A machine vision-based electric vehicle shock absorber defect detection system

Abstract

The application discloses a kind of electric vehicle shock absorber defect detection systems based on machine vision, it is related to machine vision detection technical field, including: structured light profile solving module, and three-dimensional profile is solved to multiple visual angle image and is collected;Shafting profile tensor configuration module, and the shafting profile tensor is encapsulated to structural position and is gathered;Improved wavelet neural operator configuration module, replaces wavelet coefficient convergence unit and embeds deduction structure;Shafting tensor convergence module, executes wavelet domain tensor contraction and projects reconstruction profile;Assembly deviation disassembly module, and actual profile is pasted and disassembles parity deviation;Defect result generation module, and output defect detection result are gathered to continuous deviation.This application is processed by improving wavelet neural operator network model division to handle axial displacement, circumferential eccentricity and radial deformation, reduce aliasing of different assembly deviation in same profile residual, improve electric vehicle shock absorber assembly defect positioning accuracy and determination stability.

Description

Technical Field

[0001] This invention relates to the field of machine vision inspection technology, and in particular to a machine vision-based defect detection system for electric vehicle shock absorbers. Background Technology

[0002] With the continuous increase in electric vehicle production, shock absorbers, as key components affecting the overall vehicle's driving stability and safety, are facing increasingly stringent requirements for assembly quality inspection. Electric vehicle shock absorbers typically consist of a rod, cylinder, spring, and mounting connections. During production, defects such as rod misalignment, cylinder bulges or dents, spring eccentricity, and misaligned mounting lugs may occur. Current inspection methods primarily employ manual visual inspection, 2D camera inspection, or conventional 3D point cloud fitting. Manual visual inspection is heavily influenced by experience and fatigue, making it difficult to adapt to continuous production line cycles. 2D visual inspection relies mainly on grayscale, edge, and texture changes, and is easily affected by metallic reflections, spring obstruction, and surface oil contamination, resulting in insufficient ability to identify minor assembly deviations. While conventional 3D point cloud fitting can capture contour deviations, it often mixes axial offset, circumferential eccentricity, and local radial deformation into the same residual result, making it difficult to accurately distinguish defect types and locations.

[0003] Therefore, how to provide a machine vision-based electric vehicle shock absorber defect detection system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a machine vision-based electric vehicle shock absorber defect detection system. This invention uses structured light multi-view images to solve the three-dimensional contour data of the shock absorber and constructs the shock absorber shaft system contour tensor according to the rod contour, cylinder contour, spring contour, and installation connection contour. This allows the shock absorber assembly contour to be uniformly expressed in terms of structural position, axial position, circumferential position, and radial height. Simultaneously, this invention replaces the wavelet coefficient pooling unit with a low-rank shaft system wavelet tensor pooling operator in the wavelet neural operator network model. This allows axial extension changes, circumferential offset changes, and radial height abrupt changes to be processed in the wavelet domain through directional contraction. Furthermore, the invention sequentially subtracts the axial and circumferential explained components through an orthogonal subtraction structure of the shaft system residuals, resulting in a more concentrated preservation of local bulges, depressions, warping, and other radial deformation features by the radial residual tensor contraction. Therefore, the present invention can reduce the superposition of axial offset, circumferential eccentricity and radial deformation in ordinary contour residuals, improve the positioning accuracy of rod body skew, cylinder deformation, spring eccentricity and installation connection misalignment, reduce the interference of metal reflection, spring obstruction and overall assembly offset on defect judgment, and improve the stability and sorting reliability of online defect detection of electric vehicle shock absorbers.

[0005] A machine vision-based electric vehicle shock absorber defect detection system according to an embodiment of the present invention includes:

[0006] The structured light contour calculation module acquires multi-view structured light images of the shock absorber of the electric vehicle under test, analyzes the stripes and calculates the three-dimensional contour to form the three-dimensional contour data of the shock absorber.

[0007] The shaft system profile tensor construction module gathers the three-dimensional profile data according to the structural position of the shock absorber, arranges the gathered profile along the axial direction, and forms the shock absorber shaft system profile tensor.

[0008] An improved wavelet neural operator configuration module was constructed, and an improved wavelet neural operator network model was built. The contour lifting network, wavelet kernel integration layer and contour projection network were retained. The wavelet coefficient pooling unit in the wavelet kernel integration layer was replaced with a low-rank axis wavelet tensor pooling operator, and an axis residual orthogonal subtraction structure was embedded in the low-rank axis wavelet tensor pooling operator.

[0009] The shaft system tensor merging module connects the shock absorber shaft system profile tensor to the wavelet kernel integration layer via a profile lifting network. The low-rank shaft system wavelet tensor merging operator performs shaft system tensor contraction in the wavelet domain and deducts the axial and circumferential explained components. The profile projection network then forms the shock absorber assembly profile reconstruction field.

[0010] The assembly deviation disassembly module re-attaches the three-dimensional contour data of the shock absorber to the shock absorber assembly contour reconstruction field, disassembles the co-position deviation between the actual contour and the reconstructed contour, and forms the shock absorber assembly deviation field.

[0011] The defect result generation module delineates the defect area of ​​the shock absorber assembly according to the continuous distribution position of the deviation in the shock absorber assembly deviation field, and generates the defect detection results of the electric vehicle shock absorber.

[0012] Optionally, the process of forming the three-dimensional contour data of the shock absorber includes:

[0013] Synchronous acquisition of structured light projection images of the shock absorber of the electric vehicle under test at different shooting angles is performed, and multi-view images corresponding to the same detection position are grouped into the same contour calculation group.

[0014] Centerline positioning is performed on the structured light stripes in the contour solution group to remove isolated bright spots and discontinuous noise formed by metal reflection, and neighborhood patching is performed on short stripe gaps caused by the damper surface to form a set of structured light stripe centerlines.

[0015] Based on the camera calibration relationship and the structured light projection calibration relationship, the pixel positions in the set of structured light stripe center lines are converted into spatial contour points, and pose alignment is performed on the multi-view spatial contour points under the same detection position to form the three-dimensional contour data of the shock absorber.

[0016] Optionally, the formation of the damper shaft profile tensor includes:

[0017] The three-dimensional contour data of the shock absorber is aggregated along the assembly structure to the structural positions corresponding to the rod contour, cylinder contour, spring contour and mounting connection contour.

[0018] Using the extension direction from the mounting end to the movable end of the shock absorber as the axial alignment direction, axial section cutting is performed on the aggregated three-dimensional contour data;

[0019] The contour points distributed around the axis of the shock absorber within the same axial section are arranged in circumferential order. For contour points with closed circumferential structures, a closed arrangement is performed, and the radial height of the contour points relative to the axial arrangement direction is attached to the corresponding circumferential position.

[0020] The circumferential arrangement results corresponding to adjacent axial sections are stacked sequentially along the axial direction, and the structural position, axial position, circumferential position and radial height are encapsulated as the damper shaft profile tensor.

[0021] Optionally, the construction of the improved wavelet neural operator network model includes:

[0022] The contour lifting network, wavelet kernel integration layer and contour projection network in the wavelet neural operator network are retained. The output of the contour lifting network is connected to the input of the wavelet kernel integration layer. The wavelet domain processing path of the wavelet kernel integration layer is connected to the contour projection network through the low-rank axis wavelet tensor pooling operator and the axis residual orthogonal subtraction structure.

[0023] The wavelet coefficient pooling unit in the wavelet kernel integration layer is replaced with a low-rank axis wavelet tensor pooling operator, and the low-rank axis wavelet tensor pooling operator is embedded in the wavelet domain processing path of the wavelet kernel integration layer.

[0024] The low-rank axial wavelet tensor pooling operator is configured with axial low-rank wavelet kernels, circumferential low-rank wavelet kernels and radial low-rank wavelet kernels. The axial low-rank wavelet kernel is attached to the axial wavelet coefficient position, the circumferential low-rank wavelet kernel is attached to the circumferential wavelet coefficient position, and the radial low-rank wavelet kernel is attached to the radial wavelet coefficient position.

[0025] An orthogonal subtraction structure for axis residuals is embedded within the tensor contraction path of the low-rank axis wavelet tensor converging operator. The input of the orthogonal subtraction structure for axis residuals is connected to the tensor contraction output of the low-rank axis wavelet tensor converging operator, and the output of the orthogonal subtraction structure for axis residuals is connected to the contour projection network.

[0026] Optionally, the step of inputting the shock absorber shaft system profile tensor into a wavelet kernel integration layer via a profile lifting network, and then performing shaft system tensor contraction in the wavelet domain by a low-rank shaft system wavelet tensor merging operator, includes:

[0027] The contour lifting network slides adjacent section windows along the axial arrangement direction of the contour tensor of the shock absorber shaft system. Differential convolution and normalization mapping are performed on the contour center displacement, circumferential contour difference and radial height change in the adjacent section windows, respectively. The mapped local contour response is then pasted back to the corresponding axial position to form the lifting contour feature.

[0028] The wavelet kernel integration layer configures axial wavelet kernels, circumferential wavelet kernels, and radial wavelet kernels on the lifted contour features. The axial wavelet kernel slides along the axial position and performs low-frequency preservation and high-frequency difference on the lifted contour features at adjacent axial positions to form axial approximation coefficients and axial detail coefficients.

[0029] The circumferential wavelet kernel slides along the circumferential arrangement direction of the same axial section, performs circumferential low-frequency preservation and circumferential high-frequency difference on the lifting profile features of adjacent circumferential positions, forming circumferential approximation coefficients and circumferential detail coefficients. The radial wavelet kernel slides along the radial height position, performs local low-frequency preservation and local high-frequency difference on the radial height change, forming radial approximation coefficients and radial detail coefficients.

[0030] The low-rank axial wavelet tensor pooling operator arranges the axial approximation coefficients and axial detail coefficients in the low-rank position to form the axial wavelet tensor, and arranges the circumferential approximation coefficients and circumferential detail coefficients in the circumferential position and low-rank position to form the circumferential wavelet tensor, and arranges the radial approximation coefficients and radial detail coefficients in the radial height position and low-rank position to form the radial wavelet tensor.

[0031] The low-rank axial wavelet tensor pooling operator calls trainable axial low-rank wavelet kernels, circumferential low-rank wavelet kernels, and radial low-rank wavelet kernels. It multiplies the axial low-rank wavelet kernel with the axial wavelet tensor bit by bit and accumulates them in the same position to form the axial tensor contraction result. It multiplies the circumferential low-rank wavelet kernel with the circumferential wavelet tensor bit by bit and accumulates them in the same position to form the circumferential tensor contraction result.

[0032] The radial low-rank wavelet kernel is multiplied rank-by-rank with the radial wavelet tensor and accumulated in the same position to form the radial tensor shrinkage result. The axial tensor shrinkage result, the circumferential tensor shrinkage result, and the radial tensor shrinkage result are then connected to the shaft system residual orthogonal subtraction structure.

[0033] Optionally, the step of subtracting the axial and circumferential explained components and forming the shock absorber assembly contour reconstruction field via a contour projection network includes:

[0034] The shaft system residual orthogonal subtraction structure receives the axial tensor contraction results, circumferential tensor contraction results, and radial tensor contraction results, and expands the axial tensor contraction results into axial explained components along the axial position;

[0035] In the circumferential tensor contraction results and the radial tensor contraction results, the co-position wavelet components that have the same axial position as the axially interpreted components are located, and the corresponding co-position wavelet components are subtracted one by one to form the circumferential tensor contraction results after axial subtraction and the radial tensor contraction results after axial subtraction.

[0036] The shaft system residual orthogonal subtraction structure expands the circumferential tensor contraction result after axial subtraction into circumferential explained components along the circumferential position.

[0037] In the radial tensor shrinkage result after axial subtraction, the co-position wavelet component with the same circumferential position as the circumferential interpreted component is located, and the corresponding co-position wavelet component is subtracted bit by bit to form the radial residual tensor shrinkage result.

[0038] The results of shrinking the axial explained component, circumferential explained component, and radial residual tensor are superimposed in the same position according to the axial position, circumferential position, and radial height position to form the wavelet domain contour reconstruction result;

[0039] The contour projection network projects the wavelet domain contour reconstruction results back to the shock absorber contour space, forming the shock absorber assembly contour reconstruction field.

[0040] Optionally, the formation of the shock absorber assembly deviation field includes:

[0041] The assembly deviation disassembly module locates the corresponding actual contour points in the three-dimensional contour data of the shock absorber according to the axial and circumferential positions in the shock absorber assembly contour reconstruction field, and then pastes the actual radial height of the corresponding actual contour points back to the corresponding reconstruction positions.

[0042] The assembly deviation decomposition module performs a difference decomposition between the actual radial height and the reconstructed radial height at the same axial position and the same circumferential position, forming a radial height deviation.

[0043] The assembly deviation disassembly module continuously arranges radial height deviations along the axial position to form axial deviation segments, and organizes radial height deviations along the circumferential arrangement direction of the same axial section to form circumferential deviation segments.

[0044] The assembly deviation disassembly module re-attaches radial height deviation, axial deviation segments, and circumferential deviation segments to the three-dimensional contour data of the shock absorber according to the structural position of the shock absorber, forming the shock absorber assembly deviation field.

[0045] Optionally, the defect result generation module specifically includes:

[0046] The defect result generation module performs contiguous aggregation on the continuous deviation positions in the shock absorber assembly deviation field, and classifies the deviation areas that extend continuously along the axial position into the axial offset defect candidate area, the deviation areas that are circumferentially offset along the same axial section into the circumferential eccentricity defect candidate area, and the deviation areas with concentrated radial height abrupt changes into the radial deformation defect candidate area.

[0047] The defect result generation module back-attaches the axial offset defect candidate area, circumferential eccentricity defect candidate area and radial deformation defect candidate area to the shock absorber structure location, and delineates the rod body skew area, cylinder deformation area, spring eccentricity area and installation connection misalignment area.

[0048] The defect result generation module configures defect type and defect level markers according to the location of the delineated defect area, the continuous length of the deviation, and the radial height deviation distribution, and encapsulates the defect area location, defect type markers, and defect level markers into the defect detection results of the electric vehicle shock absorber.

[0049] The beneficial effects of this invention are:

[0050] This invention uses structured light multi-view images to calculate the three-dimensional contour data of the shock absorber, and further constructs the shock absorber shaft system contour tensor according to the rod contour, cylinder contour, spring contour, and mounting connection contour. This transforms the originally dispersed three-dimensional point cloud of the shock absorber into a regular tensor with structural position, axial position, circumferential position, and radial height. Therefore, the detection process no longer relies solely on two-dimensional grayscale edges or ordinary point cloud fitting residuals, but can preserve the spatial continuity of the shock absorber assembly contour under a unified shaft system structure, improving the basis for identifying assembly defects such as rod misalignment, spring eccentricity, and mounting connection misalignment.

[0051] This invention introduces a low-rank axial wavelet tensor merging operator into the wavelet neural operator network model. It performs tensor contraction on the wavelet coefficients in the wavelet kernel integration layer according to the axial, circumferential, and radial directions, respectively, allowing axial extension changes, circumferential offset changes, and radial height abrupt changes to be expressed separately in the wavelet domain. Compared to ordinary wavelet coefficient merging methods, this invention reduces the mixing of contour changes in different directions in the same response, resulting in a clearer internal model representation of the overall bending, circumferential eccentricity, and local bulges and depressions of the shock absorber.

[0052] This invention embeds an orthogonal subtraction structure for axis system residuals within a low-rank axis wavelet tensor consolidation operator. It first subtracts the explained axial components, then the explained circumferential components, resulting in a more concentrated preservation of local radial deformations by reducing the radial residual tensor contraction. The resulting shock absorber assembly deviation field can more accurately distinguish between rod misalignment, cylinder deformation, spring eccentricity, and installation connection misalignment. It reduces interference from metal reflections, spring obstruction, and overall assembly offset on the identification of local defects, improving the defect location accuracy and sorting reliability of electric vehicle shock absorbers in online inspection. Attached Figure Description

[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0054] Figure 1 This is a schematic diagram of the overall structure of a machine vision-based electric vehicle shock absorber defect detection system proposed in this invention.

[0055] Figure 2 This is a schematic diagram of the improved wavelet neural operator network model of a machine vision-based electric vehicle shock absorber defect detection system proposed in this invention.

[0056] Figure 3 This is a flowchart illustrating the assembly deviation analysis and defect result generation of a machine vision-based electric vehicle shock absorber defect detection system proposed in this invention. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0058] refer to Figure 1 , Figure 2 and Figure 3 A machine vision-based defect detection system for electric vehicle shock absorbers includes:

[0059] The structured light contour calculation module acquires multi-view structured light images of the shock absorber of the electric vehicle under test, analyzes the stripes and calculates the three-dimensional contour to form the three-dimensional contour data of the shock absorber.

[0060] The shaft system profile tensor construction module gathers the three-dimensional profile data according to the structural position of the shock absorber, arranges the gathered profile along the axial direction, and forms the shock absorber shaft system profile tensor.

[0061] An improved wavelet neural operator configuration module was constructed, and an improved wavelet neural operator network model was built. The contour lifting network, wavelet kernel integration layer and contour projection network were retained. The wavelet coefficient pooling unit in the wavelet kernel integration layer was replaced with a low-rank axis wavelet tensor pooling operator, and an axis residual orthogonal subtraction structure was embedded in the low-rank axis wavelet tensor pooling operator.

[0062] The shaft system tensor merging module connects the shock absorber shaft system profile tensor to the wavelet kernel integration layer via a profile lifting network. The low-rank shaft system wavelet tensor merging operator performs shaft system tensor contraction in the wavelet domain and deducts the axial and circumferential explained components. The profile projection network then forms the shock absorber assembly profile reconstruction field.

[0063] The assembly deviation disassembly module re-attaches the three-dimensional contour data of the shock absorber to the shock absorber assembly contour reconstruction field, disassembles the co-position deviation between the actual contour and the reconstructed contour, and forms the shock absorber assembly deviation field.

[0064] The defect result generation module delineates the defect area of ​​the shock absorber assembly according to the continuous distribution position of the deviation in the shock absorber assembly deviation field, and generates the defect detection results of the electric vehicle shock absorber.

[0065] In this embodiment, the three-dimensional contour data of the shock absorber is formed, including:

[0066] The structured light projection images of the shock absorber of the electric vehicle under test at different shooting angles are simultaneously acquired, and the multi-view images corresponding to the same detection position are grouped into the same contour calculation group. Specifically, the simultaneous acquisition of structured light projection images of the shock absorber of the electric vehicle under test at different shooting angles is performed as follows:

[0067] The shock absorber of the electric vehicle under test is placed on a rotating inspection station or a fixed multi-camera inspection station, ensuring that the axial extension direction of the shock absorber is aligned with the reference axis of the inspection station. A structured light projector projects line or stripe structured light onto the surface of the shock absorber, and a multi-view camera acquires structured light projection images from different circumferential angles. The camera acquisition trigger signal and the structured light projection trigger signal use the same synchronization pulse, ensuring that images from different shooting angles correspond to the same inspection posture. After acquisition, the images are grouped according to the inspection position number, shooting angle number, and acquisition trigger number, with multi-view images corresponding to the same inspection position grouped into the same contour calculation group.

[0068] Centerline positioning is performed on the structured light stripes in the contour solution group to remove isolated bright spots and discontinuous noise caused by metallic reflection. Neighborhood patching is also performed on short stripe gaps caused by the damper surface to form a set of structured light stripe centerlines. Specifically, centerline positioning of the structured light stripes in the contour solution group is performed as follows:

[0069] Gray-level normalization and background brightness subtraction are performed on the structured light projected image. The gray-level distribution is scanned along the normal direction of the structured light stripes. The local gray-level peak positions and the pixels in the neighboring pixels of the peaks are delineated as candidate stripe regions. Gray-level centroid localization or sub-pixel peak fitting is performed on the candidate stripe regions to determine the stripe center point. The adjacent stripe center points are connected in series along the stripe extension direction to form a single-view stripe center line.

[0070] Specifically, the removal of isolated bright spots and intermittent noise caused by metallic reflection includes:

[0071] Read the area of ​​connected bright spots, the location of bright spot boundaries, and the continuity relationship of adjacent fringe center points in the candidate fringe region. If the area of ​​the bright spot region is less than the threshold for the area of ​​the continuous fringe region, and the bright spot boundary is not continuous with the center point of the adjacent fringe, the bright spot region is marked as an isolated bright spot and deleted. If the length of the noise segment is less than one-tenth of the length of the adjacent effective fringe, and there are no continuous fringe directions at both ends of the noise segment, the noise segment is removed from the fringe center line.

[0072] Specifically, for short stripe gaps caused by the obstruction of the damper's curved surface, neighbor-to-neighbor patching is performed:

[0073] Locate the break point in the center line of the stripes from a single viewpoint, and read the tangential direction and break length of the stripes on both sides of the break point. If the break length is less than 5 times the average width of the stripes at the same viewpoint, and the angle between the tangential directions of the stripes on both sides is less than 20°, the break point is identified as a short stripe gap. Perform neighborhood extension along the tangential direction of the center points of the stripes on both sides of the break point, and fill in the missing center point at the midpoint between the two extension trajectories to form a set of structured light stripe center lines.

[0074] Based on the camera calibration relationship and the structured light projection calibration relationship, the pixel positions in the set of structured light stripe centerlines are converted into spatial contour points, and pose alignment is performed on the multi-view spatial contour points at the same detection position to form the three-dimensional contour data of the shock absorber. Specifically, the conversion of pixel positions in the set of structured light stripe centerlines into spatial contour points is as follows:

[0075] An imaging ray corresponding to the centerline pixel position is generated based on the camera's intrinsic parameters, and the corresponding structured light projection plane is determined according to the structured light projection calibration relationship. The intersection point of the imaging ray and the structured light projection plane is calculated, and the coordinates of the intersection point are used as spatial contour points. When multiple structured light fringes are projected simultaneously, the corresponding structured light projection plane is matched according to the fringe number, so that the centerlines of different fringes are converted into spatial contour points respectively.

[0076] Specifically, pose alignment is performed on multi-view spatial contour points at the same detection location, as follows:

[0077] Spatial contour points from different viewpoints are transformed to the same detection station coordinate system according to camera extrinsic parameters, and local registration is performed using overlapping contour regions between adjacent viewpoints. Point pairs with the same name within the overlapping region are retained if the distance between them is less than the contour alignment error threshold; otherwise, they are marked as mismatched and discarded. After pose alignment is completed, the spatial contour points from multiple viewpoints are merged into the same shock absorber coordinate system to form the three-dimensional contour data of the shock absorber.

[0078] In this embodiment, forming the damper shaft system profile tensor includes:

[0079] The three-dimensional contour data of the shock absorber is aggregated along the assembly structure to the corresponding structural positions of the rod contour, cylinder contour, spring contour, and mounting connection contour. Specifically, the aggregation of the three-dimensional contour data of the shock absorber to the corresponding structural positions along the assembly structure is as follows:

[0080] The spatial contour points in the 3D contour data of the shock absorber are read, and their structural positions are assigned according to their positional relationships in the axial direction, outer diameter range, and local curvature changes of the shock absorber. Spatial contour points located in the slender cylindrical region at the center of the shock absorber with stable radial height changes are assigned to the rod contour; spatial contour points located on the surface of the outer cylindrical shell and continuously distributed along the axial direction are assigned to the cylinder contour; spatial contour points that undulate periodically along the circumference and are distributed around the outer side of the cylinder are assigned to the spring contour; spatial contour points located in mounting holes, connecting lugs, end connecting seats, or welded connection areas are assigned to the mounting connection contour. For spatial contour points located at the boundary of adjacent structures, their structural position assignment is determined according to their distance from the centerline of the adjacent structure contour and the continuity of local curvature.

[0081] Using the extension direction from the mounting end to the movable end of the shock absorber as the axial alignment direction, axial section trimming is performed on the converged three-dimensional contour data. Specifically, the axial section trimming of the converged three-dimensional contour data is as follows:

[0082] Using the center of the end face or the center of the mounting hole of the shock absorber as the axial starting reference and the direction of the movable end as the positive axial direction, an axial reference line for the shock absorber is established. The aggregated three-dimensional contour data is projected onto this axial reference line to obtain the axial position corresponding to each spatial contour point. The three-dimensional contour data of the shock absorber is segmented according to the axial sampling interval. The axial sampling interval can be determined according to the structured light measurement resolution and the length of the shock absorber, for example, 0.2 mm to 1 mm; spatial contour points within the same axial sampling interval are grouped into the same axial section.

[0083] The contour points distributed around the axis of the shock absorber within the same axial section are arranged in circumferential order. Contour points with a closed loop structure are arranged in a closed loop, and the radial height of the contour points relative to the axial arrangement direction is connected to the corresponding circumferential position. Specifically, arranging the contour points distributed around the axis of the shock absorber within the same axial section in circumferential order involves:

[0084] Within each axial section, the intersection of the damper's axial baseline and the current axial section is used as the section center. The circumferential angle and radial distance of each profile point within the section relative to the section center are calculated. The profile points within the section are arranged in ascending order of circumferential angle, and the radial distance is used as the radial height of that profile point. For profile points in the rod, cylinder, and spring profiles that can form a continuous circumferential distribution around the damper axis, a closed arrangement is executed based on the circumferential angle. For profile points in the installation connection profile that do not possess a complete circumferential structure, only the circumferential sequence and structural position markings are retained; a closed arrangement is not mandatory.

[0085] The circumferential arrangement results corresponding to adjacent axial sections are stacked sequentially along the axial direction, and the structural position, axial position, circumferential position, and radial height are encapsulated as a damper shaft profile tensor. Specifically, the circumferential arrangement results corresponding to adjacent axial sections are stacked sequentially along the axial direction.

[0086] The circumferential arrangement results of each axial section are read sequentially from the installation end to the moving end according to the axial position. Contour points with the same or closest circumferential positions in adjacent axial sections are arranged into the same circumferential column, ensuring that the circumferential contour changes between different axial sections maintain a corresponding continuity. For missing circumferential positions caused by shading, reflection removal, or discontinuity in the installation connection structure, empty or low-confidence markers are retained, and contour points from adjacent structural positions are not directly used for replacement. After stacking, the structural position is encapsulated as the structural dimension, the axial position as the first spatial dimension, the circumferential position as the second spatial dimension, and the radial height as the contour numerical dimension, forming the shock absorber shaft system contour tensor. This shock absorber shaft system contour tensor preserves the contour relationships of the shock absorber in axial extension, circumferential arrangement, and radial undulation, serving as the input object for the improved wavelet neural operator network model.

[0087] In this embodiment, an improved wavelet neural operator network model is constructed, including:

[0088] The contour lifting network, wavelet kernel integration layer, and contour projection network in the wavelet neural operator network are retained. The output of the contour lifting network is connected to the input of the wavelet kernel integration layer. The wavelet domain processing path of the wavelet kernel integration layer is connected to the contour projection network via a low-rank axis wavelet tensor merging operator and an axis residual orthogonal subtraction structure. Specifically, the contour lifting network, wavelet kernel integration layer, and contour projection network in the wavelet neural operator network are retained as follows:

[0089] The contour lifting network, located at the input of the improved wavelet neural operator network model, receives the shock absorber axis contour tensor and converts the structural position, axial position, circumferential position, and radial height in the shock absorber axis contour tensor into lifted contour features with a fixed number of channels. The contour lifting network can consist of one-dimensional convolutional, two-dimensional convolutional, or linear mapping layers, with the number of channels set to 32, 64, or 128 to convert the original contour height data into a feature representation suitable for wavelet domain processing. Following the contour lifting network, the wavelet kernel integration layer receives the lifted contour features and performs wavelet domain expansion in the axial, circumferential, and radial directions. The contour projection network, located at the model output, receives the reconstructed result after wavelet domain processing and projects the reconstructed result back into the shock absorber contour space, allowing the output result to be compared with the three-dimensional contour data of the shock absorber at the same axial and circumferential positions.

[0090] The wavelet coefficient pooling unit within the wavelet kernel integration layer is replaced with a low-rank axial wavelet tensor pooling operator. This low-rank axial wavelet tensor pooling operator is embedded into the wavelet domain processing path of the wavelet kernel integration layer. Specifically, the wavelet coefficient pooling unit and the low-rank axial wavelet tensor pooling operator are:

[0091] The wavelet coefficient pooling unit is the original pooling location within the wavelet kernel integration layer where different wavelet coefficients are aggregated. Ordinary wavelet coefficient pooling units typically pool coefficients within the wavelet domain using a unified channel, easily mixing overall axial offset, circumferential eccentricity, and radial local deformation of the shock absorber into the same contour response. After replacing the wavelet coefficient pooling unit, the low-rank axis wavelet tensor pooling operator no longer directly mixes wavelet coefficients into a single path result. Instead, it assigns wavelet coefficients to axial, circumferential, and radial wavelet tensors respectively, based on the shock absorber's axial, circumferential, and radial height positions. The "low-rank" in the low-rank axis wavelet tensor pooling operator indicates that each type of wavelet tensor is not directly pooled using a complete high-dimensional parameter matrix, but rather undergoes parameter compression and rank-by-rank operations using a smaller number of low-rank bits. Low-rank bits can be understood as the internal order position during tensor pooling; the number of low-rank bits can be set to 8, 16, or 32 depending on the model size and is used as a network hyperparameter in the training configuration.

[0092] The low-rank axial wavelet tensor pooling operator is configured with axial low-rank wavelet kernels, circumferential low-rank wavelet kernels, and radial low-rank wavelet kernels. The axial low-rank wavelet kernel is attached to the axial wavelet coefficient position, the circumferential low-rank wavelet kernel is attached to the circumferential wavelet coefficient position, and the radial low-rank wavelet kernel is attached to the radial wavelet coefficient position. Specifically, the axial low-rank wavelet kernel, circumferential low-rank wavelet kernel, and radial low-rank wavelet kernel are as follows:

[0093] The axial low-rank wavelet kernel is a trainable kernel parameter corresponding to the axial wavelet coefficients in the low-rank axial wavelet tensor confluence operator. The positions of the axial wavelet coefficients are arranged in axial order from the damper mounting end to the moving end. After being connected along the axial wavelet coefficient positions, the axial low-rank wavelet kernel mainly bears the overall offset and extension changes between adjacent axial sections. The circumferential low-rank wavelet kernel is a trainable kernel parameter corresponding to the circumferential wavelet coefficients. The positions of the circumferential wavelet coefficients are arranged in circumferential order within the same axial section. After being connected along the circumferential wavelet coefficient positions, the circumferential low-rank wavelet kernel mainly bears the circumferential eccentricity and circumferential distribution changes within the same axial section. The radial low-rank wavelet kernel is a trainable kernel parameter corresponding to the radial wavelet coefficients. The positions of the radial wavelet coefficients are arranged according to the radial height changes. After being connected along the radial wavelet coefficient positions, the radial low-rank wavelet kernel mainly bears local bulges, depressions, and abrupt height changes. The three low-rank wavelet kernels are all trainable parameters within the low-rank axis wavelet tensor pooling operator, and are updated synchronously with the contour reconstruction error and defect localization error during training.

[0094] An orthogonal subtraction structure for axisymmetric residuals is embedded within the tensor contraction path of a low-rank axisymmetric wavelet tensor merging operator. The input of this orthogonal subtraction structure is connected to the tensor contraction output of the low-rank axisymmetric wavelet tensor merging operator, and the output is connected to a contour projection network. Specifically, the orthogonal subtraction structure for axisymmetric residuals is as follows:

[0095] The tensor contraction path is the internal computational path for the axial, circumferential, and radial tensor contraction results output by the low-rank axial wavelet tensor merging operator. After the axial residual orthogonal subtraction structure is embedded in this tensor contraction path, it does not change the original order of the contour lifting network, wavelet kernel integration layer, and contour projection network. Instead, it performs directional subtraction on the three types of tensor contraction results after the low-rank axial wavelet tensor merging operator outputs them. The input of the axial residual orthogonal subtraction structure receives the axial, circumferential, and radial tensor contraction results. Internally, the axial tensor contraction result is first treated as the axial interpreted component, ensuring that wavelet components at the same axial position in the circumferential and radial tensor contraction results no longer repeatedly carry axial offsets. Then, the circumferential tensor contraction result after axial subtraction is treated as the circumferential interpreted component, ensuring that wavelet components at the same circumferential position in the radial tensor contraction result no longer repeatedly carry circumferential eccentricity. The result after subtraction is fed into the contour projection network, which then converts the wavelet domain result into a shock absorber assembly contour reconstruction field.

[0096] In the above structure, the contour lifting network, wavelet kernel integration layer, and contour projection network are retained structures of the wavelet neural operator network; the low-rank axial wavelet tensor confluence operator, axial low-rank wavelet kernel, circumferential low-rank wavelet kernel, radial low-rank wavelet kernel, and axial residual orthogonal subtraction structure are newly configured internal structures of the model in this embodiment. This structural relationship enables the improved wavelet neural operator network model to handle axial extension, circumferential arrangement, and radial height changes separately in the wavelet domain, avoiding the mixing of overall shock absorber offset, circumferential eccentricity, and local radial deformation in the same wavelet response.

[0097] In this embodiment, the shock absorber shaft system profile tensor is fed into a wavelet kernel integration layer via a profile lifting network. A low-rank shaft system wavelet tensor pooling operator performs shaft system tensor contraction in the wavelet domain, including:

[0098] The contour lifting network slides adjacent section windows along the axial alignment direction of the shock absorber shaft system contour tensor. Differential convolution and normalization mapping are performed on the contour center displacement, circumferential contour difference, and radial height change within the adjacent section windows, respectively. The mapped local contour responses are then pasted back to the corresponding axial positions to form the lifted contour features. Specifically, the contour lifting network slides adjacent section windows along the axial alignment direction of the shock absorber shaft system contour tensor as follows:

[0099] Sliding windows are set sequentially according to the axial positions in the shock absorber shaft profile tensor, with each sliding window covering 3 to 7 consecutive axial sections. When the current sliding window moves to the i-th axial section, the circumferential arrangement points and radial height values ​​of the current axial section and the adjacent axial sections are read. For each axial section within the same sliding window, the profile center position of each axial section is first calculated by averaging the spatial coordinates of the corresponding profile points at each circumferential position within the same section; then, the profile center positions of adjacent axial sections are subtracted coordinate by coordinate to obtain the profile center displacement. This profile center displacement is used to represent the overall offset change between adjacent axial sections, such as center drift caused by rod skewing, cylinder axial bending, or offset of the installation connection area.

[0100] Specifically, differential convolution is performed on the circumferential contour differences within adjacent cross-sectional windows, as follows:

[0101] Within the same axial section, the radial height values ​​of adjacent circumferential positions are read in circumferential order. The radial height value of the current circumferential position is subtracted from the radial height values ​​of the previous and next circumferential positions to obtain the forward and backward circumferential differences. For rod, cylinder, and spring profiles with closed loop structures, the first and last circumferential positions participate in differential convolution. For areas in the installation connection profile that do not have closed loop structures, differential convolution is only performed within the circumferential arrangement interval where adjacent profile points actually exist. The differential convolution kernel can be a convolution kernel of length 3. The convolution kernel reads the radial height values ​​of the previous, current, and next positions sequentially in the circumferential arrangement direction and outputs the circumferential profile difference response at the current position.

[0102] Specifically, differential convolution and normalization mapping are performed on the radial height changes within adjacent cross-sectional windows, as follows:

[0103] The radial height values ​​of adjacent axial sections at the same axial and circumferential positions are read. The radial height value of the current axial section is subtracted from the radial height values ​​of the previous and next axial sections to obtain the axial adjacent height difference. Then, the radial height difference between the current circumferential position and the adjacent circumferential position within the same axial section is read to obtain the circumferential adjacent height difference. The profile lifting network combines the profile center displacement, circumferential profile difference response, and radial height change response into a local profile response, and performs maximum and minimum normalization processing on the local profile response to ensure that the response of each channel falls within the numerical range of 0 to 1. The normalized local profile response is then backfitted to the damper shaft system profile tensor according to the corresponding axial position, circumferential position, and structural position to form the lifted profile feature.

[0104] The wavelet kernel integration layer configures axial, circumferential, and radial wavelet kernels on the lifted contour features. The axial wavelet kernel slides along the axial position, performing low-frequency preservation and high-frequency difference on the lifted contour features at adjacent axial positions, forming axial approximation coefficients and axial detail coefficients. The circumferential wavelet kernel slides along the circumferential arrangement direction of the same axial section, performing circumferential low-frequency preservation and circumferential high-frequency difference on the lifted contour features at adjacent circumferential positions, forming circumferential approximation coefficients and circumferential detail coefficients. The radial wavelet kernel slides along the radial height position, performing local low-frequency preservation and local high-frequency difference on the radial height variation, forming radial approximation coefficients and radial detail coefficients. Specifically, the wavelet kernel integration layer configures axial, circumferential, and radial wavelet kernels on the lifted contour features as follows:

[0105] The wavelet kernel integration layer is configured with a low-frequency kernel and a high-frequency kernel for axial position, circumferential position, and radial height position, respectively. The low-frequency kernel uses a weighted average of adjacent positions to preserve gradual changes, while the high-frequency kernel uses a difference method of adjacent positions to preserve abrupt changes. When the axial wavelet kernel slides along the axial position, the lifting profile features of continuous axial positions are read at the same circumferential and structural positions. The weighted average of the lifting profile features of adjacent axial positions is used to obtain the axial approximation coefficient, and the difference between the lifting profile features of the current axial position and the adjacent axial positions is arranged as the axial detail coefficient. The axial approximation coefficient is used to reflect the overall extension trend of the rod or cylinder along the axial direction, while the axial detail coefficient is used to reflect local bending or abrupt changes in the axial direction.

[0106] When the circumferential wavelet kernel slides along the circumferential arrangement direction of the same axial section, it reads the lifting profile features of consecutive circumferential positions within the same axial location. The weighted average of the lifting profile features at adjacent circumferential positions yields the circumferential approximation coefficient, and the difference between the lifting profile features at the current circumferential position and the adjacent circumferential positions is arranged as the circumferential detail coefficient. For profiles with closed loop structures, the circumferential wavelet kernel continues to slide between the first and last circumferential positions; for installation connection profiles, the circumferential wavelet kernel only slides within the actual circumferential arrangement interval. When the radial wavelet kernel slides along the radial height position, it reads the radial height change response at the same axial and circumferential positions. The weighted average of the responses at adjacent height positions yields the radial approximation coefficient, and the difference between the responses at the current height position and the adjacent height positions is arranged as the radial detail coefficient. The radial detail coefficient is used to preserve height abrupt changes corresponding to cylinder bulges, local depressions, weld protrusions, or local warping at the mounting end.

[0107] The low-rank axisymmetric wavelet tensor consolidation operator arranges the axial approximation coefficients and axial detail coefficients according to their low-rank positions to form an axial wavelet tensor, arranges the circumferential approximation coefficients and circumferential detail coefficients according to their circumferential positions and low-rank positions to form a circumferential wavelet tensor, and arranges the radial approximation coefficients and radial detail coefficients according to their radial height positions and low-rank positions to form a radial wavelet tensor. Specifically, the low-rank axisymmetric wavelet tensor consolidation operator arranges the wavelet tensors as follows:

[0108] The low-rank axial wavelet tensor pooling operator first determines the number of low-rank bits, which can be set to 8, 16, or 32. For axial approximation coefficients and axial detail coefficients, the coefficients of each axial section are read in axial position order. The approximation coefficients and detail coefficients at the same axial position are written into the corresponding low-rank bits and arranged in axial position order to form an axial wavelet tensor. For circumferential approximation coefficients and circumferential detail coefficients, the coefficients of each circumferential position within the same axial section are read in circumferential arrangement. The approximation coefficients and detail coefficients at the same circumferential arrangement position are written into the corresponding low-rank bits and arranged in circumferential position and low-rank bits to form a circumferential wavelet tensor. For radial approximation coefficients and radial detail coefficients, the height variation coefficients are read according to the radial height position. The approximation coefficients and detail coefficients at the same radial height position are written into the corresponding low-rank bits and arranged in radial height position and low-rank bits to form a radial wavelet tensor. When the number of coefficients exceeds the number of low-rank positions, they are compressed to the corresponding low-rank positions by weighting adjacent coefficients; when the number of coefficients is less than the number of low-rank positions, the missing low-rank positions are filled with zero values ​​or low-confidence filler values.

[0109] The low-rank axis wavelet tensor aggregation operator calls trainable axial, circumferential, and radial low-rank wavelet kernels. It multiplies the axial low-rank wavelet kernel with the axial wavelet tensor bit-by-bit and accumulates the results in the same position to form the axial tensor contraction result. Similarly, it multiplies the circumferential low-rank wavelet kernel with the circumferential wavelet tensor bit-by-bit and accumulates the results in the same position to form the circumferential tensor contraction result. The radial low-rank wavelet kernel is then multiplied with the radial wavelet tensor bit-by-bit and accumulated in the same position to form the radial tensor contraction result. Finally, the axial, circumferential, and radial tensor contraction results are integrated into the axis residual orthogonal subtraction structure. Specifically, the low-rank axis wavelet tensor aggregation operator calls trainable axial, circumferential, and radial low-rank wavelet kernels as follows:

[0110] The axial low-rank wavelet kernels, circumferential low-rank wavelet kernels, and radial low-rank wavelet kernels are all arranged according to their low-rank positions, with each low-rank position having a trainable kernel weight. When multiplying the axial low-rank wavelet kernel with the axial wavelet tensor, the axial wavelet tensor values ​​at each low-rank position are read one by one at the same axial location, and multiplied with the kernel weights at the same low-rank position in the axial low-rank wavelet kernel. The products of each low-rank position are accumulated within the same axial location to obtain the axial tensor contraction value at that axial location. The axial tensor contraction values ​​at all axial locations are arranged in axial order to form the axial tensor contraction result.

[0111] When multiplying a circumferential low-rank wavelet kernel with a circumferential wavelet tensor, the circumferential wavelet tensor values ​​at each low-rank position within the same circumferential arrangement are read sequentially and multiplied by the kernel weight at the same low-rank position in the circumferential low-rank wavelet kernel. The products of each low-rank position are accumulated within the same circumferential arrangement position to obtain the circumferential tensor contraction value at that position. The circumferential tensor contraction values ​​at all circumferential arrangement positions are arranged in circumferential order to form the circumferential tensor contraction result. When multiplying a radial low-rank wavelet kernel with a radial wavelet tensor, the radial wavelet tensor values ​​at each low-rank position within the same radial height position are read sequentially and multiplied by the kernel weight at the same low-rank position in the radial low-rank wavelet kernel. The products of each low-rank position are accumulated within the same radial height position to obtain the radial tensor contraction value at that radial height position. The radial tensor contraction values ​​at all radial height positions are arranged in radial height order to form the radial tensor contraction result. The three types of tensor shrinkage results retain their respective axial position, circumferential arrangement position, and radial height position, and are all incorporated into the shaft system residual orthogonal subtraction structure.

[0112] In this embodiment, after deducting the axial and circumferential explained components, a shock absorber assembly contour reconstruction field is formed via a contour projection network, including:

[0113] The shaft system residual orthogonal subtraction structure receives axial tensor contraction results, circumferential tensor contraction results, and radial tensor contraction results. The axial tensor contraction results are expanded along the axial position into axial interpreted components. Specifically, the expansion of the axial tensor contraction results into axial interpreted components along the axial position is as follows:

[0114] The contraction values ​​corresponding to each axial position in the axial tensor contraction results are read and rearranged according to the axial order from the damper mounting end to the moving end. Axial positions can be set at intervals of 0.2mm to 1mm. In a damper sample with a length of 280mm, the number of axial positions can be configured from 280 to 1400. For each axial position, the axial tensor contraction value is backfilled to the circumferential arrangement position and radial height position under the same axial section, so that each circumferential position within the same axial section shares the corresponding axial contraction value. The circumferential arrangement positions can be set at angular intervals of 1° to 5°, for example, 72, 180, or 360 circumferential positions per axial section. The axial interpreted component consists of the axial position index, axial contraction value, and corresponding structural position, used to bear the contour changes corresponding to overall bending, axial offset, and section center drift.

[0115] In the circumferential tensor contraction results and radial tensor contraction results, the isotopic wavelet components with the same axial position as the axially interpreted components are located, and the corresponding isotopic wavelet components are subtracted bit by bit to form the circumferential tensor contraction results after axial subtraction and the radial tensor contraction results after axial subtraction. Specifically, the isotopic wavelet components with the same axial position as the axially interpreted components are:

[0116] The corresponding axial section is searched in the circumferential and radial tensor contraction results according to the axial position index. When the difference between the axial position in the circumferential or radial tensor contraction result and the axial position of the interpreted axial component does not exceed half an axial sampling interval, the corresponding component is regarded as a co-position wavelet component; for example, when the axial sampling interval is 0.5 mm, components with an axial position difference of no more than 0.25 mm are included in the co-position subtraction. Before subtraction, the interpreted axial component is adjusted to the same number of channels as the circumferential and radial tensor contraction results through linear mapping. The number of channels can be set to 32, 64, or 128. After channel alignment is completed, bitwise subtraction is performed at the same axial position; the subtraction ratio can be 1.0, or it can be set to a trainable ratio between 0.8 and 1.0 in the early stage of training to avoid excessive subtraction of axial components. After subtraction, the circumferential tensor contraction result and the radial tensor contraction result after axial subtraction are formed; the original contraction value of the component that cannot match the axial position is retained and a low confidence alignment mark is configured.

[0117] The orthogonal subtraction structure for shaft system residuals expands the circumferential tensor contraction result after axial subtraction into circumferential explained components along the circumferential position. Specifically, the expansion of the circumferential tensor contraction result after axial subtraction into circumferential explained components along the circumferential position is as follows:

[0118] The contraction values ​​corresponding to each circumferential position in the circumferential tensor contraction result after axial subtraction are read and rearranged according to the circumferential order within the same axial section. For rod profiles, cylinder profiles, and spring profiles, the circumferential positions are arranged in a closed loop from 0° to 360°; when the circumferential sampling interval is 2°, 180 circumferential positions are configured within one axial section. For mounting connection profiles, the circumferential range of actual profile points is retained; for example, the mounting ear area can occupy a circumferential range of 30° to 90° without forcibly completing a complete closed loop. The circumferential interpreted component consists of the circumferential position index, the circumferential contraction value, and the corresponding axial position, used to bear the profile changes corresponding to circumferential eccentricity, circumferential non-uniformity, and local circumferential offset.

[0119] In the radial tensor shrinkage result after axial subtraction, the co-position wavelet components with the same circumferential position as the circumferential interpreted components are located, and the corresponding co-position wavelet components are subtracted bit by bit to form the radial residual tensor shrinkage result. The formation of the radial residual tensor shrinkage result is specifically as follows:

[0120] The radial shrinkage component is located in the radial tensor shrinkage result after axial subtraction according to the circumferential position index. When the radial shrinkage component and the circumferential interpreted component are in the same axial section and the circumferential position difference does not exceed half a circumferential sampling interval, the radial shrinkage component is treated as a co-position wavelet component; for example, when the circumferential sampling interval is 2°, components with a circumferential position difference of no more than 1° are included in the subtraction. The shaft system residual orthogonal subtraction structure adjusts the circumferential interpreted component to the same number of channels as the radial shrinkage component, and then performs bitwise subtraction at the same circumferential position. After the subtraction is completed, the radial variation not covered by the axial interpreted component and the circumferential interpreted component is retained to form the radial residual tensor shrinkage result. The radial residual tensor shrinkage result mainly retains radial height abrupt changes, such as cylinder bulges, local depressions, weld protrusions, and mounting end warping; in specific implementations, radial height abrupt changes can be marked according to the height difference between adjacent positions being greater than 0.1mm to 0.5mm.

[0121] The axial interpreted component, circumferential interpreted component, and radial residual tensor shrinkage results are co-positioned according to their axial, circumferential, and radial height positions to form the wavelet domain contour reconstruction result. Specifically, the co-positioning is as follows:

[0122] Using the axial, circumferential, and radial height positions in the damper shaft profile tensor as overlay references, the interpreted axial components are first written into the corresponding axial sections, then the interpreted circumferential components are written into the corresponding circumferential positions, and finally the radial residual tensor contraction result is written into the corresponding radial height position. If all three types of components exist at the same position, the wavelet domain reconstruction value is synthesized by adding them one by one at the same position; if one type of component is missing at a certain position, a 0 value is used as a placeholder, and a low confidence marker is retained. The number of channels in the wavelet domain profile reconstruction result can be maintained at 64 or 128 channels, and the arrangement of the axial, circumferential, and structural positions is consistent with that of the damper shaft profile tensor.

[0123] The contour projection network projects the wavelet domain contour reconstruction result back into the shock absorber contour space, forming the shock absorber assembly contour reconstruction field. Specifically, the contour projection network projects the wavelet domain contour reconstruction result back into the shock absorber contour space as follows:

[0124] The contour projection network reads the reconstructed values ​​from the axial and circumferential positions in the wavelet domain contour reconstruction results, and converts the wavelet domain channels into radial height reconstruction values ​​through one to three linear mapping layers or convolutional mapping layers. For example, when the input channel is 64, it can first be mapped to 32 channels, and then mapped to one radial height output channel. The radial height reconstruction values ​​are backfilled into the shock absorber contour space according to the structural position, axial position, and circumferential position, forming reconstructed contour points that are on the same coordinate reference as the three-dimensional contour data of the shock absorber. All reconstructed contour points are rearranged according to the rod contour, cylinder contour, spring contour, and installation connection contour to form the shock absorber assembly contour reconstruction field. Each reconstructed contour point in the shock absorber assembly contour reconstruction field carries at least the structural position, axial position, circumferential position, and radial height reconstruction values, and enters the decomposition stage of the co-position deviation between the actual contour and the reconstructed contour.

[0125] In this embodiment, the formation of the shock absorber assembly deviation field includes:

[0126] The assembly deviation disassembly module locates the corresponding actual contour points in the three-dimensional contour data of the shock absorber according to the axial and circumferential positions in the shock absorber assembly contour reconstruction field, and then pastes the actual radial height of the corresponding actual contour points back to the corresponding reconstruction positions. Specifically, locating the corresponding actual contour points in the three-dimensional contour data of the shock absorber involves:

[0127] The structural, axial, and circumferential positions of each reconstructed contour point in the shock absorber assembly contour reconstruction field are read, and actual contour points with consistent structural positions, similar axial positions, and similar circumferential positions are searched in the shock absorber's 3D contour data. The matching range for axial positions can be half an axial sampling interval; for example, when the axial sampling interval is 0.5 mm, actual contour points with an axial position difference of no more than 0.25 mm are included in the candidate range. The matching range for circumferential positions can be half a circumferential sampling interval; for example, when the circumferential sampling interval is 2°, actual contour points with a circumferential position difference of no more than 1° are included in the candidate range. When multiple actual contour points exist within the candidate range, the actual contour point with the smallest spatial distance from the reconstructed contour point is selected as the corresponding actual contour point. When no actual contour point exists within the candidate range, a void mark is retained at the corresponding reconstruction position, and actual contour points at adjacent structural positions are not used as substitutes. After completing the co-positioning, the actual radial height of the co-positioned actual contour point is written to the corresponding reconstruction position, so that the actual contour and the reconstructed contour are compared under the same structural, axial, and circumferential positions.

[0128] The assembly deviation breakdown module performs a difference breakdown between the actual radial height and the reconstructed radial height at the same axial and circumferential positions, forming a radial height deviation. Specifically, the radial height deviation is as follows:

[0129] Read the actual radial height and reconstructed radial height at the same structural position, axial position, and circumferential position. Subtract the reconstructed radial height from the actual radial height to obtain the signed radial height deviation. A positive radial height deviation indicates that the actual contour bulges outward relative to the reconstructed contour; a negative radial height deviation indicates that the actual contour is concave inward relative to the reconstructed contour; a radial height deviation close to 0 indicates that the actual contour and the reconstructed contour are basically consistent at that position. For positions with missing marks, the radial height deviation is not calculated, but the missing marks are retained. To reduce the influence of minor jitter in structured light measurement, deviations with absolute values ​​less than the measurement noise tolerance can be set to 0. The measurement noise tolerance can be determined based on the structured light measurement accuracy, for example, between 0.03 mm and 0.10 mm.

[0130] The assembly deviation disassembly module continuously arranges radial height deviations along the axial position to form axial deviation segments, and then organizes the radial height deviations along the circumferential arrangement direction of the same axial section to form circumferential deviation segments. Specifically, the axial deviation segments are formed as follows:

[0131] Under the same structural location and the same circumferential position, radial height deviations are arranged in order from the mounting end to the moving end along the axial position. When the deviation signs of adjacent axial positions are consistent and the absolute values ​​of the deviations are all greater than the measurement noise tolerance, the adjacent axial positions are grouped into the same axial deviation segment. When there are missing measurement marks, reversed deviation signs, or absolute deviation values ​​lower than the measurement noise tolerance between adjacent axial positions, the axial deviation segment is broken at that position. To avoid isolated noise forming false segments, the axial deviation segment can be required to cover at least three consecutive axial sampling positions, or have a continuous length of not less than 1 mm. The axial deviation segment records the starting axial position, ending axial position, corresponding circumferential position, maximum deviation value, and average deviation value.

[0132] Specifically, the circumferential deviation segment is formed as follows:

[0133] Within the same structural location and axial section, the radial height deviation of each circumferential position is read according to the circumferential arrangement direction. If the deviation signs of adjacent circumferential positions are consistent and the absolute values ​​of the deviations are all greater than the measurement noise tolerance, the adjacent circumferential positions are grouped into the same circumferential deviation segment. For rod profiles, cylinder profiles, and spring profiles, the first and last circumferential positions maintain a closed connection; for installation connection profiles, circumferential deviation segments are only organized within the circumferential interval where the profile points actually exist. A circumferential deviation segment can record the starting circumferential position, ending circumferential position, axial section, maximum deviation value, and average deviation value. If a circumferential deviation segment spans multiple adjacent axial sections, the segment correlation between adjacent axial sections is retained to facilitate the next step of delineating continuous defect areas.

[0134] The assembly deviation breakdown module patches the radial height deviation, axial deviation segments, and circumferential deviation segments back onto the shock absorber's three-dimensional contour data according to the shock absorber's structural position, forming a shock absorber assembly deviation field. Specifically, the back-patching of the shock absorber's structural position onto the three-dimensional contour data is as follows:

[0135] Based on the structural positions of the rod profile, cylinder profile, spring profile, and mounting connection profile, each radial height deviation is written back to the corresponding actual profile point position in the shock absorber's 3D profile data. A deviation symbol, absolute value of the deviation, and missing measurement mark are then attached to this actual profile point position. For positions where axial deviation segments have already formed, the axial segment number, starting axial position, and ending axial position are recorded; for positions where circumferential deviation segments have already formed, the circumferential segment number, starting circumferential position, and ending circumferential position are recorded. After the backtracking is completed, each profile point in the shock absorber's 3D profile data not only retains its original spatial coordinates but also carries the corresponding radial height deviation and the information of its associated axial and circumferential deviation segments, thus forming a shock absorber assembly deviation field. This shock absorber assembly deviation field serves as the processing object for the defect result generation module to delineate the assembly defect areas of the shock absorber.

[0136] In this embodiment, the defect result generation module is specifically as follows:

[0137] The defect result generation module performs contiguous aggregation on continuous deviation positions in the shock absorber assembly deviation field. Deviation regions extending continuously along the axial direction are grouped into the axial offset defect candidate region; deviation regions offset circumferentially along the same axial section are grouped into the circumferential eccentricity defect candidate region; and deviation regions with concentrated radial height abrupt changes are grouped into the radial deformation defect candidate region. Specifically, the contiguous aggregation on continuous deviation positions in the shock absorber assembly deviation field involves:

[0138] Read the radial height deviation, axial deviation, and circumferential deviation segments from the shock absorber assembly deviation field, and eliminate deviation points whose absolute values ​​are less than the measurement noise tolerance. The measurement noise tolerance can be taken as 0.03 mm to 0.10 mm. Establish adjacency relationships for the remaining deviation points according to their structural position, axial position, and circumferential position. When adjacent deviation points are in the same structural position, have the same deviation sign, and are continuous in position, they are classified into the same continuous deviation region. Continuous deviation regions that are continuously distributed along the axial direction and cover multiple axial sections are classified into the axial offset defect candidate region; continuous deviation regions that are concentrated in the same axial section and offset along the circumferential direction are classified into the circumferential eccentricity defect candidate region; continuous deviation regions where the change in radial height deviation at adjacent positions exceeds the height abrupt change threshold are classified into the radial deformation defect candidate region. The height abrupt change threshold can be taken as 0.1 mm to 0.5 mm.

[0139] The defect result generation module re-attaches the axial offset defect candidate area, circumferential eccentricity defect candidate area, and radial deformation defect candidate area to the damper structure location, delineating the rod misalignment area, cylinder deformation area, spring eccentricity area, and installation connection misalignment area. Specifically, the re-attachment of the defect candidate areas to the damper structure location is as follows:

[0140] Read the structural location markers corresponding to the defect candidate areas, and delineate the defect candidate areas located on the rod outline and extending continuously along the axial direction as rod deviation areas; delineate the defect candidate areas located on the cylinder outline and exhibiting local outward convexity, inward concavity, or radial abrupt change as cylinder deformation areas; delineate the defect candidate areas located on the spring outline and exhibiting circumferential or radial offset as spring eccentricity areas; delineate the defect candidate areas located on mounting holes, connecting lugs, end connecting seats, or welded connection areas as installation connection misalignment areas.

[0141] The defect result generation module configures defect type and defect level labels according to the location of the delineated defect area, the continuous length of the deviation, and the radial height deviation distribution. It then encapsulates the defect area location, defect type, and defect level labels into a defect detection result for an electric vehicle shock absorber. Specifically, the configuration of the defect type and defect level labels is as follows:

[0142] Defect type markings are determined based on the structural location and deviation morphology of the defect area, including rod misalignment markings, cylinder deformation markings, spring eccentricity markings, and installation connection misalignment markings. Defect level markings are determined based on the maximum radial height deviation, the continuous length of the deviation, and the structural location of the defect area; for example, a minor level is assigned when the maximum radial height deviation is less than 0.2mm and the continuous length is less than 5mm; a moderate level is assigned when the maximum radial height deviation is between 0.2mm and 0.5mm or the continuous length is between 5mm and 15mm; and a severe level is assigned when the maximum radial height deviation is greater than 0.5mm, the continuous length is greater than 15mm, or the defect area is located at an installation connection. After marking, the defect area location, defect type markings, and defect level markings are packaged into the electric vehicle shock absorber defect detection results.

[0143] Example 1: To verify the feasibility of this invention in practice, it was applied to the online assembly and inspection production line of an electric vehicle shock absorber manufacturer. This production line mainly produces front fork shock absorbers and rear spring shock absorbers. The inspection objects include the rod, cylinder, spring, and mounting lug. The shock absorber length ranges from 260mm to 340mm, the cylinder outer diameter ranges from 28mm to 46mm, and the spring outer diameter ranges from 38mm to 62mm. The single-piece inspection cycle is approximately 10 seconds. Common defects include rod misalignment, partial bulges in the cylinder, cylinder dents, spring eccentricity, and mounting lug misalignment. Traditional two-dimensional visual inspection methods mainly rely on grayscale difference, edge contours, and local appearance texture to judge defects. In scenarios involving metallic reflection, spring obstruction, slight cylinder deformation, and complex mounting end structures, it is easy to misjudge normal reflective areas as defects and to miss minor assembly deviations.

[0144] During the data preparation phase, the testing station was equipped with one set of line structured light projectors and four industrial cameras. The industrial cameras were positioned at 0°, 90°, 180°, and 270° circumferentially around the shock absorber under test. The structured light projected image resolution was 2448×2048. The platform continuously collected production line testing data for five days, obtaining a total of 2380 shock absorber samples, including 1740 normal samples and 640 defective samples. Among the defective samples, there were 152 samples with rod misalignment, 176 samples with cylinder deformation, 184 samples with spring eccentricity, and 128 samples with installation connection misalignment. After data acquisition, centerline positioning is performed on the structured light stripes to remove isolated bright spots and discontinuous noise caused by metal reflection. Neighborhood patching is performed on short stripe gaps caused by the damper's curved surface. Subsequently, the pixel positions of the stripe centerline are converted into spatial contour points according to the camera calibration relationship and the structured light projection calibration relationship. Pose alignment is performed on the spatial contour points of multiple views under the same detection position to form the three-dimensional contour data of the damper.

[0145] In the shaft profile tensor construction stage, the three-dimensional profile data of the shock absorber is aggregated into the rod profile, cylinder profile, spring profile, and mounting connection profile according to the shock absorber assembly structure. The axial sampling interval is set to 0.5 mm, and the circumferential sampling interval is set to 2°, with a maximum of 180 circumferential positions configured for each axial section. For the rod profile, cylinder profile, and spring profile, a closed arrangement is performed in circumferential order; for non-closed structures such as mounting lugs, mounting holes, and end connectors, only the actually collected circumferential intervals are retained, and the complete circumferential profile is not forcibly completed. After completing the axial section segmentation and circumferential arrangement, the structural position, axial position, circumferential position, and radial height are encapsulated into the shock absorber shaft profile tensor.

[0146] In the model processing stage, the shock absorber shaft system profile tensor is input into the improved wavelet neural operator network model. The number of channels in the profile lifting network is set to 64, and the number of low-rank bits in the low-rank shaft system wavelet tensor merging operator is set to 16. The wavelet kernel integration layer generates approximation coefficients and detail coefficients along the axial, circumferential, and radial directions, respectively. The low-rank shaft system wavelet tensor merging operator calls the axial low-rank wavelet kernel, circumferential low-rank wavelet kernel, and radial low-rank wavelet kernel to perform rank-by-rank multiplication and in-place accumulation, forming axial tensor contraction results, circumferential tensor contraction results, and radial tensor contraction results. The shaft system residual orthogonal subtraction structure first expands the axial tensor contraction result into the axial interpreted components and subtracts the in-place axial components from the circumferential and radial tensor contraction results; then, it expands the circumferential tensor contraction result after axial subtraction into the circumferential interpreted components and subtracts the in-place circumferential components from the radial tensor contraction result after axial subtraction, obtaining the radial residual tensor contraction result. The contour projection network projects the wavelet domain contour reconstruction results back to the shock absorber contour space, forming the shock absorber assembly contour reconstruction field.

[0147] During the defect result generation stage, the three-dimensional contour data of the shock absorber is pasted back to the shock absorber assembly contour reconstruction field. The difference between the actual radial height and the reconstructed radial height at the same axial position and the same circumferential position is calculated to form the shock absorber assembly deviation field. Positions with an absolute value of radial height deviation less than 0.05 mm are treated as measurement noise; positions with a continuous coverage length greater than 1.5 mm and consistent deviation signs are classified into continuous deviation areas. Deviation areas extending continuously along the axial position are classified into axial offset defect candidate areas, deviation areas circumferentially offset along the same axial section are classified into circumferential eccentricity defect candidate areas, and deviation areas with concentrated radial height abrupt changes are classified into radial deformation defect candidate areas. After the defect candidate areas are pasted back to the corresponding structural positions, the rod skew area, cylinder deformation area, spring eccentricity area, and installation connection misalignment area are delineated, and defect type and defect level labels are configured according to the defect area location, continuous deviation length, and radial height deviation distribution.

[0148] In the comparative experiment, the system of this invention was compared with a traditional two-dimensional appearance inspection system and a common three-dimensional point cloud fitting inspection system. The traditional two-dimensional appearance inspection system mainly judges defects based on image grayscale differences and edge contours, without constructing three-dimensional contour data of the shock absorber. The common three-dimensional point cloud fitting inspection system judges defects based on the fitting residual between the three-dimensional point cloud and the standard model, but it does not construct the shock absorber shaft system contour tensor, nor does it separate the axial explained components, circumferential explained components, and radial residual components in the wavelet domain. 500 shock absorbers continuously produced on the same production line were selected as test samples, of which 96 samples were manually verified as defective and 404 were normal. Regarding the defect detection rate, the traditional two-dimensional appearance inspection system was 82.3%, the common three-dimensional point cloud fitting inspection system was 88.5%, and the system of this invention was 96.9%. Regarding the false detection rate, the traditional two-dimensional appearance inspection system was 9.8%, the common three-dimensional point cloud fitting inspection system was 7.1%, and the system of this invention was 3.4%. Regarding the average error in defect location, the traditional two-dimensional appearance inspection system has an error of 1.42 mm, the ordinary three-dimensional point cloud fitting inspection system has an error of 0.96 mm, and the system of this invention has an error of 0.38 mm. Regarding the number of items requiring manual re-inspection, the traditional two-dimensional appearance inspection system has an error of 71 items, the ordinary three-dimensional point cloud fitting inspection system has an error of 54 items, and the system of this invention has an error of 23 items.

[0149] As can be seen from Example 1, this invention can generate three-dimensional contour data of a shock absorber using structured light multi-view images, and further construct a shock absorber shaft system contour tensor, so that the axial arrangement, circumferential distribution, and radial height variation of the rod, cylinder, spring, and mounting connection area can be expressed in the same tensor structure. The improved wavelet neural operator network model performs directional tensor contraction on the wavelet coefficients through a low-rank shaft system wavelet tensor convergence operator, and reduces the interference of axial offset and circumferential eccentricity on the radial local deformation judgment through orthogonal subtraction of shaft system residuals, making the shock absorber assembly deviation field more closely match the actual defect location. Compared with traditional two-dimensional appearance inspection systems and ordinary three-dimensional point cloud fitting inspection systems, this invention's system can improve the defect detection rate, reduce the false detection rate and defect location error, and reduce the number of manual re-inspections, making it suitable for online inspection, sorting, and rework scenarios of electric vehicle shock absorbers.

[0150] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine vision-based defect detection system for electric vehicle shock absorbers, characterized in that, include: The structured light contour calculation module acquires multi-view structured light images of the shock absorber of the electric vehicle under test, analyzes the stripes and calculates the three-dimensional contour to form the three-dimensional contour data of the shock absorber. The shaft system profile tensor construction module gathers the three-dimensional profile data according to the structural position of the shock absorber, arranges the gathered profile along the axial direction, and forms the shock absorber shaft system profile tensor. An improved wavelet neural operator configuration module was constructed, and an improved wavelet neural operator network model was built. The contour lifting network, wavelet kernel integration layer and contour projection network were retained. The wavelet coefficient pooling unit in the wavelet kernel integration layer was replaced with a low-rank axis wavelet tensor pooling operator, and an axis residual orthogonal subtraction structure was embedded in the low-rank axis wavelet tensor pooling operator. The shaft system tensor merging module connects the shock absorber shaft system profile tensor to the wavelet kernel integration layer via a profile lifting network. The low-rank shaft system wavelet tensor merging operator performs shaft system tensor contraction in the wavelet domain and deducts the axial and circumferential explained components. The profile projection network then forms the shock absorber assembly profile reconstruction field. The assembly deviation disassembly module re-attaches the three-dimensional contour data of the shock absorber to the shock absorber assembly contour reconstruction field, disassembles the co-position deviation between the actual contour and the reconstructed contour, and forms the shock absorber assembly deviation field. The defect result generation module delineates the defect area of ​​the shock absorber assembly according to the continuous distribution position of the deviation in the shock absorber assembly deviation field, and generates the defect detection results of the electric vehicle shock absorber.

2. The machine vision-based electric vehicle shock absorber defect detection system according to claim 1, characterized in that, The data used to form the three-dimensional profile of the shock absorber includes: Synchronous acquisition of structured light projection images of the shock absorber of the electric vehicle under test at different shooting angles is performed, and multi-view images corresponding to the same detection position are grouped into the same contour calculation group. Centerline positioning is performed on the structured light stripes in the contour solution group to remove isolated bright spots and discontinuous noise formed by metal reflection, and neighborhood patching is performed on short stripe gaps caused by the damper surface to form a set of structured light stripe centerlines. Based on the camera calibration relationship and the structured light projection calibration relationship, the pixel positions in the set of structured light stripe center lines are converted into spatial contour points, and pose alignment is performed on the multi-view spatial contour points under the same detection position to form the three-dimensional contour data of the shock absorber.

3. The machine vision-based electric vehicle shock absorber defect detection system according to claim 1, characterized in that, The tensor forming the shaft profile of the shock absorber includes: The three-dimensional contour data of the shock absorber is aggregated along the assembly structure to the structural positions corresponding to the rod contour, cylinder contour, spring contour and mounting connection contour. Using the extension direction from the mounting end to the movable end of the shock absorber as the axial alignment direction, axial section cutting is performed on the aggregated three-dimensional contour data; The contour points distributed around the axis of the shock absorber within the same axial section are arranged in circumferential order. For contour points with closed circumferential structures, a closed arrangement is performed, and the radial height of the contour points relative to the axial arrangement direction is attached to the corresponding circumferential position. The circumferential arrangement results corresponding to adjacent axial sections are stacked sequentially along the axial direction, and the structural position, axial position, circumferential position and radial height are encapsulated as the damper shaft profile tensor.

4. The machine vision-based electric vehicle shock absorber defect detection system according to claim 1, characterized in that, The construction of the improved wavelet neural operator network model includes: The contour lifting network, wavelet kernel integration layer and contour projection network in the wavelet neural operator network are retained. The output of the contour lifting network is connected to the input of the wavelet kernel integration layer. The wavelet domain processing path of the wavelet kernel integration layer is connected to the contour projection network through the low-rank axis wavelet tensor pooling operator and the axis residual orthogonal subtraction structure. The wavelet coefficient pooling unit in the wavelet kernel integration layer is replaced with a low-rank axis wavelet tensor pooling operator, and the low-rank axis wavelet tensor pooling operator is embedded in the wavelet domain processing path of the wavelet kernel integration layer. The low-rank axial wavelet tensor pooling operator is configured with axial low-rank wavelet kernels, circumferential low-rank wavelet kernels and radial low-rank wavelet kernels. The axial low-rank wavelet kernel is attached to the axial wavelet coefficient position, the circumferential low-rank wavelet kernel is attached to the circumferential wavelet coefficient position, and the radial low-rank wavelet kernel is attached to the radial wavelet coefficient position. An orthogonal subtraction structure for axis residuals is embedded within the tensor contraction path of the low-rank axis wavelet tensor converging operator. The input of the orthogonal subtraction structure for axis residuals is connected to the tensor contraction output of the low-rank axis wavelet tensor converging operator, and the output of the orthogonal subtraction structure for axis residuals is connected to the contour projection network.

5. The machine vision-based electric vehicle shock absorber defect detection system according to claim 1, characterized in that, The step of connecting the shock absorber shaft system profile tensor to the wavelet kernel integration layer via a profile lifting network, and then performing shaft system tensor contraction in the wavelet domain by a low-rank shaft system wavelet tensor merging operator, includes: The contour lifting network slides adjacent section windows along the axial arrangement direction of the contour tensor of the shock absorber shaft system. Differential convolution and normalization mapping are performed on the contour center displacement, circumferential contour difference and radial height change in the adjacent section windows, respectively. The mapped local contour response is then pasted back to the corresponding axial position to form the lifting contour feature. The wavelet kernel integration layer configures axial wavelet kernels, circumferential wavelet kernels, and radial wavelet kernels on the lifted contour features. The axial wavelet kernel slides along the axial position and performs low-frequency preservation and high-frequency difference on the lifted contour features at adjacent axial positions to form axial approximation coefficients and axial detail coefficients. The circumferential wavelet kernel slides along the circumferential arrangement direction of the same axial section, performs circumferential low-frequency preservation and circumferential high-frequency difference on the lifting profile features of adjacent circumferential positions, forming circumferential approximation coefficients and circumferential detail coefficients. The radial wavelet kernel slides along the radial height position, performs local low-frequency preservation and local high-frequency difference on the radial height change, forming radial approximation coefficients and radial detail coefficients. The low-rank axial wavelet tensor pooling operator arranges the axial approximation coefficients and axial detail coefficients in the low-rank position to form the axial wavelet tensor, and arranges the circumferential approximation coefficients and circumferential detail coefficients in the circumferential position and low-rank position to form the circumferential wavelet tensor, and arranges the radial approximation coefficients and radial detail coefficients in the radial height position and low-rank position to form the radial wavelet tensor. The low-rank axial wavelet tensor pooling operator calls trainable axial low-rank wavelet kernels, circumferential low-rank wavelet kernels, and radial low-rank wavelet kernels. It multiplies the axial low-rank wavelet kernel with the axial wavelet tensor bit by bit and accumulates them in the same position to form the axial tensor contraction result. It multiplies the circumferential low-rank wavelet kernel with the circumferential wavelet tensor bit by bit and accumulates them in the same position to form the circumferential tensor contraction result. The radial low-rank wavelet kernel is multiplied rank-by-rank with the radial wavelet tensor and accumulated in the same position to form the radial tensor shrinkage result. The axial tensor shrinkage result, the circumferential tensor shrinkage result, and the radial tensor shrinkage result are then connected to the shaft system residual orthogonal subtraction structure.

6. The machine vision-based electric vehicle shock absorber defect detection system according to claim 1, characterized in that, The process of deducting the axial and circumferential explained components and forming a shock absorber assembly profile reconstruction field via a profile projection network includes: The shaft system residual orthogonal subtraction structure receives the axial tensor contraction results, circumferential tensor contraction results, and radial tensor contraction results, and expands the axial tensor contraction results into axial explained components along the axial position; In the circumferential tensor contraction results and the radial tensor contraction results, the co-position wavelet components that have the same axial position as the axially interpreted components are located, and the corresponding co-position wavelet components are subtracted one by one to form the circumferential tensor contraction results after axial subtraction and the radial tensor contraction results after axial subtraction. The shaft system residual orthogonal subtraction structure expands the circumferential tensor contraction result after axial subtraction into circumferential explained components along the circumferential position. In the radial tensor shrinkage result after axial subtraction, the co-position wavelet component with the same circumferential position as the circumferential interpreted component is located, and the corresponding co-position wavelet component is subtracted bit by bit to form the radial residual tensor shrinkage result. The results of shrinking the axial explained component, circumferential explained component, and radial residual tensor are superimposed in the same position according to the axial position, circumferential position, and radial height position to form the wavelet domain contour reconstruction result; The contour projection network projects the wavelet domain contour reconstruction results back to the shock absorber contour space, forming the shock absorber assembly contour reconstruction field.

7. The machine vision-based electric vehicle shock absorber defect detection system according to claim 1, characterized in that, The formation of the shock absorber assembly deviation field includes: The assembly deviation disassembly module locates the corresponding actual contour points in the three-dimensional contour data of the shock absorber according to the axial and circumferential positions in the shock absorber assembly contour reconstruction field, and then pastes the actual radial height of the corresponding actual contour points back to the corresponding reconstruction positions. The assembly deviation decomposition module performs a difference decomposition between the actual radial height and the reconstructed radial height at the same axial position and the same circumferential position, forming a radial height deviation. The assembly deviation disassembly module continuously arranges radial height deviations along the axial position to form axial deviation segments, and organizes radial height deviations along the circumferential arrangement direction of the same axial section to form circumferential deviation segments. The assembly deviation disassembly module re-attaches radial height deviation, axial deviation segments, and circumferential deviation segments to the three-dimensional contour data of the shock absorber according to the structural position of the shock absorber, forming the shock absorber assembly deviation field.

8. The machine vision-based electric vehicle shock absorber defect detection system according to claim 1, characterized in that, The defect result generation module is specifically as follows: The defect result generation module performs contiguous aggregation on the continuous deviation positions in the shock absorber assembly deviation field, and classifies the deviation areas that extend continuously along the axial position into the axial offset defect candidate area, the deviation areas that are circumferentially offset along the same axial section into the circumferential eccentricity defect candidate area, and the deviation areas with concentrated radial height abrupt changes into the radial deformation defect candidate area. The defect result generation module back-attaches the axial offset defect candidate area, circumferential eccentricity defect candidate area and radial deformation defect candidate area to the shock absorber structure location, and delineates the rod body skew area, cylinder deformation area, spring eccentricity area and installation connection misalignment area. The defect result generation module configures defect type and defect level markers according to the location of the delineated defect area, the continuous length of the deviation, and the radial height deviation distribution, and encapsulates the defect area location, defect type markers, and defect level markers into the defect detection results of the electric vehicle shock absorber.

Images