Three-point welding dynamic imaging stabilization detection method based on motion compensation modeling

Abstract

The application discloses a three-point welding dynamic imaging stabilization detection method based on motion compensation modeling, relates to the technical field of photovoltaic energy monitoring, and comprises the following steps: acquiring sound and vibration signals, picture time marker signals and jig lifting displacement signals in the whole process of compressor shell jacking, performing time synchronization and feature fusion on the acquired signals, and generating a resonance distribution band for constructing a dynamic reference model for subsequent time sequence analysis. Through multi-source signal synchronous modeling and light machine closed-loop regulation, the jacking action and the imaging process are consistent at the millisecond level. Through the accurate control of exposure and vibration phase by using the staggered anchor point, the drift sketch, the baseline ring and the alignment key, the pixel drift and the brightness fluctuation are significantly reduced, the false alarm rate and the missed detection rate are reduced by 73% and 67% respectively, and the high precision and high stability of the three-point welding dynamic detection are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic energy monitoring, in particular to a three-point welding dynamic imaging stabilization detection method based on motion compensation modeling. BACKGROUND

[0002] The three-point welding dynamic imaging stabilization detection based on motion compensation modeling refers to that, in the process of compressor shell carrying, jacking and rotating, first, a motion model combined with geometry and optics is used to describe the workpiece posture jitter and height drift (for example, about 25mm stroke change and about 3.5 degree angle change caused by jacking) in real time, the two-dimensional pixel displacement and three-dimensional point cloud pose deviation are solved by the robust features (hole edge contour, welding bump curvature, jig reference target) in the welding point neighborhood and the optical flow field, and then the dithering registration and perspective normalization are performed on each frame of image and point cloud based on the above displacement and deviation. At the same time, the exposure and high reflection suppression photometric normalization are performed combined with the illumination matrix (mainly ring illumination, stripe side light and polarization light switching), so that the welding point area maintains constant field of view, scale and gray distribution in continuous imaging. Finally, the two-dimensional and three-dimensional fusion measurement and defect identification (hole diameter, inner recess depth, welding slag expansion, melt-through topography, etc.) are performed on the stabilized three-point welding area, and the threshold is dynamically normalized according to the posture and incident angle, so as to realize the detection of three-point welding with high consistency, high robustness and low false alarm and missing alarm under online working condition.

[0003] The prior art has the following disadvantages:

[0004] In the prior art, three-point welding online detection usually relies on a jacking mechanism to lift the workpiece to the imaging height, and then a vision system performs synchronous imaging and motion compensation. However, when the mechanical frequency of the air cylinder jacking and the inherent resonance frequency point of the machine structure are close, micro-amplitude resonance vibration will be caused, which makes the imaging device and the jig vibrate slightly. Such mechanical vibration is difficult to be identified and corrected in real time in the time synchronization mechanism of the prior art, resulting in sub-millisecond drift of the baseline of the calculation time of the image acquisition frame trigger signal and the motion compensation model. When the drift is accumulated for multiple frames, the imaging timing of the three-point welding area is slightly misaligned, causing the images of each welding point to be collected at different poses. Such misaligned signal is often mistaken for a height difference or abnormality of the welding point in the prior art, which eventually causes mutual interference between the detection results of the three-point welding, resulting in feature mismatch, depth fitting distortion, false alarm and missing detection, etc. serious consequences, which seriously affects the stability and reliability of the detection system.

[0005] The above information disclosed in the background section is only used to enhance the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] The application aims to provide a three-point welding dynamic imaging stabilization detection method based on motion compensation modeling to solve the problems in the background art.

[0007] To achieve the above-mentioned purpose, the application provides the following technical solution: a three-point welding dynamic imaging stabilization detection method based on motion compensation modeling, comprising the following steps:

[0008] Obtain the sound and vibration signals, picture time marker signals and jig lifting displacement signals in the whole process of lifting the compressor shell, time synchronize and feature fuse the obtained signals to generate a resonance distribution band, which is used to construct a dynamic reference model for subsequent time sequence analysis;

[0009] Based on the resonance distribution band, continuously track the highlight drift points in the edge area of the imaging picture, extract the starting position of the time jitter, establish a time anchor point column, and provide a time index for the workpiece pose analysis in the lifting process;

[0010] According to the time anchor point column, perform segmented analysis along the lifting displacement direction to obtain the pose deformation trend, draw a pose drift chain, and form a complete time sequence drift sketch for describing the diffusion characteristics of the equipment vibration in the time domain;

[0011] According to the time sequence drift sketch, establish a stable baseline ring outside the three-point welding area, extract a synchronous stable section from the baseline ring to generate an alignment key, and use the alignment key to establish a time unification reference for multi-point image acquisition;

[0012] Synchronize the light source incident angle and camera trigger interval by using the alignment key, perform multi-point picture on-screen dynamic fusion through light-mechanical linkage, make the image acquisition process and the motion compensation modeling process consistent in time, and form a unified time domain synchronization window;

[0013] Control in the synchronization window, adjust the pose period by using a time slot rudder, and weaken the resonance peak energy by using a phase buffer curtain, so as to establish a closed-loop stable synchronization relationship between the lifting action and the imaging process, thereby completing the dynamic imaging stabilization detection of the three-point welding area.

[0014] Preferably, the resonance distribution band generation step is as follows:

[0015] During the lifting of the compressor shell, the acoustic sensor, acceleration sensor and grating ruler are arranged to synchronously collect the whole process signals, and a unified time reference device is used to realize time synchronization;

[0016] Register the sound and vibration signals, picture time marker signals and displacement signals according to the time sequence, establish a unified time axis with exposure time as the reference, realize the time alignment of multiple source signals, and realize the time alignment of multiple source signals;

[0017] The key feature parameters of vibration energy, sound pressure energy and displacement rate are extracted based on the synchronization signal to identify the time period of resonance occurrence and energy peak value;

[0018] The time period determined as resonance is fused in time sequence to form a resonance distribution band, and a dynamic reference model is generated for subsequent time sequence analysis.

[0019] Preferably, the time anchor point column establishment process is as follows:

[0020] With the resonance distribution band as a time reference, time correlation and section division are performed on all image frames to determine the resonance stage of each image frame;

[0021] Edge detection regions are divided in the time-calibrated key frames, and the gray scale and coordinate changes of highlighted pixel points in each edge band are continuously tracked and recorded;

[0022] The brightness and position changes of the highlighted region in the continuous frames are compared, and when the drift amount or gray scale change exceeds the threshold, the jitter starting position is determined and a jitter detection point is established;

[0023] The jitter detection points are fused in time sequence to form a time anchor point column, which is combined with the jacking displacement data and the resonance distribution band to generate a time index chain, providing a time reference for pose analysis.

[0024] Preferably, the time sequence drift sketch generation steps are as follows:

[0025] The time and space correlation is established based on the time anchor point column and the jig displacement data to form a time and displacement mapping table;

[0026] The travel is segmented and analyzed along the jacking displacement direction, and the anchor point distribution density, drift direction and drift amplitude are used as segmentation feature parameters to identify the attitude change law of different sections;

[0027] Based on the segmentation analysis results, the drift direction and displacement are connected in time sequence on the time axis to construct a pose drift chain representing the attitude change trajectory;

[0028] The drift chains of each displacement section are spliced in time sequence to draw a time sequence drift sketch to describe the diffusion characteristics of device vibration in the time dimension.

[0029] Preferably, the drift direction and displacement are continuously connected to form an attitude trajectory polyline with the jig lifting direction as the vertical axis and the image drift direction as the horizontal axis, and the curvature change of the polyline reflects the synchronization relationship between attitude change and resonance energy release in the jacking process, which is used to represent the dynamic evolution characteristics of vibration deformation.

[0030] Preferably, the alignment key generation steps are as follows:

[0031] According to the time sequence drift sketch, a three-point welding detection area is positioned in space and partitioned for stability, a region with vibration energy lower than the average of total energy is identified and a stable baseline ring is formed;

[0032] Feature parameters of a pixel region in the stable baseline ring are extracted, and imaging stability of the baseline ring is judged by brightness standard deviation, gray scale change rate and pixel offset average;

[0033] A time period that meets the brightness stability and displacement stability at the same time is extracted as a synchronous stability section by longitudinal analysis of the change of the stable baseline ring over time;

[0034] In the synchronous stability section, the brightness average, the gray scale distribution barycenter coordinates and the pixel displacement median are calculated to generate an alignment key to establish a time unified reference for multi-point image acquisition.

[0035] Preferably, the synchronous window generation step is as follows:

[0036] The alignment key is used as a time reference to time-synchronize the light source illumination unit and the camera trigger device, so that the light source trigger signal and the camera exposure signal correspond on a unified time axis;

[0037] According to the gray scale distribution barycenter parameter in the alignment key, dynamic adjustment of the light source incident angle is performed to maintain a stable corresponding relationship between the illumination direction and the workpiece posture;

[0038] On the basis of synchronization of the light source incident angle, the camera trigger interval is precisely controlled to make the exposure action completely overlap the optical steady-state window;

[0039] Through light-mechanical linkage, multi-point screen fusion is realized in a unified time window to generate time-consistent image input;

[0040] The time mapping relationship of the light source angle, the camera exposure and the jig displacement is established based on the synchronous window to form a unified time domain synchronization table for motion compensation modeling.

[0041] Preferably, in the synchronous window, the time slot rudder adjustment is used to adjust the posture period, and the phase buffer curtain is used to weaken the resonance peak energy to establish a closed-loop synchronization relationship between the jacking action and the imaging process, and the steps are as follows:

[0042] In the unified time domain synchronization window, the time baseline of light-mechanical action is established to accurately map the mechanical jacking displacement curve and the camera exposure time axis to form a phase difference correction curve;

[0043] According to the phase difference result, the time slot rudder mechanism is used to fine-tune the jacking action period to make the periodic peak value of the mechanical posture change overlap the center of the exposure window;

[0044] After the time overlap is realized in the jacking cycle, the phase buffer curtain is used to cut the peak of the cylinder driving force output process to reduce the resonance energy concentration;

[0045] By establishing an optical and mechanical action closed-loop feedback path between the time slot rudder and the phase buffer curtain, real-time regulation and control within the synchronization window is realized.

[0046] After the closed-loop synchronous control is stable, dynamic imaging stabilization detection of the three-point welding area is performed.

[0047] In the above technical solution, the technical effects and advantages provided by the present application are as follows:

[0048] The present application establishes a real-time dynamic reference model through time synchronization of acoustic vibration signals, picture time markers and jig displacement; then, through the time error anchor point column and the time sequence drift sketch, the time domain diffusion law of mechanical vibration is quantitatively described; on this basis, the stable baseline ring and the alignment key realize the unified time reference of multi-point images, so that the light source incidence angle and the camera trigger interval remain consistent within milliseconds; finally, the time slot rudder and the phase buffer curtain complete the closed-loop coordination of jacking action and exposure imaging within the synchronization window, which weakens the influence of resonance peaks from the time and energy dimensions. According to actual tests, the present application reduces the time difference between exposure and jacking displacement peak time from 0.83 milliseconds to 0.04 milliseconds, reduces the average pixel drift from 0.46 pixels to 0.18 pixels, converges the brightness fluctuation from 3.9 gray levels to 1.0 gray levels, reduces the false positive rate and the false negative rate by 73% and 67% respectively, and improves the measurement repeatability of hole diameter and welding point depth by more than 50%. The present application fundamentally solves the problem of inconsistency between mechanical vibration and imaging time sequence in three-point welding online detection, realizes high consistency, high precision and high robustness detection under dynamic working conditions, and provides quantifiable and reproducible imaging stabilization support for real-time intelligent detection of compressor shell welding quality. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0050] Figure 1 The method flowchart of the three-point welding dynamic imaging stabilization detection method based on motion compensation modeling of the present application. DETAILED DESCRIPTION

[0051] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations, however, can be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the concept of example implementations to those skilled in the art.

[0052] The present application provides a three-point welding dynamic imaging stabilization detection method based on motion compensation modeling as shown in the following steps: Figure 1 The three-point welding dynamic imaging stabilization detection method based on motion compensation modeling as shown in the following steps:

[0053] The sound and vibration signals, picture time marker signals, and jig lifting displacement signals of the whole process of the compressor shell jacking are acquired, the acquired signals are time-synchronized and feature-fused to generate a resonance distribution band, which is used to construct a dynamic reference model for subsequent time sequence analysis;

[0054] To accurately identify the resonance behavior generated in the jacking process of the compressor shell, a dynamic reference model for subsequent time calibration and posture compensation is constructed, and a technical route of multi-source signal acquisition and fusion modeling is adopted to synchronously acquire, register, extract features, and model the distribution of sound and vibration signals, picture time marker signals, and jig lifting displacement signals. The specific implementation steps are as follows:

[0055] During the jacking process of the compressor shell, acoustic, vibration, and displacement acquisition devices are arranged to acquire multi-source signals throughout the process. The sound and vibration signals are acquired by fixedly installing a piezoelectric acceleration sensor with a measurement range of ±5g and a capacitive sound pressure sensor with a sensitivity of 50mV / Pa above the jacking drive cylinder. The acceleration signal sampling frequency is set to 20000Hz, and the sound pressure signal sampling frequency is set to 10000Hz, so as to simultaneously cover the low-frequency resonance and aerodynamic noise characteristics of the mechanical structure. The picture time marker signal is generated by the trigger interface of the industrial camera, the frame rate of the industrial camera is set to 120 frames per second, and the exposure start time is output to the external timer with microsecond precision. The jig lifting displacement signal is measured by a grating ruler fixed on the jacking guide rail, with a resolution of 0.01mm, to record the real-time change of the jacking displacement throughout the stroke. To reduce interference, during the acquisition process, the acoustic sensor is fixed with double-layer rubber shock pads, the acceleration sensor is connected to the cylinder end face by screw fastening, and the grating ruler is synchronized with the jig through a magnetic base to ensure the mechanical consistency of the sampling synchronization. During acquisition, the sound and vibration signals, image trigger signals, and displacement signals are all connected to a unified high-precision time reference device through a time synchronization interface, with a time accuracy controlled within 10 microseconds, thereby ensuring that the time baselines of all signals are completely consistent.

[0056] The acquired multi-source signals were registered and synchronized in chronological order. Due to differences in sampling frequencies among the various signals, a unified timeline was first established using the exposure time of the industrial camera as the central reference. The timestamps of the acoustic and displacement signals were linearly interpolated to correspond to the camera exposure time, ensuring that the exposure time of each frame could be matched with a unique acoustic amplitude and displacement increment. For example, during the lifting process, when the exposure time of the 40th frame of the camera was 248 milliseconds, the corresponding peak acceleration of the acoustic signal in the 245-250 millisecond range was 0.82g, the sound pressure signal amplitude was 89.3dB, and the jig displacement change relative to the previous frame was 0.36 mm. By time-registering all frames, a synchronization timetable was formed, consisting of frame number, acoustic amplitude, sound pressure intensity, and displacement increment. This timetable reflects the vibration intensity and jig motion state corresponding to each frame, ensuring strict temporal consistency of the acoustic, mechanical, and optical information at each moment in subsequent analysis.

[0057] After obtaining the time-aligned synchronization signal data, key physical quantities characterizing the resonance features were extracted. Energy statistical analysis was performed on the acoustic vibration signal throughout the entire lifting period, calculating the root mean square value of acceleration in every 50-millisecond interval. The average acoustic energy of the sound pressure signal was calculated for the same intervals, resulting in vibration energy curves and sound pressure energy curves varying with time. Simultaneously, the displacement rate between adjacent frames was calculated from the fixture displacement signal, and a displacement response curve was plotted in millimeters per second. By superimposing the three curves, the correlation between vibration energy and displacement response can be visually observed. When the lifting displacement reaches approximately 22 mm, the acoustic vibration curve shows its first main peak, with a peak energy of 0.87 g² / Hz, corresponding to a sound pressure of 91.6 dB, and a displacement rate of 3.2 mm / s. This location can be considered the first structural resonance point. When the displacement reaches approximately 47 mm, a second main peak appears, with a peak energy of 0.94 g² / Hz, a sound pressure of 93.8 dB, and a displacement rate of 3.4 mm / s, indicating the occurrence of the second resonance. By comparing the ratio of acoustic vibration energy to displacement response, the resonance response ratio is extracted as a quantitative indicator. When this ratio exceeds 2.5, it is determined to be a resonance period. The above process not only identifies the specific location and intensity of resonance but also reveals the coupling relationship between vibration energy and displacement velocity at different stages.

[0058] All time periods identified as resonance were merged chronologically to generate a complete resonance distribution band, and a dynamic reference model was established for subsequent time-series analysis. The specific steps were as follows: First, the start time, end time, duration, and peak intensity of each resonance segment were recorded. Then, adjacent resonance segments were merged based on a time interval of less than 0.05 seconds to obtain a continuous resonance time band. Taking a compressor housing lifting time of 3.0 seconds as an example, the analysis results showed that the resonance distribution band was concentrated in three segments: 0.8 to 1.1 seconds, 1.9 to 2.3 seconds, and 2.7 to 2.9 seconds. The peak energies at these three locations were 2.6 times, 3.2 times, and 2.8 times the reference energy, respectively. Corresponding this distribution band to the frame number of an industrial camera, it was determined that frame 96 was in the second main peak segment, and frame 138 was at the end of the third main peak segment. Further matching the time coordinates in the resonance distribution band with the fixture displacement signal yielded a mapping curve between lifting height and resonance intensity, allowing for correction using time and displacement as dual parameters in subsequent motion compensation. This dynamic reference model not only reveals the temporal variation of resonant energy during the jacking process, but also provides a time scale for subsequent staggered anchor point identification, enabling imaging stabilization analysis to be carried out in the temporal context of physical vibration.

[0059] Based on the resonance distribution zone, the bright drift points are continuously tracked in the edge area of ​​the imaging screen to extract the starting position of the time misalignment and establish a time misalignment anchor point column to provide a time index for workpiece pose analysis during the lifting process.

[0060] To accurately identify time shifts caused by resonance and establish a traceable temporal index during the compressor housing lifting process, based on the constructed resonance distribution band, continuous tracking and temporal anchoring are performed using the high-brightness drift features at the edges of the imaging image. This forms a series of time-shifted anchor points for pose analysis, providing precise temporal basis for subsequent motion compensation and attitude correction. The specific steps are as follows:

[0061] Using the obtained resonance distribution band as a time reference, all acquired image frames were temporally correlated and segmented during the lifting process. Specifically, the camera exposure time was precisely matched with the time axis of the resonance distribution band to determine the resonance stage of each image frame. Taking a total lifting time of 3.2 seconds as an example, the resonance distribution band showed obvious resonance peaks at 0.8 to 1.1 seconds, 1.9 to 2.3 seconds, and 2.8 to 3.0 seconds, corresponding to three energy concentration intervals of mechanical vibration, respectively. After time mapping, frames 96 to 130 corresponded to the first resonance segment, frames 195 to 230 to the second resonance segment, and frames 285 to 305 to the third resonance segment. Subsequently, representative keyframes were selected from each resonance segment for subsequent optical drift analysis. This process ensures that each detected image frame corresponds temporally to a specific acoustic energy peak, enabling subsequent drift detection to be based on the actual vibration state, rather than performing image difference comparison without a physical reference, thus avoiding detection errors caused by misalignment due to time drift in existing technologies.

[0062] Image region segmentation and edge feature extraction were performed on time-calibrated keyframes. During the lifting process of the compressor housing, the imaging image includes the upper surface of the housing, the weld joint neighborhood, the fixture structure, and background lighting. The edge regions, being located at the boundary of the field of view, are most sensitive to slight posture fluctuations; therefore, they were selected as drift detection areas. Specifically, each frame was divided into four edge bands: top, bottom, left, and right, with each band set to 5% of the image height or width. Within each edge band, pixel grayscale values ​​were scanned line by line, and consecutive pixels with grayscale values ​​greater than 240 were selected as highlight areas. Taking the 210th frame as an example, a highlight band with grayscale values ​​ranging from 252 to 255 and a length of 18 pixels appeared at the upper edge of the image. In the adjacent 211th frame, this highlight band shifted 3 pixels to the right, increasing its length to 22 pixels. Similarly, a region with a brightness peak of 254 was detected at the left edge and shifted downwards by 2 pixels in the adjacent frame. By recording the coordinates of the highlight distribution in the four edge bands of each frame, the spatial distribution matrix of the highlight areas can be obtained. This matrix describes the brightness changes and positional drift at the edges of each frame's image, providing spatial image information for subsequent temporal misalignment identification. Unlike traditional methods that only perform optical flow tracking in the center region of the solder joint, this implementation uses brightness changes at the image edges as the detection basis, resulting in higher drift sensitivity and stability.

[0063] A temporal series comparison of the spatial distribution matrix of the highlighted areas at the edges of each frame is performed to identify the starting time of brightness changes and determine the location of the time shift. The drift amounts Δx and Δy are calculated by comparing the changes in the center coordinates of the same area in adjacent frames. A frame is marked as the starting frame of the time shift when the drift amount in any direction exceeds 2 pixels or the average brightness change exceeds 10 gray levels. Taking the second resonance segment as an example, from frame 198 to frame 206, the average edge brightness gradually increases from 243 to 254. Frame 204 shows a brightness change of 11 gray levels, and the center coordinates of the upper edge brightness band shift 3 pixels to the right. Therefore, frame 204 is determined as the starting position of the time shift. Subsequently, the drift threshold is re-detected every 5 frames, forming multiple time shift detection points. A total of 4 time shift starting frames are detected within this resonance interval: frames 204, 212, 219, and 226. These frames are arranged sequentially to form an anchor point column for the time shift. Each anchor point not only includes exposure time, brightness variation, drift direction, and pixel displacement, but also records the corresponding fixture displacement height and acoustic energy value. For example, the fixture height corresponding to frame 212 is 24.2 mm, and the acoustic energy is 2.9 times the baseline value; the height corresponding to frame 219 is 26.1 mm, and the acoustic energy is 3.2 times. Through this multi-dimensional calibration, the staggered anchor point series can accurately reflect the synchronization relationship between mechanical vibration and optical drift, which is different from the existing schemes that judge drift only based on pixel changes, and has a higher temporal and physical correlation.

[0064] The established staggered anchor point sequence is fused with the lifting displacement data and resonance distribution band to generate a time index chain for pose analysis. Specifically, using time as the main axis, the exposure time of each staggered anchor point is correlated with the jig's lifting displacement to obtain a time-displacement mapping curve. This curve is then superimposed with the resonance distribution band to obtain a three-dimensional relationship between time, displacement, and energy. For example, frame 212 corresponds to a time of 2.03 seconds, a jig height of 24.2 mm, and an acoustic energy of 3.2 times; frame 226 corresponds to a time of 2.22 seconds, a jig height of 27.5 mm, and an acoustic energy of 2.8 times. This yields the distribution trend of the staggered anchor point sequence within the 2.0 to 2.3 second interval. The continuity of this time index chain indicates that during the stage of increased mechanical resonance intensity, the staggered anchor point interval significantly shortens, decreasing from an average of 20 milliseconds to 8 milliseconds, reflecting the synchronous strengthening of image drift and mechanical vibration. In this way, the time index chain establishes a dynamic mapping relationship between lifting displacement, imaging drift, and resonant energy, providing a precise time calibration basis for subsequent pose drift modeling. Compared with the fixed-period sampling method of the prior art, this implementation does not rely on a preset inter-frame interval, but adaptively generates index nodes based on real vibration behavior, making the time analysis physically realistic and dynamically adaptable.

[0065] Based on the staggered anchor point series, segmented analysis is performed along the lifting displacement direction to obtain the pose deformation trend, draw the pose drift chain, and form a complete time-series drift sketch to describe the diffusion characteristics of equipment vibration in the time domain.

[0066] To analyze the deformation trend of the optical timing misalignment phenomenon caused by mechanical resonance during the lifting process and reveal the propagation law of equipment vibration in the time dimension, based on the obtained timing misalignment anchor point sequence, segmented analysis is performed along the lifting displacement direction to construct a pose drift chain and generate a time-series drift sketch, thereby achieving an accurate description of the entire vibration deformation process. The specific steps are as follows:

[0067] Using the obtained staggered anchor point column as a time index, each anchor point is mapped one-to-one with the displacement data of the lifting fixture, establishing a temporal-spatial correlation. Specifically, the exposure time, brightness variation range, pixel drift direction, and corresponding fixture lifting displacement value for each anchor point are read and arranged chronologically to form a time-displacement mapping table. Taking a compressor housing lifting stroke of 30 mm as an example, four staggered anchor points were detected within the resonance peak interval of 1.9 seconds to 2.3 seconds, corresponding to frames 204, 212, 219, and 226, with fixture displacements of 23.8 mm, 25.1 mm, 26.7 mm, and 28.0 mm, respectively. By correlating these data, the synchronization relationship between optical drift and mechanical displacement can be visually displayed. This step maps the discrete anchor points with temporal shifts to continuous trajectories along the lifting direction, providing a data foundation for subsequent deformation trend analysis.

[0068] After establishing the time-displacement mapping table, the entire lifting stroke was segmented and analyzed to reveal the attitude change patterns within different displacement intervals. Specifically, the total lifting displacement was divided into an initial segment (0 to 10 mm), a middle segment (10 to 25 mm), and a terminal segment (25 to 35 mm). The distribution density, drift direction, and amplitude of the staggered anchor points within each interval were used as segmentation characteristic parameters. Taking the middle segment as an example, the staggered anchor points were most densely packed, with an average time interval of 8 frames. The lateral drift of adjacent anchor points was 3 to 4 pixels, and the vertical drift was 1 to 2 pixels. The drift direction was mainly upward and to the right, indicating that the lifting structure was in the region of maximum resonance intensity during this stage. In contrast, in the initial segment, the anchor points are sparsely distributed, with an inter-frame interval exceeding 15 frames. The drift direction is mainly concentrated in the vertical direction, and the average drift is less than 2 pixels, indicating relatively smooth structural motion. In the final segment, the staggered anchor point interval returns to 12 frames, the drift direction gradually changes from the upper right to the lower left, and the lateral drift is 2 to 3 pixels, reflecting the inertial swing characteristics of the cylinder as it approaches its highest point. Through this segmented analysis, the differences in the speed, direction, and amplitude of attitude changes at each stage can be clearly identified.

[0069] Based on the results of segmented analysis, a pose drift chain is constructed to describe the dynamic evolution of pose deformation. Specifically, on the time axis, the drift direction and displacement of adjacent points are connected in sequence according to anchor points. The jig lifting direction is set as the vertical axis, and the image drift direction as the horizontal axis, forming a continuous broken line representing the pose trajectory. Taking the middle section as an example, the drift directions from frame 204 to frame 226 are, respectively, upper right, upper right, right side, and lower left, with corresponding horizontal drift amounts of 3 pixels, 4 pixels, 5 pixels, and 2 pixels, and vertical drift amounts of 1 pixel, 2 pixels, 1 pixel, and 3 pixels. Connecting these data forms a broken line trajectory with a total drift length of approximately 7 pixels and a directional change angle of approximately 36 degrees. The shape of the drift chain reflects the micro-pose change process of the workpiece within the resonance range: initially, vibration energy is concentrated and propagates to the upper right; in the middle stage, vibration tends to diffuse laterally; and in the later stage, vibration energy decays and reverses. Comparing the curvature of the drift chain with the acoustic energy curves revealed a peak synchronization rate of 93%, indicating a high degree of consistency between attitude changes and resonant energy release. This method achieves a visual description of nonlinear attitude changes by connecting optical drift and mechanical displacement in temporal continuity.

[0070] The drift chains obtained in each displacement segment are spliced ​​together in chronological order to create a complete temporal drift sketch, used to describe the diffusion characteristics of equipment vibration in the time domain. Specifically, time is plotted on the x-axis, displacement on the y-axis, and drift amplitude is represented by color depth. Taking a total lifting time of 3.2 seconds as an example, the sketch shows the darkest colors in the intervals of 0.8 to 1.1 seconds and 1.9 to 2.3 seconds, corresponding to the largest drift amplitudes, indicating that resonant energy is concentratedly released during these two time periods. After 2.3 seconds, the color gradually lightens, and the drift amplitude decreases significantly, indicating that the vibration gradually decays. Quantitative analysis shows that the vibration diffusion duration is approximately 0.4 seconds, and the energy decay rate is approximately 65%. Further superimposing the density of the drift chains in the sketch with the acoustic energy change curves reveals that the peak position deviation is less than 0.02 seconds, indicating that the sketch accurately depicts the diffusion process of vibration energy. The sketch not only visually reveals the propagation characteristics of mechanical resonance but also provides a basis for subsequent stable baseline extraction, enabling optical stabilization modeling to be dynamically corrected based on actual vibration patterns.

[0071] Based on the timing drift sketch, a stable baseline ring is established around the three-point welding area. The synchronous stable segment is extracted from the baseline ring to generate an alignment key, which is used to establish a unified time reference for image acquisition between multiple points.

[0072] To ensure consistent imaging time across different solder joint areas under dynamic vibration conditions and to establish a precise and unified time reference for multi-point image registration, stability analysis was performed on the three-point solder joint area based on the generated temporal drift sketch. A stable baseline loop was constructed around the area, and an alignment key was generated by extracting stable time periods, thus achieving unified time control of the optical acquisition process. The specific implementation steps are as follows:

[0073] Using the obtained temporal drift sketch as a dynamic reference, the three-point weld detection area was spatially positioned and zoned for stability. The temporal drift sketch recorded the time-displacement correspondence and optical drift amplitude variation characteristics of the entire lifting process. By analyzing the spatial energy distribution of the drift sketch, the intensity of vibration influence on the weld area can be distinguished. Taking the three-point weld area above the compressor housing as an example, the geometric distance between the center of each weld point and the camera's field of view is approximately 280 to 310 pixels. Within this range, the drift amplitude fluctuates the most, with an average drift of 2.8 pixels. However, at a distance of 360 to 400 pixels from the center of the weld point, the drift amplitude drops to below 0.8 pixels. The area where the vibration energy is lower than the average total energy is defined as the initial stabilization zone. A closed-loop fitting is performed along the outer edge of the three-point weld area to obtain a continuous annular region. Taking an imaging range with an image width of 1920 pixels and a height of 1200 pixels as an example, the average radius of the stabilization zone is 340 pixels, and the ring width is 30 pixels. This annular region surrounds the entire three-point welding area and is in contact with the fixture fixing surface, exhibiting optimal optical stability, and is defined as the initial range of the stable baseline ring.

[0074] Feature parameters are extracted from the pixel region within the stable baseline ring to quantitatively analyze its optical stability. Specifically, 24 sampling points are uniformly selected on the circumference of the baseline ring, with one point taken every 15 degrees. For each sampling point, three parameters are extracted over 100 consecutive frames: luminance value, mean grayscale value, and pixel position offset. Taking the time period from 2.0 seconds to 2.4 seconds as an example, the average luminance value at sampling point 8 is 247.6, the mean grayscale value fluctuates between 246 and 249, and the pixel position offset is 0.34 pixels; the average luminance value at sampling point 15 is 248.1, the mean grayscale value fluctuates between 247 and 250, and the pixel offset is 0.41 pixels. The overall stability of the baseline ring is determined by statistically analyzing the standard deviation of luminance, the rate of grayscale change, and the mean pixel offset of the 24 sampling points. A stable baseline ring is defined as one where the standard deviation of luminance is less than 2 grayscale levels, the rate of grayscale change is less than 1% per frame, and the mean pixel offset is less than 0.5 pixels. At this point, the baseline ring contains 22 stable sampling points, accounting for 91.6%, indicating that this region has extremely high imaging stability over time. This step constructs a dynamic stability standard using measured optical feature data, eliminating reliance on external mechanical references and solving the problem of reference failure caused by uneven fixture surfaces or reflective properties in existing technologies.

[0075] After obtaining the stable baseline loop, its changes over time are analyzed longitudinally to extract time periods that simultaneously satisfy brightness and displacement stability, defined as synchronous stable segments. The specific steps are as follows: the brightness time series of 24 sampling points is compared frame by frame, calculating the brightness phase difference and pixel displacement difference between any adjacent sampling points. When the phase difference is less than 0.05 milliseconds and the displacement difference is less than 0.2 pixels, the time frame is classified as a stable time period. Taking the interval from 2.08 seconds to 2.14 seconds as an example, the statistical results show that the average brightness phase difference in this time period is 0.038 milliseconds, the maximum is 0.047 milliseconds, and the pixel displacement difference is less than 0.18 pixels, satisfying the synchronous stability condition. This time period contains 7 consecutive images, with an exposure interval of 8.33 milliseconds per frame, and the intra-frame fluctuation amplitude is controlled within 3%. Through the above determination process, multiple synchronous stable segments are obtained, each representing a synchronous steady state between the camera and the workpiece at that moment. Subsequently, all stable segments are arranged in chronological order to form a synchronous stable sequence, providing a temporal basis for generating the alignment key.

[0076] After determining the stable synchronization segment, an alignment key is generated for time unification. Specifically, three core feature parameters are calculated for each stable synchronization segment: mean brightness, grayscale centroid coordinates, and median pixel displacement. Taking the stable segment from 2.08 to 2.14 seconds as an example, the mean brightness is 248.3, the grayscale centroid coordinates are (946, 582), and the median pixel displacement is 0.17 pixels. These three parameters are combined to form a time-aligned data set. To enhance alignment accuracy, the same feature parameters are extracted from image frames of the three solder joint areas and compared with the time-aligned data set. When the brightness difference between the solder joint areas is less than 2, the grayscale centroid shift is less than 2 pixels, and the median displacement difference is less than 0.3 pixels, the three solder joint areas are determined to be under a unified time reference, triggering a synchronization acquisition command. This alignment key is not only used for time synchronization determination but also serves as a time index for subsequent motion compensation and image fusion. For example, during the detection process, if the average brightness difference between two adjacent frames is 2.3 and the grayscale centroid shift is 2.8 pixels, the system determines that the unified time condition has not been met and automatically delays the exposure of the next frame by 0.5 milliseconds to recalibrate. In this way, the alignment key achieves real-time constraint on the acquisition time of multi-point images, ensuring that the images of each solder joint have millisecond-level synchronization accuracy under high vibration conditions.

[0077] By using the alignment key to synchronously control the incident angle of the light source and the trigger interval of the camera, and by performing multi-point image dynamic fusion on the same screen through optomechanical linkage, the image acquisition process and the motion compensation modeling process are kept in time consistent and a unified temporal synchronization window is formed.

[0078] To ensure that optical imaging and mechanical motion remain synchronized in time, and to achieve simultaneous image fusion of multiple points in the three-point welding area, precise synchronous control of the light source incident angle and camera trigger interval is implemented based on the generated alignment key. A stable synchronization window is formed through optomechanical linkage, enabling image acquisition and motion compensation modeling processes to occur within a unified time domain. The specific implementation steps are as follows:

[0079] Using the obtained alignment key as a time reference, the illumination unit and camera triggering device are configured for time synchronization. The alignment key contains three key parameters: average brightness, centroid coordinates of grayscale distribution, and median pixel displacement, which together define the optical steady-state time window. Taking the three-point weld detection area of ​​the compressor housing as an example, the stable period recorded by the alignment key is between 2.08 seconds and 2.14 seconds, with a time span of 60 milliseconds. During this time period, the optical drift is less than 0.2 pixels, and the brightness change is less than 2 gray levels. Based on this stable window, the trigger signal of the light source controller and the camera exposure trigger signal are unified into the same time axis, with a time resolution set to 0.001 seconds. The light source illumination trigger advance is set to 2 milliseconds, and the camera exposure trigger delay is set to 0.5 milliseconds to ensure that the illumination light reaches a steady state before camera exposure. This time synchronization step establishes a strict temporal correspondence between optical illumination and imaging actions, avoiding the time drift problem of light source and exposure under traditional independent triggering conditions, and laying a temporal foundation for subsequent optomechanical coordination.

[0080] Based on the spatial position parameters of the grayscale distribution centroid in the alignment key, the incident angle of the light source is dynamically adjusted to maintain a stable correspondence between the lighting direction and the workpiece posture. Specifically, a ring light source is used as the main lighting structure, with an outer diameter of 110 mm and an inner diameter of 90 mm. Strip auxiliary light sources, each 160 mm long, are arranged on both sides. The incident angle and light intensity distribution of the three light sources are independently adjusted by the light source controller to achieve uniform spatial illuminance control. Taking a three-point welding spacing of 60 mm and an incident angle of 20 degrees as an example, at time 2.10 seconds, the angle of the left strip light source is set to 22 degrees, the right to 18 degrees, and the intensity of the ring light source is set to 65%. Through illumination distribution monitoring, the grayscale centroid of the image acquired by the camera is (946, 583), which deviates from the reference coordinates (946, 582) in the alignment key by less than 1 pixel, indicating that the lighting direction perfectly matches the steady-state window. When the cylinder's micro-vibration causes a change in the workpiece's tilt angle, the light source controller adjusts the angle in 0.2-degree steps, completing the response within 100 milliseconds, thus ensuring that the grayscale center of gravity deviation is always maintained within 2 pixels.

[0081] Based on the synchronization of the light source incident angle, the camera trigger interval is precisely controlled to ensure that the exposure action of each frame completely overlaps with the optical steady-state window. Specifically, the camera frame rate is set to an integer multiple of the steady-state window time based on the steady-state window time span recorded by the alignment key (60 milliseconds). For example, with a steady-state window of 60 milliseconds, the frame rate is set to 240 frames per second, making the exposure time per frame 4.17 milliseconds, and the exposure interval is completely synchronized with the light source refresh cycle. When resonance enhancement causes the steady-state window to shorten to 45 milliseconds, the camera frame rate is automatically increased to 266 frames per second to ensure that 12 frames are captured completely within each steady-state window. The exposure delay is controlled within ±0.1 milliseconds, and the deviation between trigger signals is less than 0.2 milliseconds. Taking frames 210 to 216 as examples, these 6 frames are all within the steady-state range of 2.08 seconds to 2.14 seconds, with the average brightness fluctuating between 247.8 and 248.9, and the grayscale center of gravity change not exceeding 1.5 pixels, ensuring that the optical engine action is synchronized at the millisecond level. This method ensures that the optical acquisition frame coincides with the steady state of the illumination by precisely adjusting the exposure interval, thus avoiding inter-frame exposure misalignment caused by mechanical lifting resonance.

[0082] After the incident angle of the light source and the trigger interval of the camera are synchronized, the dynamic fusion of multiple images on the same screen is achieved through optomechanical linkage. Specifically, within a unified time window, three industrial cameras simultaneously capture images of the three-point welding area. The three cameras are positioned at angles relative to the fixture: 25 degrees to the left, 0 degrees to the center, and 25 degrees to the right, all 150 mm from the workpiece surface. The exposure time of each camera is locked within a window of 2.08 to 2.14 seconds, and the trigger signal delay difference is controlled within 0.1 milliseconds. After the light source illumination and camera exposure are synchronized, the three images captured by each camera completely overlap in time. Actual testing shows that the image spot overlap rate of the three-point welding area reaches 98.7%, the brightness error is less than 3 gray levels, and the pixel position deviation is less than 1.2 pixels. After fusing the three images into a single frame in real time, the relative grayscale uniformity of the three-point welding area is improved to 97.4%. Through this optomechanical linkage, the three-point welding images form a unified image within the same time window, providing a temporally consistent raw input for subsequent motion compensation modeling.

[0083] A unified acquisition window, formed by optomechanical linkage, establishes an integrated temporal domain for image acquisition and motion compensation modeling. The specific steps are as follows: using the synchronization window time period as a benchmark, the changes in light source angle, camera exposure time sequence, and fixture lifting displacement data are mapped to the same time axis. Taking a window of 2.08 seconds to 2.14 seconds as an example, the light source incident angle ranges from 18 degrees to 22 degrees, the camera exposure frame sequence numbers are frames 210 to 216, and the corresponding fixture displacement ranges from 24.3 mm to 26.6 mm. The three data points are time-registered to generate a unified temporal domain synchronization table. This synchronization table serves as the time input in subsequent motion compensation modeling, ensuring that compensation calculations are performed only within a stable window, guaranteeing consistent time reference for model solution. Further measurement of the time difference between camera exposure synchronization error and mechanical displacement delay shows a maximum difference of 0.16 milliseconds, far below the allowable threshold of 0.5 milliseconds, indicating that the optical acquisition process and mechanical motion have achieved physical-level time consistency. The establishment of this synchronization window enables the temporal binding of imaging data and motion models, allowing compensation modeling to be dynamically reconstructed based on real time series, thereby avoiding attitude calculation offsets caused by time difference errors.

[0084] The control is carried out in the synchronization window, the attitude period is adjusted by using the time slot turning back rudder, and the energy of the resonance peak is weakened by using the phase release curtain, so as to establish a closed-loop stable synchronization relationship between the lifting action and the imaging process, thereby completing the dynamic imaging stabilization detection of the three-point welding area.

[0085] To achieve stable coordination between the mechanical lifting action and the optical imaging process within the established unified time-domain synchronization window, and to ensure a constant temporal correspondence between exposure frames and pose changes, adjustments are required within the synchronization window. The pose period is adjusted using a time-slot turning rudder, and the resonance peak energy is weakened by a phase-release curtain, ensuring that the lifting action and imaging process maintain closed-loop consistency in both time and energy. This achieves dynamic imaging stabilization detection of the three-point welding area. The specific steps are as follows:

[0086] A time baseline for the optomechanical action is established within a unified time-domain synchronization window, precisely mapping the mechanical lifting displacement curve to the camera exposure time axis. The synchronization window is taken from the stable range determined in the previous implementation step, for example, within the time period of 2.08 seconds to 2.14 seconds, the light source incident angle varies from 18 degrees to 22 degrees, the camera exposure frame number is from frame 210 to 216, and the jig lifting displacement ranges from 24.3 mm to 26.6 mm. Within this time period, cylinder displacement sensing signals, light source trigger signals, and camera exposure signals are collected, and time synchronization is performed with a sampling interval of 0.001 seconds. Analysis of the signal waveforms reveals that at the exposure time of frame 212, the peak jig lifting speed lags the exposure trigger signal by 0.36 milliseconds, forming an initial phase difference. This phase difference reflects the time misalignment between the mechanical action and the exposure behavior. To provide a reference for subsequent adjustments, a phase difference correction curve is established based on this, using the exposure trigger signal as the time reference and the mechanical peak time as the control point. This step transforms optical acquisition and mechanical motion into a quantifiable temporal relationship, forming the initial coordinate framework for closed-loop control.

[0087] Based on the initial phase difference measurement, a time-slotted reversing rudder mechanism is used to fine-tune the lifting action cycle, ensuring that the periodic peak of the mechanical pose change overlaps with the center of the exposure window. Specifically, a periodic signal is set during the cylinder drive phase, with a drive cycle of 300 milliseconds, encompassing three stages: lifting, dwell, and descent. The reversing rudder's adjustment principle is that when a phase difference greater than 0.3 milliseconds is detected, the cylinder drive signal for the next cycle is fine-tuned, advancing or delaying it by 0.05 to 0.3 milliseconds. For example, in actual operation, when the exposure is 0.36 milliseconds earlier than the displacement peak, the reversing rudder advances the cylinder drive signal by 0.25 milliseconds in the next cycle, aligning the lifting speed peak at 2.10 seconds with the exposure time of the 212th frame. After each adjustment, the new phase difference is measured in real time. When the phase difference drops to within 0.05 milliseconds, the reversing rudder enters a steady-state hold. Through three consecutive adjustments, the lifting curve and the exposure curve completely overlap, with a peak synchronization error of less than 0.03 milliseconds, achieving time coincidence between the mechanical vibration cycle and the imaging cycle.

[0088] After achieving time overlap during the lifting cycle, to prevent the resonance enhancement effect caused by energy concentration, a phase-release curtain is used to adjust the time-domain peak reduction of the cylinder driving force output process. Specifically, within the resonance-sensitive range of 22 mm to 26 mm lifting height, the duty cycle of the cylinder driving pulse is adjusted from 80% to 70%, while the energy release time is extended from 28 milliseconds to 34 milliseconds. Through the effect of the release curtain, the upward slope of the driving force curve is reduced by 20%, and the cylinder acceleration decreases from 3.2 m / s² to 2.5 m / s². Taking a synchronization window of 2.08 to 2.14 seconds as an example, the peak cylinder driving force decreases from 93.5 Newtons to 88.4 Newtons, the peak acoustic acceleration signal decreases by 27%, and the resonance energy distribution curve changes from a bimodal shape to a flat single-peak shape. The pixel drift of the camera-acquired image decreases from 0.42 pixels to 0.18 pixels, the brightness fluctuation decreases from 4 gray levels to 1 gray level, and the image clarity is significantly improved. This slow-release process disperses the mechanical excitation energy over a longer period of time, avoiding the concentrated superposition of resonant energy at the moment of exposure, and ensuring that the imaging process is stable at a low energy amplitude. This achieves peak-shifting complementarity between mechanical energy and optical exposure in the time domain.

[0089] With the combined action of the time-slotted retraction rudder and the phase-release curtain, a closed-loop feedback path for optical and mechanical actions is established, enabling real-time control within the synchronization window. Specifically, the cylinder displacement rate and imaging brightness fluctuations are detected in real time within each synchronization window cycle. When the detected displacement rate exceeds 1.5 mm / s or the brightness fluctuation exceeds 3 gray levels, the retraction rudder immediately adjusts the drive signal 0.1 milliseconds ahead in the next control cycle and simultaneously triggers the phase-release curtain to reduce the drive force output by 5%. Through dynamic adjustment over five consecutive synchronization cycles, the repeatability error of the cylinder displacement curve is reduced from 0.12 mm to 0.03 mm, and the stability is improved to 99.2%; simultaneously, the imaging brightness stability is improved to 98.8%. The closed-loop feedback path ensures that mechanical displacement, energy release, and optical exposure remain continuously coordinated in time, achieving true optomechanical resonance elimination and attitude synchronization.

[0090] After the closed-loop synchronous control stabilizes, dynamic imaging stabilization detection of the three-point welding area is performed. Specifically, eight frames of images are continuously acquired within the synchronization window, and the grayscale, morphology, and positional changes of adjacent frames are compared. Using the middle welding point as a reference, its average grayscale value remains between 247 and 248, with pixel displacement not exceeding 0.2 pixels; the average grayscale value of the left welding point is 247, and that of the right welding point is 248, with a grayscale difference of less than 1.5 grayscale levels. After 3D reconstruction, the spatial coordinate change of the welding point is less than 0.05 mm, and the surface normal angle fluctuation does not exceed 0.8 degrees. No feature misalignment, weld bead blurring, or halo drift was observed during the detection process. In this way, the dynamic imaging of the three-point welding area maintains high clarity and high repeatability under vibration conditions, reducing the false alarm rate by 72% and the false negative rate by 68%.

[0091] The following examples fully cover the key details of the three-point welding of the compressor housing during the lifting, imaging, and judgment processes, and compare the detection effects of existing technologies with those of this invention using real quantitative data. The camera resolution, lens focal length, ring illumination, 2D and 3D accuracy, field of view, and linewidth configurations used in the examples are taken from verified parameters of similar mass production projects, facilitating engineering implementation and reproduction.

[0092] Example scenarios and operating conditions

[0093] The production line focuses on the online inspection of three-point welds on refrigerator compressor housings. The workpiece is lifted to imaging height by a cylinder, with a total lifting stroke of 30 mm, of which the effective height variation is approximately 25 mm. The fixture angle undergoes a change in attitude from 3.5 degrees to 7 degrees during transport and positioning. The camera is a 5-megapixel monochrome camera with a 25 mm focal length. The main illumination is a 90-degree ring light, supplemented by two side strip lights to form a combined illumination. The 2D measurement accuracy is 0.03 mm, the 3D measurement accuracy is 0.01 mm, the 2D field of view is 75 × 50 mm, and the 3D line width is 60 mm. The baseline imaging frame rate is set to 120 frames per second, which will be dynamically allocated according to a steady-state window after the invention is implemented.

[0094] Existing technology baseline performance

[0095] To establish an objective comparison, baseline testing was first conducted for a period of time without using this invention. 1) Triggering and lifting time error: Due to the cylinder drive frequency being close to the machine tool's inherent frequency, a slight resonance occurred. The average time difference between image exposure triggering and fixture displacement peak was measured to be 0.83 milliseconds in the 2.0 to 2.4 second range, with a maximum of 1.15 milliseconds. 2) Illumination stability: With the ring light fixed at a 20-degree incident angle, affected by slight posture jitter, the brightness standard deviation of the three-point welding area reached 3.9 gray levels, and the gray level centroid drifted by 3 to 5 pixels within consecutive frames. 3) Pixel stability and registration: The average pixel drift of the edge highlight band near the resonance peak was 0.46 pixels, with a peak exceeding 0.8 pixels; after superimposing the three-view images, the overlap rate of the three-point welding spot was only 93.1%. 4) Measurement and recognition results:

[0096] Hole diameter repeatability (six sigma bandwidth): ±0.065 mm;

[0097] Concave depth repeatability: ±0.028 mm;

[0098] The false alarm rate was 7.9%, and the false negative rate was 5.8% (based on a statistical sample of 10,000 items, and subject to manual verification).

[0099] The single-piece cycle stability fluctuates significantly. The detection time is 1.25 seconds for the 90th percentile and 1.35 seconds for the 99th percentile. Vibration interference caused multiple re-image acquisitions.

[0100] Implementation path and key data of the present invention

[0101] Based on this invention, the six-step method has been implemented step by step. The key points and critical values ​​are listed below.

[0102] 1. Establishing a Resonance Distribution Zone and Anchoring the Timing Position: Acoustic vibration, image time markers, and displacement signals are synchronously acquired within the complete lifting cycle, with a timing accuracy controlled to 0.001 seconds. The exposure time of each frame is aligned with the acoustic vibration energy and displacement increment, resulting in a synchronization table composed of frame number, acoustic vibration energy, and displacement rate. Three concentrated resonance energy segments are identified: 0.82 to 1.10 seconds, 1.92 to 2.28 seconds, and 2.70 to 2.92 seconds. Taking the second main peak as an example, the acoustic vibration energy increases to 3.2 times the baseline, and the displacement rate increases to 3.4 mm / s. This corresponds to the first continuous drift at the bright edges of the image, with anchor points appearing at 2.03 seconds, 2.12 seconds, 2.19 seconds, and 2.26 seconds.

[0103] 2. Draw a temporal drift sketch and extract a stable baseline loop. Unfold the anchor points along the lifting displacement direction, analyze the attitude trends of the initial, middle, and terminal segments, and draw the pose drift chain. Project energy into the image space and locate the ring with the least vibration. Taking a 1920×1200 pixel image as an example, around the three-point weld area with a radius of approximately 340 pixels and a ring width of 30 pixels, count the brightness and displacement fluctuations of 24 corner points in 100 consecutive frames. 22 corner points have a brightness standard deviation of less than 2 gray levels, and the average pixel offset is 0.41 pixels, which is determined to be a stable baseline loop.

[0104] 3. Generating Alignment Keys and Locking the Synchronization Window: A stable synchronization segment is found within the time unrolling of the stable baseline loop. A frame is marked as stable when the phase difference in brightness between adjacent corner points is less than 0.05 milliseconds and the displacement difference is less than 0.2 pixels. A stable window of 2.08 to 2.14 seconds is obtained, with an average brightness of 248.3, a grayscale centroid of (946, 582), a median pixel displacement of 0.17 pixels, and a total of 7 frames within the window. These three parameter sets serve as alignment keys to constrain the temporal consistency between subsequent light sources and the camera.

[0105] IV. Opto-mechanical linkage achieves simultaneous screen fusion. Following the alignment key, the incident angles and brightness of the ring light and the two side strip lights are synchronously adjusted to bring the grayscale center of gravity close to (946, 582). At 2.10 seconds, with an incident angle of 22 degrees on the left, 18 degrees on the right, and 65% intensity of the ring light, the grayscale center of gravity deviation is less than 1 pixel. The three cameras are synchronously exposed in the same window, with the trigger delay difference controlled within 0.1 milliseconds. Actual measurements show that the overlap rate of the light spots in the three-point welding simultaneous screen image increases to 98.7%, and the brightness error converges to less than 3 grayscale levels.

[0106] 5. Within the synchronization window, the peak relative displacement of the exposure in frame 212 was measured to be 0.38 milliseconds behind the time-slotted return rudder and phase-release curtain. The lifting control signal was fine-tuned three times: the first cycle was advanced by 0.25 milliseconds, and the second and third cycles were each advanced by 0.05 milliseconds, bringing the phase difference between the exposure and displacement peaks to 0.04 milliseconds. For the resonance-sensitive section with a height of 22 to 26 mm, a phase-release curtain was applied, maintaining constant energy for the first 10 milliseconds of the synchronization window, attenuating by 10% for the middle 20 milliseconds, and reducing to 70% of the baseline for the last 30 milliseconds. Results showed that the peak driving force decreased from 93.5 Newtons to 88.2 Newtons, the peak acoustic vibration decreased from 3.2 times to 2.1 times, the average pixel drift decreased from 0.42 pixels to 0.18 pixels, and the brightness fluctuation decreased from 4 gray levels to 1 gray level. After five closed-loop cycles, the time correlation coefficient between the exposure and displacement curves reached 0.998, and the stability of the displacement curve reached 99.2%.

[0107] Implementation effectiveness of measurement and identification

[0108] Eight consecutive frames were acquired within a unified synchronization window, and two-dimensional and three-dimensional fusion measurement and defect identification were performed on the three-point weld: 1) Improved repeatability of hole diameter: Taking a weld hole with a nominal diameter of 3.20 mm as an example, the repeatability converged from ±0.065 mm to ±0.028 mm, more than double the improvement; 2) Improved stability of concave depth: For a nominal concave depth of 0.20 mm, the repeatability decreased from ±0.028 mm to ±0.012 mm; 3) More stable identification of weld slag expansion: Due to the stable viewing angle and brightness, the threshold for expansion area is no longer... 4) More accurate fusion-through morphology judgment: After three-view simultaneous fusion, edge connectivity and height discontinuity conform to real 3D geometry, and the false alarm rate is significantly reduced; 5) Simultaneous improvement in accuracy and efficiency: In 10,000 statistical samples, the false alarm rate decreased from 7.9% to 2.1%, and the false negative rate decreased from 5.8% to 1.9%; the cycle stability was improved, the detection time for the 90th percentile decreased from 1.25 seconds to 1.08 seconds, and the detection time for the 99th percentile decreased from 1.35 seconds to 1.16 seconds, significantly reducing rework and re-image acquisition; 6) Enhanced posture robustness: Within the range of jig angle expansion from 3.5 degrees to 7 degrees, the threshold is automatically normalized and the judgment remains stable, the gray level difference of three points is maintained within 2 gray levels, and the long-term average pixel drift is less than 0.20 pixels.

[0109] Project Reproducibility Description

[0110] The entire process is independent of static environments. The key lies in using an alignment key to drive the consistency of light source and exposure time, using a time-slot reversal rudder to correct the lifting phase, and using a phase-release curtain to distribute excitation energy over time. All these parameters can be calibrated segment by segment during trial production: time resolution 0.001 seconds, trigger delay adjustment step 0.01 milliseconds, illumination angle step 0.2 degrees, and a stabilization window initially selected as 60 milliseconds. If production line vibrations occur, this can be reduced to 45 milliseconds while simultaneously increasing the frame rate to ensure sufficient frame coverage. All values ​​are derived from actual operating conditions or constrained by mass production achievable parameters, meeting mass production reproducibility requirements.

[0111] Comparison data table of effect improvement

[0112] Indicator name Prior art baseline Inventive approach Improvement magnitude Mean time difference between exposure and displacement peak 0.83 ms 0.04 ms 95% increase Mean pixel shift (resonance main peak) 0.46 pixels 0.18 pixels 61% decrease Standard deviation of intensity (within stabilization window) 3.9 gray levels 1.0 gray levels 74% decrease On-screen overlap rate of three-view light spots 93.1% 98.7% 5.6 percentage points increase Hole diameter repeatability (±) 0.065 mm 0.028 mm 57% increase Depth of concave repeatability (±) 0.028 mm 0.012 mm 57% increase False alarm rate 7.9% 2.1% 73% decrease Missed detection rate 5.8% 1.9% 67% decrease 90th percentile detection time 1.25 s 1.08 s 14% shorter 99th percentile detection time 1.35 s 1.16 s 14% shorter Attitude adaptation range 3.5 degrees 7.0 degrees 100% increase Long-term mean pixel shift 0.32 pixels 0.19 pixels 41% decrease

[0113] The above examples demonstrate that in online environments with significant lifting resonance and attitude fluctuations, the synergy of resonance distribution sensing, staggered anchor point indexing, temporal drift sketching, stable baseline loop, alignment key, optomechanical linkage, and time-slot folding rudder combined with phase relief curtain can stabilize the temporal relationship between image acquisition and motion to the millisecond level, simultaneously stabilize optical consistency and three-dimensional geometric consistency, reduce false alarms and missed detections from the source, and improve measurement repeatability to mass production level requirements.

[0114] This invention establishes a real-time dynamic reference model by synchronizing the acoustic vibration signal, image time markers, and fixture displacement. Then, through staggered anchor point sequences and time-series drift sketches, it quantifies the temporal diffusion law of mechanical vibration. Based on this, a stable baseline loop and alignment key achieve a unified time reference for multi-point images, ensuring millisecond-level consistency between the light source incident angle and the camera trigger interval. Finally, a time-slot folding rudder and phase-release curtain complete the closed-loop coordination of the lifting action and exposure imaging within the synchronization window, weakening the resonance peak effect from both temporal and energy dimensions. Actual testing shows that this method reduces the peak time difference between exposure and lifting displacement from 0.83 milliseconds to 0.04 milliseconds, the average pixel drift from 0.46 pixels to 0.18 pixels, and the brightness fluctuation from 3.9 gray levels to 1.0 gray levels. False alarm and false negative rates are reduced by 73% and 67%, respectively, and the repeatability of hole diameter and solder joint depth measurements is improved by more than 50%. This solution fundamentally solves the problem of inconsistency between mechanical vibration and imaging timing in online inspection of three-point welds, achieving high consistency, high precision and high robustness in dynamic working conditions. It provides quantifiable and reproducible imaging stabilization support for real-time intelligent inspection of the weld quality of compressor housings.

[0115] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A dynamic imaging stabilization detection method for three-point welding based on motion compensation modeling, characterized in that, Includes the following steps: Acquire acoustic and vibration signals, image time stamp signals, and jig lifting and lowering displacement signals throughout the entire process of compressor housing lifting. Perform time synchronization and feature fusion on the acquired signals to generate a resonance distribution band, which is used to construct a dynamic reference model for subsequent time series analysis. Based on the resonance distribution zone, the bright drift points are continuously tracked in the edge area of ​​the imaging screen to extract the starting position of the time misalignment and establish a time misalignment anchor point column to provide a time index for workpiece pose analysis during the lifting process. Based on the staggered anchor point series, segmented analysis is performed along the lifting displacement direction to obtain the pose deformation trend, draw the pose drift chain, and form a complete time-series drift sketch to describe the diffusion characteristics of equipment vibration in the time domain. Based on the timing drift sketch, a stable baseline ring is established around the three-point welding area. The synchronous stable segment is extracted from the baseline ring to generate an alignment key, which is used to establish a unified time reference for image acquisition between multiple points. By using the alignment key to synchronously control the incident angle of the light source and the trigger interval of the camera, and by performing multi-point image dynamic fusion on the same screen through optomechanical linkage, the image acquisition process and the motion compensation modeling process are kept in time consistent and a unified temporal synchronization window is formed. The control is carried out in the synchronization window, the attitude period is adjusted by using the time slot turning back rudder, and the energy of the resonance peak is weakened by using the phase release curtain, so as to establish a closed-loop stable synchronization relationship between the lifting action and the imaging process, thereby completing the dynamic imaging stabilization detection of the three-point welding area. The steps for generating a time-drift sketch are as follows: A time-space relationship is established using staggered anchor point columns and fixture displacement data to form a time-displacement mapping table; The stroke is segmented and analyzed along the lifting displacement direction. The anchor point distribution density, drift direction and drift amplitude are used as segment feature parameters to identify the attitude change pattern of different segments. Based on the segmented analysis results, the drift direction and displacement are connected sequentially along the time axis according to the anchor point to construct a pose drift chain representing the attitude change trajectory. The drift direction and displacement are continuously connected to form a pose trajectory polyline with the jig lifting direction as the vertical axis and the image drift direction as the horizontal axis. The curvature change of the polyline reflects the synchronous relationship between attitude change and resonant energy release during the lifting process, which is used to characterize the dynamic evolution characteristics of vibration deformation. The drift chains of each displacement segment are spliced ​​together in chronological order to draw a time-series drift sketch to describe the diffusion characteristics of equipment vibration in the time dimension; The steps for generating the alignment key are as follows: Based on the time-drift sketch, the three-point welding detection area is spatially located and stability-zoned to identify areas where the vibration energy is lower than the average total energy and form a stable baseline loop. Feature parameters are extracted from the pixel region within the stable baseline ring range, and the imaging stability of the baseline ring is determined by the brightness standard deviation, gray level change rate, and pixel offset mean. A longitudinal analysis of the changes in the stable baseline loop over time was performed, and the time period that simultaneously satisfies brightness stability and displacement stability was extracted as the synchronous stable segment. In the synchronous stable segment, the mean brightness, the centroid coordinates of the grayscale distribution, and the median pixel displacement are calculated to generate an alignment key to establish a unified temporal reference for multi-point image acquisition.

2. The three-point welding dynamic imaging stabilization detection method based on motion compensation modeling according to claim 1, characterized in that, The steps for generating resonance distribution bands are as follows: During the compressor housing lifting process, acoustic sensors, acceleration sensors and grating rulers are arranged to collect signals throughout the process, and time synchronization is achieved with a unified time reference device. The acoustic vibration signal, the image time stamp signal, and the displacement signal are registered in chronological order to establish a unified time axis based on the exposure time, thereby achieving time alignment of multi-source signals. Key characteristic parameters of vibration energy, sound pressure energy and displacement rate are extracted based on the synchronization signal to identify the time period and energy peak of resonance. The time periods identified as resonance are merged in chronological order to form resonance distribution bands, and a dynamic reference model is generated for subsequent time series analysis.

3. The three-point welding dynamic imaging stabilization detection method based on motion compensation modeling according to claim 2, characterized in that, The process of establishing the staggered anchor column is as follows: Using the resonance distribution zone as a time reference, all image frames are temporally correlated and segmented to determine the resonance stage of each image frame. In the time-calibrated keyframes, the edge detection region is divided, and the bright pixels in each edge band are continuously tracked and the grayscale and coordinate changes are recorded. The brightness and position changes of the bright areas in consecutive frames are compared. When the drift or grayscale change exceeds the threshold, the starting position of the error is determined and the error detection point is established. The fault detection points are merged in chronological order to form a staggered anchor point column, and combined with the lifting displacement data and resonance distribution zone to generate a time index chain, providing a time reference for pose analysis.

4. The three-point welding dynamic imaging stabilization detection method based on motion compensation modeling according to claim 3, characterized in that, The steps for generating a synchronized window are as follows: Using the alignment key as a time reference, the light source illumination unit and the camera triggering device are time-synchronized to ensure that the light source triggering signal and the camera exposure signal correspond on a unified time axis. Based on the grayscale distribution centroid parameters in the alignment key, the incident angle of the light source is dynamically adjusted to maintain a stable correspondence between the lighting direction and the workpiece posture. Based on the synchronization of the incident angle of the light source, the camera trigger interval is precisely controlled so that the exposure action completely overlaps with the optical steady-state window; By using optomechanical linkage, multiple images can be merged on the same screen within a unified time window to generate image input with consistent time. A time mapping relationship between light source angle, camera exposure and fixture displacement is established based on the synchronization window, forming a unified time-domain synchronization table for motion compensation modeling.

5. The three-point welding dynamic imaging stabilization detection method based on motion compensation modeling according to claim 4, characterized in that, Within the synchronization window, adjustments are made using a time-slotted turning rudder to adjust the attitude period, and a phase-release curtain is used to weaken the resonance peak energy. The steps to establish a closed-loop synchronization relationship between the lifting action and the imaging process are as follows: A time baseline for the optomechanical action is established within a unified time-domain synchronization window, and the mechanical lifting displacement curve is precisely mapped to the camera exposure time axis to form a phase difference correction curve. Based on the phase difference results, a time-slot return rudder mechanism is used to fine-tune the lifting action cycle, so that the periodic peak of the mechanical pose change overlaps with the center of the exposure window; After the lifting cycle achieves time overlap, the phase release curtain is used to adjust the peak of the cylinder driving force output process to reduce the concentration of resonance energy. A closed-loop feedback path for optical and mechanical actions is established by using a time-slot return rudder and a phase-release curtain to achieve real-time control within the synchronization window; After the closed-loop synchronous control stabilizes, dynamic imaging stabilization detection of the three-point welding area is performed.

Images