Motor slot wedge embedding quality detection method based on machine vision

Abstract

The application discloses a motor slot wedge embedding quality detection method based on machine vision, and particularly relates to the technical field of quality detection; image collection is carried out on the motor stator along the ring direction, a quasi-three-dimensional structure model of the slot wedge embedding area is constructed, and micro convex and indentation features are identified; a micro gap contour line between the slot wedge and the stator slot side wall is extracted, a embedding trend curve is calculated in combination with a machine learning algorithm; pseudo defects are judged and filtered based on the trend and defect features, and a real feature data set is generated; a multi-dimensional slot wedge embedding quality state model is further constructed; clustering identification is carried out based on model parameter distribution, quality grade scores are output, and structure failure probability is predicted; when there are multiple score abnormal slot wedges, the assembly batch, assembly personnel or tool state thereof are traced back; the method can realize automatic and high-precision detection of the slot wedge embedding quality, has a process tracing function, and is suitable for intelligent quality control links in the motor manufacturing process.

Description

Technical Field

[0001] This invention relates to the field of quality inspection technology, and specifically to a machine vision-based method for quality inspection of motor slot wedge installation. Background Technology

[0002] With the continuous development of the motor industry, motor manufacturing processes are becoming increasingly complex, especially in the installation of slot wedges. The quality of the slot wedges directly affects the overall performance and service life of the motor. As an important component of the motor stator coils, the installation quality of the slot wedges plays a crucial role in the motor's operational stability and thermal management performance. Traditional methods for inspecting the installation quality of slot wedges mainly rely on manual inspection or simple mechanical testing. These methods are not only inefficient but also easily affected by human factors, leading to inaccurate and inconsistent test results.

[0003] In existing technologies, traditional manual inspection methods cannot accurately obtain key quality parameters such as the installation position, angle, and tightness of motor slot wedges in real time. Especially in high-speed production environments, the error rate of manual operation is relatively high, making it difficult to meet the high-efficiency operation requirements of the production line. In addition, since even small errors in the slot wedge installation process often affect the working efficiency and reliability of the motor, how to accurately and automatically inspect the quality of slot wedge installation has become an urgent technical problem to be solved in motor manufacturing.

[0004] Therefore, how to automatically identify and evaluate the installation quality of motor slot wedges using high-precision inspection technology has become a key technical issue in motor manufacturing. To address this problem, machine vision-based inspection methods have emerged. Machine vision systems can acquire image information of motor slot wedges in real time and extract their geometric features, position, and orientation through image processing technology, thereby determining whether their installation is compliant. This method not only improves inspection efficiency but also avoids errors from manual operation, achieving accurate and rapid inspection of slot wedge installation quality. Summary of the Invention

[0005] The purpose of this invention is to provide a machine vision-based method for detecting the quality of motor slot wedge installation, in order to address the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a machine vision-based method for detecting the quality of motor slot wedge installation, comprising:

[0007] S100: Image acquisition is performed on the stator of the motor along the circumferential direction to obtain a continuous image sequence containing multiple slot wedge embedding areas;

[0008] S200 performs image stitching and registration on the image sequence to construct a quasi-three-dimensional structural model of the motor stator slot wedge region, and uses depth information to identify the micro-protrusions and indentation features on the slot wedge surface;

[0009] S300: Extract the micro-gap contour line between the slot wedge and the stator slot sidewall, and calculate the slot wedge embedding trend curve by combining machine learning algorithm;

[0010] S400 determines whether there are false defects in the slot wedge based on the identified micro protrusions and indentation features on the surface of the slot wedge and the slot wedge embedding trend curve. It then removes or corrects misjudged areas through a false defect filtering algorithm, generating an optimized set of real feature data.

[0011] S500, based on the real feature data set, extract the attitude parameters, tightness index, position deviation and trend characteristics of each slot wedge region and map them to the quality risk space to construct a multi-dimensional slot wedge embedding quality status model;

[0012] S600 uses a clustering algorithm to classify and identify all slot wedges based on the parameter distribution of the multidimensional embedded quality state model of slot wedges, outputs the quality level score of each slot wedge, and predicts the probability of structural failure during its future use.

[0013] S700: When multiple slot wedge quality grade scores are abnormal, trace back to the assembly batch, assembly personnel, or tool status.

[0014] Preferably, in step S200, image registration is performed based on the extraction of motor stator slot edge points, slot wedge tip center points, and stator tooth contour feature points in the image.

[0015] Preferably, the constructed quasi-three-dimensional structural model of the motor stator slot wedge region is point cloud data in (x, y, z) format, where the z-axis represents the height value of the slot wedge surface relative to the reference plane. The quasi-three-dimensional structural model is obtained by a multi-view stereo matching algorithm combined with structured light assistance.

[0016] Preferably, in step S300, the extraction of the micro-gap contour includes:

[0017] Multiple vertical sections were collected along the wedge insertion direction, and the minimum gap distance between the left and right edges of the wedge and the inner wall of the stator slot was calculated to form a depth difference matrix. The first derivative of each micro-gap contour line was calculated and polynomial fitting was performed to extract the slope change rate, curvature change and root mean square error of the fitting residual as trend features.

[0018] Preferably, based on the feature parameters of the fitted curve, the K-means clustering algorithm is used to classify the embedded sections and form a slot wedge embedding trend curve, wherein the trend score of each point is calculated based on the curve fitting residual, slope and curvature weighted average.

[0019] Preferably, in step S400, the sliding window analysis and connected component processing method is used to extract the micro-protrusion features and indentation features on the surface of the slot wedge, calculate their corresponding height values, areas and spatial positions, and make a one-to-one correspondence with the slot wedge embedding trend curve at the corresponding embedding position to determine their trend consistency.

[0020] Preferably, by comparing the height change direction of the defect area with the derivative direction of the slot wedge embedding trend curve, when there is an inconsistency in the trend and the defect is distributed in a discrete state in space, the defect is judged as a pseudo-defect, and the defect area is repaired by the Laplace boundary expansion algorithm.

[0021] Preferably, in S500, the input vector Q of the slot wedge multidimensional embedding quality state model includes: slot wedge tilt angle, left and right height difference, tightness index, position deviation, maximum misalignment, average trend slope, trend fluctuation amplitude and trend stability. The state vector is used to train the support vector machine model to achieve risk scoring after dimensionality reduction processing by principal component analysis.

[0022] Preferably, in step S600, a density-based spatial clustering algorithm is used to cluster and classify all slot wedge state vectors, and the slot wedge quality level score is calculated based on the cluster center characteristics. The score range is 0 to 100 points, and it is divided into four quality levels: A, B, C and D.

[0023] Preferably, if two or more slot wedges in a motor stator have a score below 70, the traceability process is triggered, and the slot wedge assembly batch number, assembly personnel identification code and assembly tool and equipment number are associated respectively. Significance analysis and statistical comparison are performed to output potential anomaly source information.

[0024] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0025] 1. This invention achieves full-coverage, non-contact, and high-precision automated inspection of slot wedge fitting quality through high-precision image acquisition, quasi-3D reconstruction, and multi-dimensional feature extraction. Compared with traditional methods that rely on manual visual inspection or contact inspection, this invention solves the problem that traditional methods are difficult to detect micro-deformation, shallow damage, and abnormal assembly trends.

[0026] 2. This invention constructs a multi-dimensional slot wedge mounting quality status model and introduces clustering identification and failure probability prediction algorithms to achieve quantitative analysis and early warning from structural data to quality level, while also possessing the ability to trace the process source of abnormal mounting conditions. This method, for the first time, correlates the surface defect characteristics of the slot wedge with the mounting trend curve, effectively distinguishing between pseudo-defects and real defects, significantly improving the accuracy of defect identification and the robustness of the system, and providing reliable technical support for intelligent quality control and closed-loop traceability in motor manufacturing. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0028] Figure 1 This is a flowchart of the machine vision-based motor slot wedge embedding quality inspection method of the present invention.

[0029] Figure 2 This is a flowchart of the pseudo-defect identification method of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] For examples, please refer to Figure 1 , 2 As shown in this embodiment, the machine vision-based motor slot wedge embedding quality inspection method includes:

[0032] S100: Image acquisition is performed on the motor stator along the circumferential direction to obtain a continuous image sequence containing multiple slot wedge embedding areas.

[0033] The image acquisition process targets the stator of the motor to be inspected, acquiring visual images of its slot wedge mounting areas. To fully cover all slot wedge areas of the stator and ensure image continuity and spatial consistency, this embodiment employs a multi-angle, high-precision circumferential imaging method, specifically including the following operations:

[0034] Tooling structure design: The motor stator to be tested is horizontally mounted on a rotary worktable with a rotary drive mechanism. The rotary worktable can achieve 360° continuous or intermittent rotation with a rotation accuracy better than ±0.1°. The worktable is equipped with positioning fixtures for precise alignment, ensuring that the stator maintains a stable posture during rotation without axial or radial offset.

[0035] Image acquisition component: Several (e.g., 3 or 4) high frame rate industrial cameras are evenly distributed on the arc-shaped slide rails along the outer ring of the stator. The industrial cameras have a resolution of 12 megapixels or higher, equipped with telecentric lenses and ring light sources to reduce image distortion and shadow interference. The radial distance of the cameras can be adjusted via an electric slide to accommodate variations in the size of stators for different motor specifications.

[0036] Image acquisition process: As the stator slowly rotates, each industrial camera continuously acquires image data unfolding along the circumference at a set frame rate (e.g., 30 frames / second). To ensure the spatial continuity of the image data and the accuracy of subsequent processing, the acquisition system synchronously records the rotation angle information corresponding to each frame. The resulting image sequence ultimately forms a continuous slot wedge embedding image sequence covering the entire circumference of the stator.

[0037] Data preprocessing: Before entering the subsequent image stitching and feature extraction process, the acquired image sequence will first undergo preliminary preprocessing, including distortion correction, illumination equalization and noise suppression, in order to improve the edge sharpness and image consistency of the slot wedge region.

[0038] The image acquisition scheme described in this embodiment can ensure that the wedge-mounted areas of each slot of the motor stator are acquired without blind spots, omissions, high resolution, and high consistency, providing high-quality basic data for subsequent machine vision-based 3D modeling, feature extraction, and quality analysis.

[0039] S200 performs image stitching and registration on the image sequence to construct a quasi-three-dimensional structural model of the motor stator slot wedge region, and uses depth information to identify minute protrusions and indentation features on the slot wedge surface.

[0040] To achieve image consistency in the spatial coordinate system, this embodiment employs a feature point-based registration method, specifically including:

[0041] Extract structural feature points from each image, including the edge points on both sides of the stator slot, the center point of the top of the slot wedge, and the outline of the stator core teeth;

[0042] The SIFT (Scale Invariant Feature Transform) algorithm is used to match feature points between adjacent image pairs; the RANSAC (Random Sample Consensus) algorithm is used to remove outlier matching points; and the affine transformation matrix is ​​calculated based on the matching results to perform registration and correction on the images, so that all images are accurately aligned in a unified coordinate system.

[0043] After registration, the images are stitched together using the following method:

[0044] Following the stator rotation angle sequence, all images are stitched together sequentially to form a complete circumferential unfolded image. During the stitching process, image fusion algorithms (such as multiple exposure fusion and Poisson fusion) are used to perform grayscale transition processing on the image edges to eliminate brightness differences and boundary gaps. The final stitched image represents the circumferential length of the stator in the horizontal dimension and corresponds to the embedding depth of the slot wedge in the vertical dimension, with a resolution better than 100μm / pixel.

[0045] After obtaining the fully unfolded image, and combining the camera calibration parameters (including intrinsic and extrinsic parameters) completed in step S100, this embodiment introduces the following depth estimation and structure modeling techniques:

[0046] The depth value of each pixel is estimated using structured light-assisted or multi-view stereo matching (MVS) algorithms. The depth information is combined with the 2D image stitching to generate a quasi-3D structural model of the motor stator slot wedge embedding area. The model is stored in the form of (x,y,z) point cloud. The z-axis represents the height difference relative to the reference plane, which can reflect the subtle deformations such as bulging and sinking of the slot wedge. The system accuracy can reach ±30μm. The model supports comparison and analysis with CAD reference models to quantify structural differences.

[0047] It should be noted that, to enhance the stability and anti-reflection capability of depth estimation, this embodiment preferably uses a structured light-assisted reconstruction method. Specifically, during the stator rotation process, a set of infrared stripe projectors are synchronously triggered to project coded stripe patterns onto the surface of the stator slot wedge at a fixed frequency, while images are acquired by industrial cameras installed at multiple viewing angles.

[0048] The encoded pattern employs a grayscale phase-shift encoding scheme, including three frames of phase-shift maps, with continuously varying grayscale distribution, effectively addressing different surface reflectivity conditions. The stripe pattern is individually bound to each camera acquisition channel. Through a pre-calibrated projection-imaging geometry, pixel coordinates are mapped into three-dimensional space using a reverse mapping method, thereby obtaining a high-density depth point cloud.

[0049] This method effectively improves the accuracy of depth measurement under conditions of metallic reflection, high light spots, or curved surface deformation, with an overall accuracy of ±30 micrometers.

[0050] Based on the constructed quasi-3D model, detailed identification is performed on the surface of the slot wedge, including but not limited to the following types:

[0051] Micro-bulge identification: Small bulges with a height greater than a set threshold (e.g., 50μm) are detected through three-dimensional curvature analysis, which may be caused by foreign matter inclusions or springback of slot wedge material during assembly;

[0052] Indentation identification: Extract local areas where there is a sharp drop in the z-axis direction. The indentation depth ranges from 50 to 150 μm and the shape is a local depression, which is often caused by improper contact of assembly tools or deformation of the groove.

[0053] Edge burr detection: Analyze the high-frequency fluctuation area of ​​the groove wedge edge segment. If obvious "sawtooth abrupt change" appears, it is judged as a burr defect, which affects the installation stability.

[0054] Information such as the coordinates, size, shape, and distribution direction of the defect area will be uniformly encoded and used as a standard structural feature set to be passed into subsequent steps for trend analysis and quality assessment.

[0055] S300 extracts the micro-gap contour line between the slot wedge and the stator slot sidewall, and calculates the slot wedge embedding trend curve using machine learning algorithms.

[0056] In the acquired quasi-3D structural model, each 3D point contains (x, y, z) coordinate information, where the z-axis represents the height of the slot wedge surface relative to the reference plane. To accurately obtain the spatial clearance between the slot wedge and the stator slot sidewall, each pair of slot sidewall contour points and slot wedge edge points are first matched along the slot axial direction, and their minimum distance in the z-axis direction is calculated as the spatial clearance value.

[0057] For each section perpendicular to the length of the slot wedge, the distance from the left slot wall point to the left edge of the slot wedge is extracted and denoted as dL(i), and the distance from the right slot wall point to the right edge of the slot wedge is denoted as dR(i), where i is the slot section number, ranging from 1 to N (N is the number of sections). The depth difference data point set of the slot wedge embedding region is constructed using (i, dL(i), dR(i)). All data points are arranged into a two-dimensional matrix D, where the rows represent the section positions and the columns represent the left and right gaps, called the depth difference matrix.

[0058] To avoid assembly tolerances and noise affecting the accuracy of the results, a Gaussian filtering algorithm was used to smooth the z-axis coordinates in the point cloud before gap calculation. The filtering window was set to 5 points and the standard deviation σ was set to 0.8.

[0059] In the depth difference matrix D, the left and right gap data on each cross section are extracted to form the left gap sequence L={dL(1),dL(2),...,dL(N)} and the right gap sequence R={dR(1),dR(2),...,dR(N)}, respectively. These two sets of data are the original gap variation curves of the slot wedge in the insertion direction (i.e., along the slot axis).

[0060] To further analyze the installation stability of the slot wedge in different sections, the entire installation length was divided into several equally spaced segments, each with a length of 5 mm. For the gap data within each segment, its local subsequences were extracted separately, such as L1, L2, ..., Lm and R1, R2, ..., Rm, where m is the number of segments.

[0061] The first derivative is calculated for each subsequence to determine the upward, downward, or fluctuating trend of the gap curve. If the derivative change is continuously positive (or negative) and the rate of change exceeds the set threshold of 0.05 mm / mm, it is marked as a trend change segment.

[0062] For each local gap subsequence obtained in the previous step, a polynomial fitting is performed using the least squares method. The fitting adopts a third-order polynomial form, namely: , where x is the position coordinate of the slot wedge insertion direction, and f(x) is the gap value.

[0063] After fitting, the following structural features of each curve segment are extracted:

[0064] Rate of change of slope (i.e., derivative function) (maximum absolute value)

[0065] Curvature change (i.e.) (range of variation)

[0066] The root mean square error (RMSE) of the fitting residuals is used to determine whether the curve segment has consistency.

[0067] Based on the above characteristics, the K-means clustering algorithm was used to divide the gap curves into three categories:

[0068] The classification includes: stable embedded section (slope close to 0); intermittent loose section (gap gradually increases); and tight interference section (gap gradually decreases). This classification characterizes the assembly trend of the slot wedge within the slot, providing structural input features for quality assessment.

[0069] By integrating the classification labels, fitting parameters, and position indices of all segments on the left and right sides, a complete slot wedge embedding trend curve is constructed. The x-axis represents the embedding direction, and the y-axis represents the local gap change trend score (the score is calculated by weighting the fitting residual, slope, and curvature, and the value range is 0–1).

[0070] Let the final fitting trend curve be T(x). If the score of T(x) is greater than 0.7 in the continuous region and the trend is monotonically increasing or monotonically decreasing, it is judged that there is a potential slot wedge rotation offset or slippage trend in the region. If T(x) shows periodic fluctuations, it indicates that there may be a periodic compression section of the slot wedge caused by interference assembly.

[0071] This trend curve will be used as an input feature for subsequent slot wedge quality status modeling and will provide graphical visual results for reference in the installation process debugging.

[0072] S400 determines whether there are false defects in the slot wedge based on the identified micro-protrusions and indentation features on the surface of the slot wedge and the slot wedge embedding trend curve. It then removes or corrects misjudged areas through a false defect filtering algorithm, generating an optimized set of real feature data.

[0073] First, in the quasi-three-dimensional structural model established in step S200, the height undulation data of the groove wedge surface along the z-axis is analyzed, and regions with abrupt height changes are extracted as candidate defect points. The specific extraction method is as follows:

[0074] Using the sliding window analysis method, a 5×5 pixel window is used to scan the block by block in the three-dimensional height matrix;

[0075] If the difference between the height value of a certain pixel and the average height value of its surrounding pixels exceeds a set threshold (positive value), then... micrometers, negative direction If the value is in micrometers, then mark the point as a suspected tiny bump or indentation;

[0076] Connectivity analysis is performed on adjacent suspected defect points to form defect regions.

[0077] For each defective region, extract the following parameters:

[0078] Height value H: The difference between the maximum and minimum values ​​on the z-axis of this region;

[0079] Area A: Counts the number of pixels within a connected region and converts it into physical area (in square millimeters);

[0080] Spatial position P: Records the coordinate position (in millimeters) of this area in the slot wedge insertion direction. This position corresponds one-to-one with the position axis of the slot wedge insertion trend curve.

[0081] Finally, a set of defect parameters containing location, height, and area is obtained, denoted as {P(i),H(i),A(i)}, where i is the defect number.

[0082] Based on the obtained defect parameter set, its position P(i) is paired with the corresponding position x=P(i) in the slot wedge embedding trend curve T(x) generated in step S300 to determine the consistency of their changing trends. The judgment criteria are as follows:

[0083] For a small protrusion feature, if its height change is positive (i.e. higher than the surrounding area), then the gradient direction of the feature should be consistent with the slope direction of the embedding trend curve T(x) at that position, i.e., both rising or falling at the same time.

[0084] For indentation features, if the height change is negative (i.e., lower than the surrounding area), then the direction of the indentation should be consistent with the direction of the gap change in the trend curve. The calculation method is as follows:

[0085] Let T′(P(i)) be the local derivative of the embedding trend curve at P(i);

[0086] Let H′(i) be the direction of change of defect height. If H(i) is positive, then H′(i) = +1; if H(i) is negative, then H′(i) = -1.

[0087] If T′(P(i)) and H′(i) have the same sign, they are considered to have the same trend; otherwise, they are considered to have different trends.

[0088] Spatial distribution continuity analysis was performed on all extracted defect features:

[0089] First, sort all defects by position P(i) and calculate the distance ΔP between adjacent defects;

[0090] If ΔP is greater than the set continuity threshold (e.g., 3 mm) and the height value does not show a gradual trend (i.e., the height direction is opposite or the change amplitude changes abruptly), it is judged as a discrete distribution.

[0091] At the same time, when the direction of H′(i) of this feature is inconsistent with the direction of the trend curve T′(P(i)), it is judged as "trend mismatch".

[0092] Defects that meet both the criteria of "discrete distribution" and "inconsistent trend" are classified as non-structural pseudo-defects, which may be caused by the following factors:

[0093] Lens reflections or localized highlights during the imaging process; abrupt changes in texture caused by surface oil; disturbances caused by surface deposits or dust; and misjudgments of contours due to image stitching errors.

[0094] After completing the false defect identification, a false defect filtering algorithm is executed to clean up the dataset. The steps are as follows:

[0095] Remove the pixel locations of all false defects from the original 3D height data;

[0096] The removed regions are reconstructed and compensated using an edge interpolation algorithm. The interpolation method is Laplacian boundary extension to maintain surface continuity. The remaining, unremoved defect features constitute a new feature set, denoted as . That is, the optimized set of true structural features; among which, This indicates the installation position of the i-th slot wedge, which can be a specific coordinate point of the slot wedge or a calibration value of the installation position. This represents the surface height value of the i-th slot wedge, which can be the height of the slot wedge surface relative to the reference plane. This represents the area of ​​the i-th slot wedge, typically used to indicate the effective area covered by the wedge surface. This set serves as the data input for subsequent wedge quality status modeling, enhancing the robustness and accuracy of the evaluation model. Through the above processing, the misleading influence of non-structural interference factors on defect detection results can be significantly reduced.

[0097] S500, based on the real feature data set, extract the attitude parameters, tightness index, position deviation and trend characteristics of each slot wedge region and map them to the quality risk space to construct a multi-dimensional slot wedge mounting quality status model.

[0098] After completing the false defect filtering in step S400, the set of true structural feature data of the slot wedge embedding region is obtained, denoted as... This data includes the embedding location, height, and area information of each defect. Based on this dataset, this step further extracts several key structural and installation features for each slot wedge region, including the following four types of parameters:

[0099] The slot wedge attitude is used to describe the spatial orientation and attitude changes of the slot wedge in the slot. The extraction process is as follows:

[0100] A linear reference model is established by fitting the point cloud of the upper edge of the slot wedge. The fitting equation is y=kX+b; where X represents the position between the slot wedge and the stator slot sidewall, y represents the height or deviation of the slot wedge, k is the slope of the straight line, representing the rate of change between the slot wedge and the stator slot sidewall, and b is the intercept of the straight line, representing the initial deviation or initial installation state between the slot wedge and the stator slot wall.

[0101] Calculate the angle θ between the centerline of the slot wedge and the stator reference axis, which is expressed as the slot wedge inclination angle in degrees.

[0102] The height difference Δh between the left and right edges of the groove wedge is used to characterize its lateral torsion.

[0103] The attitude parameters are ultimately represented as a vector S=[θ,Δh].

[0104] The tightness of the slot wedge reflects the degree of compression between the slot wedge and the slot wall. The specific calculation method is as follows:

[0105] On the left and right sides of the groove wedge, the gap value on a vertical section is extracted every 1 mm, and is denoted as dL(j) and dR(j) respectively, where j is the section number;

[0106] The overall average clearance is calculated as μd = (dL + dR) / 2. The tightness index F is defined as F = 1 / (1 + μd), where dL represents the minimum clearance between the left slot wall and the slot wedge, dR represents the minimum clearance between the right slot wall and the slot wedge, and μd represents the overall average clearance, which is the average distance between the slot wedge and the stator slot wall.

[0107] If the F value is close to 1, it indicates that the slot wedge is tightly fitted; if the F value is much lower than 1, it indicates that there is a risk of loosening. The tightness index is normalized for stators of different sizes and then used for subsequent risk assessment modeling.

[0108] Positional deviation reflects the error between the actual installation position and the theoretical position of the slotted wedge, and is calculated using the following method:

[0109] Align the upper edge line of the slot wedge with the standard design model, and record the offsets Δx and Δy of each matching point in the x and y directions;

[0110] Calculate the maximum deviation This indicates the possible misalignment or runout of the slot wedge during installation; all deviation values ​​are represented by the vector P=[δx,δy,δmax], where δx represents the deviation of the slot wedge in the horizontal direction (usually the radial direction of the stator slot), that is, the difference in horizontal distance between the center of the slot wedge and the center of the stator slot. δy represents the deviation of the slot wedge in the vertical direction (usually the axial direction of the stator slot), that is, the difference in vertical distance between the center of the slot wedge and the center of the stator slot.

[0111] The trend characteristics are derived from the slot wedge embedding trend curve T(x) constructed in step S300, from which the following structural features are extracted:

[0112] The average slope α of the trend in the embedding direction reflects the overall embedding uniformity;

[0113] The local maximum fluctuation amplitude β represents the periodic offset or resistance that may exist during the assembly process;

[0114] The trend stability σt is obtained by calculating the first-order variance of the trend derivative.

[0115] Trend characteristics are represented by the vector T=[α,β,σt].

[0116] The above structural features are integrated to form a multi-dimensional feature vector Q=[S,F,P,T] for slotted wedges, i.e., Q=[θ,Δh,F,δx,δy,δmax,α,β,σt]. To achieve quantitative judgment of quality level, each slotted wedge is mapped to a multi-dimensional quality risk space based on this vector. The modeling process is as follows:

[0117] Principal component analysis was used to reduce the dimensionality of all Q vectors and extract key feature combinations.

[0118] A support vector machine classification algorithm is introduced for training, and slot wedges with known quality levels in historical samples are used as the training set to construct a discrimination boundary.

[0119] Project the Q vector of the current slot wedge onto this space and output its risk score R, which ranges from 0 to 1. The closer it is to 1, the more stable the quality.

[0120] Finally, an embedded quality state model M(Q) is constructed for the prediction and evaluation of the quality state of each slot wedge.

[0121] The model supports real-time updates and performs self-learning optimization by comparing with actual fault data to improve prediction accuracy.

[0122] To improve the computational efficiency of mounting quality state modeling, this embodiment uses principal component analysis (PCA) algorithm to reduce the dimensionality of multidimensional features after constructing the state vector Q.

[0123] Specifically, covariance matrix decomposition is performed on Q=[θ,Δh,F,δx,δy,δmax,α,β,σt] to calculate eigenvalues ​​and contribution rates. Based on the principle that the cumulative variance contribution rate is greater than or equal to 95%, the first k principal components (usually 4-5 dimensions) are retained to construct the embedding quality state model.

[0124] This process preserves the main differences between samples, reduces the complexity of the classification model, and improves the model's stability.

[0125] S600 uses a clustering algorithm to classify and identify all slot wedges based on the parameter distribution of the multi-dimensional embedded quality state model of the slot wedges, outputs the quality level score of each slot wedge, and predicts the probability of structural failure during its future use.

[0126] In step S500, the structural characteristic parameters of each slot wedge have been integrated into an embedding state vector Q=[θ,Δh,F,δx,δy,δmax,α,β,σt], and a multidimensional embedding quality state model M(Q) has been established. This embodiment uses the set of state vectors of all slot wedges as input, and performs cluster analysis, rating, and failure risk prediction on them. The specific method is as follows:

[0127] Collect the multidimensional embedded state vectors Q1, Q2, ..., Qn of all slot wedges on the stator of each motor, where n is the total number of slot wedges, and form them into a sample set Qall.

[0128] To eliminate the influence of differences in the dimensions of different features, all sample vectors are first standardized. Z-score standardization is used:

[0129] For each feature dimension, calculate its overall mean μ and standard deviation σ;

[0130] For each slot wedge, perform a normalization transformation on the value x in that dimension: x′=(x−μ) / σ; where,

[0131] x′ represents the standardized data points, resulting in a mean of 0 and a standard deviation of 1.

[0132] The resulting set of standardized vectors is used as input for subsequent clustering.

[0133] Density-Based Spatial Clustering of Applications with Noise (DBSCAN) was used to perform cluster analysis on the normalized vector set.

[0134] Set the neighborhood radius ε (e.g., 0.8) and the minimum number of samples minPts (e.g., 10) as hyperparameters of the algorithm;

[0135] During the clustering process, slot wedges with similar embedding state characteristics are grouped into the same cluster, while outlier slot wedges at the edge are automatically identified as outliers.

[0136] The clustering results divide all slot wedges into K categories, each representing a typical embedding quality feature pattern.

[0137] For each category, a weighted score is calculated based on its average tightness index F, positional deviation δmax, and trend fluctuation σt. The quality scoring function Score(Q) is defined as follows: Score(Q) = 0.4 × F + 0.3 × (1 − δmax / δmax_max) + 0.3 × (1 − σt / σt_max); where δmax_max and σt_max are the maximum values ​​of all samples used for normalization. To ensure dynamic adaptability of the scoring, δmax_max and σt_max are derived from the statistical maximum values ​​of the current batch (i.e., all slot wedges within the same stator or the same work cycle). This method allows for adjustment of the evaluation benchmark based on actual assembly conditions, avoiding distortion caused by historical extreme values. When batch dimensions change significantly or processes change, the system will trigger a scoring benchmark update process to ensure consistency and horizontal comparability of the scores.

[0138] The scoring results are set from 0 to 100 points, and are divided into the following levels:

[0139] 90-100 points: Grade A (Excellent installation quality); 70-89 points: Grade B (Acceptable installation quality); 50-69 points: Grade C (Poor installation quality, re-inspection recommended); 0-49 points: Grade D (Serious installation problems, rework recommended).

[0140] To achieve structural reliability prediction, in this embodiment, based on the slot wedge scoring results and the failure annotation data in historical samples, a probability prediction model is trained to predict the structural failure probability under future operating conditions.

[0141] Using the logistic regression method, the input variables are the state vectors Q of each slot wedge and the score Score(Q), and the output is the failure probability p, with a value range from 0 to 1.

[0142] The model training samples are sourced from data of typical failure slot wedges such as structural looseness, shedding, and thermal deformation that occurred during the existing service life.

[0143] For each current slot wedge, calculate its failure probability p. If p≥0.6, mark it as a high-risk slot wedge.

[0144] Each slot wedge will output its clustering category number, scoring level (A - D), and the structural failure probability p. The output results are used to guide subsequent re-inspection on the production line, adjustment of process parameters, and closed-loop control of assembly quality. High-risk slot wedges will be recorded in the quality traceability database as the basis for assembly optimization.

[0145] Through the processing of this step, the automated evaluation, hierarchical identification, and reliability prediction of the slot wedge installation quality are achieved, solving the problem that traditional manual methods cannot simultaneously consider multi-dimensional structural features and future failure trends, and significantly improving the quality control ability and product stability in the motor production process.

[0146] S700, when there are multiple abnormal scoring results of slot wedge quality, trace back to its assembly batch, assembly personnel, or tool status.

[0147] In step S600, the installation quality score Score(Q) is calculated for each slot wedge and classified into four categories: A, B, C, and D according to the level. When there are two or more slot wedges in a certain motor stator rated as C or D level, it is regarded that there is abnormal installation behavior for this workpiece. This step takes the abnormal scoring as the trigger condition and performs multi-dimensional traceability analysis to trace the potential root causes of assembly abnormalities, including assembly batch, personnel, and tool status. The specific method is as follows:

[0148] Represent the set of scoring results in each motor stator as Score_Set={s1, s2,..., sn}, where n is the total number of slot wedges in this stator.

[0149] Set the abnormal scoring threshold T = 70 points. If there is an abnormal scoring quantity M≥2, that is, satisfying: Count(si∈Score_Set|si<T)≥2, then trigger the traceability operation of this step.

[0150] Each slot wedge is assigned an assembly batch code before installation. The batch number (BID) is stored in the unique identifier of the slot wedge and recorded by scanning during assembly. The traceability operation includes the following processes:

[0151] Extract the batch number set B={BID1,BID2,...,BIDm} of all abnormal slot wedges; if there is a duplicate number BID* in set B and its coverage exceeds 50% of the number of abnormal slot wedges, then BID* is initially judged to be a suspected abnormal batch.

[0152] The batch number was marked as a "quality concern batch," and a horizontal batch comparison was performed to check whether there were any discrepancies between the corresponding raw material warehousing records, storage time, and outbound time.

[0153] To enhance the statistical basis for batch anomaly identification, this embodiment performs hypothesis testing on the batch distribution of scoring anomaly slot wedges. First, a batch-anomaly association table is constructed, and the anomaly frequency corresponding to each batch ID (BID) is statistically analyzed. Then, Fisher's exact test is performed on the major batch BIDs, with a significance level set at α=0.05. If the p-value is less than this threshold, the BID is considered to have a statistical association with the current anomaly slot wedge and is further marked as a "potentially anomaly batch." This method effectively avoids misjudgments caused by sample fluctuations and improves the confidence level of anomaly batch identification.

[0154] Before each stator is put into operation, the operator swipes their card to record their employee ID. The system then binds the operator's ID (PID) to the slot wedge installation time range. Traceability operations include:

[0155] Find the operator IDs within the current stator assembly period and form a personnel set P={PID1,PID2,...};

[0156] Search for the occurrence rate of slot wedge anomalies in nearly 10 stators operated by this operator;

[0157] If the anomaly rate of a certain person's PID* exceeds 30% within the statistical period, it will be marked as "operation behavior pending review".

[0158] The assembly process can be further verified by combining assembly videos or monitoring records to check whether there are any non-standard behaviors in the assembly process.

[0159] The tools used in the slot wedge installation process (such as slot wedge stamping devices or guide clamps) must have their equipment number (TID) and current status parameters recorded each time they are used. The traceability method is as follows:

[0160] By matching the timestamp of the slot wedge installation with the equipment work log, the set of tool IDs T={TID1,TID2,...} used by the abnormal slot wedge was identified.

[0161] Check if the equipment has any records of maintenance, status alarms, or overdue calibration in the current or adjacent shifts;

[0162] If the equipment corresponding to TID* has experienced more than two punching force deviations or frequent process switching in the past 24 hours, it will be marked as "unstable equipment status".

[0163] The final traceability result table includes: suspected abnormal batch number BID*; high-risk operator ID PID*; abnormal status tool number TID*; abnormal slot wedge number and corresponding score list.

[0164] The traceability results will be automatically archived and submitted to the quality analysis stage, and will also serve as a basis for adjusting production instructions. If multiple stator products appear in the same batch or from the same personnel, it will trigger a strategy of stricter quality inspection or batch suspension.

[0165] By using the above methods, the abnormality of the slot wedge structure is linked to the data of people, materials and machines in the production process, realizing a complete closed loop from "local defect identification" to "upstream source investigation", which greatly improves the systematicness and foresight of motor assembly quality management.

[0166] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A machine vision-based method for detecting the quality of motor slot wedge insertion, characterized in that: The method comprises the following steps: S100, image acquisition of the motor stator along the circumferential direction is performed to obtain a continuous image sequence containing a plurality of slot wedge embedding regions; S200, image stitching and registration are performed on the image sequence to construct a quasi-three-dimensional structure model of the motor stator slot wedge region, and the depth information is used to identify the micro convex and indentation features on the surface of the slot wedge; S300, the micro gap profile between the slot wedge and the stator slot side wall is extracted, and a slot wedge embedding trend curve is calculated by combining a machine learning algorithm; S400, according to the identified micro convex and indentation features on the surface of the slot wedge and the slot wedge embedding trend curve, it is judged whether the slot wedge has a pseudo defect, and a pseudo defect filtering algorithm is used to remove or correct the misjudgment area to generate a real feature data set after optimization; S500, based on the real feature data set, the posture parameters, fastening degree index, position deviation and trend feature of each slot wedge region are mapped to the quality risk space to construct a slot wedge multi-dimensional embedding quality state model; S600, based on the parameter distribution of the slot wedge multi-dimensional embedding quality state model, a clustering algorithm is used to classify and identify all the slot wedges, the quality level score of each slot wedge is output, and the structural failure probability of the slot wedge in the future use process is predicted; S700, when there are multiple slot wedge quality level score abnormalities, the assembly batch, assembly personnel or tool state is traced back.

2. The machine vision-based motor slot wedge installation quality detection method according to claim 1, characterized in that: In the S200, the image registration is based on the extraction of the motor stator slot edge points, slot wedge top center points and stator tooth profile feature points in the image.

3. The machine vision-based motor slot wedge installation quality detection method according to claim 2, characterized in that: The constructed quasi-three-dimensional structure model of the motor stator slot wedge region is point cloud data in the format of (x, y, z), and the z-axis represents the height value of the slot wedge surface relative to the reference plane. The quasi-three-dimensional structure model is obtained by a multi-view stereo matching algorithm combined with a structured light auxiliary method.

4. The machine vision-based motor slot wedge installation quality detection method of claim 1, wherein: In the S300, the extraction of the micro gap profile includes: A plurality of vertical sections are collected along the slot wedge embedding direction, the minimum gap distance between the left and right edges of the slot wedge and the inner wall of the stator slot is calculated respectively, and a depth difference matrix is formed. The first derivative of each segment of the micro gap profile is calculated and polynomial fitting is performed, and the slope change rate, curvature change and fitting residual root mean square error are extracted as trend features.

5. The machine vision-based motor slot wedge installation quality detection method according to claim 4, characterized in that: Based on the feature parameters of the fitting curve, a K-means clustering algorithm is used to classify the embedding sections to form a slot wedge embedding trend curve, and the trend score of each point is calculated based on the curve fitting residual, slope and curvature.

6. The machine vision-based motor slot wedge installation quality detection method of claim 1, wherein: In the S400, the micro convex feature and indentation feature on the surface of the slot wedge are extracted by using a sliding window analysis and connected domain processing method, the corresponding height value, area and spatial position are calculated, and one-to-one correspondence is performed with the slot wedge embedding trend curve at the corresponding embedding position, which is used to judge the trend consistency.

7. The machine vision-based motor slot wedge installation quality detection method according to claim 6, characterized in that: By comparing the height change direction of the defect area with the derivative direction of the slot wedge embedding trend curve, when the trend is inconsistent and the defect distribution is in a discrete state in space, the defect is determined as a pseudo defect, and a Laplace boundary expansion algorithm is used for defect area repair.

8. The machine vision-based motor slot wedge installation quality detection method of claim 1, wherein: In the S500, the input vector Q of the slot wedge multi-dimensional embedding quality state model includes: slot wedge inclination angle, left and right height difference, fastening degree index, position deviation, maximum misalignment, trend slope average, trend fluctuation amplitude and trend stability. After principal component analysis dimension reduction processing, the state vector is used for training the support vector machine model to realize risk scoring.

9. The machine vision-based motor slot wedge installation quality detection method of claim 1, wherein: In the S600, the density-based spatial clustering algorithm is used for clustering classification of all slot wedge state vectors, and the slot wedge quality grade score is calculated according to the clustering center characteristics. The score interval is 0-100 points, and the quality grade is divided into four types: A, B, C and D.

10. The machine vision-based motor slot wedge installation quality detection method of claim 9, wherein: If there are two or more slot wedges in a certain motor stator with scores below 70 points, the traceability process is triggered, and the slot wedge assembly batch number, assembly personnel identity code and assembly tool equipment number are associated respectively. Significant analysis and statistical comparison are performed, and potential abnormal source information is output.

Images