Image extraction method and prediction method for lipid tent morphological parameters

Abstract

The application provides a kind of image extraction method and prediction method of grease nipple shape parameter, and the image extraction method of grease nipple shape parameter includes the following steps: S1.image preprocessing;S2.extract the center line of the representative raceway trend as the raceway reference line;S3.grease nipple instance segmentation is carried out to generate binary segmentation mask;S4.get the inclination angle θ1 and θ2, calculate the distance parameter H and width parameter L1, L2;S5 is through the connected domain analysis statistics each observation area ROI in grease nipple instance quantity.The prediction method of grease nipple shape parameter includes: I.grease nipple shape parameter under different working conditions is obtained and training data set is constructed, sample includes working condition parameter vector and grease nipple shape parameter vector;II.neural network model is constructed, input parameter vector is working condition parameter vector, and output parameter vector is grease nipple shape parameter vector;III.after training, neural network model receives working condition parameter vector and outputs the prediction result of grease nipple shape parameter vector.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of grease filament morphological parameter data processing, in particular to a grease filament morphological parameter image extraction method and a grease filament morphological parameter prediction method based on working condition parameters. BACKGROUND

[0002] Grease has been widely used in bearing structures in mechanical systems. Most grease bearings work in a grease-starved state. The composition and behavior of the lubricating film in the contact area under the grease-starved state have an important influence on the film thickness and bearing life. The contact area under the grease-starved state will experience a process of backflow replenishment to increase the film thickness. This process mainly relies on the backflow of grease filaments under shear to replenish the film thickness. Therefore, during the operation of grease-lubricated bearings, the grease filament morphology under the grease-starved state is a key indicator reflecting the lubricating film replenishment capacity and the bearing failure risk. The grease filament is a standardized term for describing the shear morphology of the grease in the contact area. The continuous fibrous structure formed by the orientation of the thickener fibers under the shear effect of the rolling and sliding contact area is used to intuitively reflect the shear behavior and lubrication state of the grease. The grease filament morphology has a continuous trunk structure and can have short branches, but the trunk runs through the whole body. One end (root) of the grease filament is rooted in the raceway surface, and the other end (free end) extends outward to the contact area. The grease filament morphology is affected by the working conditions and will show different orientations, thicknesses and distribution rules. It widely exists in the visualized research (such as optical microscope and fluorescence microscope observation) and lubrication mechanism analysis (such as shear rheological property research) of bearing grease lubrication.

[0003] However, the micro-morphology data processing of the grease filament still faces the following core problems: the current measurement method has limitations, the quantitative standard is missing, the scene adaptation is insufficient, the image interference factors are not handled well, and the existing tools have limitations; The above problems lead to a lack of accurate quantitative basis for the correlation analysis of grease filament morphology data and working condition parameters and friction performance, which directly restricts the reliability of grease lubrication mechanism research and bearing life prediction model. SUMMARY

[0004] The present application aims to overcome the shortcomings of the prior art and provide a grease filament morphological parameter image extraction method and a grease filament morphological parameter prediction method based on working condition parameters, which aims to solve the problems of ambiguous parameter definition and non-uniform extraction method in existing methods, so that the grease filament data of different researchers and under different experimental conditions are comparable, the semi-automatic or even automatic extraction of multiple grease filament morphological parameters is realized, and the purpose of outputting the predicted grease filament morphology under the input working condition parameters is realized by constructing the grease filament morphological parameter database and the grease filament morphology prediction model. At the same time, it provides rich data resources for grease lubrication mechanism research, bearing life prediction model establishment, etc.

[0005] To achieve the above purpose, the present application realizes the following technical solutions:

[0006] An image extraction method of grease streak morphological parameters, characterized by comprising the following steps:

[0007] S1. Obtain the grease streak distribution image of the bearing raceway. After pretreatment of the image, highlight the grease streak contour through image enhancement technology;

[0008] S2. Apply a straight line detection algorithm to the raceway area to extract the center line representing the raceway trend as the raceway reference line;

[0009] S3. Perform grease streak instance segmentation on the pretreated image to generate a binary segmentation mask;

[0010] S4. Skeletonize the binary segmentation mask of a single grease streak to obtain the grease streak backbone skeleton line. Fit the backbone skeleton line to obtain the backbone axis. Calculate the inclination angles θ1 and θ2 of the backbone axis of the grease streaks on the outer ring and the inner ring of the raceway with respect to the raceway reference line, as well as the distance parameter H and the width parameters L1 and L2. The distance parameter H is the vertical straight line distance between the grease skeleton root points closest to the raceway reference line on the outer ring and the inner ring of the raceway. The width parameters L1 and L2 are the mask widths measured in the normal direction of the backbone axis at the grease root on the outer ring and the inner ring of the raceway, respectively;

[0011] S5. With the raceway reference line as the reference, set rectangular observation regions ROI at predetermined distances on the outer ring and the inner ring of the raceway, respectively. Through connected component analysis, count the number of grease instances M1-Mn in each observation region ROI. n n is the number of rectangular observation regions ROI.

[0012] Preferably, in step S1, the method for obtaining the raceway grease streak distribution image is as follows: place the grease streak to be observed under a microscope with a magnification ≥1000 times and a resolution ≥0.1 μm. The outer ring and the inner ring of the raceway each contain at least one group of grease streaks in the image.

[0013] Preferably, in step S1, the image pretreatment method includes a filtering step, which uses Gaussian filtering or median filtering for denoising.

[0014] Preferably, in step S3, an improved U-Net model is used for grease streak instance segmentation of the pretreated image.

[0015] Preferably, the improved U-Net model has an SE attention mechanism module integrated in the encoder, and the decoder uses bilinear interpolation upsampling and fuses skip layer connections.

[0016] Preferably, in step S5, a plurality of observation lines are symmetrically arranged on the inner and outer sides of the raceway radial direction based on the raceway reference line, a rectangular observation region ROI is arranged at the plurality of observation lines, and the number of fat lip instances measured at the plurality of observation lines is respectively recorded as M1-M n .

[0017] Preferably, the distance a between the two observation lines adjacent to the outer ring or the inner ring of the raceway is H a / k, wherein H a = (H1+H2+…+H i ) / i, i is the number of all fat lips in the current image, H a is the average value of the height of all fat lips in the current image, and 1 / k is a preset proportion coefficient.

[0018] A fat lip morphology parameter prediction method based on the above fat lip morphology parameter image extraction method, comprising the following steps:

[0019] I. Obtain fat lip images under different working conditions, obtain fat lip morphology parameters, and construct a training data set. The samples of the training data set include a working condition parameter vector (SR, F, T) and a fat lip morphology parameter vector (θ1, θ2, H, L1, L2, M1-M n ), wherein SR is the slide roll ratio, F is the load, and T is the temperature;

[0020] II. Construct a neural network model trained based on the above training data set. The input parameter vector of the neural network model is the working condition parameter vector, and the output parameter vector is the fat lip morphology parameter vector.

[0021] III. The trained neural network model receives the working condition parameter vector and outputs the prediction result of the fat lip morphology parameter vector.

[0022] Preferably, the neural network model comprises:

[0023] an input layer for inputting the working condition parameter vector;

[0024] a hidden layer provided with at least two layers and adopting a nonlinear activation function;

[0025] an output layer for outputting the fat lip morphology parameter vector;

[0026] and in the forward propagation process of the neural network model, the error between the predicted value and the true fat lip morphology parameter vector is calculated through a loss function, and the gradient of the model parameters with respect to the loss function is solved by automatic differentiation to update the parameters.

[0027] Preferably, the training of the neural network model adopts an Adam optimizer, and is based on an early stopping strategy. When the loss of the validation set does not decrease continuously for N rounds, the training is terminated to avoid overfitting.

[0028] The technical scheme provided by the present application has the following beneficial effects compared with the prior art:

[0029] 1. The image extraction method of the fat whisker morphology parameters provided by the present application realizes semi-automatic or even automatic extraction of multiple morphology parameters of the fat whisker, can comprehensively and accurately obtain parameters such as the inclination angle, the distance, the width and the number through multi-dimensional space sampling and multi-algorithm collaborative processing, has higher precision than traditional manual measurement and simple image processing methods, lays a reliable data foundation for in-depth research on the morphology of the fat whisker, can adapt to complex fat whisker morphology analysis under different working conditions, effectively processes interference such as fat whisker overlap and background noise, improves the robustness and applicability of data processing, and widens the application range of data processing of the fat whisker morphology;

[0030] 2. The physical meaning and extraction method of each parameter are defined, a unified data processing standard of the fat whisker morphology is established, the problems of ambiguous parameter definition and non-uniform extraction method in the prior art are solved, and the fat whisker data under different research personnel and different experimental conditions are comparable;

[0031] 3. The fat whisker morphology parameter prediction method based on working condition parameters provided by the present application can predict the fat whisker morphology parameters under the condition of known working condition parameters by constructing a fat whisker morphology parameter database and a neural network model, can realize standardized management and efficient query of data, provides rich data resources for research on the mechanism of fat lubrication, establishment of a bearing life prediction model and the like, and helps the development of related fields. DETAILED DESCRIPTION

[0032] In order to more clearly illustrate the technical scheme in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiment or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0033] Figure 1 It is a schematic diagram of the raceway distribution;

[0034] Figure 2 It is a schematic diagram of each fat whisker morphology parameter in the image extraction method of the fat whisker morphology parameters of the present application;

[0035] Figure 3 It is a neural network model training and verification curve diagram;

[0036] Figure 4 It is a prediction accuracy diagram of three parameters of the neural network model;

[0037] Figure 5A prediction accuracy map of the neural network model parameter distance parameter H;

[0038] Figure 6 A prediction accuracy map of the neural network model parameter tilt angle θ1;

[0039] Figure 7 A prediction accuracy map of the neural network model parameter tilt angle θ2;

[0040] Wherein, 1-experimental disc, 2-raceway, 3-raceway outer ring, 4-raceway inner ring. DETAILED DESCRIPTION

[0041] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0042] As shown in Figure 1 , the initial distribution of grease on the experimental disc 1 is as follows: after the experimental disc 1 is installed on the MTM ball-disc tester, the grease is uniformly coated on the protrusions at the edge of the disc, and then the tester is started, the ball and the disc are driven by independent motors, and the grease at the protrusions is pushed to both sides by the ball during movement, and the formed track is called the raceway 2, and the raceway 2 on both sides simulates the raceway outer ring 3 and the raceway inner ring 4 in the bearing, and the grease must be distributed radially along the raceway 2 to the raceway outer ring 3 and the raceway inner ring 4.

[0043] An image extraction method of a grease whisker morphology parameter, comprising the following steps:

[0044] S1. Obtain the grease whisker distribution image of the raceway 2, and highlight the grease whisker profile through image enhancement technology after pre-processing the image.

[0045] Specifically, the image shown in Figure 1 is placed under a microscope that can observe the grease whisker morphology, the magnification of the microscope is ≥1000 times, and the resolution is ≥0.1 μm, and the grease whisker on the surface of the raceway 2 is imaged and observed in multiple regions and multiple levels so as to contain as many grease whiskers as possible and at the same time contain the observation regions of the raceway 2, the raceway outer ring 3 and the raceway inner ring 4 on the experimental disc 1, each region covers a large enough range in the horizontal direction to contain a certain number of grease whiskers, so that the detailed distribution of the grease whiskers and the distribution on the raceway outer ring 3 and the raceway inner ring 4 can be known through one picture, the grease whisker image obtained through infrared imaging technology is clear in the imaging process, and the profile, direction and distribution of the grease whisker can be accurately distinguished, thereby reducing the influence of poor imaging quality on subsequent data processing.

[0046] Moreover, in the image preprocessing process, a filtering step is also included, including using Gaussian filtering or median filtering for denoising. After the image preprocessing is completed, the image enhancement technology is used to highlight the fat filament profile, that is, the image enhancement technology of contrast stretching and edge enhancement is used to highlight the profile features of the fat filament, so that the fat filament and the background are clearly distinguished, which is convenient for subsequent feature extraction.

[0047] S2. Apply Hough transform to the raceway 2 region to detect straight lines and extract the center line representing the raceway 2 direction as the raceway reference line Z0.

[0048] For the preprocessed image, the Hough transform is applied to the raceway 2 region to detect straight lines and extract the center line representing the raceway 2 direction as the raceway reference line Z0. Although the raceway 2 is circular as a whole, in the microcosmic local part, the enlarged local raceway 2, raceway inner ring 4 and raceway outer ring 3 can be regarded as straight line segments. Specifically, it includes raceway 2 region extraction, highlighting the edge profile of raceway 2 through edge detection (such as Canny operator), and then using image segmentation technology (such as threshold-based segmentation or U-Net segmentation algorithm mentioned in this paper) to extract the region of interest containing only raceway 2. Then, the Hough straight line transform is applied to the extracted raceway 2 region. This transform can convert the straight line detection problem in image space to the voting problem in parameter space. All edge points in the raceway 2 region are traversed, and the support degree of each straight line in the parameter space is counted. The highest support degree of several straight lines is selected. If there are multiple straight lines representing the raceway 2 direction (such as the edge straight lines on both sides of the raceway 2), the geometric center of these straight lines is calculated to obtain a center line representing the raceway 2 direction, which is defined as the raceway reference line Z0. If only one straight line is detected, and the straight line can accurately represent the raceway 2 direction, it is directly used as the raceway reference line Z0.

[0049] S3. Fat filament instance segmentation is performed on the preprocessed image to generate a binary segmentation mask.

[0050] Specifically, in this embodiment, the preprocessed image is segmented into beard instances based on the improved U-Net model to generate a binary segmentation mask. The encoder of the improved U-Net model integrates an SE (Squeeze-and-Excitation) attention mechanism module. This module adaptively calibrates the weights of the feature channels through a "compression-excitation" mechanism: First, the "compression" step performs global average pooling on each channel of the feature map, compressing it into a channel descriptor; then, the "excitation" step learns the nonlinear relationship between each channel through a fully connected layer that includes dimensionality reduction and restoration, and generates a weight vector; finally, the weight vector is multiplied with the original feature map channel by channel to enhance important features and suppress minor features. In the beard segmentation task, this mechanism can effectively highlight the texture and edge features of beards, suppress background noise, and improve the recognition and segmentation accuracy of small, blurry, or overlapping beards. The SE attention mechanism module can enhance the effective features of the beard region and suppress background interference by learning the weights of the feature channels and spatial locations.

[0051] The decoder of the improved U-Net model employs bilinear interpolation upsampling and fusion skip connections. These skip connections concatenate the high-resolution feature maps from each stage of the encoder with the upsampled feature maps from the corresponding stage of the decoder, thereby fusing shallow details such as edges with deep semantic information to improve segmentation accuracy, particularly for separating fine and overlapping whiskers. For regions with overlapping whiskers, an image segmentation algorithm based on the deep learning-based U-Net model is used to separate the overlapping parts, accurately identify the morphology of individual whiskers, and complete the image preprocessing.

[0052] S4. The binary segmentation mask skeletonization of a single beard is used to obtain the main skeleton line of the beard. The main axis is obtained by linear fitting of the main skeleton line through principal component analysis or least squares method. The tilt angles θ1 and θ2 of the main axis of the beard of the outer ring 3 and inner ring 4 of the raceway with the raceway reference line Z0 are calculated. The distance parameter H and the width parameters L1 and L2 are calculated. The distance parameter H is the vertical straight-line distance between the beard skeleton of the outer ring 3 and inner ring 4 of the raceway and the nearest root point of the raceway reference line Z0. The width parameters L1 and L2 are the mask widths measured along the normal direction of the main axis of the beard root of the outer ring 3 and inner ring 4 of the raceway, respectively.

[0053] Specifically, such as Figure 2 As shown, the tilt angle θ1 is the angle between the main axis of the grease on the outer ring 3 of the raceway and the raceway 2, and the tilt angle θ2 is the angle between the grease in the middle of the inner ring 4 of the raceway and the raceway 2. Both are obtained by taking the average value of repeated tests. This parameter reflects the orientation characteristics of the grease on the surface of the raceway 2. Under different entrainment speeds, the tilt angle θ1 will show obvious changes and can be used to analyze the morphological adjustment of the grease under shear force. Therefore, its size is related to the working condition parameters.

[0054] The distance parameter H is the straight-line distance between the projection points of the two points on the surface of raceway 2: the root point closest to the raceway baseline Z0 in the outer ring 3 grease whisker skeleton and the root point closest to the raceway baseline Z0 in the inner ring 4 grease whisker skeleton. The unit is μm. This parameter reflects the distribution range of grease whiskers on the surface of raceway 2. Its value is related to the generation and migration of grease whiskers and can reflect the coverage of grease whiskers around the contact area. It is of great significance for evaluating the formation range of the lubricating film. The distance parameter H is also related to temperature. For example, for a certain grease, the base oil viscosity and thickener structural stability of the grease reach equilibrium at a temperature of 60℃. The measurement repeatability error of the distance parameter H can be controlled within ±10μm, which is better than ±25μm at 40℃.

[0055] The width parameters L1 and L2 of the main stem of the grease whiskers in the outer raceway 3 and inner raceway 4 are the mask widths measured along the normal direction of the main stem (i.e., the part that coincides with the main stem axis). L1 is the main stem width of the grease whiskers in the outer raceway 3, and L2 is the main stem width of the grease whiskers in the inner raceway 4. That is, multiple measurement points are evenly selected on the main stem of the grease whiskers, and the width in the direction perpendicular to the main stem axis is measured. The average value is taken as the final width parameter, and the unit is μm. This parameter is related to the structure and distribution of the thickener in the grease. Different thickener contents and dispersion states will lead to differences in L1 and L2, which are used to study the influence of the thickener on the grease whisker morphology.

[0056] S5. Using the raceway baseline Z0 as a reference, set rectangular observation areas (ROIs) at predetermined distances on the outer ring 3 and inner ring 4 of the raceway. Count the number of whisker instances (M1 to M) within each ROI using connected component analysis. n .

[0057] Specifically, taking the raceway baseline Z0 as the reference, multiple sets of observation lines are symmetrically set on both the inner and outer sides of the raceway 2 in the radial direction (i.e., the inner raceway 4 and the outer raceway 3). Rectangular observation areas (ROIs) are set at the multiple sets of observation lines, and the number of whisker instances measured by the multiple sets of observation lines are recorded as M1 to M2 respectively. n The distance a = H between two adjacent observation lines on the outer ring 3 or the inner ring 4 of the raceway a / k, where H a =(H1+H2+…+H i ) / i, where i is the number of all lipid whiskers in the current image, i.e., i = M1 + M2 + … M n H a This is the average height of all whiskers in the current image, and 1 / k is a preset scaling factor.

[0058] In this embodiment, as Figure 2As shown, observation lines Z1 to Z3 were set at distances of 200 μm, 300 μm, and 400 μm above the raceway baseline Z0, respectively. The number of tendrils on each line Z1 to Z3 was observed. The observation area ROI was a 100 μm wide rectangular band formed by ±50 μm above and below the observation lines. The number of tendrils was denoted as M1, M2, and M3, respectively. Next, observation lines Z4 to Z6 were set at distances of 200 μm, 300 μm, and 400 μm below Z0, respectively. The number of tendrils on each line Z4 to Z6 was observed. The number of lipid whiskers was determined by constructing observation regions (ROIs) and labeling the number of whiskers as M4, M5, and M6. This yields a number of whisker images with parameters M1 to M6. The images were taken at the same resolution, such as 0.1 μm, or processed to achieve the same resolution. These parameters reflect the spatial distribution density of the lipid whiskers at different locations perpendicular to the raceway baseline Z0, and can demonstrate the distribution characteristics of the lipid whiskers in the direction perpendicular to the raceway 2 surface. This provides data support for studying the migration and replenishment mechanisms of the lipid whiskers.

[0059] The sampling interval is matched with the average height gradient of the whiskers. The height of a single whisker specifically refers to the projected length from the farthest endpoint of the whisker skeleton to its root point in the direction perpendicular to the raceway baseline Z0. In this embodiment, the average height H of all whiskers in the current image is... a =1000μm, 1 / k is 0.1, and the sampling interval is 100μm, that is, the observation lines are equally spaced at 100μm intervals. This can accurately capture the quantitative change trend of the quantitative parameters M1 to M6. Compared with the 200μm interval, the data fitting accuracy is improved by 30%. By setting sampling lines with an interval of 100μm at multiple locations, a three-dimensional quantitative model of the spatial distribution of lipid whiskers is constructed, which can capture the change in the number of lipid whiskers and provide direct data support for analyzing the radial hierarchical replenishment mechanism of lipid whisker migration.

[0060] Experiments have shown that when the suction speed is below 500 mm / s, the grease whisker morphology is prone to disordered distribution due to insufficient shear force, leading to increased measurement errors in the tilt angles θ1 and θ2 between the main axis and the raceway baseline Z0. When the suction speed is above 2000 mm / s, the grease whisker is prone to breakage, and the counting deviation of the quantity parameters M1 to M6 exceeds 20%. The suction speed range of 500-2000 mm / s can ensure that the grease whisker maintains its intact shape and reflects the significant influence of the working conditions on its orientation.

[0061] The axial space sampling strategy of the application carries out regional space sampling, is used for solving the problem of non-uniformity of lipid whisker distribution, is convenient for observing the distribution density of lipid whisker, completes the multi-reference surface image collection of the picture, basically completes the quantification of the basic characteristics of the lipid whisker and determines the distribution of the lipid whisker, combines the extraction of the geometric parameters of the lipid whisker in the picture and the statistics of the spatial distribution, finally, all the parameters in the extraction of the geometric parameters and the statistics of the spatial distribution are combined to obtain the characteristic parameters of the lipid whisker, and the quantification of the lipid whisker characteristics is also completed.

[0062] The film thickness of the contact area under the shearing effect is the main function of the lipid whisker, but the traditional research on the grease lubrication performance and the lipid whisker micro morphology is highly dependent on the expensive and time-consuming SRV testing machine and the optical microscope observation experiment, the formation and evolution mechanism of the lipid whisker is complex, is affected by the nonlinear coupling of various working condition parameters, it is difficult to establish an accurate physical model or an empirical formula, the technical prediction accuracy of the existing empirical formula and the simplified model is limited, especially the prediction ability of the complex and variable working conditions and the micro morphology is insufficient, at present, there is a lack of an efficient and high-precision method which can simultaneously predict the lipid whisker micro morphology under a specific working condition, in view of the limitations of the high cost, low efficiency, narrow application range, insufficient precision and the inability to correlate the micro and macro of these methods, therefore, through the brief description of the existing experimental methods, developing an efficient and accurate lipid whisker morphology prediction method becomes a problem to be solved in the current lubrication engineering technical field.

[0063] In order to solve the above problems, the application further discloses a lipid whisker morphology parameter prediction method based on working condition parameters, comprising the following steps:

[0064] I. Obtain the lipid whisker image under different working conditions, obtain the lipid whisker morphology parameter and construct a training data set, the sample of the training data set includes a working condition parameter vector (SR, F, T) and a lipid whisker morphology parameter vector (θ1, θ2, H, L1, L2, M1~M n ), wherein, SR is the slide roll ratio, F is the load, and T is the temperature.

[0065] Specifically, the data acquisition is completed by the MTM friction and wear testing machine: based on the testing machine, a grease lubrication working condition simulation platform is built, for different formula grease samples (covering differences in base oil viscosity, thickener type and consistency), by adjusting the slide roll ratio SR, the axial load F of the bearing, the temperature T of the experimental disc 1 and other key working condition parameters, the actual working environment of the grease lubrication bearing is simulated, after the working condition is stable, the grease lubrication sample in different areas of the bearing raceway inner ring 4 and raceway outer ring 3 is collected, the original micro morphology of the lipid whisker under the grease shortage state is ensured to be completely retained, and representative basic materials are provided for subsequent observation.

[0066] The sample observation is achieved by means of a high-resolution grease whipper microscope: the sample obtained by the testing machine under different working conditions is placed on the microscope stage, the high-resolution imaging capability is used to microscopically observe the grease whipper in multiple regions and multiple levels of the inner and outer rings of the raceway, and clear grease whipper images are obtained, which can accurately present the profile, direction, distribution characteristics of the grease whipper and the relationship with the surrounding grease matrix, and provide intuitive and accurate original data for subsequent extraction of multi-dimensional grease whipper morphological parameters such as inclination angle, distance, width and number of grease whippers.

[0067] Subsequently, the obtained sample is placed under a microscopic infrared microscope for observation, and the basic characteristics of the grease whipper are observed to obtain clear and high-quality grease whipper photos. The observation parameters of the microscope for observing the grease whipper morphology are preferably: magnification 1000-2000 times, resolution 0.1 μm, and imaging frame rate ≥10 fps. When the magnification is lower than 1000 times, the grease whipper with a diameter of ≤10 μm cannot be clearly identified, resulting in omission of the width parameter extraction; when the magnification is higher than 2000 times, the single field coverage is too small (≤50 μm x 50 μm), and it is difficult to reflect the spatial distribution density of the grease whipper. The magnification range of 1000-2000 times can simultaneously consider the identification of small components and the observation of macroscopic distribution.

[0068] Subsequently, the grease whipper pictures obtained are processed to obtain grease whipper morphological parameters, which are combined with different working condition parameters to form a structured database with MTM testing machine loading conditions and grease whipper characteristic quantization, facilitating the modeling of a neural network for predicting grease whipper morphology in the future.

[0069] II. A neural network model trained based on the training data set is constructed, and the input parameter vector of the neural network model is a working condition parameter vector, and the output parameter vector is a grease whipper morphological parameter vector.

[0070] Since the input data is structured working condition parameters rather than image data, MLP can efficiently handle the nonlinear relationship between continuous numerical characteristics, and residual connection (ResNet block) is introduced to solve the gradient vanishing problem of deep network, adapt to the complex mapping relationship of the grease whipper evolution process; in terms of network structure design, a basic neural network model is designed, which includes an input layer with a working condition parameter vector as input, two hidden layers with a ReLU activation function, and an output layer with a grease whipper morphological parameter vector as output; in the forward propagation process of the model, the error between the predicted value and the true grease whipper morphological parameter vector is calculated by the MSE loss function, and the gradient of the model parameters is solved by automatic differentiation, and then the parameters are updated; the model training adopts the Adam optimizer, and based on the early stopping strategy, the training is terminated when the validation set loss does not decrease for N consecutive rounds to avoid overfitting.

[0071] The mean square error (MSE) is selected as the core loss function, and the adaptability is reflected in the following aspects: the lipid whisker morphological parameter is a continuous physical quantity, and the prediction belongs to the regression task attribute, which is highly consistent with the mathematical characteristics of MSE, and the loss value can directly reflect the prediction accuracy of the physical parameter, and the experimental error of micro measurement directly corresponds to the physical meaning, the MSE is weighted by the square operation, and the extreme abnormal value with a prediction deviation of > 10 um is punished more, which can force the model to preferentially correct the result deviating from the true value, and compared with the mean absolute error (MAE), the prediction stability can be better guaranteed; at the same time, the MSE is continuous and second-order differentiable in the definition domain, and the smoothness can be efficiently cooperated with the Adam optimizer, so that the gradient change is continuous in the training process, the loss function is avoided to be shocked, the hidden layer parameters are stably converged within 300 rounds, the convergence speed is improved by 20% compared with the model using the SmoothL1Loss, and the final validation set loss is lower; in addition, the decomposability of the MSE supports the construction of a weighted comprehensive loss function, which can independently calculate the weighted fusion of each task loss, retain the relevance of the lipid whisker features and the sliding ratio, temperature and load, and balance the learning priority through the weight, so as to solve the problem that the single task loss ignores the parameter correlation characteristics.

[0072] The model training adopts the Adam optimizer, which can adapt to the sparse gradient and small batch training scene through the fusion of momentum and adaptive learning rate, and can cooperate with the batch normalization processing, perfectly match the complex nonlinear mapping requirement of the lipid whisker prediction model, and is significantly better than the traditional optimizer in convergence speed, stability and generalization ability, thereby providing key support for realizing high-precision prediction.

[0073] In the data preparation stage, 3 sigma criterion is used to identify and eliminate abnormal measurement values, and the training set is reasonably divided, and then the data is standardized / normalized, the normalization processing adopts Z-score standardization, and the data preparation work is completed.

[0074] III. The trained neural network model receives the working condition parameter vector and outputs the prediction result of the lipid whisker morphological parameter vector.

[0075] In order to fully illustrate the actual application effect and prediction accuracy of the present application, a database is constructed by using the method in this embodiment, and the neural network model in the prediction module is trained, in order to verify the generalization ability of the model, a new group of working condition parameters which are not trained by the model is inputted for prediction.

[0076] Figures 3-7 The fitting effect of the neural network model is described, Figure 3 The training and verification loss curve is shown, Figure 3The abscissa of the figure is the model training epoch, and the value range is 0-150; the ordinate is the mean square error (MSE), and the value range is 0-1.0; the figure contains two curves, which are respectively a 'training loss curve' and a 'validation loss curve', wherein the training loss curve reflects the error change trend of the model on the training data set, and the validation loss curve reflects the error change trend of the model on the independent validation data set; from the figure Figure 3 It can be seen that, with the increase of the training epoch, the training loss curve and the validation loss curve both show a gradual downward trend, and when the training epoch reaches about 100, both curves tend to be stable, and the final MSE value is less than 0.2; at the same time, the final difference between the training loss and the validation loss is less than 0.1, and there is no obvious deviation; the above characteristics prove that the neural network model constructed by the application can effectively learn the mapping relationship between the fat filament morphology parameters and the input features, and there is no overfitting or underfitting phenomenon, the model training process is stable, and has reliable learning ability and generalization basis.

[0077] Figures 4-6 The prediction accuracy of each parameter is shown, Figure 4 The prediction accuracy of the distance parameter H, the inclination angles θ1 and θ2, Figure 5 The prediction accuracy of the distance parameter H is shown, Figure 6 and Figure 7 The prediction accuracy of the inclination angles θ1 and θ2 is shown respectively; Figures 4-7 The abscissa of the figure is the actual measured value of the fat filament morphology parameter, and the ordinate is the predicted value output by the model; Figure 4 The black dotted line in the figure and Figures 5-7 The red dotted line in the figure is an 'ideal prediction line', which satisfies the relationship 'predicted value = actual value', and is used as a reference for the prediction accuracy; from the figure Figures 4-7 It can be seen that all the scatter points (including three types of scatter points corresponding to the distance parameter H, the inclination angles θ1 and θ2) are densely distributed on both sides of the ideal prediction line, and there is no obvious deviation trend; among them, more than 90% of the scatter points have a vertical distance from the ideal prediction line less than 10% of the actual value range, and there is no systematic deviation; the above characteristics prove that the neural network model of the application has excellent comprehensive prediction ability for the three core parameters of the fat filament morphology, and the overall consistency of the prediction result and the actual measured value is strong, which can meet the precision requirement of the automatic extraction of the fat filament morphology parameters.

[0078] Next, the prediction situation is described:

[0079] The input working condition is: slip-roll ratio SR = 100%, load F = 10 N, and temperature T = 60℃.

[0080] The model prediction result and the experimental measurement comparison (part of the key parameters) are shown in Table 1.

[0081] Table 1: Comparison of model prediction results and experimental measurements

[0082]

[0083] From the case results, the prediction method provided by the application realizes high-precision prediction of the lipid whisker morphology parameters under new working conditions, the relative error of most key parameters is less than 5%, which proves that the application can not only automatically extract the lipid whisker morphology parameters, but also accurately establish the mapping relationship from the working condition to the morphology, and provides a reliable tool for the quantitative analysis and prediction of the lipid lubrication behavior.

[0084] The above examples are only used to illustrate the technical solutions of the application, rather than limit them; although the application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements will not make the essence of the corresponding technical solutions deviate from the protection scope of the technical solutions of the embodiments of the application.

Claims

1. A method for image extraction of a lipid tubular morphology parameter, characterized by, The method comprises the following steps: S1. Obtain the grease whisker distribution image of the bearing raceway. After preprocessing the image, highlight the grease whisker contour through image enhancement technology; S2. Apply a straight line detection algorithm to the raceway area to extract the center line representing the raceway trend as the raceway reference line; S3. Perform grease whisker instance segmentation on the preprocessed image to generate a binary segmentation mask; S4. Skeletonize the binary segmentation mask of a single grease whisker to obtain the grease whisker stem skeleton line. Fit the stem skeleton line to obtain the stem axis. Calculate the inclination angles θ1 and θ2 of the stem axis of the grease whiskers on the outer ring and the inner ring of the raceway with respect to the raceway reference line, as well as the distance parameter H and the width parameters L1 and L2. The distance parameter H is the vertical straight line distance between the grease whisker skeleton of the outer ring and the inner ring of the raceway and the nearest root point of the raceway reference line. The width parameters L1 and L2 are the mask widths of the grease whisker roots on the outer ring and the inner ring of the raceway measured in the normal direction of the stem axis. S5. Take the raceway reference line as the reference, set rectangular observation regions ROI at predetermined distances on the raceway outer ring and raceway inner ring respectively, and count the number of lipid droplet instances M1-Mn in each observation region ROI through connected domain analysis n , n is the number of rectangular observation regions ROI; In the step S5, a plurality of groups of observation lines are symmetrically arranged on the inner and outer sides of the raceway in the radial direction of the raceway based on the raceway reference line, rectangular observation regions ROI are arranged at the plurality of groups of observation lines, and the numbers of fat thread instances measured at the plurality of groups of observation lines are respectively recorded as M1-Mn. n .

2. The image extraction method of lip thread shape parameters according to claim 1, characterized in that, In step S1, the grease whisker distribution image is obtained as follows: place the grease whisker to be observed under a microscope with a magnification of ≥1000 times and a resolution of ≥0.1 μm. The outer ring and the inner ring of the raceway each contain at least one group of grease whiskers.

3. The image extraction method of lip thread shape parameters according to claim 1, characterized in that, In step S1, the image preprocessing method includes a filtering step, which uses Gaussian filtering or median filtering for denoising.

4. The image extraction method of lip thread shape parameters according to claim 1, characterized in that, In step S3, an improved U-Net model is used for grease whisker instance segmentation of the preprocessed image.

5. The image extraction method of lip thread shape parameters according to claim 4, characterized in that, The improved U-Net model integrates an SE attention mechanism module in the encoder and uses bilinear interpolation upsampling and fusion of skip layer connections in the decoder.

6. The image extraction method of lip thread shape parameters according to claim 1, characterized in that, The distance a between two adjacent observation lines of the raceway outer ring or raceway inner ring is H a / k, wherein H a = (H1+H2+…+H i ) / i, i is the number of all the lipid whiskers in the current image, H a is the average value of the height of all the lipid whiskers in the current image, and 1 / k is a preset proportional coefficient.

7. A method for predicting a lip filament form parameter, based on the image extraction method for the lip filament form parameter according to any one of claims 1 to 6, characterized by, The method comprises the following steps: I. Obtain the image of the lipid whisker under different working conditions, obtain the morphological parameters of the lipid whisker, and construct a training data set. The sample of the training data set includes a working condition parameter vector (SR, F, T) and a lipid whisker morphological parameter vector (θ1, θ2, H, L1, L2, M1~M n ), wherein SR is the slip ratio, F is the load, and T is the temperature; II. Construct a neural network model trained based on the above training data set. The input parameter vector of the neural network model is the working condition parameter vector, and the output parameter vector is the grease whisker morphology parameter vector; III. The trained neural network model receives the working condition parameter vector and outputs the prediction result of the grease whisker morphology parameter vector.

8. The method according to claim 7, wherein The neural network model comprises: An input layer for inputting the working condition parameter vector; A hidden layer with at least two layers using a nonlinear activation function; An output layer for outputting the grease whisker morphology parameter vector; During the forward propagation process of the neural network model, the error between the predicted value and the true grease whisker morphology parameter vector is calculated through a loss function, and the gradient of the model parameters with respect to the loss function is solved through automatic differentiation to update the parameters.

9. The method according to claim 8, wherein The training of the neural network model uses the Adam optimizer, and based on the early stopping strategy, the training is terminated when the validation set loss does not decrease for N consecutive rounds to avoid overfitting.

Images