Bearing steel ball surface stress detection and stress concentration early warning system and method

Abstract

The application discloses a bearing steel ball surface stress detection and stress concentration early warning system and method, belongs to the technical field of bearing detection, and comprises the following steps: acquiring the photoelectric scattering intensity and array eddy current impedance sequence of the steel ball surface, constructing a bimodal feature tensor and performing spatial alignment; calculating the cross covariance of the gradients of two heterogeneous signals, extracting a joint mutation feature reflecting the coupling relationship between the surface topography and the internal stress, and generating a fusion stress response atlas; inputting the fusion stress atlas into an adaptive defect segmentation network based on reinforcement learning, extracting a defect area by dynamically adjusting a threshold value by an intelligent agent; and calculating the edge curvature and stress concentration intensity index of the defect, and outputting a graded early warning signal through a multi-level early warning model. The application improves the accuracy of quantitative detection of weak stress concentration hidden dangers, and realizes dynamic graded early warning.

Description

Technical Field

[0001] This application relates to the field of bearing testing technology, specifically to a system and method for detecting stress on the surface of bearing steel balls and for early warning of stress concentration. Background Technology

[0002] As a core rotating component, the surface integrity of bearing steel balls directly determines the service life and reliability of the bearing. With increasingly stringent requirements for bearing fatigue life in industrial applications, quality inspection of steel balls is no longer limited to identifying macroscopic geometric defects (such as cracks, pits, and scratches). Research shows that the core cause of early fatigue failure in bearings is often localized stress concentration areas located on or near the surface. These areas are highly susceptible to fatigue spalling under alternating loads.

[0003] Current automated inspection technologies typically focus only on abrupt changes in a single physical quantity, tending to categorically classify any area with signal amplitude exceeding a threshold as a defect. This approach has significant technical limitations: it cannot effectively distinguish between non-fatal shallow surface textures and fatal stress concentration hazards.

[0004] In actual working conditions, many stress concentration points with high fatigue risk often exhibit signal characteristics of a specific coupling state between surface morphology and internal impedance, rather than simply a strong signal amplitude. Existing detection methods struggle to extract this coupling characteristic reflecting the degree of stress concentration from complex background noise and lack a dynamic evaluation mechanism capable of quantifying the intensity of stress concentration in defects. This leads to situations where, when faced with subtle but potentially fatal stress concentration hazards, existing technologies not only miss detections due to excessively high threshold settings, allowing steel balls with fatigue risks to enter the assembly process, but also result in "one-size-fits-all" scrapping due to the inability to quantify the risk level, causing significant production waste. How to achieve high-precision capture and quantitative classification and early warning of stress concentration states on the surface of bearing steel balls is a pressing technical challenge for the precision bearing manufacturing industry.

[0005] In view of this, this application proposes a system and method for detecting stress on the surface of bearing steel balls and for early warning of stress concentration. Summary of the Invention

[0006] To achieve the above objectives, this application provides a system and method for detecting and predicting stress concentration on the surface of bearing steel balls, the specific technical solution of which is as follows:

[0007] Firstly, this application provides a method for detecting surface stress and providing early warning of stress concentration on bearing steel balls, including:

[0008] The bearing steel ball is scanned to obtain the photoelectric scattering intensity sequence and the array eddy current impedance real part sequence on the surface of the bearing steel ball. The photoelectric scattering intensity sequence and the array eddy current impedance real part sequence are aligned in spatial coordinates to construct a dual-mode feature tensor of photoelectric and eddy current.

[0009] Physical features are fused to the dual-modal feature tensor, the cross-covariance of the photoelectric scattering intensity gradient and the real part gradient of the eddy current impedance is calculated, the joint abrupt change feature reflecting the coupling relationship between surface morphology and internal stress is extracted, and the fused stress response spectrum is generated.

[0010] The fused stress response map is input into the adaptive defect segmentation network, and the defect boundary search is defined as a sequential decision process. The segmentation threshold is dynamically adjusted by the agent in the map feature space.

[0011] The fused stress response map is binarized based on the optimal segmentation threshold output by the agent to obtain a binarized image. Isolated foreground regions in the binarized image are extracted as the locations of flaw detection defects. The edge curvature and stress concentration intensity index of the isolated foreground regions are calculated.

[0012] A multi-level early warning model is constructed. The stress concentration intensity index of the flaw detection location is input into the multi-level early warning model, and an early warning signal is output according to the dynamic early warning interval in which the stress concentration intensity index is located.

[0013] Preferably, the bearing steel ball is spirally scanned, the photoelectric sensor emits a laser beam and receives the light signal and converts it into a photoelectric scattering intensity sequence, and the arrayed eddy current probe excites an alternating eddy current field and captures the impedance change to form an arrayed eddy current impedance real part sequence.

[0014] Establish a spherical coordinate system, use an encoder to record the rotation and revolution angle sequences, and construct a spatial mapping function to transform the two sequences to a unified spherical coordinate system;

[0015] A latitude and longitude grid is divided on the surface of the spherical coordinate system, and a bilinear interpolation algorithm is used to resample to the grid nodes to complete the spatial coordinate alignment.

[0016] The aligned features are standardized and then stitched together according to the two-dimensional spatial topology of the latitude and longitude grid to generate a three-dimensional bimodal feature tensor containing photoelectric and eddy current spatial distribution data.

[0017] Preferably, the spatial distribution data of photoelectric scattering intensity and the spatial distribution data of the real part of array eddy current impedance are separated from the dual-modal feature tensor;

[0018] In the two-dimensional spatial topology of the latitude and longitude grid, the spatial gradient operator based on the central difference method is used to calculate the partial derivatives in the longitude and latitude directions respectively, and the magnitude of the photoelectric scattering intensity gradient and the magnitude of the real part gradient of the eddy current impedance are calculated.

[0019] Set a local sliding window of fixed size on the latitude and longitude grid, and traverse it pixel by pixel according to the preset step size;

[0020] Within each local sliding window, extract all photoelectric scattering intensity gradient magnitudes and eddy current impedance real part gradient magnitudes within the window, and calculate the cross covariance value.

[0021] Preferably, the cross covariance value is used as a weight adjustment factor, and the photoelectric scattering intensity gradient magnitude and the real part gradient magnitude of the eddy current impedance on each pixel node are multiplied by the cross covariance value at the center of the corresponding local sliding window to obtain the joint mutation feature value of the pixel node.

[0022] The joint mutation feature values ​​on all pixel nodes are rearranged according to the original spatial topology of the latitude and longitude grid;

[0023] The rearranged joint mutation eigenvalue matrix is ​​subjected to global extremum normalization, and all joint mutation eigenvalues ​​are linearly mapped to a preset gray level range to form a single-channel two-dimensional gray level image as a fused stress response spectrum.

[0024] Preferably, the global gray-level histogram vector of the fused stress response map is extracted, the mean gray-level difference between the foreground and background regions under the current candidate segmentation threshold is calculated, and the global gray-level histogram vector, the mean gray-level difference, and the current candidate segmentation threshold are concatenated to form a high-dimensional state vector.

[0025] Define the action space, which includes three discrete actions: increasing, decreasing, and keeping the current segmentation threshold unchanged.

[0026] A reward function based on the contrast of defective regions is constructed. Each time the agent performs an action, the candidate segmentation threshold is updated. The interactive environment calculates the mean difference and spatial variance of grayscale between the foreground and background regions based on the updated candidate segmentation threshold, and calculates the contrast reward value to feed back to the agent.

[0027] Preferably, the agent and the interactive environment engage in trial and error interaction, storing the state transition tuple containing the high-dimensional state vector, discrete actions, contrast reward value, and the high-dimensional state vector at the next time step into the experience replay pool; randomly sampling batches of state transition tuples to calculate the mean squared error loss to optimize the weight parameters of the deep Q network;

[0028] During the defect boundary search phase, the candidate segmentation threshold is initialized as the global grayscale average value of the map. The agent calculates the expected cumulative reward value of discrete actions based on the high-dimensional state vector and selects the action corresponding to the maximum value to adjust the threshold.

[0029] When the agent continuously performs actions that keep the threshold unchanged for a preset number of times or reaches the maximum limit of steps, it stops making decisions and outputs the finally locked candidate segmentation threshold as the optimal segmentation threshold.

[0030] Preferably, the gray value of the current pixel feature is compared with the optimal segmentation threshold, and then reset to the target foreground value or the background zero value to obtain a binarized image;

[0031] Morphological opening and closing operations are performed sequentially on the binarized image. The set of connected non-zero pixels is extracted and defined as an isolated foreground region. The geometric center coordinates of the isolated foreground region are mapped back to three-dimensional space as the location of the flaw detection defect.

[0032] An eight-neighbor boundary tracing algorithm is used to construct the edge contour pixel coordinate sequence of isolated foreground regions, and a Gaussian smoothing kernel is used for one-dimensional convolution smoothing to obtain continuous edge coordinate functions.

[0033] The edge curvature is obtained by calculating the first and second derivatives of the pixels on the edge contour based on the continuous edge coordinate function, and the maximum value among all edge curvature values ​​is extracted as the maximum edge curvature feature.

[0034] Preferably, the total number of pixels contained in the isolated foreground region is extracted, and the total number of pixels is multiplied by the actual physical area represented by a single pixel to obtain the absolute physical area of ​​the isolated foreground region.

[0035] The isolated foreground region is mapped back to the fused stress response map. The fused stress response feature values ​​of all pixels within the coverage area of ​​the isolated foreground region are extracted, and the sum of all fused stress response feature values ​​is calculated. This sum is defined as the region cumulative stress response equivalent.

[0036] We assign corresponding weighting coefficients to the maximum edge curvature feature, absolute physical area, and regional cumulative stress response equivalent, and then perform nonlinear coupling calculations to construct the stress concentration intensity index of the isolated foreground region.

[0037] Preferably, the multi-level early warning model dynamically updates the stress concentration strength index of historical normal bearing steel balls according to the first-in-first-out principle, uses the updated stress concentration strength index as the early warning signal, and calculates the mean and standard deviation.

[0038] The boundary threshold of the multi-level dynamic early warning set is calculated by combining multiple preset sensitivity adjustment coefficients, and multiple dynamic early warning intervals are divided. The stress concentration intensity index of the flaw detection defect location is compared with the boundary threshold to determine the interval it falls into, and the corresponding early warning signal is output through the state mapping machine.

[0039] A bearing steel ball surface stress detection and stress concentration early warning system is used to implement the bearing steel ball surface stress detection and stress concentration early warning method, including: a dual-modal feature tensor construction module, a feature fusion and map generation module, a reinforcement learning adaptive segmentation module, a defect extraction and strength quantization module, and a multi-level dynamic hierarchical early warning module;

[0040] The dual-modal feature tensor construction module scans the bearing steel ball to obtain the photoelectric scattering intensity sequence and the array eddy current impedance real part sequence on the surface of the steel ball. It then aligns the photoelectric scattering intensity sequence and the array eddy current impedance real part sequence in spatial coordinates to construct a dual-modal feature tensor of photoelectric and eddy current.

[0041] The feature fusion and map generation module performs physical feature fusion on the dual-modal feature tensor, calculates the cross covariance of the photoelectric scattering intensity gradient and the real part gradient of the eddy current impedance, extracts the joint abrupt change feature reflecting the coupling relationship between surface morphology and internal stress, and generates a fused stress response map.

[0042] The reinforcement learning adaptive segmentation module inputs the fused stress response map into the adaptive defect segmentation network, defines the defect boundary search as a sequential decision process, and dynamically adjusts the segmentation threshold in the map feature space through the agent.

[0043] The defect extraction and intensity quantization module performs binarization processing on the fused stress response map based on the optimal segmentation threshold output by the intelligent agent, extracts isolated foreground regions in the binarized image as the locations of flaw detection defects, and calculates the edge curvature and stress concentration intensity index of the isolated foreground regions.

[0044] The multi-level dynamic hierarchical early warning module constructs a multi-level early warning model. The stress concentration intensity index at the location of the flaw detection defect is input into the multi-level early warning model, and the multi-level early warning model outputs the corresponding early warning signal according to the dynamic early warning interval in which the intensity index is located.

[0045] The beneficial effects of this application are:

[0046] This application addresses the issues of fragmented information and spatial misalignment from single-sensor data by constructing a bimodal feature tensor. It achieves deep integration of surface optics and subsurface electromagnetic properties, providing a high-fidelity, spatially rigorous, multi-dimensional data foundation for subsequent high-precision detection.

[0047] This application utilizes cross-covariance to extract joint mutation features, effectively suppressing single-mode noise and pseudo-defect interference. It transforms obscure signals into high signal-to-noise ratio spectra, significantly enhancing the feature representation and identification of small stress concentration regions.

[0048] This application introduces a reinforcement learning agent for dynamic optimization, eliminating the reliance on fixed thresholds based on human experience. It achieves adaptive locking of weak defect boundaries under varying background noise levels, improving segmentation robustness in complex environments.

[0049] This application achieves a leap from qualitative discovery to quantitative assessment by constructing a stress concentration intensity index. The stress concentration intensity index comprehensively reflects the geometric sharpness and physical destructive force of defects, providing a scientific and quantitative numerical basis for accurate classification and determination.

[0050] This application constructs a multi-level dynamic early warning model, balancing the tolerance for production quality fluctuations with the defect interception rate. The tiered mechanism avoids misjudgments caused by applying fixed thresholds across the board, significantly improving engineering practicality and flexibility.

[0051] This application's technical solution achieves high-precision, fully automated detection of bearing steel balls, solving the problem of missing subtle defects. Its unique quantitative grading mechanism constructs a closed-loop quality control system, effectively intercepting potentially defective products and significantly improving the service reliability and safety of high-end bearings. Attached Figure Description

[0052] Figure 1 A schematic diagram of a method for detecting and predicting stress concentration on the surface of bearing steel balls;

[0053] Figure 2 A flowchart for constructing the bimodal feature tensor in this application;

[0054] Figure 3 This is a schematic diagram of the multi-level dynamic early warning model in this application;

[0055] Figure 4 This is a structural diagram of a bearing steel ball surface stress detection and stress concentration early warning system. Detailed Implementation

[0056] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0057] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0058] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of this application. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0059] Example 1

[0060] Reference Figures 1 to 3 This is the first embodiment of the present application, such as Figure 1 As shown, a method for detecting surface stress and providing early warning of stress concentration on bearing steel balls is provided.

[0061] Step 1: Scan the bearing steel ball to obtain the photoelectric scattering intensity sequence and the array eddy current impedance real part sequence on the surface of the bearing steel ball. Align the photoelectric scattering intensity sequence and the array eddy current impedance real part sequence in spatial coordinates to construct a dual-mode feature tensor of photoelectric and eddy current.

[0062] A composite scanning method using a photoelectric sensor and an arrayed eddy current probe is employed to obtain the photoelectric scattering intensity sequence and the real part sequence of the arrayed eddy current impedance on the surface of the bearing steel ball. Specifically, a composite scanning hardware mechanism is constructed, comprising a precision rotating stage, a photoelectric sensor, and an arrayed eddy current probe. The bearing steel ball to be inspected is placed on the precision rotating stage, which is equipped with mutually perpendicular rotation and revolution drive shafts. Through the coordinated control of dual-axis servo motors, the bearing steel ball is driven to continuously rotate while maintaining a fixed revolution angular velocity, thereby forming a uniform and dead-angle-free spiral scanning trajectory on the surface of the steel ball.

[0063] During the movement of the bearing steel ball, a photoelectric sensor emits a monochromatic collimated laser beam of a specific wavelength onto the surface of the bearing steel ball. This specific wavelength includes either a 632.8nm helium-neon laser wavelength, which is highly sensitive to scattering from the microscopic morphology of the metal surface, or a 532nm green light wavelength. The photoelectric detection element inside the sensor receives the light signals reflected and scattered from the bearing steel ball surface in real time and converts these signals into continuous voltage signals. After sampling by an analog-to-digital converter, a discrete sequence of photoelectric scattering intensity is formed. Simultaneously, an array-type eddy current probe is placed close to the bearing steel ball surface in a non-contact manner. Multiple miniature excitation coils inside the array-type eddy current probe are energized with high-frequency alternating current, generating an alternating eddy current field on and near the surface of the bearing steel ball. The receiving coil array inside the array-type eddy current probe captures the impedance change caused by the reaction of the alternating eddy current field on the receiving coil array in real time, extracts the real part of the impedance change, and forms a sequence of the real part of the array eddy current impedance. By driving the bearing steel ball to perform omnidirectional spiral motion without blind spots through a precision rotating stage, and combining the synchronous high-frequency sampling of photoelectric sensors and arrayed eddy current probes, the physical state information of the bearing steel ball surface and near surface can be captured comprehensively and precisely, providing high-fidelity raw heterogeneous data support for subsequent stress and defect analysis.

[0064] After obtaining the photoelectric scattering intensity sequence and the array eddy current impedance real part sequence, their spatial coordinates are aligned. Due to the spatial angle between the physical installation positions of the photoelectric sensor and the array eddy current probe in the composite scanning hardware, and the differences in their sampling frequencies and response times, the physical coordinates of the bearing steel ball surface corresponding to the photoelectric scattering intensity data and the array eddy current impedance real part data acquired at the same sampling moment do not coincide.

[0065] To address the issue of non-coincidence of physical coordinates on the surface of the bearing steel balls, a spherical coordinate system with the center of the bearing steel ball as the origin is established. A high-precision encoder built into a precision rotating stage is used to record the rotation and revolution angle sequences of the bearing steel balls in real time. These rotation and revolution angle sequences are then used as reference spatial coordinate sequences to construct spatial mapping functions for photoelectric sensors and arrayed eddy current probes, respectively.

[0066] The mathematical expression for the spatial mapping function of the photoelectric sensor is: ;in, Indicates the current sampling time. Indicates the current sampling time The actual spatial coordinate vector of the sampling point of the photoelectric sensor in spherical coordinates. Indicates the current sampling time The reference space coordinate vector recorded by the high-precision encoder. This represents the rotation transformation matrix of the photoelectric sensor relative to the spherical coordinate system. The translation transformation vector of the photoelectric sensor relative to the spherical coordinate system is represented by the rotation transformation matrix, and the operation between the matrix and the reference space coordinate vector is performed by matrix multiplication.

[0067] Similarly, the mathematical expression for the spatial mapping function of the array-type eddy current probe is: ;in, Indicates the current sampling time The actual spatial coordinate vector of the sampling points of the arrayed eddy current probe in spherical coordinates. This represents the rotation transformation matrix of the arrayed eddy current probe relative to the spherical coordinate system. The translation transformation vector of the arrayed eddy current probe relative to the spherical coordinate system is represented by the rotation transformation matrix and the translation transformation vector are operated by matrix multiplication.

[0068] By employing the spatial mapping functions of the photoelectric sensor and the arrayed eddy current probe, the time-dimensional photoelectric scattering intensity sequence and the arrayed eddy current impedance real part sequence are transformed to a unified spherical coordinate system. Subsequently, a uniform latitude and longitude grid is divided on the surface of the unified spherical coordinate system. Using a bilinear interpolation algorithm, the photoelectric scattering intensity data and the arrayed eddy current impedance real part data, transformed to the spherical coordinate system, are resampled to each grid node of the uniform latitude and longitude grid, thus completing the spatial coordinate alignment of the photoelectric scattering intensity sequence and the arrayed eddy current impedance real part sequence.

[0069] For example, if the longitude division step size of the latitude and longitude grid is set to... The latitude division step size is set to Then, each is calculated using the bilinear interpolation algorithm. The interpolated photoelectric data and interpolated eddy current data correspond to the grid nodes. By establishing a spatial mapping model and performing gridded resampling, the spatial mapping model is a geometric transformation model based on rigid body kinematics, using the rotation and revolution angle sequences collected by the encoder as input variables. Specifically, the rotation matrix is ​​constructed using trigonometric functions based on the real-time collected rotation and revolution Euler angles, rotating the vectors in the reference coordinate system to a unified viewpoint; the translation vector is used to compensate for the fixed spatial offset distance between the physical installation center of the sensor and the center of the steel ball. Through this linear algebraic operation of matrix multiplication and vector addition, a unified spherical coordinate system three-dimensional coordinate is output, which can eliminate the spatial misalignment error caused by the difference in the physical position of multiple sensors, ensuring a strict one-to-one correspondence between photoelectric data and eddy current data in microscopic space, and greatly improving the physical accuracy of heterogeneous data fusion.

[0070] After spatial coordinate alignment, a dual-modal feature tensor of photoelectric and eddy current is constructed. The photoelectric scattering intensity values ​​of each aligned grid node on the latitude and longitude grid are extracted as the first modal feature, and the real part values ​​of the array eddy current impedance of each aligned grid node on the latitude and longitude grid are extracted as the second modal feature. To eliminate dimensional and numerical differences between the first and second modal features, Z-score normalization is performed on both. The normalized first and second modal features are then concatenated according to the two-dimensional spatial topology of the latitude and longitude grid.

[0071] Specifically, the longitude dimension of the latitude-longitude grid is used as the first dimension of the tensor, the latitude dimension as the second dimension, and the category of the modal feature as the third dimension. This generates a three-dimensional bimodal feature tensor. The first channel plane of the bimodal feature tensor stores globally normalized spatial distribution data of photoelectric scattering intensity, while the second channel plane stores globally normalized spatial distribution data of the real part of the array eddy current impedance.

[0072] Figure 2 This is a flowchart of the construction of a bimodal feature tensor, which specifically illustrates the data acquisition and processing process for surface stress detection of bearing steel balls. First, the bearing steel balls are subjected to a composite spiral scan using a photoelectric sensor and an array of eddy current probes to obtain heterogeneous original signal sequences. Then, a bilinear interpolation algorithm is used to align the spatial coordinates of the signals and resample them to eliminate physical position deviations. Finally, the aligned photoelectric scattering intensity and eddy current impedance data are spliced ​​together in the channel dimension to construct a bimodal feature tensor containing a multidimensional spatial topology.

[0073] This step, by constructing a dual-modal feature tensor containing multi-dimensional spatial topology and multi-physics information, not only preserves the spatial geometric neighborhood relationship of the bearing steel ball surface, but also achieves a deep structural integration of surface optical scattering characteristics and subsurface electromagnetic impedance characteristics. This lays a standardized multi-dimensional data foundation for the subsequent extraction of joint abrupt change features reflecting the coupling relationship between surface morphology and internal stress.

[0074] Step 2: Perform physical feature fusion on the dual-modal feature tensor, calculate the cross-covariance of the gradient magnitude of photoelectric scattering intensity and the gradient magnitude of the real part of eddy current impedance, extract the joint abrupt change feature reflecting the coupling relationship between surface morphology and internal stress, and generate a fused stress response spectrum.

[0075] After obtaining the bimodal feature tensor constructed in step one, physical feature fusion is performed on the bimodal feature tensor. First, the spatial distribution data of the globally normalized photoscattering intensity of the first channel plane and the spatial distribution data of the real part of the globally normalized array eddy current impedance of the second channel plane are separated from the bimodal feature tensor. To capture the microscopic abrupt changes in the surface morphology of the bearing steel ball and the local distortion of the subsurface stress field, two-dimensional spatial gradient calculations are performed on the spatial distribution data of the globally normalized photoscattering intensity and the spatial distribution data of the real part of the globally normalized array eddy current impedance. In the two-dimensional spatial topology of the latitude and longitude grid, a spatial gradient operator based on the central difference method is used to calculate the partial derivatives in the longitude and latitude directions, respectively.

[0076] The mathematical expression for calculating the magnitude of the photoelectric scattering intensity gradient is: ;in, This represents the longitude coordinate index in the latitude and longitude grid. This represents the latitude coordinate index in the latitude and longitude grid. Indicates the coordinate index The globally normalized photoelectric scattering intensity value at that location. The partial derivative of photoelectric scattering intensity along the longitude direction is represented by . The partial derivative of photoelectric scattering intensity along the latitudinal direction is represented by . Indicates the coordinate index The photoelectric scattering intensity gradient amplitude is calculated at the specified location. Similarly, using the exact same central difference method and amplitude calculation logic, the real part gradient amplitude of the eddy current impedance at each coordinate index is calculated from the spatial distribution data of the real part of the globally normalized array eddy current impedance. By calculating the photoelectric scattering intensity gradient amplitude and the real part gradient amplitude of the eddy current impedance separately, the originally smooth background signal can be suppressed, while the scattering abrupt changes caused by surface scratches and pits, as well as the impedance abrupt changes caused by internal residual stress concentration, are amplified at high frequency, effectively highlighting the edge information of potential physical defects.

[0077] After calculating the magnitudes of the photoscattering intensity gradient and the real part gradient of the eddy current impedance, the cross-covariance of these two values ​​is calculated. Since actual fatigue cracks or severe stress concentration areas in bearing steel balls typically cause both surface morphology damage and changes in subsurface electromagnetic properties, it is necessary to quantify the spatial synergistic relationship between these two physical phenomena. A fixed-size local sliding window is set on the latitude and longitude grid, for example, a square area covering 5×5 pixel nodes. The local sliding window traverses the latitude and longitude grid pixel by pixel according to a preset step size, for example, a step size of 1 pixel node. The preset step size is set based on the flaw detection resolution requirements and the overlap rate of the local sliding window to ensure continuous and comprehensive coverage of all potential small stress concentration areas during the traversal. Within each local sliding window, all photoscattering intensity gradient magnitudes and real part gradients of the eddy current impedance within the window's coverage area are extracted.

[0078] The mathematical expression for cross covariance is: ;in, This represents the longitude coordinates of the center pixel of the local sliding window. This represents the latitude and longitude coordinates of the center pixel of the local sliding window. Indicates The cross covariance within a local sliding window centered on [the target]. This indicates the total number of pixel nodes contained within the local sliding window. This indicates the index of the pixel node traversal within the local sliding window. , Indicates the first [number]th [unit] within a local sliding window The magnitude of the photoelectric scattering intensity gradient of each pixel node. This represents the arithmetic mean of the magnitudes of all photoelectric scattering intensity gradients within the local sliding window. Indicates the first [number]th [unit] within a local sliding window The gradient magnitude of the real part of the eddy current impedance of each pixel node. This represents the arithmetic mean of the real part gradient magnitudes of all eddy current impedances within the local sliding window.

[0079] For example, if the size of the partial sliding window is set to If there are n pixel nodes, then the total number of pixel nodes contained within the local sliding window is 1. By calculating the cross covariance within a local sliding window, the statistical correlation between surface morphology anomalies and internal stress anomalies can be accurately measured within a local spatial range, effectively filtering out false defect signals caused solely by noise from a single sensor or harmless surface contaminants.

[0080] Based on the calculated cross-covariance, a joint abrupt change feature value reflecting the coupling relationship between surface morphology and internal stress is extracted from the dot product of the photoscattering intensity gradient magnitude and the real part gradient magnitude of eddy current impedance. The cross-covariance is used as a weighting adjustment factor to nonlinearly weight and fuse the photoscattering intensity gradient and the real part gradient of eddy current impedance. Specifically, the photoscattering intensity gradient magnitude and the real part gradient magnitude of eddy current impedance at each pixel node are multiplied by the dot product, and then the result is multiplied by the cross-covariance at the center of the corresponding local sliding window to obtain the joint abrupt change feature value for that pixel node. When both surface morphology damage and internal stress concentration exist simultaneously in a certain region, both the photoscattering intensity gradient and the real part gradient of eddy current impedance exhibit large values, and the cross-covariance shows a high positive correlation. In this case, the extracted joint abrupt change feature value will be nonlinearly amplified; conversely, if only a single mode fluctuates, the joint abrupt change feature value will be suppressed. The process of extracting the joint abrupt change feature achieves deep coupling of heterogeneous physical field information, significantly improving the signal-to-noise ratio of real stress concentration defects under complex working conditions.

[0081] The joint mutation feature values ​​on all pixel nodes are rearranged according to the original spatial topology of the latitude and longitude grid to generate a fused stress response map. To facilitate subsequent image processing and feature recognition, the joint mutation feature values ​​on all pixel nodes are rearranged according to the original spatial topology of the latitude and longitude grid, i.e., a two-dimensional joint mutation feature value matrix is ​​constructed with latitude coordinate indices as rows and longitude coordinate indices as columns. Then, global extremum normalization is performed on the rearranged joint mutation feature value matrix, linearly mapping all joint mutation feature values ​​to a preset grayscale range. The preset grayscale range is set based on the bit depth requirements of standard digital image processing, such as the 0 to 255 range for an 8-bit grayscale image. This preserves sufficient feature dynamic range while directly compatibility with subsequent machine vision-based image segmentation and morphological processing algorithms. For example, the joint mutation feature values ​​are linearly mapped to... to Within the grayscale range, a single-channel two-dimensional grayscale image is formed, which is the fused stress response map. In the fused stress response map, the grayscale value of the pixel directly maps the severity of stress concentration on and near the surface of the bearing steel ball and the overall state of physical morphological damage.

[0082] This step successfully transforms the separated optical and electromagnetic heterogeneous signals into an intuitive and high-fidelity fused stress response map through physical feature fusion and cross-covariance calculation. This not only eliminates the physical barriers between cross-modal data but also enhances the feature representation of small stress concentration areas, providing a high-quality map data foundation for subsequent automated and high-precision defect segmentation and early warning.

[0083] Step 3: Input the fused stress response map into the adaptive defect segmentation network, define the defect boundary search as a sequential decision process, and dynamically adjust the segmentation threshold in the map feature space through the agent.

[0084] The fused stress response map generated in step two is input into the adaptive defect segmentation network constructed based on reinforcement learning. In the adaptive defect segmentation network, the fused stress response map serves as the interaction environment for reinforcement learning. To initiate the sequential decision-making process, the agent's state space is first constructed in the feature space of the fused stress response map. The agent's state at any decision time is jointly constituted by the global gray-level distribution features, local texture features, and the candidate segmentation threshold of the current fused stress response map. The global gray-level histogram vector of the fused stress response map is extracted, and the mean gray-level difference between the foreground and background regions under the current candidate segmentation threshold is calculated. The mean gray-level difference is the difference between the average gray-level of all pixels greater than or equal to the current candidate segmentation threshold and the average gray-level of all pixels less than the current candidate segmentation threshold. The global gray-level histogram vector, the mean gray-level difference, and the current candidate segmentation threshold are concatenated to form a high-dimensional state vector. The process of constructing a high-dimensional state vector enables the agent to fully perceive the macroscopic distribution and microscopic details of the fused stress response spectrum, providing rich data support for subsequent threshold adjustment and effectively avoiding the segmentation failure problem of traditional fixed threshold methods under complex lighting or non-uniform stress backgrounds.

[0085] Define the agent's action space in the graph feature space. The action space contains discrete threshold adjustment instructions, specifically set as three discrete actions: increasing the segmentation threshold, decreasing the segmentation threshold, and keeping the current segmentation threshold unchanged. At each decision time step, the adaptive defect segmentation network outputs the execution probability of each discrete action in the action space based on the current high-dimensional state vector. The agent selects the corresponding discrete action to apply to the current candidate segmentation threshold based on the execution probability. If the segmentation threshold is increased, the current candidate segmentation threshold is added to a preset threshold adjustment step size, for example, set to 1 or 2 gray levels. If the segmentation threshold is decreased, the current candidate segmentation threshold is subtracted from the preset threshold adjustment step size. The preset threshold adjustment step size is set based on the balance between the search accuracy and convergence speed of the reinforcement learning agent in the graph feature space. A smaller step size can ensure the fineness of the segmentation boundary and avoid exceeding the optimal threshold. For example, if the preset threshold adjustment step size is set to 2 gray levels, when the agent executes the action of increasing the segmentation threshold, the candidate segmentation threshold will increase by 2. By defining a discrete action space and a dynamic adjustment mechanism, the adaptive defect segmentation network is endowed with the ability to explore autonomously under unknown stress distribution, enabling the threshold optimization process to flexibly adapt to the background noise changes of bearing steel balls of different batches and materials.

[0086] To guide the agent in optimizing along the direction of the optimal segmentation threshold, a reward function based on the contrast of the defect region is constructed. The adaptive defect segmentation network defines defect boundary search as a sequential decision process. After each action by the agent updates the candidate segmentation threshold, the interactive environment calculates a contrast reward value based on the updated threshold and feeds it back to the agent. The mathematical expression for the contrast reward value is: .in, Indicates the first The contrast reward value obtained by the agent at each decision time step. Indicates the first Each decision time step uses the arithmetic mean of the gray levels of all pixels within the foreground defect region defined by the current candidate segmentation threshold. Indicates the first Each decision time step uses the arithmetic mean of the grayscale values ​​of all pixels within the normal background region defined by the current candidate segmentation threshold. Indicates the first Spatial variance of pixel grayscale values ​​within the foreground defect region at each decision time step Indicates the first Spatial variance of pixel grayscale values ​​within the normal background region at each decision time step This represents a very small positive constant used to prevent the denominator from being zero. For example, the very small positive constant is set to 0.0001. In actual detection, weak stress concentration signals are often masked by background noise such as normal surface texture or material inhomogeneity, resulting in low signal-to-noise ratio (SNR) spectral features. By maximizing the contrast reward value of the defect region, the agent can be driven to find the critical segmentation point that maximizes the difference between the foreground defect region and the normal background region while ensuring the most uniformity within each region. This allows for accurate localization of weak stress concentration defect boundaries in the low SNR fused stress response spectrum.

[0087] The adaptive defect segmentation network employs a deep Q-network architecture to fit the value mapping relationship between states and actions. The deep Q-network consists of multiple fully connected neural networks. Its input is the previously constructed high-dimensional state vector, and its output is the expected cumulative reward value corresponding to each discrete action in the action space. During training, the weight parameters of the deep Q-network are updated using a temporal difference learning algorithm. The agent continuously interacts with the interactive environment in the graph feature space through trial and error, storing the collected state transition tuples in the experience replay pool. Each state transition tuple contains the current high-dimensional state vector, the executed discrete action, the obtained contrast reward value, and the next high-dimensional state vector. At each network parameter update, a batch of state transition tuples is randomly selected from the experience replay pool, and the mean squared error loss between the target Q-value and the current deep Q-network predicted Q-value is calculated. The mathematical expression for the mean squared error loss is: ,in, This represents the mean squared error loss value of the deep Q-network at the current training iteration step. This represents the set of weight parameters for a deep Q-network. This indicates the number of batch samples randomly drawn from the experience replay pool. Represents the traversal index of the batch samples and , Indicates the first The target Q value for each sample Indicates the first High-dimensional state vectors in each sample Indicates the first Discrete actions performed by the agent in each sample This indicates that the deep Q-network has parameters of The Q-values ​​of the predicted states and actions are calculated. The mean squared error loss is minimized using a gradient descent optimizer to progressively optimize the weight parameter set of the deep Q-network. The introduction of an empirical replay pool and the deep Q-network architecture breaks the temporal correlation between sequential decision data, ensuring the convergence stability of the adaptive defect segmentation network training process and enabling the agent to learn a threshold decision strategy with extremely strong generalization ability.

[0088] After the adaptive defect segmentation network completes training, it enters the actual defect boundary search phase. The fused stress response map to be detected is input into the trained adaptive defect segmentation network, and the candidate segmentation threshold is initialized to the global grayscale average value of the fused stress response map. The agent calculates the Q-value of each discrete action based on the high-dimensional state vector of the current fused stress response map through forward propagation of a deep Q-network, and selects the discrete action with the largest Q-value to dynamically adjust the segmentation threshold. The agent continuously performs sequential decisions in the feature space of the map until a preset termination condition is met. The preset termination condition is set to the agent continuously executing actions that maintain the current segmentation threshold unchanged for a preset number of times. Setting the preset number of times ensures that the agent has converged to the global optimum and avoids accidental stagnation. For example, based on the stability of the action value function, the preset number of times is set to 3 or 5 times, or the number of steps in the sequential decision-making reaches the maximum limit. When the preset termination condition is met, the adaptive defect segmentation network stops the sequential decision-making process and outputs the finally locked candidate segmentation threshold as the optimal segmentation threshold. Step 3 introduces reinforcement learning into the field of image segmentation. By utilizing the agent's autonomous exploration and dynamic adjustment in the image feature space, it eliminates the dependence on threshold setting based on human experience and achieves adaptive high-precision localization of defect boundaries under complex stress backgrounds. This greatly improves the robustness and accuracy of stress concentration defect feature extraction.

[0089] Step 4: Binarize the fused stress response map according to the optimal segmentation threshold output by the agent to obtain a binarized image. Extract the isolated foreground region in the binarized image as the location of the flaw detection defect, and calculate the edge curvature and stress concentration intensity index of the isolated foreground region.

[0090] The optimal segmentation threshold output by the agent in step three is obtained and used as a hard criterion for pixel grayscale classification. The fused stress response map to be detected is then subjected to pixel-by-pixel binarization. During binarization, each pixel in the fused stress response map is traversed, and the feature grayscale value of the current pixel is extracted. The feature grayscale value of the current pixel is compared with the optimal segmentation threshold. If the feature grayscale value of the current pixel is greater than or equal to the optimal segmentation threshold, the pixel value of the current pixel is reset to the target foreground value; if the feature grayscale value of the current pixel is strictly less than the optimal segmentation threshold, the pixel value of the current pixel is reset to the background zero value. For example, the target foreground value can be set to 255, and the background zero value can be set to 0. After the pixel-by-pixel value reset operation, the fused stress response map is converted into a binarized image containing only the target foreground value and the background zero value.

[0091] To eliminate isolated noise points caused by high-frequency electromagnetic noise or minute surface scratches in the binarized image, morphological opening and closing operations are performed sequentially on the binarized image. Morphological opening, through pixel morphological operations of erosion followed by dilation, breaks down small abnormal connections and smooths boundaries in the binarized image; morphological closing, through pixel morphological operations of dilation followed by erosion, fills in minute voids within the target foreground. After morphological processing, the set of connected non-zero pixels in the binarized image is extracted and defined as the isolated foreground region. The geometric center coordinates of the isolated foreground region are mapped back to the three-dimensional space of the original bearing steel ball as the location of the flaw detection defect. By using optimal segmentation thresholds for binarization and combining morphological filtering to extract the isolated foreground region, background interference and pseudo-defect noise in the fused stress response spectrum are effectively filtered out, ensuring extremely high accuracy and anti-interference capability in the spatial localization of flaw detection defects.

[0092] After successfully extracting the isolated foreground region, edge contour tracking is performed on the isolated foreground region to obtain the edge curvature used to quantify the degree of abrupt changes in defect morphology. An eight-neighbor boundary tracking algorithm is used to traverse the outermost pixels of the isolated foreground region, recording the two-dimensional spatial coordinates of the outermost pixels sequentially in a clockwise direction to construct an edge contour pixel coordinate sequence. To eliminate the coordinate step error caused by the discrete pixel grid, a Gaussian smoothing kernel is used to perform one-dimensional convolution smoothing on the edge contour pixel coordinate sequence to obtain a smoothed continuous edge coordinate function. Based on the continuous edge coordinate function, the edge curvature of each pixel on the edge of the isolated foreground region is calculated. The mathematical expression for edge curvature is: .in, The first line represents the edge contour of the isolated foreground region. The edge curvature value of each pixel. The function representing the smoothed continuous edge coordinates at the th... The first derivative of the x-coordinate of each pixel with respect to the arc length of the contour. The function representing the smoothed continuous edge coordinates at the th... The first derivative of the ordinate of each pixel with respect to the arc length of the contour. The function representing the smoothed continuous edge coordinates at the th... The second derivative of the x-coordinate of each pixel with respect to the arc length of the contour. The function representing the smoothed continuous edge coordinates at the th... The second derivative of the ordinate of each pixel with respect to the contour arc length is calculated. After traversing all pixels in the edge contour pixel coordinate sequence to complete the edge curvature calculation, the maximum value among all edge curvature values ​​is extracted and defined as the maximum edge curvature feature. By calculating the edge curvature of isolated foreground regions and extracting the maximum edge curvature feature, sharp morphological abrupt changes in the defect edge can be captured, providing an extremely sensitive geometric morphological indicator for subsequent evaluation of local stress concentration effects.

[0093] To comprehensively quantify the severity of flaw locations, the stress concentration intensity index of isolated foreground regions is calculated by combining their geometric morphological features and physical stress response characteristics. The total number of pixels within the isolated foreground region is extracted, and the absolute physical area of ​​the isolated foreground region is obtained by multiplying the total number of pixels by the actual physical area represented by a single pixel. Simultaneously, the isolated foreground region is mapped back to the fused stress response map generated in step two, and the fused stress response feature values ​​of all pixels within the coverage area of ​​the isolated foreground region are extracted. The sum of all fused stress response feature values ​​is calculated and defined as the regional cumulative stress response equivalent. Combining the maximum edge curvature feature, absolute physical area, and regional cumulative stress response equivalent obtained above, the stress concentration intensity index is constructed. The mathematical expression for the stress concentration intensity index is: ,in, Indicating the stress concentration intensity index of an isolated foreground region, This represents the maximum edge curvature feature of an isolated foreground region. Represents the absolute physical area of ​​the isolated foreground region. This represents the equivalent of the regional cumulative stress response in an isolated foreground region. The weighting coefficients represent the maximum edge curvature feature. The weighting coefficient represents the absolute physical area. The weighting coefficients represent the regional cumulative stress response equivalent. Each weight is set based on the physical contribution of each feature to bearing fatigue failure: the maximum edge curvature represents the extreme value of geometric stress concentration, which is highly likely to induce microcracks, and therefore has the highest weight; physical area and cumulative stress equivalent represent the diffusion range and energy of defects, and are given equal secondary weights. For example, the weighting coefficient for the maximum edge curvature feature can be set to 0.4, the weighting coefficient for the absolute physical area to 0.3, and the weighting coefficient for the regional cumulative stress response equivalent to 0.3. The process of constructing the stress concentration intensity index nonlinearly couples the sharpness of defects, the spatial diffusion range, and the internal stress distortion energy in multiple dimensions, enabling the stress concentration intensity index to truly reflect the potential destructive force of complex defects on the fatigue life of bearing steel balls.

[0094] Step four involves converting the fused stress response map into a binary image and extracting isolated foreground regions. This allows for in-depth analysis of the edge curvature and stress concentration intensity index of the isolated foreground regions, achieving a full-chain analysis from macroscopic defect localization to microscopic stress quantification. This not only accurately pinpoints the location of flawed defects but also provides scientific, multidimensional, and highly quantified severity assessment indicators for subsequent early warning classification, significantly improving the engineering practical value and reliability of bearing steel ball surface stress detection.

[0095] Step 5: Construct a multi-level early warning model. Input the stress concentration intensity index of the flaw detection defect location into the multi-level early warning model, and output an early warning signal according to the dynamic early warning interval where the stress concentration intensity index is located.

[0096] The foundation for constructing a multi-level early warning model lies in establishing a benchmark feature space reflecting the quality fluctuation state of bearing steel balls in continuous production batches. The multi-level early warning model dynamically updates the benchmark feature space by introducing a sliding window mechanism. Specifically, in the initialization phase of the multi-level early warning model, the stress concentration intensity index of a preset number of historical normal bearing steel balls is extracted to construct an initial historical intensity index sequence as the benchmark feature space. As the inspection process progresses, the multi-level early warning model introduces a sliding window mechanism. Following the first-in-first-out queue update principle, the stress concentration intensity index of the latest inspected and determined defect-free bearing steel balls slides into the window and is added to the historical intensity index sequence. In the initialization phase of the multi-level early warning model, the stress concentration intensity index of a preset number of historical normal bearing steel balls is extracted. The preset number must meet the statistical large sample requirement of Gaussian distribution to reflect the fluctuation benchmark of normal batches. For example, the preset number can be set to 500 or 1000, and the historical intensity index sequence is constructed. As the testing process continues, the multi-level early warning model, following a first-in, first-out (FIFO) queue update principle, adds the stress concentration intensity index of the latest tested and defect-free bearing steel balls to the historical intensity index sequence, while simultaneously removing the earliest collected stress concentration intensity index from the historical intensity index sequence, thus maintaining a constant length of the historical intensity index sequence. Based on the real-time updated historical intensity index sequence, the multi-level early warning model calculates the statistical mean and statistical standard deviation of the historical intensity index sequence. The multi-level early warning model uses the statistical mean as the baseline background level of the bearing steel ball surface stress in the current production batch, and the statistical standard deviation as the natural fluctuation range of the bearing steel ball surface stress in the current production batch. By calculating the statistical mean and statistical standard deviation, the multi-level early warning model can adaptively follow small drifts in processing parameters or slow changes in ambient temperature, avoiding false alarms or missed alarms caused by using fixed thresholds. This gives the multi-level early warning model extremely strong robustness to the production environment and adaptive disturbance resistance capabilities.

[0097] After obtaining the statistical mean and statistical standard deviation, the multi-level early warning model constructs dynamic early warning intervals using these parameters. The model sets multiple sensitivity adjustment coefficients with progressive relationships, and linearly superimposes the products of the statistical mean and different sensitivity adjustment coefficients with the statistical standard deviation to calculate the multi-level dynamic early warning boundary threshold. The mathematical expression for the multi-level dynamic early warning boundary threshold is as follows: ,in, This represents the first level calculated by the multi-level early warning model. Level dynamic early warning boundary threshold, It is a positive integer. The range of values ​​depends on the total number of preset warning levels, for example, , This represents the statistical mean of the historical intensity index series. This represents the statistical standard deviation of the historical intensity index series. Indicates corresponding to the first Sensitivity adjustment coefficient for Level 1 warning status.

[0098] For example, the sensitivity adjustment coefficient can be set in three levels: the first level is set to 1.5, the second level to 2.5, and the third level to 4.0. Based on the calculated multi-level dynamic early warning boundary thresholds, the multi-level early warning model divides the one-dimensional feature space into multiple continuous and non-overlapping dynamic early warning intervals.

[0099] Specifically, the multi-level early warning model defines the interval where the stress concentration intensity index is less than the first-level dynamic early warning boundary threshold as the safe clearance interval; the interval where the stress concentration intensity index is greater than or equal to the first-level dynamic early warning boundary threshold but less than the second-level dynamic early warning boundary threshold as the mild stress concentration interval; the interval where the stress concentration intensity index is greater than or equal to the second-level dynamic early warning boundary threshold but less than the third-level dynamic early warning boundary threshold as the moderate damage evolution interval; and the interval where the stress concentration intensity index is greater than or equal to the third-level dynamic early warning boundary threshold as the severe fatigue failure interval. By constructing multi-level dynamic early warning boundary thresholds and dividing dynamic early warning intervals, the multi-level early warning model achieves refined discrete classification of continuous stress characteristic variables, providing accurate stress concentration intensity index determination boundaries for subsequent differentiated early warning responses.

[0100] After completing the division of dynamic early warning intervals, the multi-level early warning model receives the stress concentration intensity index of the flaw detection defect location output in step four. The multi-level early warning model performs a step-by-step logical comparison between the stress concentration intensity index of the flaw detection defect location and the multi-level dynamic early warning boundary thresholds to determine which dynamic early warning interval the stress concentration intensity index of the flaw detection defect location falls into. If a bearing ball has multiple stress defect problems, the maximum stress concentration intensity index on the current bearing ball is used. Calculate the warning level.

[0101] A deterministic state mapping machine is integrated within the multi-level early warning model. This state mapping machine employs a piecewise constant function as its core mapping mechanism, mapping the continuous stress concentration intensity index into discrete early warning level signals. The input to the state mapping machine is the stress concentration intensity index at the location of the flaw detection defect. With multi-level dynamic early warning boundary threshold set The output of the state mapper is a discrete warning level identifier. .

[0102] The mathematical expression for the piecewise constant mapping function inside the state mapping machine is:

[0103] ;

[0104] in, This represents the piecewise constant mapping function inside the state-mapped machine. This indicates the stress concentration intensity index of the location of the flaw to be determined during flaw detection. This represents the upper limit of the first-level dynamic early warning boundary threshold, i.e., the safe passage range. This represents the upper limit of the second-level dynamic early warning boundary threshold, i.e., the range of mild stress concentration. This represents the upper bound of the third-level dynamic early warning boundary threshold, i.e., the moderate damage evolution range. This represents the discrete warning level identifier output by the state mapper; A value of 0 corresponds to a level zero normal pass status. A value of 1 corresponds to a level 1 yellow attention status. A value of 2 corresponds to a level 2 orange re-inspection status. A value of 3 corresponds to a level 3 red scrap status.

[0105] The specific execution flow of the state mapping machine is as follows: The state mapping machine receives the stress concentration intensity index from the multi-level early warning model. Then, first read the statistical mean of the historical intensity index sequence updated based on the sliding window at the current moment. Compared with statistical standard deviation It also calls the dynamic early warning boundary threshold calculation formula to refresh the level 3 dynamic early warning boundary threshold values ​​in real time: The state mapping machine performs step-by-step logical comparison operations in order from low to high level: first, it judges... Is it strictly less than If true, output immediately. The comparison process will then be terminated; if the comparison is not successful, the process will continue. Is it strictly less than If true, then output The comparison process will then be terminated; if the comparison still fails, a decision will be made. Is it strictly less than If true, then output And terminate the comparison process; if none of the above judgment conditions are met, i.e. Greater than or equal to Then the state mapping machine output The comparison process is then terminated. The above-described step-by-step logical comparison operation ensures that each stress concentration intensity index has one and only one warning level assignment, avoiding warning signal conflicts caused by overlapping intervals.

[0106] The state mapper outputs discrete warning level identifiers. Then, discrete warning level identifiers are converted into actionable warning signals for the industrial field using a pre-compiled signal encoding table. The signal encoding table defines a one-to-one correspondence between warning level identifiers and warning signals: when... When the value equals 0, the signal encoding table outputs a level 0 normal pass signal, which is indicated by a constantly lit green status light, driving the conveyor belt to continue running and placing the current bearing steel ball into the qualified product transfer channel; when When the signal encoding table equals 1, a level 1 yellow warning signal is output. This level 1 yellow warning signal is represented by a flashing yellow status light. The multi-level warning model adds the unique identifier of the current bearing steel ball and the stress concentration intensity index of the flaw detection defect location to the observation list in the quality monitoring database, and highlights the current bearing steel ball on the monitoring terminal interface, prompting operators to pay attention to the quality change trend of subsequent bearing steel balls in the same batch. When the value equals 2, the signal encoding table outputs a level 2 orange re-inspection warning signal. This signal is characterized by a constantly lit orange status light accompanied by intermittent beeping. The pneumatic sorting baffle is activated, pushing the current bearing steel ball into the offline re-inspection channel, and an automatic re-inspection work order is generated, containing a fused stress response map and the coordinates of the flaw detection defect location. When the value equals 3, the signal encoding table outputs a level 3 red scrap warning signal. The level 3 red scrap warning signal is characterized by a constantly lit red status light accompanied by a continuous audible and visual alarm, which immediately triggers the audible and visual alarm device. At the same time, it controls the high-speed rejection mechanism to force the current bearing steel ball into the scrap collection box.

[0107] The state mapping machine employs a piecewise constant function mapping mechanism. This mechanism is a purely deterministic logical operation, containing no trainable parameters or random factors. Given the same stress concentration intensity index and the same dynamic warning boundary threshold, the state mapping machine will always output a uniquely determined warning level signal, ensuring the complete reproducibility and traceability of warning results in industrial inspection scenarios. Simultaneously, because the dynamic warning boundary threshold is updated in real-time with the historical data within the sliding window, the state mapping machine's decision boundary can adaptively track the slow drift of production conditions, maintaining an optimal balance between warning sensitivity and false alarm rate without manual intervention. By converting continuous physical quantitative indicators into discrete warning level identifiers via a piecewise constant function through the state mapping machine, and then into executable warning signals and hardware drive instructions for the industrial field via a signal encoding table, the multi-level warning model achieves seamless integration from software algorithm analysis to hardware physical actions, constructing a closed-loop quality control system.

[0108] Figure 3This is a schematic diagram of a multi-level dynamic early warning model, depicting a multi-level early warning mechanism based on the stress concentration intensity index. The multi-level early warning model calculates the mean and standard deviation based on the statistical distribution of historical data, and sets dynamic boundary thresholds (T_warn, T_risk, T_fail) in combination with sensitivity coefficients. The value range of the stress concentration intensity index is divided into a safe release interval, a mild stress concentration interval, a moderate damage evolution interval, and a severe fatigue failure interval. The state mapping machine outputs corresponding green (normal), orange (re-inspection), and red (scrap) graded early warning signals according to the interval in which the defect stress concentration intensity index falls.

[0109] This step involves constructing a multi-level early warning model based on dynamically updated historical statistical data. This model transforms the stress concentration intensity index at the flaw detection defect location into multi-level early warning signals and coordinates with industrial control hardware to execute differentiated triage actions. This tiered early warning mechanism not only overcomes the limitations of traditional one-size-fits-all fixed threshold judgments, effectively balancing the tolerance for quality fluctuations and the interception rate of extreme defects in the bearing steel ball production process, but also significantly reduces the workload of manual re-inspection through automated and intelligent tiered response strategies. This fundamentally prevents bearing steel balls with serious stress concentration risks from entering the downstream assembly stage, significantly improving the overall reliability and service life of high-end bearing products.

[0110] Example 2

[0111] Reference Figure 4 The second embodiment of this application provides a bearing steel ball surface stress detection and stress concentration early warning system, the system including: a dual-modal feature tensor construction module, a feature fusion and map generation module, a reinforcement learning adaptive segmentation module, a defect extraction and strength quantization module, and a multi-level dynamic hierarchical early warning module;

[0112] The dual-modal feature tensor construction module is used to scan the bearing steel ball, obtain the photoelectric scattering intensity sequence and the array eddy current impedance real part sequence on the surface of the bearing steel ball, align the photoelectric scattering intensity sequence and the array eddy current impedance real part sequence in spatial coordinates, and construct the photoelectric and eddy current dual-modal feature tensor.

[0113] The feature fusion and map generation module is used to perform physical feature fusion on the dual-modal feature tensor, calculate the cross covariance of the photoelectric scattering intensity gradient and the real part gradient of the eddy current impedance, extract the joint abrupt change feature reflecting the coupling relationship between surface morphology and internal stress, and generate a fused stress response map.

[0114] The reinforcement learning adaptive segmentation module is used to input the fused stress response map into the adaptive defect segmentation network, define the defect boundary search as a sequential decision process, and dynamically adjust the segmentation threshold in the map feature space through the agent.

[0115] The defect extraction and intensity quantization module is used to binarize the fused stress response map according to the optimal segmentation threshold output by the agent to obtain a binarized image, extract the isolated foreground region in the binarized image as the location of the flaw detection defect, and calculate the edge curvature and stress concentration intensity index of the isolated foreground region.

[0116] The multi-level dynamic hierarchical early warning module is used to construct a multi-level early warning model. The stress concentration intensity index of the flaw detection defect location is input into the multi-level early warning model, and the multi-level early warning model outputs the corresponding early warning signal according to the dynamic early warning interval in which the intensity index is located.

[0117] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings or direct couplings or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0118] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of this application without departing from the spirit and scope of protection of the claims. All of these variations are within the protection scope of this application.

Claims

1. A method for detecting surface stress and providing early warning of stress concentration on bearing steel balls, characterized in that, include: The bearing steel ball is scanned to obtain the photoelectric scattering intensity sequence and the array eddy current impedance real part sequence on the surface of the bearing steel ball. The photoelectric scattering intensity sequence and the array eddy current impedance real part sequence are aligned in spatial coordinates to construct a dual-mode feature tensor of photoelectric and eddy current. Physical features are fused to the dual-modal feature tensor, the cross-covariance of the photoelectric scattering intensity gradient and the real part gradient of the eddy current impedance is calculated, the joint abrupt change feature reflecting the coupling relationship between surface morphology and internal stress is extracted, and the fused stress response spectrum is generated. The fused stress response map is input into the adaptive defect segmentation network, and the defect boundary search is defined as a sequential decision process. The segmentation threshold is dynamically adjusted by the agent in the map feature space. The fused stress response map is binarized based on the optimal segmentation threshold output by the agent to obtain a binarized image. Isolated foreground regions in the binarized image are extracted as the locations of flaw detection defects. The edge curvature and stress concentration intensity index of the isolated foreground regions are calculated. A multi-level early warning model is constructed. The stress concentration intensity index of the flaw detection location is input into the multi-level early warning model, and an early warning signal is output according to the dynamic early warning interval in which the stress concentration intensity index is located.

2. The method for detecting surface stress and providing early warning of stress concentration on bearing steel balls according to claim 1, characterized in that, The bearing steel ball is spirally scanned. The photoelectric sensor emits a laser beam and receives the light signal, which is converted into a photoelectric scattering intensity sequence. The array eddy current probe excites an alternating eddy current field and captures the impedance change to form an array eddy current impedance real part sequence. Establish a spherical coordinate system, use an encoder to record the rotation and revolution angle sequences, and construct a spatial mapping function to transform the rotation and revolution angle sequences to a unified spherical coordinate system; A latitude and longitude grid is divided on the surface of the spherical coordinate system, and a bilinear interpolation algorithm is used to resample to the grid nodes to complete the spatial coordinate alignment. The aligned features are standardized and then spliced ​​according to the two-dimensional spatial topology of the latitude and longitude grid to generate a dual-modal feature tensor containing photoelectric and eddy current spatial distribution data.

3. The method for detecting surface stress and providing early warning of stress concentration on bearing steel balls according to claim 1, characterized in that, The spatial distribution data of photoelectric scattering intensity and the spatial distribution data of the real part of array eddy current impedance are separated from the dual-modal characteristic tensor; In the two-dimensional spatial topology of the latitude and longitude grid, the spatial gradient operator based on the central difference method is used to calculate the partial derivatives in the longitude and latitude directions respectively, and obtain the gradient magnitude of photoelectric scattering intensity and the gradient magnitude of the real part of eddy current impedance. Set a local sliding window of fixed size on the latitude and longitude grid, and traverse it pixel by pixel according to the preset step size; Within each local sliding window, extract all photoelectric scattering intensity gradient magnitudes and eddy current impedance real part gradient magnitudes, and calculate the cross covariance value.

4. The method for detecting surface stress and providing early warning of stress concentration on bearing steel balls according to claim 3, characterized in that, Using the cross-covariance value as a weight adjustment factor, the photoelectric scattering intensity gradient magnitude and the real part gradient magnitude of the eddy current impedance at each pixel node are multiplied by the cross-covariance value at the center of the corresponding local sliding window to obtain the joint abrupt change feature value of the pixel node. The joint mutation feature values ​​on all pixel nodes are rearranged according to the original spatial topology of the latitude and longitude grid; The rearranged joint mutation eigenvalue matrix is ​​subjected to global extremum normalization, and all joint mutation eigenvalues ​​are linearly mapped to a preset gray level range to form a single-channel two-dimensional gray level image as a fused stress response spectrum.

5. The method for detecting surface stress and providing early warning of stress concentration on bearing steel balls according to claim 1, characterized in that, Extract the global gray-level histogram vector of the fused stress response map, calculate the mean gray-level difference between the foreground and background regions under the current candidate segmentation threshold, and concatenate the global gray-level histogram vector, the mean gray-level difference, and the current candidate segmentation threshold to form a high-dimensional state vector; Define the action space, which includes three discrete actions: increasing, decreasing, and keeping the current segmentation threshold unchanged. A reward function based on the contrast of defective regions is constructed. Each time the agent performs an action, the current candidate segmentation threshold is updated. The interactive environment calculates the mean difference and spatial variance of grayscale between the foreground and background regions based on the updated candidate segmentation threshold, calculates the contrast reward value, and feeds it back to the agent.

6. The method for detecting surface stress and providing early warning of stress concentration on bearing steel balls according to claim 5, characterized in that, The agent interacts with the interactive environment through trial and error, and stores the state transition tuple containing the high-dimensional state vector, discrete actions, contrast reward value and the high-dimensional state vector of the next time step into the experience replay pool. Randomly sample batches of state transition tuples to calculate the mean squared error loss in order to optimize the weight parameters of the deep Q network; During the defect boundary search phase, the current candidate segmentation threshold is initialized to the global grayscale average value of the map. The agent calculates the expected cumulative reward value of discrete actions based on the high-dimensional state vector and selects the action corresponding to the maximum value to adjust the threshold. When the agent continuously performs actions that keep the threshold unchanged for a preset number of times or reaches the maximum limit of steps, it stops making decisions and outputs the finally determined candidate segmentation threshold as the optimal segmentation threshold.

7. The method for detecting surface stress and providing early warning of stress concentration on bearing steel balls according to claim 1, characterized in that, The current pixel node feature gray value is compared with the optimal segmentation threshold, and then reset to the target foreground value or the background zero value to obtain a binarized image. Morphological opening and closing operations are performed sequentially on the binarized image. The set of connected non-zero pixels is extracted and defined as an isolated foreground region. The geometric center coordinates of the isolated foreground region are mapped back to three-dimensional space as the location of the flaw detection defect. An eight-neighbor boundary tracing algorithm is used to construct the edge contour pixel coordinate sequence of isolated foreground regions, and a Gaussian smoothing kernel is used for one-dimensional convolution smoothing to obtain continuous edge coordinate functions. The edge curvature is obtained by calculating the first and second derivatives of pixels on the edge contour based on the continuous edge coordinate function, and the maximum value among all edge curvatures is extracted as the maximum edge curvature feature.

8. The method for detecting surface stress and providing early warning of stress concentration on bearing steel balls according to claim 7, characterized in that, Extract the total number of pixels contained in the isolated foreground region, multiply the total number of pixels by the actual physical area represented by a single pixel, and obtain the absolute physical area of ​​the isolated foreground region. The isolated foreground region is mapped back to the fused stress response map. The fused stress response feature values ​​of all pixels within the coverage area of ​​the isolated foreground region are extracted, and the sum of all fused stress response feature values ​​is calculated. The sum is defined as the regional cumulative stress response equivalent. We assign corresponding weighting coefficients to the maximum edge curvature feature, absolute physical area, and regional cumulative stress response equivalent, and then perform nonlinear coupling calculations to construct the stress concentration intensity index of the isolated foreground region.

9. The method for detecting surface stress and providing early warning of stress concentration on bearing steel balls according to claim 1, characterized in that, The multi-level early warning model dynamically updates the stress concentration strength index of historical normal bearing steel balls according to the first-in-first-out principle, uses the updated stress concentration strength index as the early warning signal, and calculates the mean and standard deviation. The boundary threshold of the multi-level dynamic early warning set is calculated by combining multiple preset sensitivity adjustment coefficients, and multiple dynamic early warning intervals are divided. The stress concentration intensity index of the flaw detection defect location is compared with the boundary threshold of the multi-level dynamic early warning set to determine the dynamic early warning interval it falls into, and the corresponding early warning signal is output through the state mapping machine.

10. A bearing steel ball surface stress detection and stress concentration early warning system, used to implement the bearing steel ball surface stress detection and stress concentration early warning method according to any one of claims 1 to 9, characterized in that, include: The module includes a dual-modal feature tensor construction module, a feature fusion and map generation module, a reinforcement learning adaptive segmentation module, a defect extraction and intensity quantization module, and a multi-level dynamic hierarchical early warning module. The dual-modal feature tensor construction module is used to scan the bearing steel ball, obtain the photoelectric scattering intensity sequence and the array eddy current impedance real part sequence on the surface of the bearing steel ball, align the photoelectric scattering intensity sequence and the array eddy current impedance real part sequence in spatial coordinates, and construct the photoelectric and eddy current dual-modal feature tensor. The feature fusion and map generation module is used to perform physical feature fusion on the dual-modal feature tensor, calculate the cross covariance of the photoelectric scattering intensity gradient and the real part gradient of the eddy current impedance, extract the joint abrupt change feature reflecting the coupling relationship between surface morphology and internal stress, and generate a fused stress response map. The reinforcement learning adaptive segmentation module is used to input the fused stress response map into the adaptive defect segmentation network, define the defect boundary search as a sequential decision process, and dynamically adjust the segmentation threshold in the map feature space through the agent. The defect extraction and intensity quantization module is used to binarize the fused stress response map according to the optimal segmentation threshold output by the agent to obtain a binarized image, extract the isolated foreground region in the binarized image as the location of the flaw detection defect, and calculate the edge curvature and stress concentration intensity index of the isolated foreground region. The multi-level dynamic hierarchical early warning module is used to construct a multi-level early warning model. The stress concentration intensity index of the flaw detection defect location is input into the multi-level early warning model, and the multi-level early warning model outputs the corresponding early warning signal according to the dynamic early warning interval in which the intensity index is located.

Images