Multi-station carbon fiber preform visual inspection device, method and system

Abstract

The invention discloses a multi-station carbon fiber preform visual detection device, method and system, which are used for automatically detecting trace lines, yarn arc reduction, warp and weft density and preform height of a carbon fiber preform. The equipment comprises a detection platform, a mobile module, a bottom plate, a prefabricated body supporting plate, a control cabinet and three stations, the first station is provided with a flat light source, a 2D camera and a polarizer; the second station is provided with a strip-shaped light source, a 2D camera and a polarizer; the method comprises the steps of path planning and coincidence distribution analysis, acquisition time sequence control node determination, spectral region difference extraction, direction distribution space parameter calculation and multi-scale texture gradient feature analysis, and multi-dimensional, continuous and accurate detection of the surface features of the carbon fiber preform is achieved. According to the invention, the automation level and precision of prefabricated body detection are obviously improved, the limitation of existing single-station detection is broken through, and the method has important engineering application value.

Description

Technical Field

[0001] This invention belongs to the field of industrial vision inspection and automation equipment technology, specifically relating to a multi-station automated vision inspection device and its inspection method for carbon fiber preforms, which is particularly suitable for application scenarios of multi-angle and multi-scale inspection of the surface structure, contour features and spatial morphology of carbon fiber preforms. Background Technology

[0002] As a key intermediate in the molding process of composite carbon fiber preforms, the surface structure quality, layup orientation consistency, and geometric accuracy of carbon fiber preforms directly affect the subsequent molding quality and service performance. In actual production, carbon fiber preforms are typically characterized by complex structures, large dimensions, and distinct surface texture directions, and are susceptible to defects caused by weaving processes, layup processes, and environmental factors before molding.

[0003] Existing carbon fiber preform inspection methods mostly rely on manual visual inspection or single-station, single-vision imaging, which has the following shortcomings: (1) low inspection efficiency, making it difficult to meet the needs of mass production; (2) poor adaptability of single imaging methods to complex surface structures and directional textures; (3) the inspection results are greatly affected by lighting conditions and operator experience, resulting in insufficient stability; and (4) it is difficult to take into account both two-dimensional surface features and three-dimensional spatial morphological information during the inspection process. Therefore, there is an urgent need for a carbon fiber preform inspection equipment and method that can achieve multi-station collaboration, automated operation, and integrate multiple light sources and multiple vision sensing methods to improve inspection efficiency and the consistency and reliability of inspection results. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the present invention aims to provide a multi-station carbon fiber preform visual inspection device, method and system. By configuring different imaging methods and optical structures at multiple inspection stations, it can achieve high-precision automated visual inspection of carbon fiber preform trace lines, yarn reduction arcs and warp and weft density, improve inspection accuracy and efficiency, and overcome the technical problem of unstable identification under complex structures by traditional edge detection and single light source methods.

[0005] The technical solution adopted in this invention is: An automated visual inspection device for multi-station carbon fiber preforms includes an inspection platform, a moving module, a base plate, a carbon fiber preform support plate, and multiple inspection stations set on the inspection platform; the inspection platform is provided with a first station, a second station, and a third station arranged sequentially along the inspection path; The mobile module is fixedly installed on the testing platform and can drive the base plate to move along a preset direction; The base plate is fixedly installed on the mobile module, the carbon fiber preform support plate is fixedly installed on the base plate, and the carbon fiber preform to be tested is placed on the carbon fiber preform support plate. The system is divided into three stages: the first station is used for two-dimensional contour and texture imaging of the carbon fiber preform surface, and is equipped with a flat panel light source, a first 2D camera and a first polarizing filter; the second station is used for multi-directional enhanced imaging detection, and is equipped with multiple sets of strip light sources, a second 2D camera and a second polarizing filter; the third station is used for spatial morphology detection, and is equipped with a 3D camera and a third polarizing filter; the visual imaging unit of each station is electrically connected to the control cabinet to achieve coordinated image acquisition, processing and motion control.

[0006] Based on the above-mentioned equipment, the present invention also provides an automated visual inspection method for multi-station carbon fiber preforms, comprising the following steps: S1: Path Analysis Steps Based on the vision inspection module, the continuous displacement and velocity changes in the multi-station inspection path are analyzed. By collecting the movement distance and time data, the inspection path is divided, and the overlapping intervals of the boundary nodes of each path are compared. The acquisition sequence of the inspection nodes is adjusted to obtain the path overlap distribution characteristics. S2: Timing Control Steps Based on the path overlap distribution characteristics, the relationship between the speed change and position change of the detection module in different path segments is determined, the displacement encoder signal of the workstation slide rail bearing group is screened, and combined with the speed feedback information of the vision detection module, key data is compared step by step to locate the moment of motion state switching and obtain the acquisition timing control node. S3: Spectral Analysis Steps Based on the acquisition timing control node, the continuously acquired multi-wavelength illumination image sequence is analyzed, each frame image is divided into contour region and background region, the brightness of the region pixels is calculated and brightness equalization processing is performed, the brightness difference before and after processing is compared and the scale is unified to obtain the trend of spectral region difference. S4: Direction Resolution Steps Based on the spectral region difference trend, band regions with significant brightness changes are selected, the spatial gray-scale distribution characteristics within these regions are analyzed, the sets of principal and secondary direction vectors are calculated, and the angle changes between the principal and secondary directions are compared to obtain the spatial parameters of the direction distribution. S5: Texture Feature Extraction Steps Based on the aforementioned directional distribution spatial parameters, the target region positioning is optimized, the principal directional structural layer on the surface of the carbon fiber preform is analyzed, and the directional gradient features under different scale windows are calculated by combining multi-scale morphology processing components. Key regions are then selected to obtain multi-scale texture gradient features.

[0007] The beneficial effects of the present invention are: (1) By using multi-station division of labor for detection, the collaboration between two-dimensional imaging and three-dimensional imaging is realized, thereby improving the integrity of detection; (2) By using a multi-source and polarization imaging structure, surface reflection interference is effectively reduced, thereby improving the imaging quality; (3) By using path overlap analysis and acquisition timing control, the stability of image acquisition is improved; (4) By combining multi-band spectral analysis and directional feature analysis, the ability to identify the surface structural features of carbon fiber preforms is enhanced; (5) The detection process is automated, significantly improving detection efficiency and reducing reliance on manual labor. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the structure of the multi-station carbon fiber preform visual inspection device of the present invention; Figure 2 Top view of a multi-station carbon fiber preform visual inspection device; Figure 3 Flowchart of a multi-station visual inspection method for carbon fiber preforms; Figure 4 Here is a flowchart of the steps in S1; Figure 5 Here is a flowchart of the steps in S2; Figure 6 Here is a flowchart of the steps in S3; Figure 7 Here is a flowchart of the steps in S4; Figure 8 Here is a flowchart of the steps in S5; Figure 9 System block diagram Figure 10 This is the testing process for carbon fiber preforms.

[0009] Explanation of reference numerals in the attached drawings: 1-Detection platform; 2-Moving module; 3-Base plate; 4-Carbon fiber preform support plate; 5-Carbon fiber preform; 6-First station; 7-Second station; 8-Third station; 9-Flat panel light source; 101-First 2D camera; 102-Second 2D camera; 11-Strip light source; 12-3D camera; 131-First polarizing filter; 132-Second polarizing filter; 133-Third polarizing filter; 14-Control cabinet. Detailed Implementation

[0010] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0011] Please see Figure 1 and 2A specific embodiment of the present invention provides a multi-station visual inspection device for carbon fiber preforms, including an inspection platform 1, a moving module 2, a base plate 3, a carbon fiber preform support plate 4, a carbon fiber preform 5, a first station 6, a second station 7, a third station 8, a flat panel light source 9, a first 2D camera 101, a second 2D camera 102, a strip light source 11, a 3D camera 12, a first polarizing filter 131, a second polarizing filter 132, a third polarizing filter 133, and a control cabinet 14.

[0012] The testing platform 1 is equipped with a first station 6, a second station 7, and a third station 8; a moving module 2 is fixed on the testing platform and is used to drive the base plate 3 to move along the testing path; the moving module is installed on the base plate 3 to support the carbon fiber preform tray 4; the carbon fiber preform tray 4 is fixedly installed on the base plate and is used to place the carbon fiber preform 5; the first station is equipped with a flat panel light source 9, a first 2D camera 101, and a first polarizing filter 131; the second station is equipped with a strip light source 11, a second 2D camera 102, and a second polarizing filter 132; the third station is equipped with a 3D camera 12 and a third polarizing filter 133; the control cabinet 14 centrally controls the light source, camera, moving module, and signal acquisition.

[0013] The installation positions of the light source and camera are as follows: The first workstation features symmetrically mounted flatbed light sources at a 45° angle to the base plate, 500mm apart, and 400mm high. A first 2D camera is symmetrically positioned at the center of the workstation, with a first polarizing filter mounted directly below it. The first 2D camera at the first workstation acquires 3D information about the carbon fiber preform surface, extracts features from the tracer region, obtains the continuous contour of the tracer in the image, locates the spatial position of the tracer in different warp and weft directions, calculates the spatial distance between any two feature points within the same tracer, obtains the tracer length parameter, and calculates the spacing between adjacent warp and weft tracers based on the spatial distribution of the tracers in the warp and weft directions. Simultaneously, the number of warp and weft yarns is counted within a unit length range to calculate the warp and weft density, achieving automated visual measurement of the yarn density of the carbon fiber preform.

[0014] The second station features two symmetrical strip light sources, with heights of 200mm and 450mm respectively. A second 2D camera is located in the center, and a second polarizing filter is mounted directly below the camera. The carbon fiber preform is illuminated by the strip light sources, and the second 2D camera acquires images. The yarn contour and texture direction are analyzed to identify the yarn reduction arc region. The position, width, and shape parameters of the yarn reduction arc are extracted to achieve yarn reduction arc detection.

[0015] The 3D camera at the third workstation is fixed at the center of the workstation, and the third polarizing filter is located directly below the camera. The area for height calculation is determined using a tangent plane. Based on the tangent plane, the XLD profile curve is extracted, a planar reference is established, profile height information is obtained, and the height of each region of the carbon fiber preform is calculated, thus achieving height measurement.

[0016] Another specific embodiment of this application provides a multi-station carbon fiber preform visual inspection method, implemented based on the multi-station carbon fiber preform visual inspection device described above, with reference to... Figure 3-4 This includes the following steps: S1: Path Analysis Steps Based on the vision inspection module, the continuous displacement and velocity changes in the multi-station inspection path are analyzed. By collecting the movement distance and time data, the inspection path is divided, and the overlapping intervals of the boundary nodes of each path are compared. The acquisition sequence of the inspection nodes is adjusted to obtain the path overlap distribution characteristics. S2: Timing Control Steps Based on the path overlap distribution characteristics, the relationship between the speed change and position change of the detection module in different path segments is determined, the displacement encoder signal of the workstation slide rail bearing group is screened, and combined with the speed feedback information of the vision detection module, key data is compared step by step to locate the moment of motion state switching and obtain the acquisition timing control node. S3: Spectral Analysis Steps Based on the acquisition timing control node, the continuously acquired multi-wavelength illumination image sequence is analyzed, each frame image is divided into contour region and background region, the brightness of the region pixels is calculated and brightness equalization processing is performed, the brightness difference before and after processing is compared and the scale is unified to obtain the trend of spectral region difference. S4: Direction Resolution Steps Based on the spectral region difference trend, band regions with significant brightness changes are selected, the spatial gray-scale distribution characteristics within these regions are analyzed, the sets of principal and secondary direction vectors are calculated, and the angle changes between the principal and secondary directions are compared to obtain the spatial parameters of the direction distribution. S5: Texture Feature Extraction Steps Based on the aforementioned directional distribution spatial parameters, the target region positioning is optimized, the principal directional structural layer on the surface of the carbon fiber preform is analyzed, and the directional gradient features under different scale windows are calculated by combining multi-scale morphology processing components. Key regions are then selected to obtain multi-scale texture gradient features.

[0017] Path planning and data acquisition control (1) Path analysis and overlapping distribution characteristics: The continuous displacement and velocity curves of the carbon fiber preform on the detection path are analyzed using the visual inspection module; the velocity change trend and movement state of each segment are compared according to the acquisition time, the path division is optimized, and the overlapping distribution characteristics of the path are obtained.

[0018] (2) Acquisition timing control node Based on the characteristics of overlapping paths, analyze the relationship between speed and position: filter the displacement encoder signal of the slide rail bearing group, combine it with the module speed feedback, and locate the motion switching time; determine the acquisition timing control node to ensure that image acquisition and motion status are synchronized.

[0019] Spectral image analysis and orientation parameter acquisition (1) Spectral region difference trend: Contour and background division, pixel brightness calculation, and brightness equalization processing are performed on continuously acquired multi-wavelength illumination images; the brightness difference before and after equalization is compared, the scale is unified, and the spectral region difference trend is obtained.

[0020] (2) Spatial parameters of directional distribution: Based on the differences in spectral regions, the brightness peak regions are screened, the main and secondary directional vector sets are calculated, the vector angles are compared, and the directional offset and spatial distribution are analyzed to provide directional distribution parameters for multi-scale texture analysis.

[0021] Multi-scale texture gradient feature extraction: Target region localization is optimized based on directional distribution spatial parameters. Multi-scale morphology processing components are used to calculate the gray-level gradients in the main direction and vertical direction under each window. Regions with significant amplitude changes are screened, and local gradient differences are statistically analyzed to obtain multi-scale texture gradient features, including principal axis gradient coefficient, vertical axis gradient coefficient, and hierarchical texture characteristics.

[0022] The path overlap distribution features include the interval overlap ratio, the temporal distribution of path nodes, and the detection segment switching identifier. The acquisition timing control nodes include synchronous acquisition markers, motion control benchmarks, and trigger time indexes. The spectral region difference trends include the brightness change amplitude, spectral response information, and regional contrast coefficients. The directional distribution spatial parameters include the directional principal axis parameters, spatial distribution factors, and angle offset indices. The multi-scale texture gradient features include the principal axis gradient coefficient, the perpendicular axis gradient coefficient, and hierarchical texture characteristics.

[0023] In S1, the preform refers to the blade preform, which is an intermediate composite material blade that has not yet been finalized and is awaiting further processing; it is the object of visual inspection. Continuous displacement refers to the continuous positional change process of the visual inspection module (a moving platform with a camera and light source) on its motion path as it moves between workstations, used to cover the entire inspection line. The speed curve refers to the trajectory of the speed change over time during the movement of the inspection module, used to determine the motion state and acquisition rhythm. The path division method refers to dividing the entire inspection path into several inspection segments according to the needs of movement and acquisition, providing a spatiotemporal basis for multi-workstation image acquisition. The path boundary time refers to the time point at which each inspection segment ends, defining the separation between segments. The next path start time refers to the start time of the next inspection segment, used to connect with the previous segment. The overlapping interval refers to the overlapping part of two adjacent segments in time or path, used to achieve seamless acquisition connection and splicing. Judging and adjusting the corresponding path nodes refers to dynamically adjusting the position and time of the boundary nodes on the inspection path according to the size and distribution of the overlapping interval, making image acquisition more continuous.

[0024] In S2, the time-period speed change refers to the speed change of the visual detection module within a specific detection period, reflecting the smoothness of motion and switching nodes; the position change refers to the dynamic change of the spatial position of the detection module or blade preform over time, used for positioning and synchronous acquisition; the displacement encoder signal refers to the real-time position signal output by the displacement encoder installed in the workstation slide rail bearing assembly, used to accurately locate the current position of the module; the speed feedback signal refers to the electrical signal of the current motion speed output by the drive motor assembly of the detection module, providing a basis for speed acquisition and motion control; the step-by-step comparison refers to the sequential comparison and contrast of speed and position data according to the acquisition time axis, used for synchronous analysis; the actual time node of motion switching refers to the precise time point when the detection module changes from one motion state (such as moving to a stop) to another state, which is also the key moment for image acquisition switching.

[0025] In S3, multi-wavelength illumination refers to illuminating the blade preform with light sources of different wavelengths (color, energy) to enhance the imaging contrast of different surface structures and contours; contour region refers to the region in each frame of the detection image that corresponds to the actual outline of the blade preform, which is usually the main target area for detection; background region refers to other parts of the detection image that do not belong to the blade preform itself, which can be used as a reference for brightness comparison and data normalization; brightness equalization processing refers to adjusting the brightness distribution differences in the image caused by uneven illumination, making the brightness distribution of the entire image more uniform, which is convenient for subsequent feature analysis; brightness difference refers to the numerical difference in pixel brightness (grayscale) between the contour region and the background region, which is an important basis for judging object and background segmentation and contour feature extraction; brightness difference scale refers to the parameter after uniformly normalizing or standardizing the brightness difference, so that the data between different images or regions are comparable.

[0026] In S4, the brightness change peak band refers to the range where the brightness changes most significantly in a certain band of an image acquired under multi-wavelength illumination, reflecting the structural features of the target; spatial grayscale features refer to the distribution of pixel grayscale in space (two-dimensional image) within the brightness peak region, which is a key indicator for judging features such as contours and defects; vector set refers to multiple vector data of the main structural direction and secondary direction within the region obtained through calculation, used to describe the orientation and feature trend of the region; angle data refers to parameters such as the angle between the main direction vector and the secondary direction vector, reflecting the directionality and structural regularity of the region; the distribution of regional direction offset refers to the overall distribution of the offset of the main structural direction within the region as the spatial position changes, used to judge the consistency of shape and abnormal distribution.

[0027] In S5, target area localization refers to the precise definition of the surface or contour spatial region of the blade preform that needs to be focused on or analyzed during the detection process, providing a basis for subsequent processing; principal structural layer refers to the part of the blade preform surface with a main arrangement direction or specific structural layer, such as surface texture, process lines, etc.; multi-scale morphology processing component refers to the hardware or software module in the vision system that has the ability to analyze multiple sizes and shapes, realizing feature extraction and processing at different scales; directional gradient refers to the rate or magnitude of grayscale change in a certain region of the image in a specific direction, often used for feature discrimination such as contours and edges; amplitude comparison refers to comparing the gradient change intensity of the principal direction and the vertical direction to reveal directional characteristics and local anomalies; critical region of difference refers to the spatial block with the largest change during the gradient comparison process, which usually corresponds to defects, deformations or important detection objects.

[0028] like Figure 4 As shown, the specific steps for obtaining the path overlap distribution characteristics are as follows: S101: Based on the visual inspection module, analyze its continuous displacement and velocity curves on the multi-station prefabricated body inspection path, divide each position point into segments according to the acquisition time sequence, compare the velocity change trend and movement state change within each segment, determine the velocity fluctuation and movement continuity within each movement segment, calculate the time series connection between each segment, and obtain the movement trajectory time series group. Based on a vision inspection module, specifically a device equipped with an industrial camera and multi-band light source capable of moving between multiple workstations, the system first analyzes real-time position and velocity data acquired during its movement along the inspection path of the prefabricated object. Each position point on the movement path is recorded in real-time by an encoder, with data acquired every 10 milliseconds, forming a raw data sequence containing timestamps, position values, and velocity values. This data sequence is sorted along a time axis. Subsequently, the continuous displacement data within the entire inspection cycle is divided into multiple time segments, each approximately one second long. The velocity changes at each position point within each segment are statistically analyzed to determine whether the velocity value remains within a stable range. For example, segments with velocity changes within 0.05 meters per second are considered stable segments, while sudden increases or decreases in velocity are identified as fluctuating segments. Simultaneously, comparisons are made... The start and end times of adjacent segments are determined. If the time interval between two segments is less than 0.1 seconds, it is considered a temporally continuous segment; otherwise, it is identified as a discontinuous segment. Based on this, a data structure is constructed for each segment, which includes the start and end times, start and end positions, average speed, and speed fluctuation degree of the segment. By analyzing the slope change of the speed curve within each segment, it is determined whether the segment is in an acceleration, deceleration, or constant speed state. For example, if the speed in a segment increases from 0.3 meters per second to 0.5 meters per second and maintains an upward trend, it is defined as an acceleration segment. The connection status between segments is further statistically analyzed, and all segment information is combined into a whole motion trajectory data set. This set is rearranged according to the time axis to form a trajectory temporal group, which represents the motion behavior and path nodes of the visual inspection module in the entire multi-station prefabricated body inspection process.

[0029] S102: Based on the time sequence group of the movement trajectory, compare the end time of each segment and the start time of the subsequent segment, analyze the continuity of each group of time points and the spatial connection between path segments, filter out discontinuous nodes, optimize the distribution of the end and start times of path segments, adjust the position order of path nodes, and obtain the path overlap distribution characteristics. Based on the pairing and comparison of the end time of each segment in the aforementioned movement trajectory time series with the start time of the subsequent segment, the start and end time points of each group are extracted sequentially, and the time interval is judged. When the interval between two adjacent segments is less than 0.1 seconds, it is considered that the time connection is good; if it is greater than this value, it is judged that there is a time jump. Further analysis is conducted on the spatial connection relationship between the two segments, comparing whether there is a sudden change between the end position of the previous segment and the start position of the next segment. If the position difference is large and there is a significant directional shift, it is marked as a spatial breakpoint. This time breakpoint and the spatial abrupt change position are recorded together in the breakpoint list. For this type of breakpoint, a path node adjustment operation is performed, for example... If the start time of the next segment is advanced to connect with the previous segment, and the start position of the next segment is adjusted forward to be closer to the end position of the previous segment, in practice, if the end time of the previous segment is 1.5 seconds and the start time of the next segment is 1.8 seconds, then the latter is advanced by 0.2 seconds to reduce the time gap. At the same time, the spatial position is updated according to the midpoint between the two segments to keep the trajectory continuous. After adjustment, all path nodes are rearranged so that the time and spatial order of all segments are logical, thereby generating a new path segment sequence. The overlap duration and position distribution between each segment are extracted and summarized into path overlap distribution characteristics.

[0030] like Figure 5 As shown, the specific steps for acquiring the timing control node are as follows: S201: Based on the overlapping distribution characteristics of the paths, analyze the speed changes and corresponding position changes of each path segment, synchronously compare the speed change curve of the visual detection module during the movement process with the displacement change process, determine the relationship between speed fluctuation and position change in each time period, distinguish the characteristic interval of motion state switching, and obtain the speed-position correlation interval. By analyzing the data structure formed by the temporal overlap, spatial coincidence, and connection sequence of each detection path segment collected by the visual inspection module along the multi-station prefabricated body path, the start and end times of each path segment are read segment by segment. Simultaneously, the velocity change curves and displacement change records within each time period are extracted. The velocity curves are reconstructed along the time axis and matched one-to-one with the displacement change data. The velocity value of each sampling point is paired with its corresponding position value. The change in adjacent velocity values ​​within each segment is compared with the displacement difference of the corresponding segment to identify whether velocity fluctuations are accompanied by significant displacement increases. If the displacement increase exceeds 5 mm while the velocity increases, the segment is considered to be in an accelerating state. If the velocity decreases while the displacement change tends to stabilize, the segment is considered to be in a decelerating or near-stationary state. During the analysis, segments with a velocity change rate of 0.1 m / s or more are marked as key segments. Meanwhile, segments with a displacement change of less than 2 mm are identified as short-term jitters caused by velocity fluctuations. All path segments are traversed sequentially to construct a bivariate change sequence of velocity and position. A velocity-position relationship image is plotted in a visualization coordinate system. Based on the curve trend changes, characteristic segments under different motion states are identified, such as transitions from uniform velocity segments to acceleration segments, and from acceleration segments to short pause segments. Based on the identification of such change segments, the timestamp positions of velocity peak points, velocity inflection points, and points where position changes tend to stabilize are extracted in the time dimension. The segments where the time points are located are associated and classified to form multiple motion segments with clear velocity-position linkage characteristics, which are output as a set of velocity-position associated segments.

[0031] S202: Based on the speed-position correlation interval, filter the displacement encoder signal of the workstation slide rail bearing group, synchronously analyze the position data output by the displacement encoder and the speed feedback signal of the drive motor group of the vision detection module, compare the temporal correspondence between position change and speed change, determine the segment of motion state switching, and obtain the node feature sequence. In each identified linkage segment, the displacement encoder signal collected by the workstation slide rail bearing assembly is extracted segment by segment. The sampling period is set to one data set every 20 milliseconds. The timestamp and position data are read, and the data is truncated according to the start and end times of the speed change interval. Simultaneously, the speed feedback signal within this time period is extracted from the drive motor assembly of the vision inspection module. The speed signal is linear speed information after voltage value calibration and conversion. This speed signal and the encoder displacement signal are time-aligned. The position change and speed feedback value at the same time point are compared for each data set to analyze whether there is time lag or slow position response. If the displacement remains essentially unchanged at the same time point while the speed signal shows significant fluctuations, it is judged as signal mismatch or data drift. This process is then repeated. For each velocity-position associated segment, this type of data comparison processing is performed. For segments with high synchronicity of changes, the displacement increase corresponding to the velocity increase and the position stagnation corresponding to the velocity decrease are recorded as effective motion response segments. Then, based on the nodes in the segment where the velocity changes from rising to falling and the nodes where the velocity continues to decrease until it approaches zero, a comparison is performed item by item. The state switching behavior is marked at the node position, such as the module changing from a fast movement state to a braking state, or from slow forward movement to complete stillness. A feature node sequence is constructed based on the above marking results. Each node includes parameters such as its time point, velocity change direction, whether the displacement increment is less than 1 mm, and the velocity difference before and after it. This sequence is output as the recognition result of motion state switching, and the node feature sequence is obtained.

[0032] S203: Based on the node feature sequence, determine the velocity change trend of each motion segment, locate the feature node that changes from continuous motion to stationary state, compare the velocity decline process between nodes with the displacement termination time, identify the node with the shortest time interval, record the corresponding node as a control point, and obtain the acquisition timing control node. For each node, a velocity change trend analysis is performed, screening the velocity trends within the time window before and after each node. Velocity records are read within 200 milliseconds before and after each node, and the time it takes for the velocity to gradually decline from a positive value to near zero is calculated. It is then determined whether the displacement change synchronously converges to a stable range of less than 2 millimeters during this velocity decline. When the velocity trend transitions from a declining state to a constant zero value and the displacement remains stationary for more than 100 milliseconds, the node is considered a valid stationary node. If there is a segment between adjacent nodes where the velocity continuously decreases by more than 0.15 meters per second, and the final velocity is below 0.05 meters per second, then it is marked as a stationary node. The feature segment that transitions from a moving state to a stationary state is denoted as the segment. The time difference between the end node of the segment and the previous node is compared one by one to identify the node pair with the shortest time interval. In practical applications, if there are multiple candidate segments with time differences of 0.12 seconds, 0.08 seconds and 0.05 seconds, the node pair with a time difference of 0.05 seconds is selected as the final control point. The end node of this segment is used as the acquisition switching point. Its timestamp, position index and corresponding speed value are recorded to generate a corresponding control node set. Each node set can be mapped to the control unit of the acquisition system as the basis for triggering the start and stop of image acquisition, forming the acquisition timing control node.

[0033] like Figure 6 As shown, the specific steps for obtaining the spectral region difference trend are as follows: S301: Based on the acquisition timing control node, analyze the continuously acquired multi-wavelength illumination image sequence, determine the outline region and background region of the blade preform in each frame image, compare the pixel brightness distribution of the outline and background, distinguish the brightness difference of pixels in the region, and obtain the region brightness distribution group. Based on the module motion stop point obtained in the previous stage as the timing trigger reference point for image acquisition, at the corresponding moment of each control node, the acquired multi-wavelength image frame is read, and each pixel in the image frame is partitioned to divide the entire image into a pre-formed contour region and a background region. The blade boundary is identified based on the pixel grayscale value change trend in the contour region and the image edge contour line extraction results. During the identification process, the maximum grayscale gradient value of the region is used as the boundary identification reference, and the pixel column with grayscale change exceeding the set threshold is used as the boundary starting point. The contour region is the closed region within the boundary, and the background region is the region outside the contour. After obtaining the region range, the pixel sets of the contour region and the background region are extracted respectively. For each frame of the image, the average brightness value of the two regions is calculated separately. The distribution frequency of gray values ​​between 0 and 255 in each region is counted, and the range of regional brightness is calculated. Image frames with a range greater than 40 are marked as frames with strong brightness differences. The standard deviation of all pixel values ​​in the region is calculated and the concentration of brightness distribution is recorded. Frames with a standard deviation greater than 20 are selected as frames with balanced distribution differences. At the same time, the image frame number and timestamp are extracted to establish a brightness distribution record table corresponding to the image frame. The average brightness, range, and standard deviation are used as the three key data of the brightness characteristics of the frame contour and background region. The same processing flow is performed on all acquired image frames in the above manner to form a group of regional brightness distributions arranged in the order of the image sequence.

[0034] S302: Based on the regional brightness distribution group, calculate the pixel brightness difference at each wavelength, analyze the change process of brightness difference in the image frame sequence, use brightness equalization processing to adjust the pixel brightness of the image frame, judge the change of brightness distribution before and after equalization processing, unify the brightness change scale, and obtain a multi-band brightness response set. For each image frame, a set of contour and background region brightness parameters is established. In each image frame under each wavelength illumination condition, the corresponding average brightness and brightness range are sequentially read. The difference in pixel grayscale between the two regions under different wavelengths is calculated. All image frames are summarized according to the acquisition order. The changing trend of brightness difference between consecutive image frames is extracted from the sequence. The increase or decrease in brightness difference between adjacent frames is statistically analyzed, and a change curve is plotted for the entire sequence. When the brightness difference fluctuation between adjacent frames is greater than 10 grayscale values, it is determined as a response abrupt change segment. The starting and ending frames are marked, and the band number is recorded. Brightness equalization processing is performed on frames with abrupt brightness differences in the image frame sequence. The grayscale histogram range is unified, the maximum grayscale value is fixed to 240, and the minimum grayscale value is fixed to 10. Linear interpolation is performed on the remaining pixels to complete the pixel brightness adjustment. Before and after the brightness equalization process, the average brightness of the contour and background areas is recalculated. The difference in mean, range, and standard deviation before and after the adjustment are compared. If the range decreases by more than 30 and the standard deviation decreases by more than 15, it is marked as a valid frame for brightness equalization. The difference before and after the processing is recorded for all valid frames, and the brightness change range is unified so that the brightness response scale of all image frames under different band illumination is normalized to the same interval, forming an image frame dataset with normalized response, and outputting a multi-band brightness response set.

[0035] S303: For multi-band brightness response sets, compare the changing trends under each illumination band, analyze the response characteristics of regional brightness differences under each spectral band, classify the brightness contrast relationship in the region of the difference band, determine the brightness change trend of each region in the spectral distribution, and obtain the spectral region difference trend. In image frames with a standardized brightness response scale, brightness difference data is extracted according to different band numbers. The brightness response change curves of adjacent frames are compared band by band to identify peak brightness fluctuation frames in each band. If the brightness difference in a certain band is greater than 50, the band is judged to have prominent difference characteristics, and its band number and frame sequence information are recorded. The gray-level distribution of the contour region and the background region under the difference bands is compared in detail. The gray-level value distribution range, concentration of the main gray-level segment, and peak frequency in the contour region are classified, and the difference between the main gray-level frequency interval and the background gray-level densest interval is calculated. The difference is divided into amplitude ranges: less than 20 is a weak response range, 20 to 40 is a medium response range, and greater than 40 is a strong response range. The results are categorized to form response intensity labels, which are recorded in the image frame set corresponding to the band. The response features of all bands are summarized horizontally, and the band with the strongest response and its corresponding image frame position are marked in the summary. The corresponding spectral range is mapped according to the band number. The brightness data of bands with different response intensity features are combined and output to construct the regional brightness change trend sequence under each spectral band and classify it into spectral regional difference trends.

[0036] like Figure 7As shown, the specific steps for obtaining the spatial parameters of the directional distribution are as follows: S401: Based on the trend of spectral region differences, analyze the brightness changes of each spectral band, screen the band regions with brightness changes, determine the spatial gray-scale distribution characteristics of each region, identify the main direction and the secondary direction, calculate the spatial arrangement relationship of the main and secondary directions of pixel distribution, and summarize the spatial distribution of each direction to obtain the set of main and secondary direction vectors. Based on the spectral difference data formed by the brightness difference and response amplitude of each band constructed in the previous stage, the band images with significant brightness changes are extracted from the image frame sequence as the analysis targets. First, the entire image region is traversed in each band image, and the region with a brightness difference greater than 40 is marked as the change region. Then, a two-dimensional pixel matrix is ​​established within each change region, and the gray-level change amplitude in each row and each column is extracted. The frequency of increase and decrease of gray-level values ​​with the horizontal and vertical directions is counted. The direction with the most concentrated continuous gray-level changes is used as the main direction for determination, and the other direction with a smaller change amplitude is recorded as the secondary direction. In an image, if the concentrated change amplitude in the main direction exceeds 60 gray-level units, and the change in the secondary direction is less than 30 gray-level units, then the analysis is performed accordingly. Once the unit is identified, it can be determined as a clearly directional region. Then, based on the positioning of the main and secondary directions, the coordinate offset of each pixel relative to the image center in the main and secondary directions is extracted. The overall distribution of the offset values ​​is statistically analyzed, and the distribution map of gray-level gradients in each direction is plotted in the spatial coordinate system. It is determined whether the pixel gradients are concentrated along the main direction. If the gray-level increase or decrease in the main direction shows a linear trend, the gray-level gradient set in that direction is recorded. At the same time, the gray-level change data in the secondary direction is statistically analyzed using an equal-length sliding window. The spatial arrangement position of all pixels in the main and secondary directions is recorded as a direction vector group. Each group of data includes the starting point, direction angle, number of pixels, and gray-level distribution amplitude of that direction. The main and secondary direction vector sets are generated by analyzing each image and summarizing each band.

[0037] S402: Compare the angle data of the main and secondary directions in the set of main and secondary direction vectors, analyze the offset characteristics of the main direction in each region as the spatial position changes, determine the spatial variation law of the distribution of each main direction, and statistically analyze the spatial distribution parameters of the main direction in the region to obtain the spatial parameters of the direction distribution.

[0038] First, extract the coordinates of the starting and ending points representing the main and secondary directions from each vector pair. Convert the coordinate pairs into direction angle information. Angle calculation uses the horizontal line as a reference in the image coordinate system. Set the main direction angle as the reference value, and sequentially read the angle values ​​of each secondary direction vector and subtract them from the main direction angle value to obtain the included angle between each pair of main and secondary directions. Then, classify them according to the region number in each image. In each image frame, calculate the mean, variance, and maximum value of all included angles. If the angle change exceeds 30 degrees and the variance is greater than 10 degrees, the main direction of that region is judged to have a shift trend. Subsequently, establish a region location index for regions with significant main direction shifts in each image frame. The spatial coordinates of the region in the image are extracted, and the offset angle and spatial distribution coordinates corresponding to the main direction are recorded. After summarizing the offsets of all regions, the data are divided into regions according to the spatial distribution direction. For example, the blade preform image is divided into three segments: head, middle and tail. The average offset angle of the main direction in each segment is compared. If the average offset angle of the tail region is more than 15 degrees greater than that of the middle region, it is determined that there is an directional anomaly in the tail region. The spatial distribution parameters of the main direction of the region are generated by combining the average, variance, maximum offset value and minimum offset value of the statistical angles of each region. This parameter set includes the directional angle range, offset amplitude distribution and relative spatial position of each image partition, and the directional distribution spatial parameters are output.

[0039] like Figure 8 As shown, the specific steps for obtaining multi-scale texture gradient features are as follows: S501: Based on the spatial parameters of directional distribution, analyze the location and directional characteristics of the target area, determine the spatial distribution law of the main directional structural layer on the surface of the precast body, optimize the boundary and shape of the target area, adjust the regional positioning order in combination with directional characteristics, determine the coverage of the main directional structural layer, and obtain the spatial distribution of the structural layer. First, the entire image is divided into regular spatial grid regions. The angular statistics of the principal direction within each grid are read. The principal direction angles of all grids are screened, and adjacent grids with a principal direction difference of no more than 10 degrees are classified as the same structural block. Then, the distribution center point and edge extension range of each structural block are further analyzed, and the boundary positions of each structural block in image space are marked. Region correction operations are performed on the contours of irregular boundaries or areas overlapping with adjacent blocks. For example, parts with overly tortuous boundaries or abrupt angle changes are smoothed and redefined. Simultaneously, direction change buffers are added between different blocks to avoid judgment interference. Based on this, the actual distribution pattern of the principal direction structural layer on the prefabricated surface is identified, and the continuous spatial direction is... Consistent structural blocks are merged into a group, and their spatial coverage areas are analyzed to project onto the image, along with their position coordinates. The starting point, ending point, center point, and orientation values ​​are recorded. Based on the position index and orientation characteristics of each structural layer region, a priority analysis order is established, prioritizing regions with stronger spatial continuity and higher orientation consistency for focused localization. At the same time, the original image region localization order is adjusted, for example, regions near the upper left of the image but with larger orientation deviations are moved to the back of the order to obtain a clearer concentrated representation of the main structural layer blocks. Finally, the starting and ending positions, center coordinates, main orientation angle, and coverage area of ​​each structural layer are output in the form of vector region combinations as spatial distribution quantities of the structural layers.

[0040] S502: Based on the spatial distribution of the structural layer, calculate the grayscale gradient of the multi-scale morphology processing component of the module camera in the main direction and vertical direction within each window, analyze the gradient magnitude in each direction, determine the relationship between the main direction and the vertical direction, and obtain the directional gradient distribution. Based on the spatial distribution of the structural layers, i.e., the pre-defined and labeled principal structural layer regions, image block processing is performed within each structural region using a rectangular sliding window. Each window is set to three different sizes: 16×16 pixels, 32×32 pixels, and 64×64 pixels, sequentially covering the image content of that structural region. Within each window, all pixel grayscale values ​​are read, and the grayscale gradients in the principal and vertical directions are extracted. The principal direction grayscale gradient is obtained by summing the grayscale differences between adjacent pixels along the principal direction axis, while the vertical direction grayscale gradient is processed in the same way in the direction 90 degrees from the principal direction. During execution, the gradient results in each direction are normalized to 0. The values ​​are between 255 to facilitate comparison between multiple windows. The total gradient magnitude in the main direction and vertical direction within each window is counted, and the difference between the two directions is calculated. If the total gradient magnitude in the main direction is greater than 200 and the gradient magnitude in the vertical direction is less than 100, the window is determined to be a directionally significant region. If the gradient difference between the two directions is between 50 and 100, it is determined to be a region biased towards the main direction. Based on this, the number of windows occupied by the directionally significant regions in the entire image and their spatial distribution are counted. The gradient values, direction differences, main direction identifiers, and spatial coordinates of all windows are recorded to form a window gradient distribution dataset. The dataset is summarized using the image index as the index number and the directional gradient distribution is output.

[0041] S503: Filter spatial regions with prominent amplitude changes in directional gradient distribution, statistically analyze the gradient difference distribution in the main direction and vertical direction within each structural window, analyze the gradient difference in local regions, and summarize the gradient change law under multi-scale windows to obtain multi-scale texture gradient features. In the directional gradient distribution data of each scale window, the gradient difference between the principal direction and the vertical direction is extracted. When the difference is greater than 80, it is marked as a window with a prominent change. The spatial coordinates of the window are extracted and marked in the image to form a preliminary anomaly region map. Then, the original pixel blocks of each anomaly region window are read again, and the difference in grayscale mean of each row and column is calculated. A difference matrix is ​​built according to the row and column directions of the gradient statistics results in each direction. Points in the difference matrix that are greater than twice the average value are marked, and their continuous length in the principal and secondary directions is counted. When the continuous difference change length in the principal direction exceeds 10 pixels and in the vertical direction is less than 5 pixels, the window is classified as a principal direction abrupt change region. Significant changes in both directions are categorized into bidirectional change regions. The number and spatial distribution of each type of window are summarized, and the total number and distribution density of different types of windows in the image are statistically analyzed. Then, the distribution ratio of window types is statistically analyzed at three different sizes: 16×16, 32×32, and 64×64. When the frequency of the main direction change region is higher in small windows and lower in large windows, it is determined that the local texture difference is relatively concentrated. Conversely, it is determined that the overall texture change of the structural layer is uneven. Finally, the statistical results of all windows at multiple scales are classified, and the proportion of main direction change, the average value of bidirectional difference amplitude, and the spatial density of texture regions at each scale are extracted and combined to form a complete multi-scale texture gradient feature.

[0042] Another specific embodiment of this application provides a multi-station carbon fiber preform visual inspection system for implementing the multi-station carbon fiber preform visual inspection method described above. (Refer to...) Figure 9 It includes a path analysis module, a timing control module, a first detection module, a second detection module, and a third detection module. The path analysis module is used to analyze the displacement and velocity changes in the detection path and generate path overlap distribution characteristics; the timing control module is used to determine the acquisition timing control nodes for each station; the first detection module is used to realize the detection of the carbon fiber preform trace line and warp and weft density; the second detection module is used to realize the detection of the carbon fiber preform yarn reduction arc; and the third detection module is used to realize the measurement of the height information of the carbon fiber preform.

[0043] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-station visual inspection device for carbon fiber preforms, characterized in that, It includes a testing platform (1), a mobile module (2), a base plate (3), a carbon fiber preform pallet (4), a carbon fiber preform (5), a first workstation (6), a second workstation (7), a third workstation (8), and a control cabinet (14). The first station (6), the second station (7) and the third station (8) are set sequentially along the detection path on the detection platform (1). The moving module (2) is fixedly installed on the detection platform (1) and is used to drive the base plate (3) to move along the detection path. The base plate (3) is installed on the moving module (2). The carbon fiber preform support plate (4) is fixedly installed on the base plate (3). The carbon fiber preform (5) is placed on the carbon fiber preform support plate (4) and moves synchronously with the base plate. The first workstation (6) is equipped with a flat light source (9), a first 2D camera (101) and a first polarizing mirror (131) for imaging and detecting the trace lines and warp and weft density of the carbon fiber preform; The second workstation (7) is equipped with a strip light source (11), a second 2D camera (102) and a second polarizing mirror (132) for imaging and detecting the yarn reduction arc of the carbon fiber preform; The third workstation (8) is equipped with a 3D camera (12) and a third polarizing mirror (133) for measuring the height information of the carbon fiber preform; The control cabinet (14) is electrically connected to the mobile module (2), each station camera and light source, and is used to realize multi-station detection timing control and data acquisition.

2. The multi-station carbon fiber preform visual inspection device according to claim 1, characterized in that, The two flat light sources (9) in the first station (6) are arranged symmetrically, and the angle between the flat light source (61), the flat light source (62) and the base plate (3) is 45°. The two flat light sources are 500mm apart, and the distance from the flat light source to the base plate is 400mm. The first 2D camera (101) in the first station (6) is arranged symmetrically in the center, and the first polarizing mirror (131) is fixedly installed below the imaging optical path of the first 2D camera.

3. The multi-station carbon fiber preform visual inspection device according to claim 1, characterized in that, In the second work station (7), the strip light sources (11) are symmetrically distributed on both sides of the carbon fiber preform. The distance between the strip light sources (71) and (73) and the base plate (3) is 450mm, and the distance between the strip light sources (72) and (74) and the base plate is 200mm, so as to form multi-angle cross lighting conditions. The second 2D camera (102) in the second station (7) is located above the detection area, and the second polarizing mirror (132) is set in the imaging optical path of the second 2D camera to suppress surface reflection interference.

4. The automated visual inspection equipment for multi-station carbon fiber preforms according to claim 1, characterized in that, The 3D camera (12) in the third station (8) is fixedly installed at the center position directly above the third station, and the third polarizing mirror (133) is set in the imaging optical path of the 3D camera to reduce the influence of surface reflection on height measurement.

5. A visual inspection method for multi-station carbon fiber preforms based on the equipment described in any one of claims 1-4, characterized in that, Includes the following steps: S1: Multi-station detection path and acquisition timing establishment steps: Based on the vision detection module, the displacement and velocity information of the moving module in the multi-station detection path are acquired, the relationship between continuous displacement and velocity changes is analyzed, the detection path is segmented, the overlap of adjacent segments in time and space is compared, the order of detection nodes is adjusted, and the path overlap distribution characteristics are obtained. S2: Steps for determining the timing control node: Based on the path overlap distribution characteristics, combined with the displacement encoder signal of the workstation slide rail bearing group and the speed feedback signal of the drive motor, locate the time node when the motion state changes from moving to stationary, and determine the timing control node corresponding to each workstation. S3: First station trace line and warp and weft density detection steps: Under the acquisition timing control node, the flat panel light source of the first station and the first 2D camera are used to acquire the carbon fiber preform image, the trace line area and warp and weft structure in the image are subjected to brightness equalization and contrast enhancement processing, the trace line length and spacing are extracted, and the warp and weft density parameters are calculated. S4: Second station yarn reduction arc detection step: Under the acquisition timing control node, the second station uses a strip light source and a second 2D camera to acquire multi-directional lighting images. Through brightness equalization and regional comparison analysis, the brightness change characteristics of the yarn reduction arc area are highlighted. Combined with the regional spatial grayscale distribution, the position and shape of the yarn reduction arc are identified. S5: Third station height measurement step: Under the acquisition timing control node, the 3D camera of the third station is used to perform three-dimensional imaging of the carbon fiber preform, establish the cutting plane and extract the corresponding spatial contour curve, and calculate the height information of the carbon fiber preform through the plane reference.

6. The automated visual inspection method for multi-station carbon fiber preforms according to claim 5, characterized in that, The path overlap distribution features include the segment overlap ratio, the temporal distribution of path nodes, and the detection segment switching identifier. The acquisition timing control node includes a synchronous acquisition trigger marker, a motion control reference, and a time index.

7. The automated visual inspection method for multi-station carbon fiber preforms according to claim 5, characterized in that, In the yarn reduction arc detection step, regional brightness statistics and comparative analysis are performed on the multi-directional illumination image to screen out regions with significant brightness changes, and the continuity and direction characteristics of the yarn reduction arc are determined based on the spatial gray-scale direction distribution of the region.

8. The multi-station carbon fiber preform visual inspection method according to claim 5, characterized in that, In the height measurement step, multiple cutting planes are established on the surface of the carbon fiber preform, the corresponding spatial contour curves are extracted, and the height distribution at each position is calculated based on the reference plane.

9. A multi-station carbon fiber preform visual inspection system, used to implement the multi-station carbon fiber preform visual inspection method according to any one of claims 5-8, characterized in that, It includes a path analysis module, a timing control module, a first detection module, a second detection module, and a third detection module. The path analysis module is used to analyze the displacement and velocity changes in the detection path and generate path overlap distribution characteristics; the timing control module is used to determine the acquisition timing control nodes for each station; the first detection module is used to realize the detection of the carbon fiber preform trace line and warp and weft density; the second detection module is used to realize the detection of the carbon fiber preform yarn reduction arc; and the third detection module is used to realize the measurement of the height information of the carbon fiber preform.

Images