truncated and scaled to obtain the scaled rank. Map the scaling rank to a fixed interval; Based on prior data from fluid dynamics simulations and wind tunnel experiments, two fixed classification thresholds, including an upper threshold and a lower threshold, are set to determine the morphological classification of structural ice phase units.

5. The visual detection method for early warning of icing environments as described in claim 4, characterized in that, The process of constructing a four-dimensional physical property vector for each structural ice phase unit and calculating the fusion ice uniformity index includes: A four-dimensional physical property vector is constructed for each structural ice phase unit, including the center angle coordinates, angle span, volume estimate, and effective rank; The angular span is calculated from the angle difference between the leftmost maximum point and the rightmost maximum point within the unit. The volume estimate is obtained by multiplying the mean of the profile deformation in each structural ice phase unit by its angular span. Based on the four-dimensional physical property vector, the risk factor of each structural ice phase unit is calculated; The risk factor is composed of a weighted average of morphological risk components, volumetric risk components, and location risk components. Among them, the morphological risk component is calculated from the effective rank using the Sigmoid function; The volume risk component is obtained by normalizing the volume estimate of the current structural ice phase unit relative to the global maximum volume in the same frame; The location risk component was calculated using a linear weighted method based on normalized distance. The risk factors of all structural ice phase units in each candidate region are summed to obtain the comprehensive risk factor of each candidate region. The comprehensive risk factors of all candidate regions are then summed to obtain the total global structural risk value. The fusion icing uniformity index is calculated based on the volume estimate, risk factors, and total global structural risk.

6. The visual detection method for early warning of icing environments as described in claim 5, characterized in that, The method of comprehensively judging the icing environment based on morphological classification results and fused icing uniformity index, and providing graded early warnings based on the judgment results, includes: The mean values ​​of the real-time contour vector and the real-time grayscale field vector are calculated respectively. If the mean values ​​of the real-time contour vector and the real-time grayscale field vector are both greater than the preset contour threshold and grayscale threshold, it is determined that there are signs of icing in this frame and the continuous frame stability judgment is entered. Otherwise, it is determined to be normal and the detection of the next frame continues. If N consecutive frames are determined to show signs of icing, the icing environment is officially confirmed and a graded warning judgment is entered. If any frame does not meet the conditions during this period, the counter is reset to zero, the count is restarted, and it is determined that no icing is detected. The aforementioned graded early warning judgment refers to classifying the fusion icing uniformity index calculated in real time through preset regular thresholds and emergency thresholds, including regular early warning, enhanced early warning, and emergency early warning.

7. The visual detection method for early warning of icing environments as described in claim 6, characterized in that, The physical motion control based on the warning level includes: When a routine warning is issued, the standard de-icing mode will be executed. When an enhanced warning is detected, a zone-specific de-icing mode is executed. When an emergency warning is issued, the powerful de-icing mode will be activated. The standard de-icing mode refers to uniform heating throughout the entire system. The partitioned differentiated de-icing mode sends a global uniform heating command to the electrothermal anti-icing and de-icing system, with all heating elements operating at 100% of their rated power, while simultaneously sending an icing presence status marker to the flight control system and imposing no aerodynamic restrictions. The partitioned differential de-icing mode refers to increasing the power of the heating element in the corresponding angle range to 150% of the rated value according to the angular coordinate list of the transitional structure ice phase unit, while maintaining the standard heating power in the remaining uniform area, and sending the non-uniform icing state mark and the angle range of the non-uniform area to the flight control system at the same time. The powerful de-icing mode directs all heating elements to be heated at maximum power for emergency heating, while simultaneously transmitting the four-dimensional physical property vector of the needle-shaped ice phase unit and risk factors to the flight control system in real time for it to perform aerodynamic protection actions. During the de-icing process, the risk factors in the heated area are continuously monitored; If the risk factor in a certain region continues to decline, the current power level will be maintained. If the risk factor in a certain area increases instead of decreasing, or if the needle-like ice phase unit continues to exist at maximum power for more than the set time limit, an instantaneous overpower pulse will be triggered in that area. After all the needle-like units disappear, the heating power will be gradually restored to the standard mode. When the signs of icing disappear, the warning is lifted, a stop heating command is sent to the anti-icing system and the power outage is confirmed, a return to normal mode command is sent to the flight control system, all aerodynamic restrictions are canceled, the internal state of the system is reset, and it returns to the normal monitoring state.

8. A visual detection system for early warning of icing environments, based on the visual detection method for early warning of icing environments as described in any one of claims 1 to 7, characterized in that, include: The acquisition and extraction module is used to acquire and preprocess real-time images of the pilot icing sensing component, and extract real-time contour vectors and real-time grayscale field vectors. A clustering module is constructed to build a two-dimensional joint feature residual space based on real-time contour vectors and real-time grayscale field vectors, generate a candidate region set, construct a three-dimensional feature vector for the maxima points in the candidate region set, and use the affinity propagation clustering algorithm to cluster them to obtain structured ice phase units. The decomposition and classification module is used to generate trajectory matrices based on structured ice phase units, obtain effective rank through singular value decomposition, Shannon entropy calculation and exponential function, and perform morphological classification of structured ice phase units. The computational fusion module is used to construct a four-dimensional physical property vector for each structural ice phase unit and calculate the fusion ice uniformity index. The judgment execution module is used to make a comprehensive judgment on the icing environment based on morphological classification results and fused icing uniformity index, and to issue graded warnings based on the judgment results. Based on the warning level, physical action control is executed.

Images