Equipment appearance damage detection method based on multi-feature fusion

Abstract

The invention discloses an equipment appearance damage detection method based on multi-feature fusion, and relates to the technical field of equipment appearance damage detection, and the method comprises the following steps: constructing an equipment appearance damage image data set; carrying out seven-dimensional feature extraction by adopting a seven-dimensional feature model, and fusing to obtain a seven-dimensional feature vector; performing normalization processing on each feature of the seven-dimensional feature vector; performing damage classification and confidence solution on the normalized seven-dimensional feature vector through a Bayesian probability model of adaptive kernel density estimation to obtain a trained equipment appearance damage detection model; and carrying out actual equipment appearance damage detection by adopting the trained equipment appearance damage detection model. According to the method, the precision, efficiency and scene adaptability of equipment appearance damage detection are improved by combining the seven-dimensional feature model and the Bayesian probability model for adaptive kernel density estimation.

Description

Technical Field

[0001] This application relates to the field of equipment appearance damage detection technology, and in particular to a method for equipment appearance damage detection based on multi-feature fusion. Background Technology

[0002] During long-term service, the appearance of equipment is easily affected by various factors such as environment, load, and material aging, resulting in defects such as fine cracks, inclusions, patches, pitting, rolled-in oxide scale, and scratches. These defects not only damage the materials but may also lead to a decline in material properties, seriously affecting the quality and performance of the product. Therefore, developing efficient and accurate surface damage detection technology is crucial.

[0003] Appearance inspection technology has evolved from manual to automated and intelligent inspection. Early methods primarily relied on traditional image processing techniques, such as edge detection based on the Canny operator and threshold segmentation based on the Otsu algorithm. These methods have low computational complexity and can achieve a certain level of damage detection in simple scenarios, but they are sensitive to image quality and their performance significantly degrades in complex industrial environments such as uneven lighting or oily surfaces. With the development of deep learning technology, object detection networks such as Faster R-CNN and Mask R-CNN have been widely used in damage detection. These methods can achieve accuracy rates of over 90% under ideal conditions, but they have significant limitations: first, they rely on a large number of labeled samples, resulting in high training costs; second, they require high computational resources; and third, their model interpretability is poor, making it difficult to meet the traceability requirements of industrial fields.

[0004] In recent years, image processing-based damage detection technology has been widely used in industrial inspection due to its non-contact and high-efficiency characteristics. However, traditional methods often rely on single feature extraction techniques and have poor adaptability to multiple types and scales of damage.

[0005] In summary, the accuracy, efficiency, and scenario adaptability of existing equipment appearance damage detection technologies need to be improved. Summary of the Invention

[0006] This application provides a method for detecting equipment appearance damage based on multi-feature fusion, which aims to address the shortcomings of existing equipment appearance damage detection technologies in terms of accuracy, efficiency, and scenario adaptability.

[0007] On the one hand, this application provides a method for detecting equipment appearance damage based on multi-feature fusion, including the following steps: Step 1: Construct a dataset of images showing equipment surface damage.

[0008] Step 2: Use a seven-dimensional feature model to extract seven-dimensional features from the images in the equipment appearance damage image dataset and fuse them to obtain a seven-dimensional feature vector.

[0009] Step 3: Normalize each feature of the seven-dimensional feature vector.

[0010] Step four: The normalized seven-dimensional feature vector is classified and its confidence is calculated using a Bayesian probability model based on adaptive kernel density estimation, resulting in a trained equipment appearance damage detection model.

[0011] Step 5: Use the trained equipment appearance damage detection model to perform actual equipment appearance damage detection.

[0012] In one possible implementation, step one involves preprocessing the images in the equipment appearance damage image dataset, including grayscale conversion.

[0013] In one possible implementation, step two, the seven-dimensional feature extraction, includes: Seven features were extracted from the images in the equipment appearance damage image dataset: mean intensity, intensity standard deviation, Fourier energy, edge density, local binary mode entropy, gradient magnitude, and fractal dimension.

[0014] The combination of seven features constitutes a seven-dimensional feature vector, which is used to characterize six types of equipment appearance damage, including: fine cracks, inclusions, patches, pitted surfaces, rolled-in oxide scale, and scratches.

[0015] In one possible implementation, in step three, the normalization process employs min-max normalization.

[0016] In one possible implementation, step four includes: The image data of a single damage type in the equipment appearance damage image dataset is divided into five-fold cross operations to obtain the cross operation training set.

[0017] Calculate the KDE probability density of each sample in the corresponding validation subset of the cross-operation training set.

[0018] The average log-likelihood of the candidate bandwidth is calculated based on the KDE probability density.

[0019] Select the optimal bandwidth from all candidate bandwidths.

[0020] The posterior probability is calculated based on the optimal bandwidth, and the value of the maximum posterior probability is selected as the confidence level.

[0021] In one possible implementation, step four, the five-fold cross division operation, includes: Image data of a single damage type were randomly divided into 5 non-overlapping subsets.

[0022] Each time, one subset is selected as the validation set, and the remaining four subsets are combined into the training subset.

[0023] Repeat the process of partitioning and obtaining the validation set and training subset five times to obtain the cross-operation training set.

[0024] In one possible implementation, in step four, a lower limit for the KDE probability density is set when calculating the KDE probability density.

[0025] In one possible implementation, step four, calculating the average log-likelihood of the candidate bandwidth, includes: Taking the natural logarithm of the KDE probability density yields the single-fold log-likelihood.

[0026] The total likelihood logarithmic value is obtained by summing the log-likelihood of the single fold over 5 cycles.

[0027] The mean log-likelihood is calculated based on the total likelihood log value.

[0028] In one possible implementation, step four, calculating the posterior probability based on the optimal bandwidth, includes: Calculate the optimal KDE probability density based on the optimal bandwidth.

[0029] The joint probability is obtained based on the optimal KDE probability density and the prior probability.

[0030] The ratio of the joint probability to the evidence factor is the posterior probability, where the evidence factor is the sum of the joint probabilities of all valid types.

[0031] The equipment appearance damage detection method based on multi-feature fusion in this application has the following advantages: By combining a seven-dimensional feature model with a Bayesian probability model based on adaptive kernel density estimation, the accuracy, efficiency, and scene adaptability of equipment appearance damage detection are improved.

[0032] By extracting seven types of features, such as mean intensity, and constructing feature vectors, the system comprehensively characterizes six types of equipment appearance damage, including fine cracks. This achieves multi-dimensional complementary characterization of damage attributes, improving the completeness of damage feature representation and the ability to distinguish and identify different damage types.

[0033] Model training was completed by using five-fold cross-operation, optimal bandwidth selection, and posterior probability calculation. The classification and confidence calculation logic of the adaptive kernel density estimation Bayesian model was optimized, which improved the accuracy of damage classification and the reliability of confidence results.

[0034] By setting a lower limit threshold for the KDE probability density, calculation errors caused by excessively small density values ​​are avoided, ensuring the stable and uninterrupted operation of the algorithm and improving the robustness of the damage classification model and the stability of the detection results.

[0035] By calculating joint probability, posterior probability, and evidence factors based on optimal bandwidth, damage type determination and confidence quantification are accurately completed, improving the credibility of equipment appearance damage detection results and the scientific nature of intelligent decision-making. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 A flowchart illustrating the equipment appearance damage detection method based on multi-feature fusion provided in this application embodiment; Figure 2 Examples of six types of equipment appearance damage provided in the embodiments of this application; Figure 3 A seven-dimensional feature damage distribution histogram provided for embodiments of this application; Figure 4 This is a display diagram of the seven-dimensional feature values ​​of a single detected image provided in an embodiment of this application; Figure 5 A heat map showing the comparison of typical appearance damage and defect detection results provided in the embodiments of this application; Figure 6 Heat map of performance indicators of the appearance damage detection model of the detection and evaluation equipment provided in the embodiments of this application. Detailed Implementation

[0038] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0039] like Figure 1 As shown in the figure, this application provides a method for detecting equipment appearance damage based on multi-feature fusion, including the following steps: Step 1: Construct a dataset of images showing equipment surface damage.

[0040] Step 2: Use a seven-dimensional feature model to extract seven-dimensional features from the images in the equipment appearance damage image dataset and fuse them to obtain a seven-dimensional feature vector.

[0041] Step 3: Normalize each feature of the seven-dimensional feature vector.

[0042] Step four: The normalized seven-dimensional feature vector is classified and its confidence is calculated using a Bayesian probability model based on adaptive kernel density estimation, resulting in a trained equipment appearance damage detection model.

[0043] Step 5: Use the trained equipment appearance damage detection model to perform actual equipment appearance damage detection.

[0044] like Figure 2 As shown, specific equipment surface damage mainly includes fine cracks, inclusions, patches, pitted surfaces, rolled-in oxide scale, and scratches, each with unique visual characteristics. For example, fine cracks appear as slender linear structures, typically 0.1-0.5 mm wide, exhibiting low contrast in grayscale images with some edge continuity; inclusions are foreign objects embedded in the equipment surface, presenting irregular blocky shapes with obvious color and texture differences from the substrate, and relatively uniform brightness distribution; patches present as large, continuous areas of abnormal color, with poor texture uniformity, significant brightness variations, and relatively blurred boundaries with the surrounding area; pitted surfaces consist of numerous discrete corrosion pits, exhibiting densely distributed spot characteristics with many irregular edges; rolled-in oxide scale consists of irregular sheet-like structures with obvious texture directionality, typically presenting a layered texture; scratches appear as linear groove structures, ranging in length from a few millimeters to several centimeters, with grayscale values ​​lower than the surrounding area, and relatively clear edges but gentle gradients.

[0045] For example, in step one, the images in the equipment appearance damage image dataset are preprocessed, including grayscale conversion.

[0046] Specifically, in this embodiment, after reading the images from the equipment appearance damage image dataset, they are directly converted into grayscale images, and the three-channel color information is compressed into single-channel intensity information. This is intended to simplify computational complexity and focus on the texture and structural features related to the damage, rather than color information.

[0047] For example, in step two, the seven-dimensional feature extraction includes: Seven features were extracted from the images in the equipment appearance damage image dataset: mean intensity, intensity standard deviation, Fourier energy, edge density, local binary mode entropy, gradient magnitude, and fractal dimension.

[0048] The combination of seven features constitutes a seven-dimensional feature vector, which is used to characterize six types of equipment appearance damage, including: fine cracks, inclusions, patches, pitted surfaces, rolled-in oxide scale, and scratches.

[0049] Specifically, the mean intensity reflects the overall brightness level of the damaged area, indicating that the damage characteristics are caused by strong reflections from materials of varying brightness. For example, low intensity corresponds to indentation. The formula for calculating the mean intensity is: .

[0050] In the formula, Let be the grayscale intensity value of the i-th pixel within the damaged region; N is the total number of pixels within the damaged region. This represents the average gray value of the damaged area, typically ranging from 0 to 255.

[0051] The intensity standard deviation quantifies the dispersion and uniformity of gray values ​​within the damaged area. For example, patches of varying shades and the alternating light and dark pits exhibited by complex pitting corrosion show a higher intensity standard deviation, indicating strong contrast and unevenness within the area. Conversely, rolled-in oxide scale, single inclusions, and shallow, uniform scratches indicate uniformity and consistency within the area, resulting in a lower intensity standard deviation. The formula for calculating the intensity standard deviation is: .

[0052] In the formula, As above, this represents the mean intensity of the region; denoted as the standard deviation of the grayscale values ​​in the region.

[0053] Fourier energy reflects the relative intensity of high-frequency components in an image, such as details, edges, and texture. It is primarily used to assess the periodicity and roughness of texture. High values ​​indicate a high level of texture in the image, typical of fine cracks and rolled-in oxide scale, while low values ​​indicate a smooth image or a lack of periodicity in the texture, commonly seen in inclusions, simple scratches, and certain patches. The formula for calculating Fourier energy is: .

[0054] In the formula, L(u,v) is the logarithmic amplitude spectrum, which is used to enhance visualization and reduce the numerical dynamic range; The peripheral region of the frequency domain is usually defined as excluding the center. The rest of the area; The final characteristic value is the ratio of high-frequency energy to total energy.

[0055] Edge density quantifies the number of edge pixels per unit area, reflecting the complexity and continuity of the damage contour. High values ​​are a strong indicator of fine cracks, as they form a dense mesh-like edge. Low values ​​indicate sparse edges, seen in wide, shallow scratches, large areas of uniform inclusions, or textures rolled into oxide scale that, while periodic, may have low edge contrast. The formula for calculating edge density is: .

[0056] In the formula, E is the number of edge pixels, that is, the total number of pixels that meet the conditions; W and H are the width and height of the image; and D is the final calculated edge density, that is, the proportion of edge pixels to the total number of pixels.

[0057] Local binary mode entropy measures the randomness, complexity, and irregularity of surface texture. High entropy values ​​indicate complex, disordered, and random textures, typical of patchy and pitted surfaces. Medium entropy values ​​correspond to regular but complex textures, such as fine cracks. Low entropy values ​​indicate uniform, simple, and ordered textures, found in directional textures, uniform scratches, and single inclusions in rolled-in oxide scale. The formula for calculating local binary mode entropy is: .

[0058] In the formula, Let K be the probability of the k-th LBP mode, where K is the total number of texture types. In this embodiment, K = 256.

[0059] Gradient amplitude reflects the average intensity and sharpness of the damaged area's edge. A high gradient amplitude indicates a steep intensity change and sharp edges, commonly seen in fine cracks, deep pitting, and high-contrast patches. A low gradient amplitude indicates a gentle intensity change and blurred edges, seen in shallow scratches, rolled-in oxide scale, and inclusions that blend well with the background. The formula for calculating the gradient amplitude is: .

[0060] In the formula, coordinates The morphological gradient value at the location; H and W are the same as above, representing the height and width of the image.

[0061] Fractal dimension quantifies the geometric complexity and space-filling ability of surface structures, exhibiting scale invariance. A high fractal dimension indicates an extremely rough, complex, and irregular surface, typical of fine cracks with a network structure and pitted surfaces with complex pits and patches. A low fractal dimension indicates a relatively smooth and simple surface, seen in rolled-in oxide scale, which may have texture but high self-similarity, smooth scratches, and simple inclusions.

[0062] Based on fractal theory, the fractal dimension is typically calculated using box counting at four scales (2, 4, 8, and 16 pixels). First, the multi-scale box count is calculated using the total count at a single scale, and then a linear fit is performed. and The regression slope is then used to obtain the fractal dimension. The formula for the fractal dimension is: .

[0063] In the formula, s is the box counting scale (in the code). ); For multi-scale box counting, b is the intercept of the linear fit.

[0064] By combining the above seven features, a seven-dimensional feature model of equipment appearance damage can be constructed, which comprehensively represents the characteristics of appearance damage of different types of equipment and provides strong feature support for subsequent classifiers.

[0065] To verify the effectiveness of the proposed method in intelligent damage detection and improve detection efficiency, 1619 damage images from a constructed equipment appearance damage image dataset were analyzed. These included 270 images of fine cracks, 268 images of inclusions, 274 images of patches, 276 images of pitted surfaces, 265 images of rolled-in oxide scale, and 266 images of scratches. The main characteristics and numerical ranges of the seven feature values ​​are shown in Table 1. Statistical analysis of the seven-dimensional feature parameters yielded the distribution histogram. Figure 3 As shown.

[0066] Table 1. Main characteristics and numerical ranges of typical equipment appearance damage types.

[0067] Analysis Table 1 and Figure 3 It can be observed that although the seven-dimensional features of the six typical appearance damages differ in numerical values ​​and distribution characteristics, it is difficult to effectively distinguish the appearance damage type by relying on a single feature. However, it may be easier to carry out damage type identification by fusing multiple features.

[0068] For example, in step three, the normalization process employs min-max normalization.

[0069] Specifically, because the original feature values ​​of the extracted seven-dimensional feature vector vary greatly in scale and range (e.g., the average gray value ranges from 0 to 255, while the Fourier energy ranges from 0 to 1), directly using these features for classification would lead to the large-range features dominating the classification process. Therefore, all successfully extracted features need to be normalized before being stored in the database. The formula for min-max normalization is: .

[0070] In the formula, X is the original eigenvalue. and These are the minimum and maximum values ​​of this feature in all samples of the current database, respectively.

[0071] The normalized features are stored together with the original features in the database to provide standardized input for damage classification.

[0072] For example, step four includes: The image data of a single damage type in the equipment appearance damage image dataset is divided into five-fold cross operations to obtain the cross operation training set.

[0073] Calculate the KDE probability density of each sample in the corresponding validation subset of the cross-operation training set.

[0074] The average log-likelihood of the candidate bandwidth is calculated based on the KDE probability density.

[0075] Select the optimal bandwidth from all candidate bandwidths.

[0076] The posterior probability is calculated based on the optimal bandwidth, and the value of the maximum posterior probability is selected as the confidence level.

[0077] Specifically, damage classification and confidence level calculation are the core decision-making steps of intelligent appearance damage detection methods. The core objective is to determine the damage type of the appearance image of the equipment to be inspected based on the seven-dimensional features (mean intensity, intensity standard deviation, Fourier energy, etc.) extracted and normalized in the early stage, and to determine the reliability of the determination result, i.e., the confidence level, through an adaptive kernel density estimation Bayesian probability model.

[0078] For example, in step four, the five-fold cross division operation includes: Image data of a single damage type were randomly divided into 5 non-overlapping subsets.

[0079] Each time, one subset is selected as the validation set, and the remaining four subsets are combined into the training subset.

[0080] Repeat the process of partitioning and obtaining the validation set and training subset five times to obtain the cross-operation training set.

[0081] Specifically, in this embodiment, a single damage type image data X is randomly divided into 5 non-overlapping subsets. (Each group has approximately the same number of samples), and one subset is selected as the validation set each time. The remaining four subsets are combined into the training subset. ,Right now Repeat the above process 5 times to obtain the complete training set for cross-operation.

[0082] For example, in step four, a lower limit for the KDE probability density is set when calculating the KDE probability density.

[0083] Specifically, the KDE probability density function uses a standard function called a "kernel" and a "stacking" method to construct a smooth, continuous "density" function from a finite number of discrete data points, thereby "estimating" the true but unknown population probability distribution. The formula for the KDE probability density function is: .

[0084] In the formula, The standardized vector (7-dimensional) of the i-th training sample. h is a candidate bandwidth used to characterize the distribution pattern of this type of damage (initially 0.1, and subsequently optimized by iterating from 0.1 to 2.0). For feature dimensions; The kernel function value for the i-th training sample is calculated using the following formula: .

[0085] In the formula, For standardized vectors of Norm (Euclidean distance), calculated as follows: .

[0086] To prevent errors in subsequent logarithmic calculations or divisions due to excessively small density values, a lower limit for the KDE probability density is set. ,Right now: .

[0087] For example, in step four, calculating the average log-likelihood of the candidate bandwidth includes: Taking the natural logarithm of the KDE probability density yields the single-fold log-likelihood.

[0088] The total likelihood logarithmic value is obtained by summing the log-likelihood of the single fold over 5 cycles.

[0089] The mean log-likelihood is calculated based on the total likelihood log value.

[0090] Specifically, in this embodiment, the natural logarithm of the KDE probability density is taken. The log-likelihood of a single fold is obtained. By summing the log-likelihood of 5 iterations, the total log-likelihood can be obtained, as follows: .

[0091] To eliminate the influence of sample size, the bandwidth was obtained. The average fit, the average log-likelihood ,Right now: .

[0092] In the formula, is the total log-likelihood; N is the total number of samples for a single damage type.

[0093] Specifically, in this embodiment, for all candidate bandwidths Traverse within the range (0.1-2.0) and select... The maximum bandwidth is the optimal bandwidth. .

[0094] For example, in step four, calculating the posterior probability based on the optimal bandwidth includes: Calculate the optimal KDE probability density based on the optimal bandwidth.

[0095] The joint probability is obtained based on the optimal KDE probability density and the prior probability.

[0096] The ratio of the joint probability to the evidence factor is the posterior probability, where the evidence factor is the sum of the joint probabilities of all valid types.

[0097] Specifically, in this embodiment, the damage type with the highest posterior probability is selected for prediction, and the value of the highest posterior probability is chosen as the confidence level. The posterior probability is: .

[0098] In the formula, For type The joint probability; This is an evidence factor. , Through optimal bandwidth The calculated optimal KDE probability density; This is the prior probability, which can be determined by the number of training samples. With the total number of training samples The ratio calculation. Evidence factor. The sum of the joint probabilities of all valid types can be obtained from... Calculation, where This is a lower threshold to prevent the sum from being too small (approaching 0) and causing subsequent division by zero errors.

[0099] Specifically, in this embodiment, the aforementioned 1619 images are used as the training set, and another 180 images are selected as the test set. The data is divided into training and test sets in a 9:1 ratio; the data source is equipment metal surface damage collection, the background is removed, only the damage is displayed, and each image is labeled with the damage type. The training set is labeled with the correct names to provide reliable samples for feature statistics and classifier training. During the training phase, the range of seven features is calculated for each type of reliable sample, and the data is input into the database to indicate the range of each type of damage. Subsequently, the prior probability and optimal bandwidth of each type of damage are calculated.

[0100] Taking a single defect image as an example, the seven-dimensional feature values ​​of the image are first obtained and displayed within the range of feature values ​​after training. Then, the Bayesian probability is calculated to perform typical damage classification and confidence score determination.

[0101] Figure 4 It is a seven-dimensional feature value visualization of a single equipment appearance damage image, containing 7 sub-images, each corresponding to one of the seven features. The feature distribution of the six types of damage training sets is displayed with colored scatter dots. The core purpose is to intuitively compare the feature fit between the new sample and the training set. The feature values ​​all fall within the reasonable distribution range of the training set, which verifies the effectiveness of feature extraction and provides intuitive data support, demonstrating the accurate representation capability of the seven-dimensional feature model for damage characteristics.

[0102] The proposed method performed well overall, achieving a comprehensive accuracy of 97.77%, by validating a seven-dimensional model on 180 test images. Of the 180 test samples, the system correctly identified 176, with only 4 samples misclassified. Specific results are as follows... Figure 5 As shown.

[0103] Figure 5 The data shows that one inclusion defect was misidentified as a pitting surface defect, one pitting surface defect was misidentified as an inclusion defect, and two patch defects were incorrectly identified as fine crack defects.

[0104] Meanwhile, to ensure the dynamic updating of the detection model, newly detected images will be manually reviewed. Samples that pass the review will be added to the database and used to re-optimize bandwidth and calculate probabilities for the next detection, thereby further improving classification accuracy and adaptability in practical applications.

[0105] The equipment appearance damage detection model obtained by the method of this application performs well in the three evaluation indicators of accuracy, precision and recall, and has strong overall classification ability, but there is still room for optimization in some damage types.

[0106] First, in terms of overall performance, the accuracy of all damage types is higher than 98.89%, with the two types of rolled-in oxide scale and scratches achieving a perfect accuracy of 100%. The precision and recall are also 100%, indicating that the model has no false negatives or false negatives in the identification of these two types of damage, and the classification effect is ideal.

[0107] However, the performance of other damage types showed some differences. The recall rate for the fine crack type was 100%, indicating that all samples truly belonging to this type were successfully identified with no false negatives; however, its precision was 94.12%, relatively low, suggesting that a small number of samples from other types were misclassified as fine cracks, resulting in false positives. Conversely, the precision rate for the patch type was 100%, meaning that all samples correctly identified by the model as belonging to this type were correct; however, its recall rate of 91.30% was the lowest among the six types, implying that approximately 8.7% of true patch samples were not identified, indicating some false negatives.

[0108] The performance of the model for inclusions and pitted surfaces was relatively balanced, with precision and recall both exceeding 95% and being close in value. This indicates that the model's recognition ability for these two types is stable, and false positives and false negatives are well controlled. The performance index heatmap of the equipment appearance damage detection model is shown below. Figure 6 As shown.

[0109] Figure 6 The recognition performance of six types of equipment appearance damage was evaluated based on three core indicators: accuracy, precision, and recall. The results fully validated that the detection model based on multi-feature fusion has efficient and stable multi-type damage recognition capabilities. The multi-feature fusion strategy is effective and practical, and the recognition performance of most damage types meets the needs of industrial inspection. However, the training set size for each type of damage is only about 270 images, which can easily lead to overfitting due to small sample sizes.

[0110] Table 2 shows the comparative experimental results of the method in this application with neural network models and Ant Forest algorithm: Table 2. Comparative experimental results of the method in this application with neural network models and Ant Forest algorithm.

[0111] Compared with the neural network model, the equipment appearance damage detection model of this application has achieved an absolute improvement of approximately 4.44 percentage points, 3.37 percentage points and 3.42 percentage points in the three core indicators, respectively, and the relative improvement is all above 3.5%. This indicates that the equipment appearance damage detection model of this application has made significant progress in overall classification accuracy, reduction of false alarms and reduction of false negatives.

[0112] Compared to the Ant Forest algorithm, the method in this application, while showing a smaller advantage, still maintains a stable lead across all metrics. Its overall accuracy is 1.11 percentage points higher, precision is 1.39 percentage points higher, and recall is 0.59 percentage points higher. In conclusion, the method in this application better balances the accuracy and completeness of classification.

[0113] The embodiments of this application improve the accuracy, efficiency, and scene adaptability of equipment appearance damage detection by combining a seven-dimensional feature model and a Bayesian probability model with adaptive kernel density estimation.

[0114] By extracting seven types of features, such as mean intensity, and constructing feature vectors, the system comprehensively characterizes six types of equipment appearance damage, including fine cracks. This achieves multi-dimensional complementary characterization of damage attributes, improving the completeness of damage feature representation and the ability to distinguish and identify different damage types.

[0115] Model training was completed by using five-fold cross-operation, optimal bandwidth selection, and posterior probability calculation. The classification and confidence calculation logic of the adaptive kernel density estimation Bayesian model was optimized, which improved the accuracy of damage classification and the reliability of confidence results.

[0116] By setting a lower limit threshold for the KDE probability density, calculation errors caused by excessively small density values ​​are avoided, ensuring the stable and uninterrupted operation of the algorithm and improving the robustness of the damage classification model and the stability of the detection results.

[0117] By calculating joint probability, posterior probability, and evidence factors based on optimal bandwidth, damage type determination and confidence quantification are accurately completed, improving the credibility of equipment appearance damage detection results and the scientific nature of intelligent decision-making.

[0118] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0119] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for detecting equipment surface damage based on multi-feature fusion, characterized in that, Includes the following steps: Step 1: Construct a dataset of images showing equipment surface damage; Step 2: Use a seven-dimensional feature model to extract seven-dimensional features from the images in the equipment appearance damage image dataset and fuse them to obtain a seven-dimensional feature vector; Step 3: Normalize each feature of the seven-dimensional feature vector; Step 4: The normalized seven-dimensional feature vector is classified and the confidence score is calculated using the Bayesian probability model with adaptive kernel density estimation to obtain the trained equipment appearance damage detection model. Step 5: Use the trained equipment appearance damage detection model to perform actual equipment appearance damage detection.

2. The equipment appearance damage detection method based on multi-feature fusion according to claim 1, characterized in that, In step one, the images in the equipment appearance damage image dataset are preprocessed, including grayscale conversion.

3. The equipment appearance damage detection method based on multi-feature fusion according to claim 1, characterized in that, In step two, the seven-dimensional feature extraction includes: Seven features were extracted from the images in the equipment appearance damage image dataset: mean intensity, intensity standard deviation, Fourier energy, edge density, local binary mode entropy, gradient magnitude, and fractal dimension. The combination of seven features constitutes a seven-dimensional feature vector, which is used to characterize six types of equipment appearance damage, including: fine cracks, inclusions, patches, pitted surfaces, rolled-in oxide scale, and scratches.

4. The equipment appearance damage detection method based on multi-feature fusion according to claim 1, characterized in that, In step three, the normalization process employs minimum-maximum normalization.

5. The equipment appearance damage detection method based on multi-feature fusion according to claim 1, characterized in that, Step four includes: The image data of a single damage type in the equipment appearance damage image dataset is divided into five-fold cross operations to obtain the cross operation training set. Calculate the KDE probability density of each sample in the corresponding validation subset of the cross-operation training set; The average log-likelihood of the candidate bandwidth is calculated based on the KDE probability density. Select the optimal bandwidth from all candidate bandwidths; The posterior probability is calculated based on the optimal bandwidth, and the value of the maximum posterior probability is selected as the confidence level.

6. The equipment appearance damage detection method based on multi-feature fusion according to claim 5, characterized in that, In step four, the five-fold cross division operation includes: Image data of a single damage type are randomly divided into 5 non-overlapping subsets; Each time, one subset is selected as the validation set, and the remaining four subsets are used as the training subset; Repeat the process of partitioning and obtaining the validation set and training subset five times to obtain the cross-operation training set.

7. The equipment appearance damage detection method based on multi-feature fusion according to claim 5, characterized in that, In step four, a lower limit for the KDE probability density is set when calculating the KDE probability density.

8. The equipment appearance damage detection method based on multi-feature fusion according to claim 5, characterized in that, In step four, calculating the average log-likelihood of the candidate bandwidth includes: Taking the natural logarithm of the KDE probability density yields the one-fold log-likelihood; The total log likelihood is obtained by summing the log-likelihood of the single fold over 5 cycles. The mean log-likelihood is calculated based on the total likelihood log value.

9. The equipment appearance damage detection method based on multi-feature fusion according to claim 5, characterized in that, In step four, the calculation of the posterior probability based on the optimal bandwidth includes: Calculate the optimal KDE probability density based on the optimal bandwidth; The joint probability is obtained based on the optimal KDE probability density and the prior probability; The ratio of the joint probability to the evidence factor is the posterior probability, where the evidence factor is the sum of the joint probabilities of all valid types.

Images