An online ferrography-based abrasive particle group feature extraction method

Abstract

The present application relates to the technical field of mechanical equipment fault diagnosis, and particularly relates to a kind of based on online ferrography's abrasive particle group feature extraction method, it is characterized by: including the following steps: obtaining segmented online abrasive particle image;Abrasive particle group feature extraction;Abrasive particle group feature representation;Abrasive particle group feature selection.The present application extracts the abrasive particle number, maximum abrasive particle chain length, maximum abrasive particle chain width, abrasive particle chain width mean value, abrasive particle chain length mean value, abrasive particle area mean value feature parameters based on pixel significance, calculates fractal dimension and abrasive particle coverage area index, the abrasive particle features extracted are characterized and verified by typical online abrasive particle image, eight kinds of characteristic time series under the whole life wear are constructed, the abrasive particle number distribution of different area levels under time series is proposed, and finally abrasive particle coverage area index, abrasive particle number, maximum abrasive particle chain width are selected as the representation parameters of abrasive particle group, which provides important reference index for next step of wear state recognition and prediction.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment fault diagnosis technology, specifically to a method for extracting abrasive particle group features based on online ferrography. Background Technology

[0002] Effective abrasive feature extraction is crucial for analyzing the severity of wear on mechanical equipment. Compared to traditional ferrography, online ferrography offers the advantage of real-time performance. However, several issues remain regarding feature extraction from online abrasive images: First, the abrasive particles in images obtained from online visual ferrography are not isolated but rather comprised of clusters of particles of varying sizes and quantities. Currently, online ferrography primarily uses abrasive particle concentration as an indicator of mechanical equipment wear, and the analysis combining shape, size, and other statistical characteristics of abrasive particles is insufficient. Second, there is limited research on the overall distribution of large and small abrasive particles in the images, leading to inaccurate assessments of equipment anomalies. Third, feature analysis based on wear time series is not sufficiently refined. Summary of the Invention

[0003] The technical problem this invention aims to solve is to provide a method for extracting abrasive group features based on online ferrography. This method extracts pixel-based features such as the number of abrasive particles, maximum abrasive chain length, maximum abrasive chain width, average abrasive chain width, average abrasive chain length, and average abrasive area. It calculates the fractal dimension and abrasive coverage area index, characterizes and verifies the extracted abrasive features using typical online abrasive images, constructs eight feature time series under full-life wear, proposes the distribution of the number of abrasive particles at different area levels under the time series, and finally selects the abrasive coverage area index, the number of abrasive particles, and the maximum abrasive chain width as characterization parameters of the abrasive group. This provides important reference indicators for the next step of wear state identification and prediction.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a method for extracting abrasive grain group features based on online ferrography, comprising the following steps:

[0005] Step S10: Obtain the segmented online abrasive grain image, which is a binary image based on pixel meaning;

[0006] Step S20: Abrasive grain group feature extraction: All abrasive grains in the binary image are labeled using a scanning method. The number of abrasive grains and their equivalent size based on pixels are extracted. There are five types of equivalent abrasive grain sizes, including the maximum abrasive grain chain length, the maximum abrasive grain chain width, the average abrasive grain chain length, the average abrasive grain chain width, and the average abrasive grain area. The fractal dimension, which characterizes the abrasive grain complexity of the image, is calculated using an improved box counting method. The abrasive grain coverage area index is obtained through online visual ferrography.

[0007] Step S30, Abrasive Particle Group Feature Characterization: Based on the abrasive particle group feature extraction in step S20, the four types of typical online abrasive particle images are characterized to verify the characterization ability of the extracted abrasive particle group features.

[0008] Step S40: Selection of abrasive group features: The maximum abrasive chain width is selected as the feature parameter of the abrasive equivalent size, and finally the abrasive coverage area index, the number of abrasives, and the maximum abrasive chain width are selected as the feature parameters characterizing the abrasive group.

[0009] Furthermore, in step S20, the marking steps of the scanning method are as follows: 1. Invert the colors of the binary image and measure the region attributes of the abrasive grain group using the regionprops function; 2. Mark all abrasive grains in the image in the form of a minimum rectangle, determine the starting coordinate point of the minimum rectangle, and mark the abrasive grain with the largest area among all abrasive grains with a red rectangle; 3. Extract the abrasive grain group features based on pixel meaning.

[0010] Furthermore, the number of abrasive grains is the number of rectangles marked in the binary image; the maximum abrasive grain chain length is the maximum length among the marked rectangles in the binary image; the maximum abrasive grain chain width is the maximum width among the marked rectangles in the binary image; the average abrasive grain chain length is the sum of the chain length feature values ​​of all marked rectangles in the binary image; the average abrasive grain chain width is the sum of the chain width feature values ​​of all marked rectangles in the binary image; the average abrasive grain area is the sum of the area feature values ​​of all marked rectangles in the binary image; and the abrasive grain coverage area index is the ratio of the total area of ​​the abrasive grain group to the total area of ​​the spectrum in the pixel sense after online abrasive grain image segmentation.

[0011] Furthermore, the improved box counting method is based on the original box counting method as follows: First, for the number of target pixels in the box, target pixels that do not meet the requirement of one grid number are integrated to obtain the total area, which is then divided by the area of ​​the complete grid to obtain the number of newly added boundary grids. The remaining grids are then added to obtain the overall optimal grid number. Second, the optimal fitting point is sought by using the maximum correlation coefficient between the grid size and the number of boxes covering the target grid. The slope and intercept of the curve are returned by a single curve, and the fractal dimension is calculated.

[0012] Furthermore, the four types of online abrasive images are described below:

[0013] Type a: The different abrasive grain sizes vary greatly and are unevenly distributed, with a small number of abrasive grains;

[0014] Type b: There are many abrasive grains and their distribution is complex. The small abrasive grains are piled up in a thin chain-like distribution, and the main axis of the chain length is horizontal and parallel to the main axis of magnetic force.

[0015] Type C: It is an intermediate state between Type A and Type B, showing smaller abrasive grains, with discrete individual abrasive grain distribution and no aggregation.

[0016] Type d: Individual extremely large abrasive grains appear, with obvious abrasive grain aggregation. The abrasive grain chains are thicker and longer than those of type b. The main axis of the chain length is horizontal and parallel to the main axis of magnetic force, and the spacing between the abrasive grain chains is obvious.

[0017] Furthermore, in step S30, under a 91-hour full-life wear condition, oil samples are taken at fixed time intervals to obtain online abrasive images. Eight abrasive group feature parameters are extracted from the online abrasive images under the full-life wear condition, and eight feature time series based on the full-life wear condition are constructed. The eight feature time series are fractal dimension time series, abrasive number time series, maximum abrasive chain length time series, maximum abrasive chain width time series, abrasive coverage area index time series, abrasive area mean time series, abrasive chain width mean time series, and abrasive chain length mean time series.

[0018] Furthermore, in step S40, the grey relational analysis method is used to screen the five equivalent sizes of abrasive particles using multiple features, and the feature parameters with strong characterization ability are selected based on the degree of correlation.

[0019] Furthermore, the characteristic time series under the entire wear life is divided into three segments: 0-18 hours, 18-72 hours, and 72-91 hours. Three time series under five characteristics—maximum abrasive chain width, maximum abrasive chain length, average abrasive chain length, average abrasive chain width, and average abrasive area—are selected as evaluation sequences. The correlation between the maximum abrasive chain length time series, the maximum abrasive chain width time series, the average abrasive area time series, the average abrasive chain width time series, and the average abrasive chain length time series and the abrasive coverage area index time series is calculated using the grey relational analysis method. The characteristic parameter with the best correlation with the abrasive coverage area index time series is then selected.

[0020] Furthermore, in step S30, based on the abrasive area values ​​extracted from the online abrasive image during the entire lifespan wear process, the abrasive groups in the online abrasive image are divided into ten levels according to their area size, and a distribution of the number of abrasive particles based on different area levels under the time series is proposed.

[0021] As can be seen from the above description, the abrasive grain group feature extraction method based on online ferrography provided by the present invention has the following beneficial effects:

[0022] First, feature extraction of abrasive grains was carried out. All abrasive grains in the online abrasive grain image were marked by scanning method. Feature parameters such as number of abrasive grains, maximum abrasive grain chain length, maximum abrasive grain chain width, mean abrasive grain chain width, mean abrasive grain chain length, and mean abrasive grain area were extracted based on pixel meaning. For binary abrasive grain images, the fractal dimension that characterizes the abrasive grain complexity of the online abrasive grain image was calculated using an improved box counting method.

[0023] Second, the extracted abrasive features were characterized and verified using typical online abrasive images. The extracted features can independently reflect the overall trend of wear, indicating that online abrasive images at different times exhibit different characteristics. Eight feature time series under full-life wear were constructed, and the feature trends under the time series illustrate the changes in the wear severity of mechanical equipment. The distribution of the number of abrasive particles of different area levels under the time series was proposed, providing a more accurate indicator for the analysis of wear severity based on time series.

[0024] Third, to further reduce feature redundancy, a grey relational analysis method was used for multi-feature screening. Based on the degree of correlation, the maximum abrasive chain width was finally selected as the feature for the equivalent size of the abrasive particles. At the same time, to enhance the complementarity of feature parameters, the abrasive coverage area index, the number of abrasive particles, and the maximum abrasive chain width were finally selected as the characterization parameters of the abrasive particle group, providing important reference indicators for the next step of wear state identification and prediction. Attached Figure Description

[0025] Figure 1 Parts (a), (b), (c), and (d) in the image are all abrasive group markers in the online abrasive image.

[0026] Figure 2 A schematic diagram for calculating the fractal dimension using the box counting method.

[0027] Figure 3 Parts (a), (b), (c), and (d) in the image are all online abrasive particle image segmentation diagrams.

[0028] Figure 4 Part (a) in the figure is the fitting plot of the fractal dimension calculation of the online abrasive image with a value of 1.236.

[0029] Figure 4 Part (b) is the fitting plot of the fractal dimension calculation of the online abrasive image with a value of 1.1252.

[0030] Figure 4 Part (c) in the figure is the fitting plot of the fractal dimension of the online abrasive image with a value of 1.3337.

[0031] Figure 4 The (d) part is the fitting plot of the fractal dimension calculation of the online abrasive image with a value of 1.7286.

[0032] Figure 5 These are typical online abrasive grain images for four different types.

[0033] Figure 6 Part (a) in the figure is a fractal dimension time series.

[0034] Figure 6Part (b) in the figure is the time series of abrasive particle number.

[0035] Figure 7 Part (a) in the figure is the long-term sequence of the maximum abrasive grain chain.

[0036] Figure 7 Part (b) in the figure is the time series of the maximum abrasive chain width.

[0037] Figure 8 Part (a) in the figure is the time series of abrasive coverage area index.

[0038] Figure 8 Part (b) in the figure is the time series of the average abrasive grain area.

[0039] Figure 9 Part (a) in the figure is the time series of the average width of the abrasive grain chain.

[0040] Figure 9 Part (b) in the figure is the time series of the average abrasive chain length.

[0041] Figure 10 This represents the distribution of the number of abrasive particles at different area levels over a time series.

[0042] Figure 11 The characteristic time series correlation distributions under the three wear stages are shown. Detailed Implementation

[0043] The present invention will be further described below through specific embodiments.

[0044] This invention is based on the research results of the Fujian Provincial Natural Science Foundation project, project number 2024J011544, titled "Wear Monitoring and Remaining Life Study of Gear Transmission System Based on Microscopic Image Information of Abrasive Particles".

[0045] like Figures 1 to 11 As shown, the present invention provides a method for extracting abrasive grain group features based on online ferrography, comprising the following steps:

[0046] Step S10: Obtain the segmented online abrasive grain image, which is a binary image based on pixel meaning;

[0047] Step S20: Abrasive grain group feature extraction: All abrasive grains in the binary image are labeled using a scanning method. The number of abrasive grains and their equivalent size based on pixels are extracted. There are five types of equivalent abrasive grain sizes, including the maximum abrasive grain chain length, the maximum abrasive grain chain width, the average abrasive grain chain length, the average abrasive grain chain width, and the average abrasive grain area. The fractal dimension, which characterizes the abrasive grain complexity of the image, is calculated using an improved box counting method. The abrasive grain coverage area index is obtained through online visual ferrography.

[0048] Step S30, Abrasive Particle Group Feature Characterization: Based on the abrasive particle group feature extraction in step S20, the four types of typical online abrasive particle images are characterized to verify the characterization ability of the extracted abrasive particle group features.

[0049] Step S40: Abrasive grain group feature selection: Select the maximum abrasive grain chain width as the feature parameter of the abrasive grain equivalent size, and finally select the abrasive grain coverage area index, the number of abrasive grains, and the maximum abrasive grain chain width as feature parameters characterizing the abrasive grain group.

[0050] like Figure 1 As shown in parts (a), (b), (c), and (d) of the image, the overall marking of the abrasive grain group is displayed. The marking steps of the scanning method are as follows: 1. Invert the colors of the binary image and measure the region attributes of the abrasive grain group using the regionprops function; 2. Mark all abrasive grains in the image in the form of a minimum rectangle, determine the starting coordinate point of the minimum rectangle, and mark the abrasive grain with the largest area among all abrasive grains with a red rectangle; 3. Extract the abrasive grain group features based on pixel meaning.

[0051] The number of abrasive grains is the number of rectangles marked in the binary image. The number of abrasive grains varies in a single abrasive grain image. The program automatically calculates the number of rectangles marked in the image; the number of rectangles in the image is the number of abrasive grains. express.

[0052] The maximum abrasive chain length is the maximum length among the marked rectangles in the binary image; the maximum abrasive chain width is the maximum width among the marked rectangles in the binary image; the average abrasive chain length is the sum of the chain length feature values ​​among the marked rectangles in the binary image; the average abrasive chain width is the sum of the chain width feature values ​​among the marked rectangles in the binary image; and the average abrasive area is the sum of the area feature values ​​among the marked rectangles in the binary image.

[0053] Set up online abrasive grayscale image Size is ( and (Representing the number of pixels in the image length and width, respectively). Since the image only contains abrasive grains and the background, abrasive grain features can be extracted from the binary abrasive grain image. The binary image expression is: In the formula, These are the high-frequency components of a binary image. , .

[0054] As defined by the matrix, 0 represents a black pixel in a binary image, and 1 represents a white pixel. By scanning the total number of pixels with a value of 1 in the entire binary image, the total area of ​​the abrasive particles in the image can be calculated. Its calculation expression is:

[0055] .

[0056] The minimum rectangle marked by the abrasive grains has a length and width that are equivalent to the chain length and chain width of a single abrasive grain's connected region, respectively. The maximum length and width of the marked rectangles are automatically selected, each equivalent to the maximum abrasive grain length. Maximum abrasive grain width Its value is calculated by counting the number of pixels with a value of 1. This represents the number of pixels with a value of 1 along the length direction. This represents the number of pixels with a value of 1 in the width direction.

[0057] Based on pixel-level abrasive grain markings, the chain length, chain width, and area of ​​each abrasive grain are calculated using the above formula, yielding the number of abrasive grains in a single abrasive grain image. Therefore, by summing the feature values ​​in the image and taking the average, the average chain length, average area, and average chain width of the abrasive grains can be calculated. Their expressions are as follows:

[0058] ; ; In the formula: This represents the total number of abrasive grains. For the nth abrasive grain value, For the nth abrasive grain value, For the nth abrasive grain value.

[0059] Fractal theory provides methodological support for explicit analysis under chaotic conditions, greatly increasing its credibility. For abrasive particle groups in an image, extracting their fractal features can express complex objects that are difficult to describe quantitatively in a quantitative way, thereby revealing the hidden laws of the system and judging the wear state of mechanical equipment. Fractal dimension is a parameter variable that characterizes fractals and is a measure of the complexity of graphics and natural features. Differences in its definition lead to different methods for calculating fractal dimension, including the criterion method, box counting method, area-perimeter method, probability density method, and cumulative distribution method.

[0060] Box counting is based on binary images and its principle is as follows: Figure 2 As shown, based on abrasive grain segmentation, rows and columns are divided into... The pixels form a side with a length of . A square grid is applied to the abrasive grains to be measured, and the number of grid cells occupying the abrasive grain region is calculated. Plot the side length scale on a log-log coordinate system. With the number of overlapping grids regression slope Its regression slope The negative value represents the fractal dimension of the desired abrasive particle population, and its expression is: .

[0061] Traditional box counting methods are prone to affecting the accuracy of fractal dimension calculation. Firstly, when the abrasive grains are not divisible by the mesh, redundant "marginal" portions are omitted, resulting in an unclear box count. Secondly, data fitting between the number of overlapping meshes and the mesh density causes calculation errors. To address these issues, the improved box counting method involves the following improvements to the original method:

[0062] 1. For the target pixel count in the box, integrate the target pixels that do not meet the requirement of one grid number, divide the total area by the area of ​​the complete grid to get the number of new boundary grids, and add the remaining grids to get the overall optimal grid number; 2. Find the best fit point by using the maximum correlation coefficient of the double logarithm of the grid size and the number of target grid boxes covered, return the slope and intercept through a first-order curve, and calculate the fractal dimension.

[0063] Calculate using an improved box counting method Figure 3 The fractal dimension of four abrasive grain images (parts (a), (b), (c), and (d)) is calculated. The initial minimum grid number is set to 1. The abrasive grain target region is extracted, and the fractal dimension is determined by the change in the edge grid number. and The maximum correlation coefficient R was used to calculate the best-fit point, with 9 fit points set. The regression slopes for different box sizes and their corresponding number of boxes were plotted on a log-log coordinate system. Figure 4 To improve the fractal dimension of the four abrasive grain images calculated using the box counting method, Figure 4 Part (a) of the text Figure 4 Part (b) of the text Figure 4 Part (c) of the middle Figure 4 The values ​​of (d) in the figure are 1.236, 1.1252, 1.3337 and 1.7286, respectively. The abrasive grain complexity is different at different wear stages. The more complex the abrasive grain, the larger the fractal value.

[0064] Drawing upon the construction approach of offline ferrographic abrasive particle concentration quantification indices, an online abrasive particle quantitative index is established. The abrasive particle coverage area index is the ratio of the total area of ​​the abrasive particle group in pixel terms after online abrasive particle image segmentation to the total area of ​​the spectrum. This ratio quantifies the abrasive particle concentration. This value can be directly obtained through online visual ferrography. For a single abrasive particle image, the expression for calculating the abrasive particle coverage area index is:

[0065] In the formula, This represents the sum of the areas of all abrasive grains in the spectrum; , These represent the length and width of the bright area in the image, respectively.

[0066] Abrasive grain images acquired at different wear time points exhibit different characteristics, such as Figure 5 As shown, four types of typical online abrasive grain images are presented, with four images for each type. The online abrasive grain images for the four types are described below:

[0067] Type a: The different abrasive grain sizes vary greatly and are unevenly distributed, with a small number of abrasive grains;

[0068] Type b: There are many abrasive grains and their distribution is complex. The small abrasive grains are piled up in a thin chain-like distribution, and the main axis of the chain length is horizontal and parallel to the main axis of magnetic force.

[0069] Type C: It is an intermediate state between Type A and Type B, showing smaller abrasive grains, with discrete individual abrasive grain distribution and no aggregation.

[0070] Type d: Individual extremely large abrasive grains appear, with obvious abrasive grain aggregation. The abrasive grain chains are thicker and longer than those of type b. The main axis of the chain length is horizontal and parallel to the main axis of magnetic force, and the spacing between the abrasive grain chains is obvious.

[0071] Eight abrasive group parameters were extracted from the four typical online abrasive images mentioned above, as shown in Table 1.

[0072] Table 1. Characteristic values ​​of abrasive particle groups

[0073] .

[0074] Different abrasive characteristic parameters have a certain characterizing ability for abrasive images at different wear time points. The characterizing ability of four typical abrasive images is analyzed as follows:

[0075] I. In images of type a and type c, type b and type d The values ​​are basically the same, but the values ​​of five characteristics—number of abrasive grains, maximum abrasive grain chain width, average abrasive grain chain width, average abrasive grain chain length, and average area—differences are significant. When the values ​​are the same, type a has significantly fewer abrasive grains than type c, but type a has a larger maximum abrasive grain chain width and a larger average abrasive grain chain width, indicating that type a has some abnormally large abrasive grains; types b and d The maximum abrasive chain lengths are similar, but the number of abrasive grains in type b is much greater than that in type d. The maximum abrasive chain width, average chain length, average chain width, and average area of ​​type d are significantly greater than those of type b, indicating that type d has obvious aggregation phenomena.

[0076] II. Analyzing the complexity of abrasive particle distribution, the fractal dimensions of the abrasive particles differ across the four image types. Type b and type d have similar fractal dimension values, with type d having a higher value than type b; types a and c have similar fractal dimension values, with type c having a higher value than type a; the fractal dimension value is related to the number of abrasive particles... This indicates that the fractal dimension of the abrasive particles is related to the abrasive particle concentration in the image;

[0077] 3. The number of abrasive grains in a single type of image is approximately the same, but the values ​​of the maximum abrasive grain chain length and the maximum abrasive grain chain width vary. The chain formation of abrasive grains can be analyzed based on the maximum abrasive grain chain length and the maximum abrasive grain chain width.

[0078] IV. The average width, length, and area of ​​the abrasive chain reflect the overall particle size of the abrasive group. These three characteristic values ​​are indirect representations of abrasive concentration, individual abnormally large abrasive particles, and abrasive chain formation.

[0079] The gear's full-life wear test lasted 91 hours from startup to complete failure. Equipment parameters were adjusted, and oil samples were taken at fixed time intervals during the 91-hour wear test to acquire online abrasive images (e.g., every two minutes). Each abrasive image corresponded to a specific time point. Eight abrasive group feature parameters were extracted from the online abrasive images during the full-life wear test, constructing eight characteristic time series based on the full-life wear test. This provides an important analytical method for online wear monitoring and analysis. The eight characteristic time series are: fractal dimension time series, abrasive number time series, maximum abrasive chain length time series, maximum abrasive chain width time series, abrasive coverage area index time series, average abrasive area time series, average abrasive chain width time series, and average abrasive chain length time series. Figure 6 Part (a) of the text Figure 6 Part (b) of the text Figure 7 Part (a) of the text Figure 7 Part (b) of the text Figure 8 Part (a) of the text Figure 8 Part (b) of the text Figure 9 Part (a) of the text Figure 9 As shown in part (b), based on the extraction of multiple parameters of the abrasive group, eight feature time series based on the wear process are constructed.

[0080] The time series of the above eight features generally exhibit a "bathtub curve" trend. In the figure, blue dots represent the points of change in feature parameter data, and red lines represent the numerical fitting curve. Due to the different types of abrasive particle group features extracted, the feature values ​​differ at different time points. The time series analysis of the above eight feature parameters is as follows:

[0081] (1) The characteristic time series under the whole life wear reflects the changes in four stages: In the first 0-18 hours, the data values ​​of various indicators are relatively high. Around the 3rd hour, the characteristic values ​​suddenly increase and then gradually decrease. By the 18th hour, the values ​​tend to stabilize, indicating that the gear is in the break-in period and the abrasive particles produced are generally large. In the first 18-66 hours, the characteristic indicators are generally stable, indicating that the gear has entered the stable wear period, mainly with small abrasive particles and occasional abnormal values. In the first 66-89 hours, the values ​​of various characteristic indicators increase significantly and show an unstable growth trend, indicating that the gear wear is constantly intensifying. In the first 89-91 hours, the various characteristic indicators show a jump increase, indicating that the wear has reached the limit.

[0082] (2) The two parameters, abrasive coverage area index and abrasive number, represent the changes in abrasive group concentration; the maximum abrasive chain length and the maximum abrasive chain width correspond to the maximum value of abrasive in the image, and the occurrence of abnormally large abrasive can be judged from these two features; during the wear process of the gear throughout its entire life, the fractal dimension changes in the range of 0-2, which is used to characterize the distribution complexity of the abrasive group;

[0083] (3) The overall trend of the constructed abrasive group characteristic time series is consistent with the gear running state under actual wear conditions. The characteristic values ​​at different time points in the characteristic time series are different. Therefore, based on the change of characteristic values ​​at time points, a combination of multiple characteristic parameters can further analyze the wear status of mechanical equipment.

[0084] For wear monitoring of mechanical equipment, the impact of large abrasive particles is more prominent. During the entire wear process of gears, the online visual ferrography instrument samples the oil every 2 minutes and acquires online abrasive images. Based on the abrasive area value extracted in step S20, the abrasive size can be basically determined, but it cannot complete the distribution of the number of abrasive particles of different area levels under the time series. Since the wear conditions reflected by abrasive particles of different sizes are different, in order to further accurately analyze the influencing factors of individual large abrasive particles, based on the abrasive area value extracted from the online abrasive images during the entire wear process, the abrasive groups in the online abrasive images are divided into ten levels according to their area size, as shown in Table 2, and the distribution of the number of abrasive particles of different area levels under the time series is proposed.

[0085] Table 2 Distribution of Abrasive Grain Area Grades

[0086] .

[0087] Since the abrasive grains in the image are composed of abrasive grains with different areas, the number of abrasive grains with different area distributions shows a certain trend on the time series axis during the gear's entire lifespan wear process, such as... Figure 10 As shown, the distribution of abrasive particles differs significantly at different wear stages, as detailed below:

[0088] (1) , , The number of abrasive grains in three area ranges is distributed throughout the time series, with the abrasive grain area in... The highest number of abrasive particles within the range indicates that abrasive particles with an area value below 100 are present throughout the entire time series, which is considered normal abrasive wear.

[0089] (2) , , The number of abrasive particles across the three area ranges gradually decreases with increasing area over the entire wear time series. During the 0-18 hour and 81-91 hour periods, the number of abrasive particles is higher across all three area ranges. The number of abrasive grains gradually decreases during the 18-63 hour wear period; it can be preliminarily determined that when the abrasive grain area value is greater than 100, it is an abnormal abrasive grain.

[0090] (3) Area in , , The abrasive particles in the range mainly appear in the two time periods of 0-9 hours and 81-91 hours. The number of abrasive particles in the three ranges is less in the 9-81 hour range, indicating that the abrasive particles in this range are abnormal abrasive particles and the wear has entered the severe wear stage.

[0091] (4) Regarding area The abrasive particles appearing in certain time periods of the 90-hour period can be used to determine the specific number of abrasive particles and thus indicate that the wear has entered the failure period.

[0092] The abrasive area is divided into 10 levels. The number of abrasive grains corresponding to different areas shows a dynamic trend in the time series. It can be known that there are specific numbers of different abrasive grain areas at each time point, reflecting the relationship between the number of large abrasive grains and small abrasive grains in the image. Therefore, the distribution of the number of abrasive grains with different areas based on the time series can provide a detailed analysis of the severity of wear, and at the same time provide an important reference for the identification and prediction of wear status.

[0093] As can be seen from step S30, each abrasive particle group characteristic parameter can be used for the preliminary judgment of the wear severity. However, for the next step of wear state identification and prediction, multiple characteristic parameters are prone to information accumulation and computational complexity. To address this issue, the above characteristic parameters need to be reasonably screened, and characteristic parameters with stronger characterization ability should be selected. Since the abrasive particle group characteristic parameters change dynamically throughout the entire wear time series, and since grey relational analysis has the characteristics of strong trend quantification and suitability for dynamic processes, grey relational analysis method is selected for feature selection.

[0094] In step S20, five equivalent size features of the abrasive grain group based on pixel meaning were extracted, including maximum abrasive grain chain length, maximum abrasive grain chain width, average abrasive grain chain length, average abrasive grain chain width, and average abrasive grain area. To reduce redundancy of multiple features, the grey relational analysis method was used to screen the five equivalent sizes of the abrasive grains. The features were sorted according to the degree of correlation. The abrasive grain coverage area index was used as a reference sequence, and the five size features were used as an evaluation sequence to select the feature parameters with strong characterization ability.

[0095] Grey relational analysis, also known as "grey relational degree", is a method for judging the degree of similarity and dissimilarity of the overall trends of different time series. It selects features by calculating the correlation degree value of feature time series. The larger the value, the better the correlation, and vice versa.

[0096] The specific steps for calculating the correlation degree are as follows:

[0097] First, determine the reference sequence reflecting the characteristics of the system's behavior and the comparison sequence affecting the system's behavior. Assume... The characteristic reference sequence is identified as the parent sequence, and the characteristic sequences of other factors are represented by... This indicates that the sub-feature sequences are represented by a matrix, as shown in the following formula:

[0098] In the formula, For the number of indicators, ;

[0099] Second, since the characteristic parameters representing abrasive particle groups have different physical meanings, the grey relational degree of the time series needs to be dimensionless to enhance the comparability between data. Commonly used dimensionless methods include the mean method, extreme value method, initial value method, and standard deviation standardization method. Because the mean method can eliminate the influence of dimensions and orders of magnitude while retaining information about the degree of difference in the values ​​of each variable in the time series, this paper chooses the mean method, whose expression is:

[0100] The dimensionless data sequence forms the following matrix: After dimensionless processing, the absolute difference between corresponding elements of each evaluated index sequence and the reference sequence is calculated one by one. The expression is:

[0101] ; Calculate the maximum difference between the target parameter and the influencing parameter. minimum difference Its expression is: ; ;

[0102] III. Calculate the elements in the comparison sequence separately. elements in the reference sequence The correlation coefficient between them is expressed as:

[0103] In the formula, The resolution coefficient, which measures the degree of association, ranges from [0,1]. When the value is small, the resolution is more obvious and the difference between association coefficients becomes larger.

[0104] IV. The correlation between reference and comparison values ​​varies significantly across time series data points, and the number of points is large. Therefore, the average correlation coefficient across the entire time series is calculated to reflect the correlation between the evaluation series and the reference series. The calculation formula is as follows:

[0105] The correlation between each evaluation sequence and the reference sequence is calculated; the closer the value is to 1, the better the correlation.

[0106] Since the gear life wear test process is divided into three stages, to improve the accuracy of the analysis, the characteristic time series under the life wear is divided into three segments: 0-18 hours, 18-72 hours, and 72-91 hours. Three time series under the five characteristics of the maximum abrasive chain width, the maximum abrasive chain length, the average abrasive chain length, the average abrasive chain width, and the average abrasive area are selected as evaluation sequences. According to the grey relational analysis method, the correlation between the maximum abrasive chain length time series, the maximum abrasive chain width time series, the average abrasive area time series, the average abrasive chain width time series, and the average abrasive chain length time series and the abrasive coverage area index time series are calculated respectively. The feature parameter with the best correlation with the abrasive coverage area index time series is selected.

[0107] For each stage of feature data, correlation degree calculation is performed. The five features—maximum abrasive chain width, maximum abrasive chain length, average abrasive chain length, average abrasive chain width, and average abrasive area—are represented by the numbers "1", "2", "3", "4", and "5," respectively. The correlation degree values ​​of the five features under three time series segments are calculated, such as... Figure 11 As shown in (a), (b), and (c), the correlation values ​​of the features under the three time series are different. The correlation between the maximum abrasive chain width and the abrasive coverage area index is the best. The correlation values ​​of the three time series are respectively , , Based on the degree of correlation analysis, the maximum abrasive chain width was ultimately selected as the equivalent size feature of the abrasive grains.

[0108] Based on the above analysis, the maximum abrasive chain width has the strongest characterization ability for the equivalent size of the abrasive group. Similarly, in step S20, the number of abrasive groups and the fractal dimension in the image were obtained. Since the overall change value of the fractal dimension during the entire wear process is small, the overall characterization of the abrasive group is not accurate enough. Therefore, the fractal dimension is not selected as a characteristic parameter to characterize the abrasive group. In summary, the abrasive coverage area index, the number of abrasives, and the maximum abrasive chain width are finally selected as characteristic parameters to characterize the abrasive group.

[0109] Wear particle characteristics are an important source of information for analyzing mechanical equipment faults, while wear particle group characteristics can comprehensively and effectively reflect the wear state. Based on wear particle segmentation, feature extraction of wear particle groups was carried out. The analysis mainly focuses on three aspects: feature extraction methods, extraction of wear particle group features, and feature selection based on correlation. The results are as follows:

[0110] First, feature extraction of abrasive grains was carried out. All abrasive grains in the online abrasive grain image were marked by scanning method. Feature parameters such as number of abrasive grains, maximum abrasive grain chain length, maximum abrasive grain chain width, mean abrasive grain chain width, mean abrasive grain chain length, and mean abrasive grain area were extracted based on pixel meaning. For binary abrasive grain images, the fractal dimension that characterizes the abrasive grain complexity of the online abrasive grain image was calculated using an improved box counting method.

[0111] Second, the extracted abrasive features were characterized and verified using typical online abrasive images. The extracted features can independently reflect the overall trend of wear, indicating that online abrasive images at different times exhibit different characteristics. Eight feature time series under full-life wear were constructed, and the feature trends under the time series illustrate the changes in the wear severity of mechanical equipment. The distribution of the number of abrasive particles of different area levels under the time series was proposed, providing a more accurate indicator for the analysis of wear severity based on time series.

[0112] Third, to further reduce feature redundancy, a grey relational analysis method was used for multi-feature screening. Based on the degree of correlation, the maximum abrasive chain width was finally selected as the feature for the equivalent size of the abrasive particles. At the same time, to enhance the complementarity of feature parameters, the abrasive coverage area index, the number of abrasive particles, and the maximum abrasive chain width were finally selected as the characterization parameters of the abrasive particle group, providing important reference indicators for the next step of wear state identification and prediction.

[0113] The above are merely some specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.

Claims

1. A method for extracting abrasive grain group features based on online ferrography, characterized in that: Includes the following steps: Step S10: Obtain the segmented online abrasive grain image, which is a binary image based on pixel meaning; Step S20: Abrasive grain group feature extraction, obtaining eight abrasive grain group feature parameters: All abrasive grains in the binary image are labeled using a scanning method. The number of abrasive grains and their equivalent size based on pixels are extracted. There are five equivalent sizes: maximum abrasive grain chain length, maximum abrasive grain chain width, average abrasive grain chain length, average abrasive grain chain width, and average abrasive grain area. An improved box counting method is used to calculate the fractal dimension characterizing the abrasive grain complexity of the image. This improved box counting method is based on the original box counting method with the following improvements:

1. For the number of target pixels in the box, target pixels that do not meet the requirement of one grid number are integrated, and the total area is divided by the area of ​​the complete grid to obtain the number of newly added boundary grids. The remaining grids are then added to obtain the overall optimal grid number.

2. The maximum correlation between the grid size and the number of boxes covering the target grid is used. The coefficients are used to find the optimal fitting point. The slope and intercept of a linear curve are returned, and the fractal dimension is calculated. The abrasive grain coverage area index is obtained through online visual ferrography. The number of abrasive grains is the number of rectangles marked in the binary image. The maximum abrasive grain chain length is the maximum length among the marked rectangles in the binary image. The maximum abrasive grain chain width is the maximum width among the marked rectangles in the binary image. The average abrasive grain chain length is the sum of the characteristic values ​​of the chain lengths of all marked rectangles in the binary image. The average abrasive grain chain width is the sum of the characteristic values ​​of the chain widths of all marked rectangles in the binary image. The average abrasive grain area is the sum of the characteristic values ​​of the areas of all marked rectangles in the binary image. The abrasive grain coverage area index is the ratio of the total area of ​​the abrasive grain group to the total area of ​​the spectrum in the pixel sense after online abrasive grain image segmentation. Step S30, Characterization of abrasive particle group features: Eight abrasive group feature parameters were extracted from four different types of online abrasive images and analyzed for characterization, verifying the characterization ability of the extracted abrasive group feature parameters for online abrasive images at different wear time points; Step S40, Selection of abrasive group features: As can be seen from step S30, each abrasive group feature parameter can be used for preliminary judgment of wear severity. In order to reduce the redundancy of multiple features, the grey relational analysis method is used to screen the five abrasive equivalent dimensions. Based on the correlation degree, the maximum abrasive chain width is selected as the feature parameter of the abrasive equivalent size. Finally, the abrasive coverage area index, the number of abrasives, and the maximum abrasive chain width are selected as the feature parameters characterizing the abrasive group.

2. The method for extracting abrasive grain group features based on online ferrography according to claim 1, characterized in that: In step S20, the marking steps of the scanning method are as follows:

1. Invert the colors of the binary image and measure the region attributes of the abrasive grain group using the regionprops function; 2. Mark all abrasive grains in the image in the form of a minimum rectangle, determine the starting coordinate point of the minimum rectangle, and mark the abrasive grain with the largest area among all abrasive grains with a red rectangle; 3. Extract the abrasive grain group features based on pixel meaning.

3. The method for extracting abrasive grain group features based on online ferrography according to claim 1, characterized in that: In step S30, under a full-life wear of 91 hours, oil samples are taken at fixed time intervals to obtain the online abrasive images. Eight abrasive group feature parameters are extracted from the online abrasive images under the full-life wear, and eight feature time series based on the full-life wear are constructed. The eight feature time series are fractal dimension time series, abrasive number time series, maximum abrasive chain length time series, maximum abrasive chain width time series, abrasive coverage area index time series, abrasive area mean time series, abrasive chain width mean time series, and abrasive chain length mean time series.

4. The method for extracting abrasive grain group features based on online ferrography according to claim 3, characterized in that: The characteristic time series under the whole life wear is divided into three segments: 0-18 hours, 18-72 hours, and 72-91 hours. Three time series under five features—maximum abrasive chain width, maximum abrasive chain length, average abrasive chain length, average abrasive chain width, and average abrasive area—are selected as evaluation sequences. The correlation degree between the maximum abrasive chain length time series, the maximum abrasive chain width time series, the average abrasive area time series, the average abrasive chain width time series, and the average abrasive chain length time series and the abrasive coverage area index time series is calculated according to the grey relational analysis method. The feature parameter with the best correlation degree with the abrasive coverage area index time series is selected.

5. The method for extracting abrasive grain group features based on online ferrography according to claim 3, characterized in that: In step S30, based on the abrasive area values ​​extracted from the online abrasive image during the entire lifespan wear process, the abrasive groups in the online abrasive image are divided into ten levels according to their area size. A distribution of the number of abrasive particles at different area levels based on the time series is proposed, providing a more accurate indicator for the analysis of wear severity based on the time series.

Images