Damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion

Abstract

The invention discloses a damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion, and relates to the technical field of insulator defect detection.The method comprises the steps that firstly, in the data acquisition stage, an insulator image is synchronously acquired, and a humidity sensor and a wind speed sensor are arranged to obtain data containing environment information; for high humidity, a de-noising and enhancing algorithm based on humidity self-adaption is adopted to process the blurred image, and when the wind speed is high, anti-shake operation is carried out on the dithered image by analyzing wind speed data. Secondly, extracting temperature gradient and humidity change rate environment-related characteristics in an environment-related characteristic extraction link; and in the subsequent feature dynamic fusion step, different schemes are used according to different environment states, and the two schemes are synthesized and weighted to obtain a final detection conclusion. The mechanism effectively guarantees the reliability of the system, and is better in the aspects of image recognition and detection of the damaged ceramic insulator.

Description

Technical Field

[0001] The present invention relates to the technical field of insulator defect detection, and in particular to a method for image recognition and detection of damaged ceramic insulators based on multi-algorithm fusion. Background Art

[0002] In the operation of the power system, ceramic insulators, as important insulating components, play a key role in support and insulation. However, ceramic insulators exposed to the outdoor environment for a long time are easily damaged by various factors, which will not only affect the stability of power transmission, but may even cause power accidents and endanger the safety of the power grid. Traditional ceramic insulator detection methods mainly rely on manual inspections. Inspectors need to use the naked eye or some simple tools to determine whether the insulator is damaged in a complex outdoor environment. This method has many disadvantages. On the one hand, the efficiency of manual inspections is extremely low, and facing the huge power line network, it takes a lot of manpower and time; on the other hand, the judgment accuracy of the human eye is limited, especially for some minor damage or in bad weather conditions, it is easy to miss or misdetect.

[0003] With the development of computer vision technology, the use of image recognition technology to detect ceramic insulators has become a new research direction. Early image recognition methods were mostly based on traditional visual features of insulator images, such as edge features, texture features, and color features for analysis and judgment. Although these methods have improved the degree of automation of detection to a certain extent, there are still obvious shortcomings. Since ceramic insulators are mostly installed outdoors, the environment they are in is complex and changeable. Environmental factors such as light, humidity, and wind speed have a huge impact on the quality of insulator images. High humidity environments may cause image blur, and strong winds may cause the captured images to shake, which will interfere with the accurate extraction and analysis of traditional visual features, thereby affecting the accuracy of the detection results. Moreover, relying solely on traditional visual features often cannot fully and accurately reflect the true state of the insulator, especially when distinguishing different types and degrees of damage. In order to overcome these problems, researchers began to explore the inclusion of environmental factors into consideration and tried to integrate multiple algorithms to improve the effect of image recognition and detection of damaged ceramic insulators. Therefore, a damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion came into being. It aims to comprehensively consider environmental information and multiple features, and through advanced algorithm fusion means, achieve more accurate, efficient and reliable insulator damage detection to ensure the safe and stable operation of the power system. Summary of the invention

[0004] The purpose of the present invention is to provide a method for image recognition and detection of damaged ceramic insulators based on multi-algorithm fusion. By synchronously collecting insulator images and equipping humidity and wind speed sensors to obtain data containing environmental information, corresponding algorithms are used to process images for high humidity and high wind speed. Then, environmental-related features are extracted to obtain temperature gradient and humidity change rate features. The subsequent dynamic feature fusion step will use different methods to fuse different results according to different environmental conditions, and then combine the two and assign weights to obtain the final detection conclusion to solve the above problems.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion, characterized by comprising the following steps:

[0007] S1: Data acquisition, while collecting insulator images, equipped with humidity sensors and wind speed sensors;

[0008] S2: Environment-related feature extraction, extracting environment-related features from comprehensive data that integrates environmental information;

[0009] S3: Dynamic feature fusion, which uses a dynamic fusion algorithm to fuse environment-related features with traditional visual features;

[0010] S4: Based on the hierarchical fusion of environmental states, a hierarchical fusion model is constructed to divide the features into different levels for fusion according to the environmental states;

[0011] S5: Establish an adaptive model update mechanism to update the model parameters and structure in real time based on the data collected by environmental sensors and the performance feedback of the model in actual applications.

[0012] In step S1, when the humidity is high, the image is blurred by collecting data and then a humidity-adaptive denoising and enhancement algorithm is used. The denoising and enhancement function used is:

[0013] I enhanced (x,y)=f H (I(x,y),H)

[0014] Among them, I(x,y) represents the original insulator image, where x is the pixel coordinate of the image in the horizontal direction, which is used to determine the horizontal position of a point in the image; y is the pixel coordinate of the image in the vertical direction, which is used to determine the vertical position of a point in the image. Through these two coordinates, any pixel point in the image can be located, and then the entire image can be processed. enhanced(x, y) is the image processed by the humidity-adaptive denoising and enhancement algorithm. It is a new image obtained by denoising and enhancing the original image I(x, y) by considering the humidity factor. The position of each pixel in the image is also represented by the x and y coordinates. H is a denoising and enhancement function based on humidity H. Its function is to dynamically adjust the denoising and enhancement parameters according to the humidity value, so as to effectively remove image blur and improve image quality under different humidity conditions. H represents the ambient humidity value measured by the humidity sensor. This value is obtained in real time and is used as the function f H It is an important input parameter used to determine how humidity-related parameters in the denoising and enhancement algorithms are adjusted.

[0015] The specific form of denoising and enhancement function is:

[0016]

[0017] Among them, a H (H) is the coefficient used to adjust the weight of the original image I(x,y) in the denoising and enhancement process; b H (H) is the same as the Laplace operator The multiplication coefficient is used to control the degree of processing based on the second-order derivative information of the image; c H (H) is a constant term used to perform an overall offset adjustment operation on the processed image. is the Laplace operator, which is used to calculate the second-order derivative of the image. For the image function I(x,y), the Laplace operator at the point (x,y) is calculated as follows:

[0018]

[0019] in, Represents the second-order partial derivative of the image function I(x,y) in the x direction, which reflects the change in the curvature of the image in the x direction. It represents the second-order partial derivative of the image function I(x,y) in the y direction, reflecting the change of the curvature of the image in the y direction. The second-order derivative information calculated by the Laplace operator helps to highlight the edge details of the image during the denoising and enhancement process, thereby effectively removing the image blur caused by humidity and improving the image quality.

[0020] The step S1 determines the situation when the wind speed is high by collecting data, and performs translation and rotation compensation operations on the image by analyzing the wind speed data V to achieve anti-shake processing, and its formula is expressed as:

[0021] I stable (x,y)=M V (V) Ijitter (x,y)

[0022] Among them, I jitter (x, y) represents the image that is shaken due to wind speed. x and y are the pixel coordinates of the image in the horizontal and vertical directions, respectively, which are used to determine the position of each pixel in the image. Due to the effect of wind speed, the image is shaken during the acquisition process, which may affect the image quality. stable (x, y) is the image after anti-shake processing. The x and y coordinates are also used to represent the position of each pixel. Through the anti-shake processing operation, the image is restored to a relatively stable state after being shaken by the wind speed, so as to perform accurate analysis and processing later. V (V) is a transformation matrix related to wind speed V, which is obtained through a series of complex calculations based on wind speed data V. Its specific form is adjusted according to the magnitude and direction of wind speed. The function of this transformation matrix is ​​to jitter (x, y) is translated and rotated to compensate for the image jitter caused by wind speed and ensure the stability of the image. V represents the wind speed value measured by the wind speed sensor, which is the environmental wind speed information obtained in real time and is used to calculate the transformation matrix M V Different wind speed values ​​will lead to different forms of the transformation matrix, thereby achieving accurate anti-shake processing for different wind speed conditions.

[0023] The M V (V) is calculated as follows:

[0024]

[0025] Among them, V x Indicates the horizontal component value of wind speed, which reflects the degree of jitter caused by wind speed on image acquisition in the horizontal direction. Different V x The value V will cause the transformation matrix to compensate the image differently in the horizontal direction. y Indicates the vertical component value of wind speed, which reflects the degree of jitter caused by wind speed on image acquisition in the vertical direction. Different V y The value will make the transformation matrix compensate the image differently in the vertical direction. m represents the number of pixels in the horizontal direction of the image. It is a basic attribute of the image and is used to determine the size of the image. V (V), the horizontal component of wind speed V x Combined with the above, the parameters required for the horizontal compensation of the image can be accurately calculated. n represents the number of pixels in the vertical direction of the image, which is also a basic attribute of the image and is used to determine the size of the image.V (V), the component V perpendicular to the wind speed y Combined with the above, the compensation operation of the image in the vertical direction can be accurately calculated.

[0026] The step S2 includes environment-related feature extraction and feature dynamic fusion; the environment-related feature extraction includes temperature gradient feature extraction and humidity change rate feature extraction;

[0027] The temperature gradient feature extraction is obtained by calculating the gradient of the temperature function. The temperature gradient vector The calculation formula is:

[0028]

[0029] in, Represents the temperature gradient vector, which describes the trend of temperature change on the two-dimensional plane of the insulator surface. This vector consists of two components, which reflect the rate of change of temperature in the x direction and the y direction respectively. By analyzing this vector, we can understand the change of temperature at different positions on the insulator surface, and then associate it with the damage state of the insulator. T(x,y) represents the temperature distribution function of the insulator surface, where x and y are the coordinates of the insulator surface on the two-dimensional plane. The position of any point on the insulator surface can be determined by these two coordinates, and T(x,y) gives the temperature value corresponding to the point. It represents the partial derivative of the temperature function T(x,y) in the x direction. It reflects the rate of change of temperature in the x direction, that is, when the x coordinate changes slightly, the corresponding speed of temperature change. By calculating this partial derivative, the trend of temperature change in the horizontal direction can be obtained. It is an important component of the temperature gradient vector. It represents the partial derivative of the temperature function T(x,y) in the y direction, which reflects the rate of change of temperature in the y direction, that is, when the y coordinate changes slightly, the corresponding speed of temperature change. By calculating this partial derivative, the trend of temperature change in the vertical direction can be obtained, which is another important component of the temperature gradient vector.

[0030] The humidity change rate feature extraction is to extract the humidity change rate feature of the insulator surface from the humidity data H, and the calculation formula is:

[0031]

[0032] in Represents the humidity change rate characteristic vector, which describes the change trend of humidity on the two-dimensional plane of the insulator surface. This vector consists of two components, which reflect the change rate of humidity in the x direction and the y direction respectively. By analyzing this vector, we can understand the change of humidity at different positions on the insulator surface, which may be related to the aging and damage of the insulator. H(x,y) represents the humidity distribution function of the insulator surface, where x and y are the coordinates of the insulator surface on the two-dimensional plane. The position of any point on the insulator surface can be determined by these two coordinates, and H(x,y) gives the humidity value corresponding to the point. In this way, the entire function describes the humidity distribution on the insulator surface. It represents the partial derivative of the humidity function H(x,y) in the x direction, which reflects the rate of change of humidity in the x direction, that is, when the x coordinate changes slightly, the corresponding change speed of humidity. By calculating this partial derivative, the change trend of humidity in the horizontal direction can be obtained, which is an important component of the characteristic vector of humidity change rate. Represents the partial derivative of the humidity function H(x,y) in the y direction.

[0033] The step S3 of dynamic feature fusion is specifically to fuse the environment-related features and the traditional visual features using a dynamic fusion algorithm, and the dynamic fusion algorithm is expressed as:

[0034]

[0035] in Represents the traditional visual feature vector, which contains the feature information obtained from the insulator image through the traditional visual feature extraction method, including edge features, texture features, and color features. These features are extracted based on the visual attributes of the image itself and are used for subsequent analysis and judgment of the insulator status. It represents the environment-related feature vector, which is a vector composed of features extracted from environmental information that may be related to the insulator state, including the temperature gradient feature vector and humidity change rate feature vector mentioned above. These environment-related features can reflect the state changes of the insulator in a specific environment from different angles. It is the fusion feature vector obtained after the dynamic fusion algorithm is processed. This vector combines the traditional visual feature vector and environment-related feature vector By integrating these two different types of features, it is expected that the state of the insulator can be described more comprehensively and accurately for subsequent classification and detection operations. t represents a time series variable, which is used to identify the situation at different times in the entire data processing process. In practical applications, environmental conditions may change over time, and the weight of feature fusion may also need to be dynamically adjusted according to the environmental conditions at different times, so t is introduced to reflect this dynamic nature. w v (t) is the traditional visual feature vector The weight function at different times t adjusts the importance of traditional visual features in fusion in real time according to the data collected by the environmental sensor, that is, it determines the importance of traditional visual features in fusion as time and environmental conditions change. The fused feature vector The proportion of w in the fusion features affects the composition of the final fusion features. e (t) is the environment-related feature vector The weight function at different times t adjusts the importance of environment-related features in fusion in real time according to the data collected by the environmental sensors, which determines The fused feature vector The proportion of the environment-related features enables the environment-related features to reasonably participate in the fusion process according to the actual situation, so as to better reflect the state of the insulator.

[0036] The S4 constructs a hierarchical fusion model, which specifically includes fusion under a stable environment state and fusion under an abnormal environment state;

[0037] The fusion under the stable environment is specifically when the environment is relatively stable, that is, when the temperature, humidity and wind speed are within the normal range, the traditional visual features are fused by a simple weighted average method to obtain a preliminary classification result, and the formula is:

[0038]

[0039] in, represents the i-th traditional visual feature vector, where i ranges from 1 to n. These traditional visual feature vectors contain various visual feature information extracted from the sub-image as described above. Different vectors may correspond to different types of visual features or the same type of features extracted from different parts of the image. v1i is the i-th traditional visual feature vector These weights are pre-set based on the factors verified by the previous experiments and are used to determine the importance of each traditional visual feature vector in the fusion process. Different weight settings will affect the formation of the final preliminary classification results. v1i is the i-th traditional visual feature vector These weights are pre-set based on the factors verified by the previous experiments and are used to determine the importance of each traditional visual feature vector in the fusion process. Different weight settings will affect the formation of the final preliminary classification results and meet That is, the sum of the weights of all traditional visual feature vectors is 1 to ensure the rationality of the weighted average. C1 is the preliminary classification result obtained by fusing the traditional visual feature vectors through a simple weighted average method. It is a result that integrates the information of various traditional visual feature vectors. It is used to make a preliminary judgment on the state of the insulator under a stable environmental state. It will be further integrated with the results of other layers to obtain the final detection conclusion.

[0040] When the environment is abnormal, that is, the temperature is too high, the humidity is too high, or the wind speed is too high, the environment-related features and the traditional visual features are fused according to the specific situation of the abnormal environment using a fusion strategy based on a support vector machine (SVM) to obtain another classification result. The fusion strategy based on SVM is as follows:

[0041] First, the environment-related feature vectors and traditional visual feature vectors are combined into a new training dataset Then use the SVM classification algorithm to classify this data set and get the classification result C2. Suppose the SVM classification function is f SVM (x), where x is the input feature vector, that is, the combined feature vector The classification result C2 is expressed as:

[0042]

[0043] in, represents the environment-related feature vector extracted under abnormal environmental conditions, represents the traditional visual feature vector extracted under abnormal environment, f SVM (x) is the classification function of the support vector machine (SVM), which is the core function based on the SVM algorithm and is used to classify the input feature vector x, that is, the combined feature vector Perform classification operations and find a hyperplane to separate samples of different categories to obtain classification results.

[0044] The S4 constructs a hierarchical fusion model, which specifically includes fusion under a stable environment state and fusion under an abnormal environment state;

[0045] The results of the fusion under the stable environment state and the abnormal environment state are fused, and different weights are assigned to the two layers of results according to the environmental state and the accuracy factors verified by the previous experiments. The final fusion formula is:

[0046] C final=w1·C1+w2·C2

[0047] Among them C final It is the final detection conclusion, which is a combination of the information of the first-level preliminary classification result C1 and the second-level classification result C2. The final judgment result on whether the insulator is damaged and the degree of damage is obtained by fusing them according to their respective weights. It is used to guide subsequent related operations. w1 is the weight of the result under the stable environment state, and w2 is the weight of the result under the abnormal environment.

[0048] The step S5 establishes an adaptive model updating mechanism, specifically to update the parameters and structure of the model in real time according to the data collected by the environmental sensors and the performance feedback of the model in actual application.

[0049] In the process of establishing the adaptive model update mechanism in step S5, when the model performance feedback shows that the detection accuracy drops continuously by more than 5%, in addition to updating the parameters and structure of the model, it should also trigger a re-evaluation of the data collection link to check whether the humidity sensor and wind speed sensor are faulty or whether the frequency of data collection is appropriate.

[0050] First, in the data collection stage, the insulator image is collected and equipped with humidity sensors and wind speed sensors to obtain data containing environmental information. When the collected data determines that the humidity is high, the humidity-adaptive denoising and enhancement algorithm is used to process the blurred image caused by humidity; if the wind speed is determined to be high, the wind speed data is analyzed to perform translation and rotation compensation operations on the jittered image to achieve anti-shake. Then enter the environment-related feature extraction link, extract the temperature gradient feature, humidity change rate feature and environment-related features from the comprehensive data of the fused environmental information, and then the dynamic feature fusion step, which will distinguish different situations according to the environmental state. In the stable environmental state, the simple weighted average method is used to fuse the traditional visual features to obtain the preliminary classification result; when the environment is abnormal, the environment-related features and traditional visual features are fused according to the specific situation using the support vector machine-based fusion strategy to obtain another classification result. Finally, the fusion results under these two different environmental states are combined and different weights are assigned to obtain the final detection conclusion.

[0051] During the whole process, an adaptive model update mechanism is also established. Based on the data collected by environmental sensors and the performance feedback of the model in actual applications, the model parameters and structure are updated in real time. When the detection accuracy drops by more than 5% continuously, in addition to updating the model parameters and structure, it will also trigger a re-evaluation of the data collection link to check whether the relevant sensors are faulty and whether the collection frequency is appropriate, so as to continuously optimize the recognition and detection effects of damaged ceramic insulator images.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] During the data collection phase, humidity sensors and wind speed sensors are equipped to fully acquire environmental information. This allows the impact of humidity, wind speed and temperature environmental factors on insulator images to be fully considered in subsequent processing, avoiding image quality degradation caused by environmental changes that affects the accuracy of recognition and detection.

[0054] By establishing humidity-adaptive denoising and enhancement algorithms and anti-shake processing mechanisms for high wind speeds, the image processing method can be automatically adjusted according to the actual environmental conditions, ensuring that relatively stable image data can be provided for subsequent feature extraction and analysis under different environmental conditions. Compared with existing technologies, it has better adaptability in complex and changeable actual environments.

[0055] The environment-related features are extracted and the dynamic fusion algorithm is used to fuse the environment-related features with the traditional visual features. This fusion of multiple types of features can more comprehensively describe the state of the insulator. Compared with the existing technology that only relies on a single type of feature, it can mine more information about whether the insulator is damaged and the degree of damage, thereby improving the accuracy of recognition and detection.

[0056] Layered fusion is performed based on environmental conditions, and different fusion strategies are adopted in stable and abnormal environmental conditions. In addition, the fusion results at different levels can be dynamically adjusted according to the environmental conditions and accuracy factors verified by previous experiments to obtain the final detection conclusion. This can better adapt to the differences in the characteristics of insulators in different environments and further improve the detection effect.

[0057] When the model performance feedback shows that the detection accuracy rate has continuously dropped by more than a certain threshold, in addition to updating the model parameters and structure, it will also trigger a re-evaluation of the data collection process to check whether the humidity sensor and wind speed sensor are faulty or whether the frequency of data collection is appropriate. This mechanism further guarantees the reliability of the entire recognition and detection system and ensures the data quality of the input model. Compared with existing technologies, it has more advantages in terms of system stability and data quality control. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of a damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion of the present invention; DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present invention will be fully described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0060] like Figure 1 As shown, a damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion is characterized by comprising the following steps:

[0061] S1: Data acquisition, while collecting insulator images, equipped with humidity sensors and wind speed sensors;

[0062] S2: Environment-related feature extraction, extracting environment-related features from comprehensive data that integrates environmental information;

[0063] S3: Dynamic feature fusion, which uses a dynamic fusion algorithm to fuse environment-related features with traditional visual features;

[0064] S4: Based on the hierarchical fusion of environmental states, a hierarchical fusion model is constructed to divide the features into different levels for fusion according to the environmental states;

[0065] S5: Establish an adaptive model update mechanism to update the model parameters and structure based on the data collected by environmental sensors and the performance feedback of the model in actual applications.

[0066] In step S1, when the humidity is high, the image is blurred by collecting data and then a humidity-adaptive denoising and enhancement algorithm is used. The denoising and enhancement function used is:

[0067] I enhanced (x,y)=f H (I(x,y),H)

[0068] Among them, I(x,y) represents the original insulator image, where x is the pixel coordinate of the image in the horizontal direction, y is the pixel coordinate of the image in the vertical direction, and I enhanced (x, y) is the image processed by humidity adaptive denoising and enhancement algorithm, f H is the denoising and enhancement function based on humidity H, where H represents the ambient humidity value measured by the humidity sensor;

[0069] The specific form of denoising and enhancement function is:

[0070]

[0071] Among them, a H (H) is the coefficient used to adjust the weight of the original image I(x,y) in the denoising and enhancement process; b H (H) is the same as the Laplace operator The multiplication coefficient; c H (H) is a constant term, is the Laplace operator, which is used to calculate the second-order derivative of the image. For the image function I(x,y), the Laplace operator at the point (x,y) is calculated as follows:

[0072]

[0073] in, Represents the second-order partial derivative of the image function I(x,y) in the x direction, which reflects the change in the curvature of the image in the x direction. It represents the second-order partial derivative of the image function I(x,y) in the y direction, reflecting the change of the curvature of the image in the y direction.

[0074] The step S1 determines the situation when the wind speed is high by collecting data, and performs translation and rotation compensation operations on the image by analyzing the wind speed data V to achieve anti-shake processing, and its formula is expressed as:

[0075] I stable (x,y)=M V (V) I jitter (x,y)

[0076] Among them, I jitter (x, y) represents the image that is shaken by wind speed, where x and y are the pixel coordinates of the image in the horizontal and vertical directions, respectively, which are used to determine the position of each pixel in the image. stable (x, y) is the image after anti-shake processing, and the position of each pixel is also represented by x and y coordinates. V (V) is a transformation matrix related to wind speed V, where V represents the wind speed value measured by the wind speed sensor;

[0077] The M V (V) is calculated as follows:

[0078]

[0079] Among them, V x Indicates the horizontal component of wind speed, V y It represents the vertical component value of wind speed, m represents the number of pixels in the horizontal direction, and n represents the number of pixels in the vertical direction.

[0080] The step S2 includes environment-related feature extraction and feature dynamic fusion; the environment-related feature extraction includes temperature gradient feature extraction and humidity change rate feature extraction;

[0081] The temperature gradient feature extraction is obtained by calculating the gradient of the temperature function. The temperature gradient vector The calculation formula is:

[0082]

[0083] in, represents the temperature gradient vector, T(x,y) represents the temperature distribution function of the insulator surface, where x and y are the coordinates of the insulator surface on the two-dimensional plane, and T(x,y) gives the temperature value corresponding to the point. represents the partial derivative of the temperature function T(x,y) in the x direction, Represents the partial derivative of the temperature function T(x,y) in the y direction;

[0084] The humidity change rate feature extraction is to extract the humidity change rate feature of the insulator surface from the humidity data H, and the calculation formula is:

[0085]

[0086] in represents the humidity change rate eigenvector, H(x,y) represents the humidity distribution function on the insulator surface, where x and y are the coordinates of the insulator surface on the two-dimensional plane, represents the partial derivative of the humidity function H(x,y) in the x direction, Represents the partial derivative of the humidity function H(x,y) in the y direction.

[0087] The step S2 of dynamic feature fusion is specifically to use a dynamic fusion algorithm to fuse the environment-related features and the traditional visual features. The dynamic fusion algorithm is expressed as:

[0088]

[0089] in Represents the traditional visual feature vector, which contains the feature information obtained from the insulator image through the traditional visual feature extraction method, including edge features, texture features, and color features. These features are extracted based on the visual attributes of the image itself and are used for subsequent analysis and judgment of the insulator state. represents the environment-related feature vector, is the fusion feature vector obtained after being processed by the dynamic fusion algorithm, t represents the time series variable, and w v (t) is the traditional visual feature vector The weight function at different times t, w e (t) is the environment-related feature vector Weight function at different times t.

[0090] The step S3 of dynamic feature fusion is specifically to fuse the environment-related features and the traditional visual features using a dynamic fusion algorithm, and the dynamic fusion algorithm is expressed as:

[0091]

[0092] in represents the traditional visual feature vector, represents the environment-related feature vector, is the fusion feature vector obtained after being processed by the dynamic fusion algorithm, t represents the time series variable, and w v (t) is the traditional visual feature vector The weight function at different times t, w e (t) is the environment-related feature vector Weight function at different times t.

[0093] The step S3 of dynamic feature fusion specifically includes fusion under a stable environment state and fusion under an abnormal environment state;

[0094] The fusion under the stable environment is specifically when the environment is relatively stable, that is, when the temperature, humidity and wind speed are within the normal range, the traditional visual features are fused by a simple weighted average method to obtain a preliminary classification result, and the formula is:

[0095]

[0096] in, represents the i-th traditional visual feature vector, where i ranges from 1 to n, and w v1i is the i-th traditional visual feature vector The weight, w v1i is the i-th traditional visual feature vector The weight of That is, the sum of the weights of all traditional visual feature vectors is 1 to ensure the rationality of weighted average. C1 is the preliminary classification result obtained by fusing traditional visual feature vectors through simple weighted average method;

[0097] When the environment is abnormal, that is, the temperature is too high, the humidity is too high, or the wind speed is too high, the environment-related features and the traditional visual features are fused according to the specific situation of the abnormal environment using a fusion strategy based on a support vector machine (SVM) to obtain another classification result. The fusion strategy based on SVM is as follows:

[0098] First, the environment-related feature vectors and traditional visual feature vectors are combined into a new training dataset Then use the SVM classification algorithm to classify this data set and get the classification result C2. Suppose the SVM classification function is f SVM (x), where x is the input feature vector, that is, the combined feature vector The classification result C2 is expressed as:

[0099]

[0100] in, represents the environment-related feature vector extracted under abnormal environmental conditions, represents the traditional visual feature vector extracted under abnormal environment, f SVM (x) is the classification function of the support vector machine (SVM), which is the core function based on the SVM algorithm and is used to classify the input feature vector x, that is, the combined feature vector Perform classification operations and find a hyperplane to separate samples of different categories to obtain classification results.

[0101] The step S3 of dynamic feature fusion specifically includes fusion under a stable environment state and fusion under an abnormal environment state;

[0102] The results of the fusion under the stable environment state and the abnormal environment state are fused, and different weights are assigned to the two layers of results according to the environmental state and the accuracy factors verified by the previous experiments. The final fusion formula is:

[0103] C final =w1·C1+w2·C2

[0104] Among them C final is the final detection conclusion, w1 is the weight of the result under a stable environment, and w2 is the weight of the result under an abnormal environment.

[0105] The step S5 establishes an adaptive model updating mechanism, specifically to update the parameters and structure of the model in real time according to the data collected by the environmental sensors and the performance feedback of the model in actual application.

[0106] In the process of establishing the adaptive model update mechanism in step S5, when the model performance feedback shows that the detection accuracy drops continuously by more than 5%, in addition to updating the parameters and structure of the model, it should also trigger a re-evaluation of the data collection link to check whether the humidity sensor and wind speed sensor are faulty or whether the frequency of data collection is appropriate.

[0107] The specific implementation method is:

[0108] S1: Data acquisition. When collecting images of insulators, humidity sensors and wind speed sensors are equipped. Specifically, when the humidity is high, the image becomes blurred. At this time, a humidity-adaptive denoising and enhancement algorithm is used. The denoising and enhancement functions used are:

[0109] I enhanced (x,y)=f H (I(x,y),H)

[0110] Among them, I(x,y) represents the original insulator image, where x is the pixel coordinate of the image in the horizontal direction, y is the pixel coordinate of the image in the vertical direction, and I enhanced (x, y) is the image processed by humidity adaptive denoising and enhancement algorithm, f H is the denoising and enhancement function based on humidity H, where H represents the ambient humidity value measured by the humidity sensor;

[0111] The specific form of denoising and enhancement function is:

[0112]

[0113] Among them, a H (H) is the coefficient used to adjust the weight of the original image I(x,y) in the denoising and enhancement process; b H (H) is the same as the Laplace operator The multiplication coefficient; c H (H) is a constant term, is the Laplace operator, which is used to calculate the second-order derivative of the image. For the image function I(x,y), the Laplace operator at the point (x,y) is calculated as follows:

[0114]

[0115] in, Represents the second-order partial derivative of the image function I(x,y) in the x direction, which reflects the change in the curvature of the image in the x direction. It represents the second-order partial derivative of the image function I(x,y) in the y direction, reflecting the change of the curvature of the image in the y direction.

[0116] By collecting data to determine the situation when the wind speed is high, the image is translated and rotated to compensate for the wind speed data V to achieve anti-shake processing. The formula is expressed as:

[0117] I stable (x,y)=M V (V) I jitter (x,y)

[0118] Among them, I jitter (x, y) represents the image that is shaken by wind speed, where x and y are the pixel coordinates of the image in the horizontal and vertical directions, respectively, which are used to determine the position of each pixel in the image. stable (x, y) is the image after anti-shake processing, and the position of each pixel is also represented by x and y coordinates. V (V) is a transformation matrix related to wind speed V, where V represents the wind speed value measured by the wind speed sensor;

[0119] The M V (V) is calculated as follows:

[0120]

[0121] Among them, V x Indicates the horizontal component of wind speed, V y It represents the vertical component value of wind speed, m represents the number of pixels in the horizontal direction, and n represents the number of pixels in the vertical direction.

[0122] S2: Environment-related feature extraction, extracting environment-related features from comprehensive data that integrates environmental information;

[0123] It includes environment-related feature extraction and feature dynamic fusion; the environment-related feature extraction includes temperature gradient feature extraction and humidity change rate feature extraction;

[0124] The temperature gradient feature extraction is obtained by calculating the gradient of the temperature function. The temperature gradient vector The calculation formula is:

[0125]

[0126] in, represents the temperature gradient vector, T(x,y) represents the temperature distribution function of the insulator surface, where x and y are the coordinates of the insulator surface on the two-dimensional plane, and T(x,y) gives the temperature value corresponding to the point. represents the partial derivative of the temperature function T(x,y) in the x direction, Represents the partial derivative of the temperature function T(x,y) in the y direction;

[0127] The humidity change rate feature extraction is to extract the humidity change rate feature of the insulator surface from the humidity data H, and the calculation formula is:

[0128]

[0129] in represents the humidity change rate eigenvector, H(x,y) represents the humidity distribution function on the insulator surface, where x and y are the coordinates of the insulator surface on the two-dimensional plane, represents the partial derivative of the humidity function H(x,y) in the x direction, Represents the partial derivative of the humidity function H(x,y) in the y direction.

[0130] The step S2 of dynamic feature fusion is specifically to use a dynamic fusion algorithm to fuse the environment-related features and the traditional visual features. The dynamic fusion algorithm is expressed as:

[0131]

[0132] in Represents the traditional visual feature vector, which contains the feature information obtained from the insulator image through the traditional visual feature extraction method, including edge features, texture features, and color features. These features are extracted based on the visual attributes of the image itself and are used for subsequent analysis and judgment of the insulator state. represents the environment-related feature vector, is the fusion feature vector obtained after being processed by the dynamic fusion algorithm, t represents the time series variable, and w v (t) is the traditional visual feature vector The weight function at different times t, w e (t) is the environment-related feature vector Weight function at different times t.

[0133] S3: Dynamic feature fusion, which uses a dynamic fusion algorithm to fuse environment-related features with traditional visual features;

[0134] Specifically, a dynamic fusion algorithm is used to fuse environment-related features and traditional visual features. The dynamic fusion algorithm is expressed as:

[0135]

[0136] in represents the traditional visual feature vector, represents the environment-related feature vector, is the fusion feature vector obtained after being processed by the dynamic fusion algorithm, t represents the time series variable, and w v (t) is the traditional visual feature vector The weight function at different times t, w e (t) is the environment-related feature vector Weight function at different times t.

[0137] S4: Based on the hierarchical fusion of environmental states, a hierarchical fusion model is constructed to divide the features into different levels for fusion according to the environmental states;

[0138] The step S4 of dynamic feature fusion specifically includes fusion under a stable environment state and fusion under an abnormal environment state;

[0139] The fusion under the stable environment is specifically when the environment is relatively stable, that is, when the temperature, humidity and wind speed are within the normal range, the traditional visual features are fused by a simple weighted average method to obtain a preliminary classification result, and the formula is:

[0140]

[0141] in, represents the i-th traditional visual feature vector, where i ranges from 1 to n, and w v1i is the i-th traditional visual feature vector The weight, w v1i is the i-th traditional visual feature vector The weight of That is, the sum of the weights of all traditional visual feature vectors is 1 to ensure the rationality of weighted average. C1 is the preliminary classification result obtained by fusing traditional visual feature vectors through simple weighted average method;

[0142] When the environment is abnormal, that is, the temperature is too high, the humidity is too high, or the wind speed is too high, the environment-related features and the traditional visual features are fused according to the specific situation of the abnormal environment using a fusion strategy based on a support vector machine (SVM) to obtain another classification result. The fusion strategy based on SVM is as follows:

[0143] First, the environment-related feature vectors and traditional visual feature vectors are combined into a new training dataset Then use the SVM classification algorithm to classify this data set and get the classification result C2. Suppose the SVM classification function is f SVM (x), where x is the input feature vector, that is, the combined feature vector The classification result C2 is expressed as:

[0144]

[0145] in, represents the environment-related feature vector extracted under abnormal environmental conditions, represents the traditional visual feature vector extracted under abnormal environment, f SVM (x) is the classification function of the support vector machine (SVM), which is the core function based on the SVM algorithm and is used to classify the input feature vector x, that is, the combined feature vector Perform classification operations and find a hyperplane to separate samples of different categories to obtain classification results.

[0146] The step S4 of dynamic feature fusion specifically includes fusion under a stable environment state and fusion under an abnormal environment state;

[0147] The results of the fusion under the stable environment state and the abnormal environment state are fused, and different weights are assigned to the two layers of results according to the environmental state and the accuracy factors verified by the previous experiments. The final fusion formula is:

[0148] Cfinal =w1·C1+w2·C2

[0149] Among them C final is the final detection conclusion, w1 is the weight of the result under a stable environment, and w2 is the weight of the result under an abnormal environment.

[0150] S5: Establish an adaptive model update mechanism to update the model parameters and structure based on the data collected by environmental sensors and the performance feedback of the model in actual applications.

Claims

1. A damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion, characterized in that: The following steps are involved: S1: Data acquisition, while collecting insulator images, equipped with humidity sensors and wind speed sensors; S2: Environment-related feature extraction, extracting environment-related features from comprehensive data that integrates environmental information; S3: Dynamic feature fusion, which uses a dynamic fusion algorithm to fuse environment-related features with traditional visual features; S4: Based on the hierarchical fusion of environmental states, a hierarchical fusion model is constructed to divide the features into different levels for fusion according to the environmental states; S5: Establish an adaptive model update mechanism to update the model parameters and structure based on the data collected by environmental sensors and the performance feedback of the model in actual applications.

2. The damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion according to claim 1 is characterized in that: In step S1, when the humidity is high, the image is blurred by collecting data and then a humidity-adaptive denoising and enhancement algorithm is used. The denoising and enhancement function used is: I enhanced (x,y)=f H (I(x,y),H) Among them, I(x,y) represents the original insulator image, where x is the pixel coordinate of the image in the horizontal direction, y is the pixel coordinate of the image in the vertical direction, and I enhanced (x, y) is the image processed by humidity adaptive denoising and enhancement algorithm, f H is the denoising and enhancement function based on humidity H, where H represents the ambient humidity value measured by the humidity sensor; The specific form of denoising and enhancement function is: Among them, a H (H) is the coefficient used to adjust the weight of the original image I(x,y) in the denoising and enhancement process; b H (H) is the same as the Laplace operator The multiplication coefficient; c H (H is a constant term, is the Laplace operator, which is used to calculate the second-order derivative of the image. For the image function I(x,y), the Laplace operator at the point (x,y) is calculated as follows: in, Represents the second-order partial derivative of the image function I(x,y) in the x direction, which reflects the change in the curvature of the image in the x direction. It represents the second-order partial derivative of the image function I(x,y) in the y direction, reflecting the change of the curvature of the image in the y direction.

3. The damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion according to claim 1 is characterized in that: The step S1 determines the situation when the wind speed is high by collecting data, and performs translation and rotation compensation operations on the image by analyzing the wind speed data V to achieve anti-shake processing, and its formula is expressed as: I stable (x,y)=M V (V)·I jitter (x,y) Among them, I jitter (x, y) represents the image that is shaken by wind speed, where x and y are the pixel coordinates of the image in the horizontal and vertical directions, respectively, which are used to determine the position of each pixel in the image. stable (x, y) is the image after anti-shake processing, and the position of each pixel is also represented by x and y coordinates. V (V) is a transformation matrix related to wind speed V, where V represents the wind speed value measured by the wind speed sensor; The M V (V) is calculated as follows: Among them, V x Indicates the horizontal component of wind speed, V y It represents the vertical component value of wind speed, m represents the number of pixels in the horizontal direction, and n represents the number of pixels in the vertical direction.

4. The damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion according to claim 1 is characterized in that: The step S2 includes environment-related feature extraction and feature dynamic fusion; the environment-related feature extraction includes temperature gradient feature extraction and humidity change rate feature extraction; The temperature gradient feature extraction is obtained by calculating the gradient of the temperature function. The temperature gradient vector The calculation formula is: in, represents the temperature gradient vector, T(x,y) represents the temperature distribution function of the insulator surface, where x and y are the coordinates of the insulator surface on the two-dimensional plane, and T(x,y) gives the temperature value corresponding to the point. represents the partial derivative of the temperature function T(x,y) in the x direction, Represents the partial derivative of the temperature function T(x,y) in the y direction; The humidity change rate feature extraction is to extract the humidity change rate feature of the insulator surface from the humidity data H, and the calculation formula is: in represents the humidity change rate eigenvector, H(x,y) represents the humidity distribution function on the insulator surface, where x and y are the coordinates of the insulator surface on the two-dimensional plane, represents the partial derivative of the humidity function H(x,y) in the x direction, Represents the partial derivative of the humidity function H(x,y) in the y direction.

5. The damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion according to claim 4 is characterized in that: The step S3 of dynamic feature fusion is specifically to fuse the environment-related features and the traditional visual features using a dynamic fusion algorithm, and the dynamic fusion algorithm is expressed as: in Represents the traditional visual feature vector, which contains the feature information obtained from the insulator image through the traditional visual feature extraction method, including edge features, texture features, and color features. These features are extracted based on the visual attributes of the image itself and are used for subsequent analysis and judgment of the insulator state. represents the environment-related feature vector, is the fusion feature vector obtained after being processed by the dynamic fusion algorithm, t represents the time series variable, and w v (t) is the traditional visual feature vector The weight function at different times t, w e (t) is the environment-related feature vector Weight function at different times t.

6. The damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion according to claim 4 is characterized in that: The step S3 of dynamic feature fusion is specifically to fuse the environment-related features and the traditional visual features using a dynamic fusion algorithm, and the dynamic fusion algorithm is expressed as: in represents the traditional visual feature vector, represents the environment-related feature vector, is the fusion feature vector obtained after being processed by the dynamic fusion algorithm, t represents the time series variable, and w v (t) is the traditional visual feature vector The weight function at different times t, w e (t) is the environment-related feature vector Weight function at different times t.

7. The damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion according to claim 1 is characterized in that: The step S4 constructs a hierarchical fusion model including fusion under a stable environment state and fusion under an abnormal environment state; The fusion under the stable environment is specifically when the environment is relatively stable, that is, when the temperature, humidity and wind speed are within the normal range, the traditional visual features are fused by a simple weighted average method to obtain a preliminary classification result, and the formula is: in, represents the i-th traditional visual feature vector, where i ranges from 1 to n, and w v1i is the i-th traditional visual feature vector The weight, w v1i is the i-th traditional visual feature vector The weight of That is, the sum of the weights of all traditional visual feature vectors is 1 to ensure the rationality of weighted average. C1 is the preliminary classification result obtained by fusing traditional visual feature vectors through simple weighted average method; When the environment is abnormal, that is, the temperature is too high, the humidity is too high, or the wind speed is too high, the environment-related features and the traditional visual features are fused according to the specific situation of the abnormal environment using a fusion strategy based on a support vector machine (SVM) to obtain another classification result. The fusion strategy based on SVM is as follows: First, the environment-related feature vectors and traditional visual feature vectors are combined into a new training dataset Then use the SVM classification algorithm to classify this data set and get the classification result C2. Suppose the SVM classification function is f SVM (x), where x is the input feature vector, that is, the combined feature vector The classification result C2 is expressed as: in, represents the environment-related feature vector extracted under abnormal environmental conditions, represents the traditional visual feature vector extracted under abnormal environment, f SVM (x) is the classification function of the support vector machine (SVM), which is the core function based on the SVM algorithm and is used to classify the input feature vector x, that is, the combined feature vector Perform classification operations and find a hyperplane to separate samples of different categories to obtain classification results.

8. The damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion according to claim 7 is characterized in that: The step S4 of constructing the hierarchical fusion model specifically includes fusion under a stable environment state and fusion under an abnormal environment state; The results of the fusion under the stable environment state and the abnormal environment state are fused, and different weights are assigned to the two layers of results according to the environmental state and the accuracy factors verified by the previous experiments. The final fusion formula is: C final =w1·C1+w2·C2 Among them C final is the final detection conclusion, w1 is the weight of the result under a stable environment, and w2 is the weight of the result under an abnormal environment.

9. The damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion according to claim 7 is characterized in that: The step S5 establishes an adaptive model updating mechanism, specifically to update the parameters and structure of the model in real time according to the data collected by the environmental sensors and the performance feedback of the model in actual application.

10. The damaged ceramic insulator image recognition and detection method based on multi-algorithm fusion according to claim 7, characterized in that: In the process of establishing the adaptive model update mechanism in step S5, when the model performance feedback shows that the detection accuracy drops continuously by more than 5%, in addition to updating the parameters and structure of the model, it should also trigger a re-evaluation of the data collection link to check whether the humidity sensor and wind speed sensor are faulty or whether the frequency of data collection is appropriate.

Images