A rolling bearing failure mode recognition method based on a dynamic gradient correction mechanism

Abstract

A rolling bearing failure mode recognition method based on a dynamic gradient correction mechanism belongs to the technical field of rolling bearing failure mode recognition, and is used to solve the problems of weak small sample generalization and similar mode confusion. The method comprises the following steps: step 1, bearing failure data set construction; step 2, multi-level failure feature extraction: geometric contour feature, texture statistical feature and frequency domain feature are extracted to construct a multi-dimensional failure feature matrix; step 3, expert rule construction and dynamic weight correction: the expert rule is defined, including a feature threshold rule and a position association rule; the back propagation gradient of each category is corrected; step 4, lightweight YOLOv11-OBB model training and reasoning: a YOLOv11-OBB model with an input size of 640*640 is adopted, and the gradient is trained in combination with the expert rule; in the reasoning stage, an expert rule filtering module is introduced to recheck the detection frame with a confidence of less than 0.5.

Description

Technical Field

[0001] This invention relates to the field of rolling bearing failure mode recognition technology, and in particular to an intelligent recognition method for rolling bearing failure modes based on an improved backpropagation gradient using an expert system, applicable to fault diagnosis of industrial equipment. Background Technology

[0002] As a core component of rotating machinery, the failure mode identification of rolling bearings is crucial for preventing major failures. Traditional methods rely on vibration signal analysis, but these suffer from difficulties in sensor installation and distinguishing between multiple coupled failures. While image recognition-based detection techniques can visually reflect the failure morphology, they face two major challenges:

[0003] 1) Sample scarcity: The high cost of obtaining real failure samples leads to underfitting of deep learning models (YOLO model mAP50 is only 38.8% with 367 samples).

[0004] 2) Pattern Similarity: Failure mechanisms such as fatigue spalling and fretting corrosion are complex and require the application of materials science knowledge for identification. Spalling and corrosion have similar failure morphologies, and traditional models lack sufficient discriminative power (difference in similarity category identification <3%).

[0005] Especially in small sample scenarios, the model struggles to learn subtle texture and geometric differences, leading to a higher misclassification rate in industrial applications. While existing YOLO models have made progress in industrial inspection, they lack the ability to distinguish between failure modes with similar morphologies (such as peeling and corrosion) in small sample scenarios. Summary of the Invention

[0006] To address the issues of weak generalization with small sample sizes and confusion between similar patterns, this invention provides a rolling bearing failure mode identification method based on a dynamic gradient correction mechanism.

[0007] The present invention discloses a rolling bearing failure mode identification method based on a dynamic gradient correction mechanism, the method comprising the following steps:

[0008] Step 1: Construction of bearing failure dataset: Collect images of rolling bearing failures and classify them into multiple defect patterns according to preset standards; mark the defect areas with rotated rectangles to generate labels compatible with the rotating target detection model;

[0009] Step 2, Multi-level failure feature extraction: Extract geometric contour features, texture statistical features and frequency domain features to construct a multi-dimensional failure feature matrix;

[0010] Step 3: Expert rule construction and dynamic weight adjustment;

[0011] Define expert rules, including feature threshold rules and location association rules;

[0012] Obtain the average weight of the batch according to expert rules. Then normalize to generate dynamic expert weights for this batch. ;

[0013] According to the formula Correct the backpropagation gradient for each category;

[0014] In the formula, This represents the gradient after correction for the j-th type of failure mode during backpropagation. This represents the original gradient of the j-th type of failure mode during backpropagation. The strength of the training gradient is determined by the intervention of expert weights;

[0015] Step 4: Lightweight YOLOv11-OBB Model Training and Inference:

[0016] A YOLOv11-OBB model with an input size of 640×640 was used, and gradient training was performed using expert rule constraints.

[0017] An expert rule filtering module is introduced during the inference phase to re-examine detection boxes with a confidence level of <0.5.

[0018] Preferably, the process of constructing the bearing failure dataset in step 1 is as follows:

[0019] Step 101: Classify the rolling bearing failure images into 15 failure modes, including 3 types of spalling morphology, 4 types of wear morphology, 4 types of corrosion morphology, 3 types of plastic deformation morphology, and fracture; among them, the 3 types of spalling morphology include slight spalling, severe flaky spalling, and multi-point spalling; the 4 types of wear morphology include slight abrasive wear, matte wear, polished wear, and adhesive wear; the 4 types of corrosion morphology include slight corrosion, severe corrosion, fretting corrosion, and pseudo-indentation; the 3 types of plastic deformation morphology include particle indentation, overload indentation, and overload deformation;

[0020] Step 102: Based on the 15 failure modes in Step 101, set 15 main labels and 8 auxiliary labels for complete bearing, bearing inner ring exterior, bearing inner ring interior, bearing outer ring exterior, bearing outer ring interior, ball, roller surface, roller end face and bearing cage. Use the rotating rectangle annotation tool to annotate the failure area, generate an XML file containing center coordinates, length, width and rotation angle, and convert it into YOLOv11-OBB format labels.

[0021] Step 103: Perform scale standardization on the image, mark the actual size information in the 5% area at the bottom right corner of the image, and realize pixel-physical size mapping through OCR recognition.

[0022] Preferably, the process of multi-level failure feature extraction in step 2 is as follows:

[0023] Step 201: Iteratively estimate the light source direction based on the Lambertian reflection model to generate a light source-free image;

[0024] Step 202: Use HED network to detect the edge of the failure area, extract the contour polygon and calculate the contour features: including contour, aspect ratio, perimeter, area, shape complexity, average curvature and internal and external color difference;

[0025] Step 203: Extract texture features through the gray-level co-occurrence matrix, including gray-level mean, contrast, energy, entropy value and inverse difference moment, and obtain frequency domain features by combining wavelet transform, including wavelet low-frequency energy, high-frequency Shannon entropy and directional contrast.

[0026] Step 204: Construct a 15-dimensional failure feature matrix, including contour features, texture features, and frequency domain features;

[0027] Images are processed in batches. Any image in a batch is labeled with n target detection boxes. The n target detection boxes of the second-to-first image in the batch are subjected to 15-dimensional feature extraction to obtain an n×15 failure feature matrix.

[0028] Preferably, the feature threshold rule is defined based on feature statistics:

[0029] 1) If the contour detection is successful, the weights for slight rust, false indentation, fracture, severe flaking, all categories of plastic deformation morphology, and all categories of wear morphology will be adjusted to +0.5, +0.3, +0.2, +0.3, -0.5, and -0.3, respectively; otherwise, the weights for particle indentation, adhesive wear, severe corrosion, slight flaking, and fracture will be adjusted to +0.6, +0.4, +0.3, -0.5, and -0.7, respectively.

[0030] 2) For features without detected contours, if the contrast is >60, adjust the micro-motion corrosion weight by +0.6; if the high-frequency energy is >5E+04 and the inverse torque is <0.45, adjust the slight abrasive wear weight by +0.8.

[0031] 3) For features with detected contours, if the aspect ratio is >5 and the area is >40, the weights for slight rust and fretting corrosion are adjusted by +1.2 and +0.8 respectively. If the shape complexity is >40 under the condition that the aspect ratio is >5 and the area is >40, the weight for false indentation is adjusted by +0.5.

[0032] If the aspect ratio is (1.8, 5.2) and the mean gray value is (90, 130), then adjust the weight of severe corrosion by +1.0.

[0033] If the aspect ratio is greater than 8, adjust the fracture weight by +1.5, and adjust the weight of all categories of peeling morphology by -0.8.

[0034] If the aspect ratio is (3,8) and the area is (12,32), the weights for severe flaking, multi-point flaking, and slight flaking are adjusted by +1.0, +0.6, and -0.5, respectively; if the aspect ratio is <3 and the area is (2,12), the weight for slight flaking is +0.8.

[0035] 4) Perform global feature analysis on all images:

[0036] If the reverse torque is greater than 100 and the directional contrast is greater than 20, the polishing wear weight and adhesive wear weight are adjusted by +1.2 and +0.7 respectively. Under the condition that the reverse torque is greater than 100 and the directional contrast is greater than 20, if the low frequency energy is higher than 5E+05, the matte wear weight is adjusted by +1.2.

[0037] If the average grayscale value is >150 and the internal and external color difference is >30, then adjust the overload indentation weight by +1.1. If the uniformity is <0.3 under the condition that the average grayscale value is >150 and the internal and external color difference is >30, then adjust the overload deformation weight by +0.8.

[0038] 5) Perform similar type exclusion for major category judgments:

[0039] If it belongs to corrosion morphology and the average curvature is >0.7 and the energy is <0.03, adjust the weight of all categories of corrosion morphology by -0.7, and adjust the weight of all categories of wear morphology by +0.4;

[0040] If the wear morphology is low and the low-frequency energy is <1E+05, then the weights for slight abrasive wear, matte wear, and multi-point spalling are adjusted to -1, -0.8, and +0.6, respectively.

[0041] Preferably, the location association rule is defined based on the manual classification method and auxiliary labels:

[0042] 1) The inner part of the bearing outer ring, the outer part of the bearing inner ring, and the curved surfaces of the balls and rollers are the load-bearing parts:

[0043] If the defect box is completely contained in the union of the connected loaded parts, then the weights of particle indentation, overload indentation, all categories of spalling morphology, and all categories of wear morphology are adjusted by +0.3, +0.3, +0.3, and +0.3, respectively.

[0044] If the union of the defect box and the connected loaded part box intersects, the overload deformation weight and fracture weight are adjusted by +0.3 and +0.3 respectively.

[0045] If the defect box and the connected loaded part box do not intersect, the micro-motion corrosion weight and the pseudo-indentation weight are adjusted by +0.3 and +0.3 respectively.

[0046] 2) If the defect frame is contained within the sphere and roller surfaces, the weights of all categories of peeling morphology and adhesive wear are adjusted by +0.3 and +0.3 respectively;

[0047] 3) If the defect box is contained within the inner raceway, adjust the fretting corrosion weight by +0.6;

[0048] 4) If the defect frame is contained outside the outer raceway, the weights for slight corrosion and severe corrosion are adjusted by +0.6 and +0.6 respectively;

[0049] 5) If the defect frame is contained within the cage, adjust the fracture weight by +0.6;

[0050] 6) If the defect box is contained outside the inner ring of the bearing, the weights for slight spalling, multi-point spalling, and all categories of wear morphology are adjusted by +0.3, +0.3, and +0.2, respectively.

[0051] 7) If the defect frame is contained inside the outer ring of the bearing, the weight of flaking and the weight of all categories of wear morphology are adjusted by +0.3 and +0.2 respectively.

[0052] Preferably, dynamic expert weights The acquisition process:

[0053] Step 301: Initialize the probability weight matrix p of the 15 failure modes into a matrix of all 1s:

[0054] , For the nth object detection bounding box of the i-th image in the batch Initial weights for each type of failure mode;

[0055] Step 302: For the n×15 failure feature matrix within the batch, dynamically adjust the category weights according to expert rules to obtain the n×15 expert weight matrix for the i-th image. Set the non-positive matrix elements of the n×15 expert weight matrix to zero, and take the average weight of the n target detection boxes to obtain the 1×15 expert weight matrix for the i-th image. , For the i-th image, the th Weights of different failure modes;

[0056] Step 303: Calculate the average expert weight for a batch by averaging the expert weights within that batch. :

[0057]

[0058] In the formula, B represents the number of images in the batch;

[0059] Step 304: Assess the average expert weights Normalize:

[0060]

[0061] In the formula, This represents the normalized average expert weight for the corresponding batch.

[0062] The average expert weight for the j-th failure mode within the batch. ;

[0063] It is the minimum value. .

[0064] Preferably, step 4 is as follows:

[0065] Step 401: Use a YOLOv11-OBB model with an input size of 640×640 and train the gradient by combining expert rule constraints.

[0066] Step 402: Introduce an expert rule filtering module during the inference phase to re-examine detection boxes with a confidence level < 0.5.

[0067] Preferably, step 101 is based on the GB / T24611-2020 standard and classifies images based on their appearance.

[0068] Preferably, the rotation annotation process described in step 102 is based on the existing Robalimg tool, and the YOLOv11-OBB format label is a format label that includes the category number and the coordinates of the four points of the rectangle.

[0069] Preferably, the process of generating the light-removed image in step 201 is as follows: a preset light source is separated from the single image light source; the normal field is estimated by surface modeling through Lambertian reflection; RANSAC fitting is performed based on the estimated normal field; the light source is updated by the obtained parameters; and the process is iterated until the difference between the light source direction and the light source vector of the previous iteration is less than the minimum value ε, thus obtaining an approximate light source direction; the normal field and reflectivity are calculated by the Lambertian reflection method, and the value of the reflectivity component can be used as the light-removed image.

[0070] The beneficial effects of this invention are:

[0071] 1. Overcome the bottleneck of small sample learning.

[0072] With 367 samples, the traditional YOLOv11l-OBB model has an mAP50 of only 38.8%, indicating underfitting.

[0073] This invention is the first to transform expert rules into gradient correction factors, breaking through the theoretical bottleneck of small sample learning; by dynamically correcting the gradient using expert rules, the overall mAP50 is improved to 40.8% (↑5.1%); the recognition effect of rare sample categories is optimized, for example, the mAP50 of sheet-like peeling is 0.852 (↑7.6%, originally 0.792), the mAP50 of dot-like peeling is 0.541 (↑4.8%, originally 0.516), and the mAP50 of matte wear is 0.468 (↑125%, originally 0.208).

[0074] 2. To resolve confusion caused by similar failure modes, expert rules are used to accurately suppress and reduce the false positive rate, and multi-dimensional feature fusion is used to enhance the distinguishability.

[0075] 3. Lightweight design.

[0076] The weight matrix operation added in this invention (with negligible computational cost) is completed before training the YOLOv11l-OBB model. The resulting weights are directly applied to gradient correction, reducing the memory required during training.

[0077] 4. It has high technical scalability, with features such as iterable rule base and cross-domain adaptability.

[0078] The rule base is iterative: when adding a new failure mode, only the feature threshold rules need to be expanded (such as adding an electrolytic erosion rule: grayscale mean >180 and high frequency energy <1E+04).

[0079] Cross-domain adaptability: The identification method of this invention is also applicable to the failure detection of mechanical components such as gears and sealing rings. Attached Figure Description

[0080] Figure 1 This is a principle block diagram of a rolling bearing failure mode identification method based on a dynamic gradient correction mechanism as described in this invention;

[0081] Figure 2 This is a schematic diagram of the intersection-union ratio and average accuracy in an example, where... Figure 2 (a) is the intersection-union ratio. Figure 2 (b) represents the average precision;

[0082] Figure 3 This is the object detection result image of the YOLOv11 model for test image 1, where... Figure 3 (a) shows the training results of the model without expert rules. Figure 3 (b) shows the training results of an expert rule model;

[0083] Figure 4 This is the object detection result image of the YOLOv11 model for test image 2, where... Figure 4 (a) shows the training results of the model without expert rules. Figure 4(b) shows the training results of an expert rule model;

[0084] Figure 5 This is the object detection result image of the YOLOv11 model for test image 3, where... Figure 5 (a) shows the training results of the model without expert rules. Figure 5 (b) shows the training results of an expert rule model;

[0085] Figure 6 This is a recognition result image of peeling and wear failure images, in which... Figure 6 (a) shows the recognition results for the peeled image. Figure 6 (b) The recognition effect of wear and tear failure images;

[0086] Figure 7 This is a graph showing the changes in mAP50 and mAP50-95 during the model training process (λ=0, 5), where... Figure 7 (a) mAP50, Figure 7 (b) mAP50-95. Detailed Implementation

[0087] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0088] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0089] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0090] Specific Implementation Method 1: The following is combined with... Figures 1 to 7 This embodiment describes a rolling bearing failure mode identification method based on a dynamic gradient correction mechanism, which includes the following steps:

[0091] Step 1: Construction of bearing failure dataset: Collect images of rolling bearing failures and classify them into multiple defect patterns according to preset standards; mark the defect areas with rotated rectangles to generate labels compatible with the rotating target detection model;

[0092] Step 2, Multi-level failure feature extraction: Extract geometric contour features, texture statistical features and frequency domain features to construct a multi-dimensional failure feature matrix;

[0093] Step 3: Expert rule construction and dynamic weight adjustment;

[0094] Define expert rules, including feature threshold rules and location association rules;

[0095] Obtain the average weight of the batch according to expert rules. Then normalize to generate dynamic expert weights for this batch. ;

[0096] According to the formula Correct the backpropagation gradient for each category;

[0097] In the formula, This represents the gradient after correction for the j-th type of failure mode during backpropagation. This represents the original gradient of the j-th type of failure mode during backpropagation. The strength of the training gradient is determined by the intervention of expert weights;

[0098] Step 4: Lightweight YOLOv11-OBB Model Training and Inference:

[0099] A YOLOv11-OBB model with an input size of 640×640 was used, and gradient training was performed using expert rule constraints.

[0100] An expert rule filtering module is introduced during the inference phase to re-examine detection boxes with a confidence level of <0.5.

[0101] The core innovation of this invention:

[0102] 1. Multi-level feature fusion: Combining contour features (aspect ratio > 5, etc.), global features (grayscale mean, high-frequency Shannon entropy, etc.) with position discrimination (auxiliary labels) to construct highly discriminative expert rules;

[0103] 2. Dynamic Weight Adjustment: Expert rules (based on feature thresholds and positional relationships) are transformed into a weight matrix. Based on highly discriminative expert rules, a condition-triggered weight adjustment mechanism is designed to dynamically adjust the backpropagation gradient. The classification probability distribution is dynamically optimized during the model inference phase.

[0104] 3. Small sample optimization: The expert rule guidance weight is reset to zero / weighted to solve the problem of insufficient generalization caused by sample scarcity.

[0105] 4. Lightweight deployment: The YOLOv11-OBB model is used, combined with expert rules to filter low-confidence detection boxes.

[0106] The method of this invention is based on Figure 1 The system architecture is implemented and consists of four modules: data construction, feature extraction, expert rules, and gradient correction.

[0107] Step 1, the process of constructing the bearing failure dataset, is as follows:

[0108] Step 101: Based on the GB / T24611-2020 standard, and based on the image morphology, the failure images of rolling bearings are classified into 15 failure modes, including 3 types of spalling morphology, 4 types of wear morphology, 4 types of corrosion morphology, 3 types of plastic deformation morphology, and fracture. Among them, the 3 types of spalling morphology include slight spalling, severe flaky spalling, and multi-point spalling; the 4 types of wear morphology include slight abrasive wear, matte wear, polished wear, and adhesive wear; the 4 types of corrosion morphology include slight corrosion, severe corrosion, fretting corrosion, and pseudo-indentation; and the 3 types of plastic deformation morphology include particle indentation, overload indentation, and overload deformation.

[0109] Step 102: Based on the 15 failure modes in Step 101, set 15 main labels and 8 auxiliary labels for complete bearing, bearing inner ring exterior, bearing inner ring interior, bearing outer ring exterior, bearing outer ring interior, ball, roller surface, roller end face and bearing cage. Use the rotating rectangle annotation tool to annotate the failure area, generate an XML file containing center coordinates, length, width and rotation angle, and convert it into YOLOv11-OBB format labels. The YOLOv11-OBB format labels are format labels that contain the category number and the coordinates of the four points of the rectangle.

[0110] Step 103: Perform scale standardization on the image, mark the actual size information in the 5% area at the bottom right corner of the image, and realize pixel-physical size mapping through OCR recognition.

[0111] Step 2, the process of multi-level failure feature extraction, is as follows:

[0112] Step 201: Iteratively estimate the light source direction based on the Lambertian reflection model to generate a light source-free image;

[0113] The process of generating a de-illuminated image is as follows: A single-image light source is separated from a preset light source; surface modeling is performed using Lambertian reflection to estimate the normal field; RANSAC fitting is performed based on the estimated normal field; the light source is updated using the obtained parameters; and this process is iterated until the difference between the light source direction and the light source vector of the previous iteration is less than the minimum value ε. The approximate light source direction is obtained, and the normal field and reflectivity are calculated using the Lambertian reflection method. The value of the reflectivity component can be used as the image without the light source.

[0114] Step 202: Use HED network to detect the edge of the failure area, extract the contour polygon and calculate the contour features: including contour, aspect ratio, perimeter, area, shape complexity, average curvature and internal and external color difference;

[0115] The algorithm uses the HED deep learning network to detect the edges of failure regions, combines the Douglas-Peucker algorithm to extract contour polygons, and employs anisotropic dilation with an elliptic kernel to enhance contour continuity. It calculates the RGB color difference between the inside and outside of the contour at equal distances, obtains the actual perimeter L' by calculating the lengths of the lines connecting multiple points, and calculates the actual area A' inside the contour using the Gaussian area formula. Finally, it calculates the contour complexity C based on the actual perimeter and area. The average curvature of the contour is calculated based on the vertex curvature of the approximate polygon. The contour's circumscribed rectangle is traversed, and the aspect ratio of the smallest circumscribed rectangle is obtained as the contour's aspect ratio.

[0116] Step 203: Extract texture features through the gray-level co-occurrence matrix, including gray-level mean, contrast, energy, entropy value and inverse difference moment, and obtain frequency domain features by combining wavelet transform, including wavelet low-frequency energy, high-frequency Shannon entropy and directional contrast.

[0117] Calculate the gray-level mean for the failure area, and calculate the contrast CON, energy ASM, entropy ENT, and inverse difference moment IDM through the gray-level co-occurrence matrix.

[0118] Step 204: Construct a 15-dimensional failure feature matrix, including contour features, texture features, and frequency domain features;

[0119] Images are processed in batches. Any image in a batch is labeled with n target detection boxes. The n target detection boxes of the second-to-first image in the batch are subjected to 15-dimensional feature extraction to obtain an n×15 failure feature matrix.

[0120] The expert rules in step 3 include feature threshold rules and location association rules;

[0121] The feature threshold rule is defined based on feature statistics:

[0122] 1) If the contour detection is successful, the weights for slight rust, false indentation, fracture, severe flaking, all categories of plastic deformation morphology, and all categories of wear morphology will be adjusted to +0.5, +0.3, +0.2, +0.3, -0.5, and -0.3, respectively; otherwise, the weights for particle indentation, adhesive wear, severe corrosion, slight flaking, and fracture will be adjusted to +0.6, +0.4, +0.3, -0.5, and -0.7, respectively.

[0123] 2) For features without detected contours, if the contrast is >60, adjust the micro-motion corrosion weight by +0.6; if the high-frequency energy is >5E+04 and the inverse torque is <0.45, adjust the slight abrasive wear weight by +0.8.

[0124] 3) For features with detected contours, if the aspect ratio is >5 and the area is >40, the weights for slight rust and fretting corrosion are adjusted by +1.2 and +0.8 respectively. If the shape complexity is >40 under the condition that the aspect ratio is >5 and the area is >40, the weight for false indentation is adjusted by +0.5.

[0125] If the aspect ratio is (1.8, 5.2) and the mean gray value is (90, 130), then adjust the weight of severe corrosion by +1.0.

[0126] If the aspect ratio is greater than 8, adjust the fracture weight by +1.5, and adjust the weight of all categories of peeling morphology by -0.8.

[0127] If the aspect ratio is (3,8) and the area is (12,32), the weights for severe flaking, multi-point flaking, and slight flaking are adjusted by +1.0, +0.6, and -0.5, respectively; if the aspect ratio is <3 and the area is (2,12), the weight for slight flaking is +0.8.

[0128] 4) Perform global feature analysis on all images:

[0129] If the reverse torque is greater than 100 and the directional contrast is greater than 20, the polishing wear weight and adhesive wear weight are adjusted by +1.2 and +0.7 respectively. Under the condition that the reverse torque is greater than 100 and the directional contrast is greater than 20, if the low frequency energy is higher than 5E+05, the matte wear weight is adjusted by +1.2.

[0130] If the average grayscale value is >150 and the internal and external color difference is >30, then adjust the overload indentation weight by +1.1. If the uniformity is <0.3 under the condition that the average grayscale value is >150 and the internal and external color difference is >30, then adjust the overload deformation weight by +0.8.

[0131] 5) Perform similar type exclusion for major category judgments:

[0132] If it belongs to corrosion morphology and the average curvature is >0.7 and the energy is <0.03, adjust the weight of all categories of corrosion morphology by -0.7, and adjust the weight of all categories of wear morphology by +0.4;

[0133] If the wear morphology is low and the low-frequency energy is <1E+05, then the weights for slight abrasive wear, matte wear, and multi-point spalling are adjusted to -1, -0.8, and +0.6, respectively.

[0134] Location association rules are defined based on manual classification methods and auxiliary labels as follows:

[0135] 1) The inner part of the bearing outer ring, the outer part of the bearing inner ring, and the curved surfaces of the balls and rollers are the load-bearing parts:

[0136] If the defect box is completely contained in the union of the connected loaded parts, then the weights of particle indentation, overload indentation, all categories of spalling morphology, and all categories of wear morphology are adjusted by +0.3, +0.3, +0.3, and +0.3, respectively.

[0137] If the union of the defect box and the connected loaded part box intersects, the overload deformation weight and fracture weight are adjusted by +0.3 and +0.3 respectively.

[0138] If the defect box and the connected loaded part box do not intersect, the micro-motion corrosion weight and the pseudo-indentation weight are adjusted by +0.3 and +0.3 respectively.

[0139] 2) If the defect frame is contained within the sphere and roller surfaces, the weights of all categories of peeling morphology and adhesive wear are adjusted by +0.3 and +0.3 respectively;

[0140] 3) If the defect box is contained within the inner raceway, adjust the fretting corrosion weight by +0.6;

[0141] 4) If the defect frame is contained outside the outer raceway, the weights for slight corrosion and severe corrosion are adjusted by +0.6 and +0.6 respectively;

[0142] 5) If the defect frame is contained within the cage, adjust the fracture weight by +0.6;

[0143] 6) If the defect box is contained outside the inner ring of the bearing, the weights for slight spalling, multi-point spalling, and all categories of wear morphology are adjusted by +0.3, +0.3, and +0.2, respectively.

[0144] 7) If the defect frame is contained inside the outer ring of the bearing, the weight of flaking and the weight of all categories of wear morphology are adjusted by +0.3 and +0.2 respectively.

[0145] The above feature threshold rules are derived from the range of feature statistics of failure points extracted from failure images. The location association rules are based on the location judgment of failure modes determined manually. The weighting of these rules is generally related to the number of feature images obtained.

[0146] The process of dynamically adjusting weights based on expert rules is as follows:

[0147] Step 301: Initialize the probability weight matrix p of the 15 failure modes into a matrix of all 1s:

[0148] , For the nth object detection bounding box of the i-th image in the batch Initial weights for each type of failure mode;

[0149] Step 302: For the n×15 failure feature matrix within the batch, dynamically adjust the category weights according to expert rules to obtain the n×15 expert weight matrix for the i-th image. Set the non-positive matrix elements of the n×15 expert weight matrix to zero, and take the average weight of the n target detection boxes to obtain the 1×15 expert weight matrix for the i-th image. , For the i-th image, the th Weights of different failure modes;

[0150] Step 303: Calculate the average expert weight for a batch by averaging the expert weights within that batch. :

[0151]

[0152] In the formula, B represents the number of images in the batch;

[0153] It is a 1×15 matrix.

[0154] Step 304: Assess the average expert weights Normalize:

[0155]

[0156] In the formula, This represents the normalized average expert weight for the corresponding batch. It is a 1×15 matrix.

[0157] The average expert weight for the j-th failure mode within the batch. ;

[0158] It is the minimum value. .

[0159] Step 304, according to the formula Correct the backpropagation gradient for each category;

[0160] The backpropagation gradients of high-probability classes are dynamically amplified. The backpropagation gradients of all 15 classes are increased according to the weight matrix, with targeted amplification of gradients for high-weight classes to accelerate model iteration and output. A confidence-weighted mechanism is used to optimize the model output.

[0161] To address the convergence and accuracy issues in training with few samples, this invention proposes a backpropagation gradient scheme that accelerates the optimization of important features by combining categories with low zero-probability classes with weighted high-probability classes. In traditional methods, inaccurate expert weights can lead to slower convergence, while excessive weights can cause gradient explosion. This invention statistically analyzes fifteen feature classes of failure modes from different categories in a database, selects appropriate expert rules, and increases the accuracy of expert weights. To prevent gradient explosion caused by inaccurate expert weights, the strength of expert weight intervention needs to be experimented with when correcting gradients using expert weights, and the optimal expert weight intervention strength is selected for model training.

[0162] Step 4, the process of training and inference of the lightweight YOLOv11-OBB model, is as follows:

[0163] Step 401: Use a YOLOv11-OBB model with an input size of 640×640 and train the gradient by combining expert rule constraints.

[0164] Step 402: Introduce an expert rule filtering module during the inference phase to re-examine detection boxes with a confidence level < 0.5.

[0165] The following is a specific example to illustrate this.

[0166] The bearing failure mode recognition system uses Windows 11 64-bit as its operating system, NVIDIA GeForce RTX 4060 (8GB) GPU, 32GB of RAM, Intel(R) Core(TM) i9 13900HX CPU @3.90GHz CPU, and PyTorch deep learning framework. Default parameters are used for training and validation. Rotated bounding boxes are used for annotation. The YOLOv11 model used for training is YOLOv11l-OBB. 367 bearing failure images are used, with 4 images per training iteration. The input image resolution is 640×640. Training lasts for 100 epochs. The expert weight intervention intensity λ is selected from six values: 0, 0.5, 2, 5, 10, and 50. Training with λ=0 indicates training without expert rules.

[0167] Commonly used evaluation metrics for object detection tasks include precision, recall, mAP50, and mAP50-95. Precision represents the proportion of correctly detected objects in all bounding boxes where the model has detected them; recall represents the detection rate in all ground truth bounding boxes; precision and recall focus on the balance between "correct" and "complete" detections.

[0168] In the object detection task results, mAP50 and mAP50-95 are obtained based on the concepts of average precision (AP) and intersection over union (IoU). IoU is the ratio of the intersection area of ​​the predicted box and the ground truth box to the union area, which reflects the degree of overlap between the predicted box and the ground truth box. Given an IoU threshold m, an IoU higher than m indicates that the predicted box is correct. For a single class, given an IoU threshold of 0.5, all predicted boxes are arranged in descending order of confidence. Boxes are selected from highest to lowest confidence, one at a time, two at a time, and so on. The Precision-Recall value is calculated for each selected box until all boxes are selected. The Precision-Recall points are connected to form a curve, and the area under the curve is numerically integrated to obtain a numerical AP. The AP for all classes is then averaged to obtain mAP50. mAP50-95 represents the AP calculated at IoU thresholds from 0.5 to 0.95 (10 thresholds in increments of 0.05). The average of these AP values ​​is then calculated. The calculation is illustrated below. Figure 2 As shown.

[0169] During training, the maximum value of mAP50-95 was used as the optimal model, and the first time mAP50 reached 0.35 was used as the convergence mark. The epoch at which convergence was reached was used to judge the convergence speed, and the epoch at which the optimal model was reached was used as the label to judge the training efficiency of the model. The optimal model evaluation index with different expert weight intervention intensities is shown in Table 1.

[0170] Table 1. Training evaluation indicators for different expert intervention weights

[0171]

[0172] The training results of the YOLOv11 model without expert rules and the training results of the YOLOv11 model with expert rules (λ=5) were compared. The trained model was then used to identify bearing failure images, and the results are as follows. Figures 3 to 5 As shown. Without the use of expert rules, similar failure characteristics, such as corrosion and flaking, can easily lead to confusion and misjudgment, such as... Figure 5 As shown.

[0173] The model trained with expert weights and an intervention strength of λ=5 performed as follows in identifying peeling and adhesive wear in Chapter 4: Figure 6 As shown, the adhesive wear and material transfer (steel material peeling) of the inner ring can be effectively identified.

[0174] Table 1 shows that the intervention of expert weights significantly improved recall, slightly increased mAP50 and mAP50-95, and decreased model prediction accuracy. There was no significant difference in convergence speed. At λ=50, all indicators decreased significantly, and gradient explosion occurred. mAP50 and mAP50-95, as indicators combining Precision, Recall, and IoU, best evaluate the overall performance of the model training. Table 1 shows that the training effect is best when λ=5. The neural network achieves a comprehensive mAP50 of 40.8% for identifying 15 types of failures, while the neural network without expert rules achieves 38.8%, an improvement of 5.1%. At λ=5, the neural network has the highest mAP50 for sheet-like peeling and overload deformation, reaching 0.852 and 0.824 respectively. The changes in mAP50 and mAP50-95 during model training are shown in the table. Figure 7 As shown, the overall accuracy of the model with expert weights is slightly higher than that of the model without expert weights.

[0175] Both models achieved mAP50 values ​​less than 0.2 for adhesive wear, minor corrosion, and particle indentation. The model without an expert system achieved an mAP50 value less than 0.2 for matte wear. These categories resulted in lower overall training accuracy. The limited training samples for particle indentation, matte wear, polished wear, and minor corrosion, coupled with the potential for misclassification due to the similarity of adhesive wear morphology to spalling and corrosion, contributed to the lower overall training accuracy. The model with expert weights demonstrated higher recognition rates for flaking spalling and overload deformation, achieving mAP50 values ​​greater than 0.8, specifically 0.852 and 0.824, respectively. In contrast, the model without expert weights only achieved an mAP50 greater than 0.8 for overload deformation, at 0.862.

[0176] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A method for identifying failure modes of rolling bearings based on a dynamic gradient correction mechanism, characterized in that, The method includes the following steps: Step 1: Construction of bearing failure dataset: Collect images of rolling bearing failures and classify them into multiple defect patterns according to preset standards; mark the defect areas with rotated rectangles to generate labels compatible with the rotating target detection model; Step 2, Multi-level failure feature extraction: Extract geometric contour features, texture statistical features and frequency domain features to construct a multi-dimensional failure feature matrix; Step 3: Expert rule construction and dynamic weight adjustment; Define expert rules, including feature threshold rules and location association rules; Obtain the average weight of the batch according to expert rules. Then normalize and generate dynamic expert weights for this batch. ; According to the formula Correct the backpropagation gradient for each category; In the formula, This represents the gradient after correction for the j-th type of failure mode during backpropagation. This represents the original gradient of the j-th type of failure mode during backpropagation. The strength of the training gradient is determined by the intervention of expert weights; Step 4: Lightweight YOLOv11-OBB Model Training and Inference: A YOLOv11-OBB model with an input size of 640×640 was used, and gradient training was performed using expert rule constraints. An expert rule filtering module is introduced during the inference phase to re-examine detection boxes with a confidence level of <0.

5.

2. The rolling bearing failure mode identification method based on dynamic gradient correction mechanism according to claim 1, characterized in that, Step 1, the process of constructing the bearing failure dataset, is as follows: Step 101: Classify the rolling bearing failure images into 15 failure modes, including 3 types of spalling morphology, 4 types of wear morphology, 4 types of corrosion morphology, 3 types of plastic deformation morphology, and fracture; among them, the 3 types of spalling morphology include slight spalling, severe flaky spalling, and multi-point spalling; the 4 types of wear morphology include slight abrasive wear, matte wear, polished wear, and adhesive wear; the 4 types of corrosion morphology include slight corrosion, severe corrosion, fretting corrosion, and pseudo-indentation; the 3 types of plastic deformation morphology include particle indentation, overload indentation, and overload deformation; Step 102: Based on the 15 failure modes in Step 101, set 15 main labels and 8 auxiliary labels for complete bearing, bearing inner ring exterior, bearing inner ring interior, bearing outer ring exterior, bearing outer ring interior, ball, roller surface, roller end face and bearing cage. Use the rotating rectangle annotation tool to annotate the failure area, generate an XML file containing center coordinates, length, width and rotation angle, and convert it into YOLOv11-OBB format labels. Step 103: Perform scale standardization on the image, mark the actual size information in the 5% area at the bottom right corner of the image, and realize pixel-physical size mapping through OCR recognition.

3. The rolling bearing failure mode identification method based on dynamic gradient correction mechanism according to claim 2, characterized in that, Step 2, the process of multi-level failure feature extraction, is as follows: Step 201: Iteratively estimate the light source direction based on the Lambertian reflection model to generate a light source-free image; Step 202: Use HED network to detect the edge of the failure area, extract the contour polygon and calculate the contour features: including contour, aspect ratio, perimeter, area, shape complexity, average curvature and internal and external color difference; Step 203: Extract texture features through gray-level co-occurrence matrix: including gray-level mean, contrast, energy, entropy value and inverse difference moment; combine wavelet transform to obtain frequency domain features: including wavelet low-frequency energy, high-frequency Shannon entropy and directional contrast. Step 204: Construct a 15-dimensional failure feature matrix, including contour features, texture features, and frequency domain features; Images are processed in batches. Any image in a batch is labeled with n target detection boxes. The n target detection boxes of the second-to-first image in the batch are subjected to 15-dimensional feature extraction to obtain an n×15 failure feature matrix.

4. The rolling bearing failure mode identification method based on dynamic gradient correction mechanism according to claim 1, characterized in that, The feature threshold rule is defined based on feature statistics: 1) If the contour detection is successful, the weights for slight rust, false indentation, fracture, severe flaking, all categories of plastic deformation morphology, and all categories of wear morphology will be adjusted to +0.5, +0.3, +0.2, +0.3, -0.5, and -0.3, respectively; otherwise, the weights for particle indentation, adhesive wear, severe corrosion, slight flaking, and fracture will be adjusted to +0.6, +0.4, +0.3, -0.5, and -0.7, respectively. 2) For features without detected contours, if the contrast is >60, adjust the micro-motion corrosion weight by +0.6; if the high-frequency energy is >5E+04 and the inverse torque is <0.45, adjust the slight abrasive wear weight by +0.

8. 3) For features with detected contours, if the aspect ratio is >5 and the area is >40, the weights for slight rust and fretting corrosion are adjusted by +1.2 and +0.8 respectively. If the shape complexity is >40 under the condition that the aspect ratio is >5 and the area is >40, the weight for false indentation is adjusted by +0.

5. If the aspect ratio is (1.8, 5.2) and the mean gray value is (90, 130), then adjust the weight of severe corrosion by +1.

0. If the aspect ratio is greater than 8, adjust the fracture weight by +1.5, and adjust the weight of all categories of peeling morphology by -0.

8. If the aspect ratio is (3,8) and the area is (12,32), the weights for severe sheet-like peeling, multi-point peeling, and slight peeling are adjusted by +1.0, +0.6, and -0.5, respectively; if the aspect ratio is <3 and the area is (2,12), the weight for slight peeling is +0.

8. 4) Perform global feature analysis on all images: If the reverse torque is greater than 100 and the directional contrast is greater than 20, the polishing wear weight and adhesive wear weight are adjusted by +1.2 and +0.7 respectively. Under the condition that the reverse torque is greater than 100 and the directional contrast is greater than 20, if the low frequency energy is higher than 5E+05, the matte wear weight is adjusted by +1.

2. If the average grayscale value is >150 and the internal and external color difference is >30, then adjust the overload indentation weight by +1.

1. If the uniformity is <0.3 under the condition that the average grayscale value is >150 and the internal and external color difference is >30, then adjust the overload deformation weight by +0.

8. 5) Perform similar type exclusion for major category judgments: If it belongs to corrosion morphology and the average curvature is >0.7 and the energy is <0.03, adjust the weight of all categories of corrosion morphology by -0.7, and adjust the weight of all categories of wear morphology by +0.4; If the wear morphology is low and the low-frequency energy is <1E+05, then the weights for slight abrasive wear, matte wear, and multi-point spalling are adjusted to -1, -0.8, and +0.6, respectively.

5. The rolling bearing failure mode identification method based on dynamic gradient correction mechanism according to claim 1, characterized in that, Location association rules are defined based on manual classification methods and auxiliary labels as follows: 1) The inner part of the bearing outer ring, the outer part of the bearing inner ring, and the curved surfaces of the balls and rollers are the load-bearing parts: If the defect box is completely contained in the union of the connected loaded parts, then the weights of particle indentation, overload indentation, all categories of spalling morphology, and all categories of wear morphology are adjusted by +0.3, +0.3, +0.3, and +0.3, respectively. If the union of the defect box and the connected loaded part box intersects, the overload deformation weight and fracture weight are adjusted by +0.3 and +0.3 respectively. If the defect box and the connected loaded part box do not intersect, the micro-motion corrosion weight and the pseudo-indentation weight are adjusted by +0.3 and +0.3 respectively. 2) If the defect frame is contained within the sphere and roller surfaces, the weights of all categories of peeling morphology and adhesive wear are adjusted by +0.3 and +0.3 respectively; 3) If the defect box is contained within the inner raceway, adjust the fretting corrosion weight by +0.6; 4) If the defect frame is contained outside the outer raceway, the weights for slight corrosion and severe corrosion are adjusted by +0.6 and +0.6 respectively; 5) If the defect frame is contained within the cage, adjust the fracture weight by +0.6; 6) If the defect box is contained outside the inner ring of the bearing, the weights for slight spalling, multi-point spalling, and all categories of wear morphology are adjusted by +0.3, +0.3, and +0.2, respectively. 7) If the defect frame is contained inside the outer ring of the bearing, the weight of flaking and the weight of all categories of wear morphology are adjusted by +0.3 and +0.2 respectively.

6. The rolling bearing failure mode identification method based on dynamic gradient correction mechanism according to claim 1, characterized in that, Dynamic expert weights The acquisition process: Step 301: Initialize the probability weight matrix p of the 15 failure modes into a matrix of all 1s: , For the nth object detection bounding box of the i-th image in the batch Initial weights for each type of failure mode; Step 302: For the n×15 failure feature matrix within the batch, dynamically adjust the category weights according to expert rules to obtain the n×15 expert weight matrix for the i-th image. Set the non-positive matrix elements of the n×15 expert weight matrix to zero, and take the average weight of the n target detection boxes to obtain the 1×15 expert weight matrix for the i-th image. , For the i-th image, the th Weights of different failure modes; Step 303: Calculate the average expert weight for a batch by averaging the expert weights within that batch. : In the formula, B represents the number of images in the batch; Step 304: Assess the average expert weights Normalize: In the formula, This represents the normalized average expert weight for the corresponding batch. The average expert weight for the j-th failure mode within the batch. ; It is the minimum value. .

7. The rolling bearing failure mode identification method based on dynamic gradient correction mechanism according to claim 1, characterized in that, Step 4 is as follows: Step 401: Use a YOLOv11-OBB model with an input size of 640×640 and train the gradient by combining expert rule constraints. Step 402: Introduce an expert rule filtering module during the inference phase to re-examine detection boxes with a confidence level < 0.

5.

8. The rolling bearing failure mode identification method based on dynamic gradient correction mechanism according to claim 2, characterized in that, Step 101 is based on the GB / T24611-2020 standard and classifies images according to their appearance.

9. The rolling bearing failure mode identification method based on dynamic gradient correction mechanism according to claim 2, characterized in that, Step 102 describes using the Rotated Rectangle Labeling Tool to label the failure area. Based on the existing RoLabelimg tool, the YOLOv11-OBB format label is a format label that includes the category number and the coordinates of the four points of the rectangle.

10. The rolling bearing failure mode identification method based on dynamic gradient correction mechanism according to claim 3, characterized in that, The process of generating the de-sourced image in step 201 is as follows: perform single-image source separation and preset source, perform surface modeling and estimate the normal field through Lambertian reflection, perform RANSAC fitting based on the estimated normal field, update the source through the obtained parameters, and iterate until the difference between the source direction and the source vector of the previous iteration is less than the minimum value ε, thus obtaining the approximate source direction, and calculate the normal field and reflectivity through the Lambertian reflection method. The value of the reflectivity component can be used as the de-sourced image.

Images