A tube sheet and pump shell machining positioning deviation compensation method based on image correction

Abstract

The application belongs to the technical field of image correction, and particularly relates to a tube plate and pump shell machining positioning deviation compensation method based on image correction, which comprises a time sequence image sequence acquisition, a deep convolutional neural network motion blur correction, a frequency domain decomposition module vibration artifact separation, a sub-pixel edge extraction and registration, an exponential decay model deformation prediction and recursive least square method parameter adaptive update, and a predicted control strategy compensation amount calculation and closed loop verification. Through the acquisition and motion blur correction processing of the time sequence image sequence, the deep convolutional neural network is used for motion blur kernel estimation and deblurring recovery of each frame of image, the edge blur effect caused by the continuous slight deformation of the workpiece in the image acquisition process can be eliminated, and the image data basis for the subsequent accurate deformation extraction is provided.

Description

Technical Field

[0001] This invention belongs to the field of image correction technology, specifically a method for compensating for the machining positioning deviation of tube sheet and pump casing based on image correction. Background Technology

[0002] Tube sheets and pump casings are pressure-bearing components in fluid machinery such as centrifugal pumps and heat exchangers. Their machining accuracy and assembly alignment quality directly determine the sealing reliability and long-term service life of the equipment. Due to the combined effects of multiple factors, including the gradual release of residual stress within the blank material, elastic deformation caused by clamping forces, and cumulative deviations in the CNC machining toolpath, the geometric shape of the actual machined workpiece generally exhibits localized micro-deformation and deviates from the design theoretical value. Vision-based machining positioning deviation correction and compensation technology is a supporting technology for ensuring the precision assembly quality of tube sheets and pump casings. This involves acquiring image data of preset reference features on the workpiece surface through an industrial vision system, and then... To address imaging distortion caused by factors such as uneven lighting, lens distortion, and workpiece surface reflection, an image correction algorithm is used to restore the distortion and remap the pixel coordinates of the original image, recovering the true projection shape of the reference features in the image plane. Then, based on the corrected image, the precise pixel coordinates of the feature points are extracted, and the mapping transformation from the image coordinate system to the world coordinate system is completed in combination with the camera calibration parameters. The global positioning deviation is obtained, and the coordinate parameters of the processing station are corrected to complete the deviation compensation. This technical approach optimizes the imaging quality through the pre-image correction stage and can meet the basic positioning correction requirements in conventional processing scenarios where the workpiece deformation is uniform and the feature shape is regular. However, this type of correction and compensation method has a technical drawback: insufficient adaptability to actual processing conditions. The deviation calculations are generally based on the geometric assumption that the workpiece is an ideal rigid body. They can only correct global positioning deviations such as overall translation and rotation, and are difficult to deal with non-uniform local deformations such as micro-warping of the sealing surface edge and local deformation of the circumferential hole caused by residual stress release. This results in a deviation between the reference features on which visual correction is based and the actual shape of the workpiece. Therefore, this invention provides a method for compensating for the positioning deviation of tube sheet and pump casing processing based on image correction. Summary of the Invention

[0003] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0004] The technical solution adopted by this invention to solve its technical problem is as follows: This invention proposes a method for compensating for the machining positioning deviation between tube sheet and pump casing based on image correction, comprising the following steps: Step 1: Acquire time-series image sequences of the tube sheet and pump casing surfaces to be tested; Step 2: Perform motion blur correction processing on the acquired time-series image sequence; Step 3: Perform vibration artifact separation processing on the time-series image sequence; Step 4: Perform subpixel edge extraction and registration of the time-series image sequence; Step 5: Extract the temporal data of processing deformation based on the temporal image sequence of adjacent frames, and construct a deformation prediction model; Step 6: Based on the deformation prediction model, predict the deformation amount at future times, calculate the compensation amount for the positioning deviation based on the prediction results, and output it. Step 7: After compensation, re-acquire time-series image sequences, compare the measured deformation values ​​with the predicted values ​​predicted by the deformation prediction model, and recursively update the parameters of the deformation prediction model.

[0005] Preferably, the steps for completing motion blur correction processing include: Each frame of the temporal image sequence is input into a pre-trained deep convolutional neural network to perform motion blur kernel estimation. Deep convolutional neural networks are built on a generative adversarial network architecture, which includes a fuzzy kernel prediction subnetwork and an image complex atom network; The blur kernel prediction subnetwork receives the current frame image and outputs the corresponding motion blur kernel parameters; The image complex atom network deblurs the current frame image based on the predicted motion blur kernel parameters, generating a corrected and clear image sequence.

[0006] Preferably, the step of separating vibration artifacts from a time-series image sequence includes: The corrected, clear image sequence is input into the frequency domain decomposition module for vibration artifact separation; The frequency domain decomposition module first performs a fast Fourier transform on multiple consecutive frames of corrected images, converting them from the spatial domain to the frequency domain; In the frequency domain, the parameters of the band-stop filter are set according to the spectral characteristics of the vibration interference. After the filtering process is completed, the image is restored to the spatial domain by inverse fast Fourier transform to generate a clean image sequence with vibration artifacts removed.

[0007] Preferably, the steps for subpixel edge extraction and registration of time-series image sequences include: Subpixel edge detection and registration are performed on adjacent frames in a clean image sequence after removing vibration artifacts. Subpixel edge detection employs a subpixel edge localization algorithm based on grayscale moments; The subpixel edge localization algorithm first uses a Gaussian filter to smooth and denoise the image; The initial edge point set is extracted using the Canney edge detection operator; The initial edge points are calculated with sub-pixel precision based on the principle of gray-scale moment invariance.

[0008] Preferably, the step of extracting the temporal data of processing deformation based on the temporal image sequence of adjacent frames and constructing a deformation prediction model includes: Based on the geometric transformation relationship between adjacent frame images, calculate the deformation increment of the tube sheet flange surface or pump casing sealing surface within the time sampling interval.

[0009] Preferably, the time-accumulated deformation increments are arranged in chronological order to construct a time-series data sequence of deformation evolution over time, and the deformation prediction model adopts an exponential decay model with a characteristic time constant.

[0010] Preferably, the deformation prediction model is estimated and updated online using the real-time recursive least squares method; The recursive least squares method employs an adaptive update strategy with a forgetting factor.

[0011] Preferably, the steps of predicting the deformation amount at future times based on the deformation prediction model, calculating the compensation amount for the positioning deviation based on the prediction results, and outputting the results include: The deformation amount at future moments is predicted based on the deformation prediction model. The predicted deformation amount is compared with the preset fit gap threshold, and the direction and magnitude of the compensation amount are calculated. The compensation amount calculation adopts a predictive control strategy, which predicts several sampling periods in advance based on the known deformation evolution trend at the current moment, and determines the compensation amount at the current moment based on the prediction results.

[0012] Preferably, the step of re-acquiring the time-series image sequence after compensation, comparing the measured deformation values ​​with the predicted values ​​predicted by the deformation prediction model, and recursively updating the parameters of the deformation prediction model includes: After the CNC machining system executes the compensation command, the measurement area of ​​the tube sheet or pump casing is re-acquired.

[0013] Preferably, the deformation measurement value generated by image acquisition is compared with the predicted value before compensation; The parameters of the deformation prediction model are updated recursively using the new measurement data.

[0014] The beneficial effects of this invention are as follows: 1. This invention acquires and processes time-series image sequences for motion blur correction. It uses a deep convolutional neural network to estimate the motion blur kernel and restore the image for each frame. This can eliminate the edge blurring effect caused by the continuous small deformation of the workpiece during image acquisition, and provide an image data basis for subsequent accurate deformation extraction. 2. This invention uses a frequency domain decomposition module to separate vibration artifacts and employs a band-stop filter to accurately filter out vibration interference within a specific frequency range, effectively separating vibration artifacts from the real deformation signal. This avoids distortion of deformation detection results based on the frame difference method due to vibration interference. 3. This invention improves the measurement accuracy of deformation calculation by using sub-pixel edge extraction and registration technology; by superimposing the exponential decay model with characteristic time constant and the time-varying driving term, and combining it with online parameter estimation using recursive least squares method, the deformation prediction model can achieve real-time adaptation to changes in working conditions such as roughing cutting parameters, initial stress state of blank and workshop ambient temperature, thus avoiding prediction errors caused by the accumulation of offline calibration parameters over time. Attached Figure Description

[0015] The invention will now be further described with reference to the accompanying drawings.

[0016] Figure 1 This is a flowchart of a method for compensating for positioning deviations in tube sheet and pump casing processing based on image correction in this invention; Figure 2 This is a flowchart illustrating the logic of preprocessing and subpixel edge registration of temporal image sequences in this invention. Detailed Implementation

[0017] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0018] Example 1: As Figure 1 and Figure 2 As shown, this invention provides a method for compensating for positioning deviations in the processing of tube sheets and pump casings based on image correction, comprising the following steps: In step one, the system acquires a time-series image sequence of the tube sheet surface to be tested; The industrial camera selected is a 16-megapixel resolution CMOS industrial camera, paired with a 35mm telecentric lens to eliminate perspective distortion. The strategy for selecting the measurement area is to mark a circular measurement area with a diameter of 600 mm on the end face of the tube sheet flange. This area covers the annular zone between the inner and outer diameters of the flange sealing surface. The image acquisition timing control module performs continuous image acquisition according to preset parameters. The acquisition cycle is set to 1 second, that is, one frame of image is acquired and stored every second. The exposure time for a single frame image is set to 2 milliseconds. This exposure time parameter was determined through a prior imaging quality assessment experiment to ensure that while obtaining a sufficient image signal-to-noise ratio, the deterioration of imaging sharpness due to thermal deformation is effectively suppressed. The total image acquisition time within the time limit is set to 24 hours, that is, 86,400 frames of images are continuously acquired to form a time-series image sequence; During image acquisition, the acquisition timestamp information of each frame is recorded synchronously for subsequent correlation analysis of time-series data. The ambient temperature sensor records the workshop temperature data in real time at a sampling frequency of 5 Hz. The temperature data is stored in alignment with the timestamp of the image sequence to provide environmental parameter input for the time-varying driving term in the subsequent deformation prediction model.

[0019] In step two, the system performs motion blur correction processing on the temporal image sequence. This processing step utilizes a pre-trained deep convolutional neural network, constructed using a generative adversarial network architecture. It comprises two functional modules: a blur kernel prediction subnetwork and an image complex atom network. The blur kernel prediction subnetwork employs an encoder-decoder symmetric structure, with the encoder consisting of four convolutional layers and one pooling layer. The first convolutional layer contains 64 convolutional kernels, each with a size of 3×3 pixels. The activation function is a modified linear unit. The input image first passes through this layer for low-level feature extraction. The second convolutional layer contains 128 convolutional kernels and continues to perform feature abstraction; The third convolutional layer contains 256 convolutional kernels to further extract intermediate semantic features; the fourth convolutional layer contains 512 convolutional kernels to complete high-level feature encoding. The pooling operation adopts max pooling with a pooling window size of 2×2 pixels and a stride of 2 pixels. It is set after the fourth convolutional layer to achieve spatial downsampling of the feature map. The decoder part contains four deconvolutional layers and one output layer. The deconvolution kernel size is uniformly 3×3 pixels. The spatial resolution of the feature map is gradually restored through deconvolution operation. The output layer is responsible for outputting the horizontal and vertical component parameters of the blur kernel. The size of the blur kernel is set to 7×7 pixels. The blur kernel parameters are normalized before output to ensure the physical reasonableness of the estimation results.

[0020] The training sample set construction process of this deep convolutional neural network is as follows: First, a total of 2,000 real-shot images of the φ800mm tube sheet flange surface under different processing conditions are collected as the basic sample library; The training set is constructed by applying different degrees of motion blur to the base samples through artificial synthesis. Specifically, the blur kernel with a size in the range of 5×5 to 9×9 pixels is used to perform convolution operation on the clear image to generate degraded image pairs containing different degrees of motion blur. The direction angle of the blur kernel is uniformly sampled in the range of 0 degrees to 180 degrees, and the blur degree parameter of the blur kernel is distributed in an arithmetic sequence in the range of 0.5 to 3.0 pixels. During training, a joint optimization strategy of perceptual loss function and adversarial loss function is adopted: the perceptual loss function extracts high-level semantic features based on the pre-trained VGG16 network and calculates the Euclidean distance between the deblurred output image and the clear reference image in the feature space. The adversarial loss function evaluates the realism of the deblurred image through the discriminator network, improving the visual quality of the results generated by the deblurring network. In the joint optimization process, the weight ratio of the two loss terms is set as 0.6 for perceptual loss and 0.4 for adversarial loss. The training epochs are set to 100, the batch size is set to 8, and the initial learning rate is set to 0.0001 and dynamically adjusted using a cosine annealing strategy.

[0021] Each frame in the time-series image sequence is processed sequentially through a blur kernel prediction sub-network and an image complex atom network: The current frame image is input into the blur kernel prediction subnetwork. The network forward inference outputs the 7×7 pixel blur kernel parameters corresponding to the current frame. These parameters include three dimensions: blur direction angle, blur length, and blur intensity. The predicted blur kernel parameters, along with the current frame image, are input into the image complex atom network. The image complex atom network estimates the original sharp image based on the predicted blur kernel parameters and outputs the deblurred sharp image using an end-to-end image-to-image translation method. The edge sharpness of each frame in the deblurred temporal image sequence is improved, and the edge modulation transfer function value is increased from the original 0.3 to over 0.85, providing a high-quality image data foundation for subsequent subpixel edge detection.

[0022] In step three, the system performs vibration artifact separation processing on the time-series image sequence. The input to the vibration artifact separation module is the corrected clear image sequence output in step two. The module contains three functional components: a frequency domain conversion unit, a band-stop filter parameter configuration unit, and a time domain recovery unit. The frequency domain conversion unit performs fast Fourier transform processing on 32 consecutive frames of corrected images to convert the images from the spatial domain to the frequency domain in order to reveal the frequency composition characteristics of the image signal; In the frequency domain, vibration interference signals exhibit energy concentration characteristics along a specific frequency axis, and their spectral distribution is highly correlated with the frequency characteristics of workshop floor vibration.

[0023] The band-stop filter parameter configuration unit sets the filter parameters according to the spectral characteristics of the vibration interference. In specific implementation, vibration environment spectrum acquisition and analysis are performed during the system startup phase. Vibration acceleration signals are continuously acquired for 10 seconds at a sampling frequency of 100 Hz by an accelerometer installed on the machine tool workbench, and the vibration spectrum distribution is obtained by performing a fast Fourier transform on the acquired signals. Assuming that the vibration spectrum analysis results in this embodiment show that the main vibration frequency is concentrated in the range of 5 Hz to 12 Hz and the energy peak occurs at 8 Hz, then the center frequency of the band-stop filter is automatically adjusted to 8 Hz and the stopband width is set to the range of 2 Hz to 3 Hz. The filter type selected is a fourth-order infinite impulse response Butterworth filter to obtain a flat passband response and steep stopband attenuation characteristics. The frequency response characteristics of the Butterworth filter exhibit a monotonically smooth attenuation trend in the passband, avoiding the image detail loss problem caused by passband ripple in Chebyshev filters.

[0024] After filtering, the frequency domain image data is restored to the spatial domain by inverse fast Fourier transform, generating a clean image sequence with vibration artifacts removed. During the inverse transform process, the recovered image is normalized to eliminate the amplitude deviation introduced by numerical calculation. The effectiveness of vibration artifact separation is verified by the following methods: During machine tool no-load operation, continuous image acquisition was performed on the same measurement area for 30 seconds. The image sequences before and after separation processing were subjected to inter-frame difference operation. Before separation processing, the root mean square value of the inter-frame displacement of adjacent frames was 0.15 pixels. After separation processing, the value was reduced to 0.02 pixels, and the vibration artifact suppression efficiency reached 87%.

[0025] In step four, the system performs subpixel edge extraction and registration processing on the time-series image sequence. The goal of this processing step is to improve the edge positioning accuracy from the pixel level to the subpixel level, providing high-precision geometric feature data for subsequent deformation calculation. The subpixel edge detection adopts a subpixel edge positioning algorithm based on gray-scale moments. The implementation process of this algorithm includes three stages: image smoothing and denoising, edge point set extraction, and subpixel position calculation.

[0026] In the image smoothing and denoising stage, a Gaussian filter is used to perform convolution operations on the input image to suppress random noise introduced during image acquisition and transmission. The kernel size of the Gaussian filter is set to 5×5 pixels, and the sigma parameter is set to 1.5 pixels. This parameter configuration is determined through the following evaluation process: Imaging experiments with different sigma parameter values ​​were conducted on tube sheet flange test pieces to evaluate the trade-off between edge positioning accuracy and noise suppression effect. When the sigma parameter is less than 1.0 pixels, the noise suppression effect is insufficient, resulting in an edge positioning standard deviation greater than 0.05 pixels; when the sigma parameter is greater than 2.0 pixels, excessive smoothing results in an edge position offset greater than 0.03 pixels. When the sigma parameter is 1.5 pixels, the edge positioning accuracy and noise robustness reach the optimal balance.

[0027] In the edge point set extraction stage, the Canni edge detection operator is used to identify edge feature points in the image. The parameter configuration for the Canni edge detection is as follows: The low threshold is set to 0.3, and the high threshold is set to 0.7. This threshold pairing is determined by analyzing the edge response characteristic curve of the tube sheet flange test piece to ensure that complete edges can be detected in high signal-to-noise ratio areas while avoiding the introduction of false edges. The minimum edge length threshold used in the edge connection process is 10 pixels, meaning that broken edges with a length of less than 10 pixels do not participate in subsequent registration processing. The initial edge point set output by edge detection is stored in the form of a coordinate sequence, and each edge point contains three attribute information: pixel coordinates, gradient magnitude, and gradient direction.

[0028] In the sub-pixel position calculation stage, the initial edge points are calculated with sub-pixel precision based on the principle of gray-level moment invariance. The principle of the gray-level moment method is to use the gray-level distribution of pixels near the edge to satisfy the condition of first-order moment invariance. By fitting the local gray-level distribution curve, the sub-pixel position of the edge is accurately determined. The fitting window size used in the sub-pixel positioning calculation is 5×5 pixels. The determination of this window size takes into account both the statistical independence of pixels near the edge and the accuracy requirements of edge direction estimation. When the fitting window is too large, it will introduce more interference from nearby edges, leading to an increase in positioning deviation. When the fitting window is too small, the statistical samples are insufficient, leading to an increase in estimation variance. After sub-pixel edge positioning processing, the edge positioning accuracy is improved from 1 pixel to 0.08 pixels, which meets the subsequent accuracy requirement of more than 0.1 pixels.

[0029] The edge registration process employs a feature point matching strategy based on an affine transformation model; Harris corner points are extracted as feature points from the sub-pixel edge point sets of two adjacent frames. The threshold of the response function for corner point detection is set to 100 to ensure that the selected feature points have good localization repeatability. The normalized cross-correlation algorithm is used for feature point matching. The matching cost matrix is ​​calculated and high-quality matching point pairs are screened through bidirectional consistency test. In the matching point pairs, the random sampling consistency algorithm is used to remove mismatched points. The number of iterations of the random sampling consistency algorithm is set to 1000 times and the inlier threshold is set to 2 pixels. After iterative optimization, the set of inliers that meet the constraints of the affine transformation model is retained. The six parameters of the affine transformation model are estimated using the least squares method based on the set of interior points, thereby calibrating the geometric transformation relationship between adjacent frames. The affine transformation model can describe two-dimensional geometric distortions such as translation, rotation, scaling, and shearing between images, and has a complete descriptive ability for in-plane deformation of the tube sheet flange surface during its aging period.

[0030] In step five, the system performs time-series extraction of processing deformation and construction of a prediction model. The goal of this step is to extract time-series data of deformation from the geometric transformation relationship between adjacent frame images obtained in step four, and to establish a prediction model that can accurately describe the deformation evolution law.

[0031] The deformation time sequence is extracted as follows: For each pair of adjacent frames in the time-series image sequence, the horizontal translation, vertical translation, and rotation transformation parameters are extracted from the affine transformation model. The rotation transformation parameters are only used for auxiliary correction in this calculation, while the horizontal and vertical translations directly reflect the displacement of the tube sheet flange surface on the image plane. These two translations are in pixels and need to be multiplied by a conversion coefficient predetermined by the camera calibration. In this embodiment, the coefficient is 0.05 mm per pixel, and the movement of one pixel in the image is equivalent to a deformation of 0.05 mm in actual physical space; After converting the horizontal and vertical translations into actual millimeters, the square root of the sum of the squares of these two orthogonal displacement components is taken, which is to calculate their composite modulus. This modulus is defined as the increase in deformation generated in the current frame relative to the previous frame. For example, if the horizontal translation of a frame is 4 pixels and the vertical translation is 3 pixels, then the converted horizontal translation is 0.2 mm and the vertical translation is 0.15 mm. Their combined modulus is 0.25 mm, which represents the amount of deformation added during this time interval. By accumulating the deformation increments corresponding to each frame in chronological order, the cumulative deformation from the initial moment to the current moment is obtained, forming a sequence of deformation changes over time. The number of data points in this sequence is exactly equal to the total number of frames in the image sequence.

[0032] The mathematical structure of the deformation prediction model adopts the superposition of an exponential decay model with a characteristic time constant and a time-varying driving term. The structural design of this model is based on the following physical considerations: the exponential decay term is used to describe the deformation decay trend caused by the relaxation of residual stress inside the workpiece. The release process of residual stress approximately follows the exponential decay law, and the characteristic time constant τ reflects the characteristic rate of stress relaxation. The mathematical structure of the deformation prediction model consists of the superposition of two physical mechanisms; The first term is an exponential decay form, used to characterize the deformation decay trend caused by the gradual release of residual stress inside the workpiece during the aging process. This release rate is determined by the inherent characteristics of the material and is concentrated in a characteristic time constant. The second term is the time-varying driving term, which reflects the additional deformation disturbance caused by the fluctuation of ambient temperature to the workpiece. This disturbance is linearly proportional to the amount of temperature change relative to the reference temperature. Its proportionality coefficient, i.e. the amplitude coefficient, integrates the combined effects of the material's thermal expansion characteristics and the overall stiffness of the component. The output deformation of the model at any given moment is equivalent to the sum of the component that decreases with time according to the exponential decay law starting from the deformation at the initial moment, and the component obtained by multiplying the amplitude coefficient by the temperature change at that moment. If the initial deformation is 0.50 mm and the characteristic time constant is 12 hours, then the attenuation term contributes approximately 0.18 mm when aging reaches 12 hours. If the temperature is 2 degrees Celsius higher than the reference temperature and the amplitude coefficient is 0.03 mm per degree Celsius, then the driving term contributes 0.06 mm, and the current total deformation is 0.24 mm. The three undetermined parameters in the model, namely the initial deformation, the characteristic time constant, and the amplitude coefficient, are not determined offline at once, but are updated online successively during the aging process using the real-time recursive least squares method. This recursive estimation mechanism obtains the current measured deformation at each new observation time, combines it with the predicted value constructed from the regression vector of the previous time, calculates the deviation between the two, and then uses the Kalman gain as a weight to use the deviation to correct the parameter estimation vector of the previous time, thereby obtaining the new parameter estimation result at the current time. The covariance matrix, which represents the uncertainty of the estimate, is also updated synchronously. The update method introduces a forgetting factor with a fixed value of 0.97. This factor balances historical information and current information by adjusting the relative weights of old and new data in the covariance matrix update. A forgetting factor of 0.97 means that historical data retains approximately 97% of its weight after each update step. The specific value is chosen based on the following trade-offs. In this embodiment, for Q345R material and its corresponding roughing process conditions, the candidate value of the forgetting factor is preset to 0.97. The system automatically selects this value based on the initially input processing condition information to ensure a balance between stability and response speed.

[0033] The parameter initialization strategy of recursive least squares method adopts the following strategy: image data collected in the first 10 minutes of the effective period is used for offline batch estimation to obtain the initial value of the parameter. This initial value serves as the starting point for online recursive estimation, ensuring that the parameter estimation converges quickly while avoiding the impact of the initial estimation bias on the prediction accuracy. The convergence of parameter estimation is judged by monitoring the change in parameter estimate. When the parameter update amount is less than the preset threshold of 0.001 for 10 consecutive iterations, the parameter estimation is considered to have converged to the steady state value.

[0034] In step six, the system calculates and outputs the compensation amount for the positioning deviation. The goal of this step is to calculate the compensation instructions that the CNC machining system needs to execute to offset the predicted deformation based on the deformation prediction model established in step five.

[0035] The predictive control strategy for compensation calculation is implemented as follows: the current time is set as the starting point, the prediction time domain is set to the range from the current time to the next eight sampling periods, i.e., the next eight seconds, and the performance index in the rolling optimization process is composed of two weighted terms: the prediction deformation deviation term and the compensation smoothing term. Among them, the predicted deformation deviation term is the sum of the squares of the deviation of the predicted deformation amount from the target value at each prediction time, and the compensation amount smoothing term is the sum of the squares of the change of the compensation amount at each step relative to the compensation amount at the previous time. The weight coefficient of the former is set to 0.8, and the weight coefficient of the latter is set to 0.2. The purpose of this weight configuration is to suppress the drastic fluctuation of the compensation amount while ensuring the accuracy of deformation compensation, and to ensure the stability of the CNC machining system. The rolling optimization process uses a sequential quadratic programming algorithm to solve the optimization problem. The constraint condition of the optimization variable is set to the absolute value of the compensation increment within each sampling period not exceeding 0.01 mm, so as to avoid the impact of sudden changes in compensation commands on the CNC system. The optimized solution output is the optimal compensation amount sequence at the current time. The first element of this sequence is taken as the compensation amount command output at the current time. The compensation output adopts an incremental output method, that is, what is sent to the CNC system is the difference between the current compensation and the compensation at the previous moment, rather than the absolute value of the compensation. This design effectively avoids the problem of cumulative error of the compensation during the communication transmission process. If the compensation amount at the previous moment was 0.15 mm, and the optimal absolute compensation amount obtained by the current optimization is 0.17 mm, then the system only sends an incremental command of positive 0.02 mm to the CNC system, and the CNC system performs this incremental adjustment based on the absolute compensation amount maintained internally. After the compensation amount is output, the CNC machining system adjusts the positioning parameters of the tool relative to the workpiece according to the compensation command. The adjustment amount is the opposite of the compensation amount, i.e., the compensation amount is subtracted to offset the deformation effect. The response delay time of the compensation adjustment is measured to be 0.3 seconds through system identification experiments. In the rolling optimization process of the predictive controller, this delay time is incorporated into the input delay compensation link of the predictive model to improve the prediction accuracy. In step seven, the system performs closed-loop verification of the compensation effect and adaptive correction of the model. This step evaluates the compensation effect by re-acquiring image data and comparing it with the predicted values, and triggers adaptive correction of model parameters when necessary. The implementation process of closed-loop verification is as follows: after the CNC machining system executes the compensation command, wait for two sampling cycles, i.e., two seconds, to ensure that the compensation effect is fully manifested; Re-acquire images of the tube sheet flange surface measurement area. The re-acquired images are processed according to the complete processing flow from step one to step six to obtain new deformation measurement values. These measurement values ​​are compared with the predicted values ​​before compensation, and the absolute deviation between the two is calculated. The error threshold is dynamically configured according to the tube sheet specifications. The 800 mm diameter tube sheet in this embodiment belongs to the medium-sized tube sheet specification, and the error threshold is set to 0.06 mm. This threshold setting follows the principle of matching the workpiece size and accuracy requirements. For medium-sized tube sheets with diameters between 800 mm and 1200 mm, the error threshold is set to a range of 0.05 mm to 0.08 mm. When the absolute deviation is greater than 0.06 mm, the deviation is determined to exceed the error threshold, triggering the adaptive correction mechanism of the model parameters. When the triggering condition of adaptive correction of model parameters is met, the system uses new measurement data to recursively update the parameters of the deformation prediction model. The specific update process is to use the new measurement data as a component of the observation value and the regression vector, and substitute it into the parameter update equation of the recursive least squares method to calculate the updated parameter estimate. After the parameters are updated, the updated parameters are saved to the model parameter storage unit and used for subsequent prediction calculations. When the absolute deviation is less than or equal to 0.06 mm, the deviation is determined to be within the error threshold range. The model parameters are kept unchanged, and the process returns to step one to continue the next round of monitoring. The complete processing flow of this embodiment achieves precise compensation for positioning deviations in tube sheet precision machining through the coordinated execution of the above seven steps; The parameter estimation of the deformation prediction model converged within 30 minutes after the start of the effective period, with the convergence value of the characteristic time constant being 4.2 hours, and the deviation from the offline calibration value being less than 5%. The compensated assembly error is stably controlled within the range of 0.10 mm to 0.13 mm, meeting the requirement of a 0.15 mm fit clearance; the model adaptive correction mechanism can be correctly triggered and effectively reduce the prediction error when the compensation deviation exceeds the threshold.

[0036] Example 2: The ambient temperature in the workshop of a certain manufacturing plant is about 8°C in winter. The effective value of the ground vibration acceleration is 0.015g. The main vibration frequency is concentrated in the range of 3 Hz to 8 Hz. The weight of a single pump casing is about 3.5 tons. The aging period after rough machining is 36 hours. Before fine machining, it is necessary to ensure that the deformation of the sealing surface is controlled within 0.12mm.

[0037] In step one, the system acquires a time-series image sequence of the pump casing sealing surface. Unlike in Example 1, the sealing surface of the pump casing is a conical sealing surface structure rather than a planar flange surface. Therefore, the selection strategy for the measurement area needs to be adjusted according to the geometric characteristics of the sealing surface. A 50mm wide annular measurement zone is marked on the conical surface of the pump casing sealing surface. The radial position of this annular zone corresponds to the sealing mating line. The image acquisition system is configured with a 20-megapixel resolution CMOS industrial camera paired with a 50mm telecentric lens to accommodate a larger measurement range. The acquisition period was set to 0.8 seconds, the exposure time was set to 3 milliseconds, and the total image acquisition time was set to 36 hours, which means continuously acquiring 162,000 frames of images.

[0038] Due to the conical geometry of the pump casing sealing surface, perspective distortion occurs in the image. Before image acquisition, a camera extrinsic calibration model needs to be established to eliminate the impact of perspective effect on the accuracy of deformation calculation. The calibration process uses a nine-point calibration method to arrange nine precisely machined marking points on the pump casing sealing surface. The tilt angle and rotation angle parameters of the camera relative to the normal of the sealing surface are solved by the calibration algorithm. These parameters are used to compensate for perspective distortion effects in subsequent deformation calculations.

[0039] In step two, the system performs motion blur correction processing on the time-series image sequence. The deep convolutional neural network architecture in this step is consistent with that in Example 1, but the composition of the training sample set has been adaptively adjusted according to the material characteristics of the pump casing sealing surface. Duplex stainless steel surfaces have good diffuse reflection properties and exhibit different grayscale distribution characteristics under different lighting angles. Therefore, the training sample set needs to collect images of the pump casing sealing surface under various lighting conditions to improve the network's adaptability to material properties. The training sample set contains 1,500 real-world images and 4,500 artificially synthesized degraded images. In addition to applying motion blur during the degradation process, specular reflection artifact simulation is also added to enhance the network's robustness to imaging characteristics of complex material surfaces.

[0040] In step three, the system performs vibration artifact separation processing on the time-series image sequence. Unlike embodiment one, the vibration environment spectrum analysis results in this embodiment show that the main vibration frequency is concentrated in the range of 3 Hz to 8 Hz. Therefore, the parameters of the band-stop filter need to be adjusted accordingly. The center frequency is set to 5.5 Hz, and the stopband width is set to the range of 1.5 Hz to 2 Hz, which is 20-30% of the main frequency range; The same filter type and order design as in Example 1 ensures the stability of the passband response and stopband attenuation characteristics.

[0041] In step four, the system performs subpixel edge extraction and registration processing of time-series image sequences. Due to the conical geometric features of the pump casing sealing surface, the transformation model for edge registration needs to be adjusted from the affine transformation model in Example 1 to the projection transformation model in order to accurately describe the pose changes in three-dimensional space. The projection transformation model contains eight independent parameters, which, in addition to describing translation and rotation in the plane, can also describe the deformation effect introduced by perspective projection; The edge registration process first extracts scale-invariant feature transformation feature points from adjacent frame images and matches them, and then estimates the eight parameters of the projection transformation model through the random sampling consensus algorithm. The subpixel edge positioning algorithm is consistent with that in Example 1, ensuring that the edge positioning accuracy meets the requirement of more than 0.1 pixels.

[0042] In step five, the system extracts the time sequence of the processing deformation and constructs a prediction model; the pump casing is made of duplex stainless steel and weighs 3.5 tons per piece. The distribution of residual stress and the relaxation rate inside it are different from those of the tube sheet. The characteristic parameters of the deformation prediction model need to be recalibrated. The initial estimate of the characteristic time constant was set at 6.5 hours, which is longer than the 4.2 hours for the tube sheet, reflecting the slow stress relaxation process of large and heavy castings. The amplitude coefficient of the time-varying drive term was set to 0.025 mm per degree Celsius. This parameter was calibrated by conducting a temperature shock experiment in a controlled temperature environment. The forgetting factor λ of the recursive least squares method is set to 0.98 to adapt to the relatively slow changes in pump casing conditions.

[0043] In step six, the system calculates and outputs the compensation amount for the positioning deviation. Considering that the pump casing is precision machined using a CNC boring and milling machine, the interface protocol for the compensation amount output is different from that in Example 1. The compensation amount is sent to the CNC system in an absolute output manner, rather than an incremental value. This is because the CNC system of the boring and milling machine adopts an absolute coordinate programming method, and the absolute compensation amount output can be directly superimposed on the workpiece coordinate system without the need for additional accumulation calculation. The predictive control strategy for calculating the compensation amount is consistent with that in Example 1. The prediction time domain is set to 8 sampling periods, and the weight coefficient configuration of the performance index function is the same as that in Example 1.

[0044] In step seven, the system performs closed-loop verification of the compensation effect and adaptive correction of the model. In this embodiment, the φ1200mm pump casing belongs to a large tube sheet specification, and the error threshold is set to 0.10 mm. This threshold setting is determined according to the specifications and dimensions of the pump casing and the requirements for the fitting clearance. For large tube sheets with diameters between 1200 mm and 1500 mm, the error threshold is set in the range of 0.08 mm to 0.12 mm. The model adaptive correction mechanism is consistent with that of Example 1, but the sensitivity parameter for triggering correction has been adjusted to adapt to the relatively slow deformation process of large workpieces. Specifically, in this embodiment, the number of triggers is adjusted to 5 consecutive times, and the effective judgment interval of the residual threshold is expanded from a single sampling time to the average value within the sliding time window. The width of the sliding window is set to 10 sampling periods. Only when the average residual value within the window continuously exceeds the error threshold of 0.10 mm and the duration exceeds 8 sampling periods will the model parameter update of the recursive least squares method be triggered, so as to avoid frequent miscorrection of the model caused by accidental residual abnormalities due to workpiece deformation response lag.

[0045] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for compensating for positioning deviations in tube sheet and pump casing machining based on image correction, comprising the following steps: Step 1: Acquire time-series image sequences of the tube sheet and pump casing surfaces to be tested; Step 2: Perform motion blur correction processing on the acquired time-series image sequence; Step 3: Perform vibration artifact separation processing on the time-series image sequence; Step 4: Perform subpixel edge extraction and registration of the time-series image sequence; Step 5: Extract the temporal data of processing deformation based on the temporal image sequence of adjacent frames, and construct a deformation prediction model; Step 6: Predict the deformation amount at future times based on the deformation prediction model, calculate the compensation amount for the positioning deviation based on the prediction results and output it. Step 7: After compensation, re-acquire time-series image sequences, compare the measured deformation values ​​with the predicted values ​​predicted by the deformation prediction model, and recursively update the parameters of the deformation prediction model.

2. The method for compensating for positioning deviations in tube sheet and pump casing machining based on image correction according to claim 1, characterized in that, The steps to complete motion blur correction include: Each frame of the temporal image sequence is input into a pre-trained deep convolutional neural network to perform motion blur kernel estimation. Deep convolutional neural networks are built on a generative adversarial network architecture, which includes a fuzzy kernel prediction subnetwork and an image complex atom network; The blur kernel prediction subnetwork receives the current frame image and outputs the corresponding motion blur kernel parameters; The image complex atom network deblurs the current frame image based on the predicted motion blur kernel parameters, generating a corrected and clear image sequence.

3. The method for compensating for positioning deviations in tube sheet and pump casing machining based on image correction according to claim 2, characterized in that, The steps for separating vibration artifacts from time-series image sequences include: The corrected, clear image sequence is input into the frequency domain decomposition module for vibration artifact separation; The frequency domain decomposition module first performs a fast Fourier transform on multiple consecutive frames of corrected images, converting them from the spatial domain to the frequency domain; In the frequency domain, the parameters of the band-stop filter are set according to the spectral characteristics of the vibration interference. After the filtering process is completed, the image is restored to the spatial domain by inverse fast Fourier transform to generate a clean image sequence with vibration artifacts removed.

4. The method for compensating for positioning deviations in tube sheet and pump casing machining based on image correction according to claim 1, characterized in that, The steps for subpixel edge extraction and registration of time-series image sequences include: Subpixel edge detection and registration are performed on adjacent frames in a clean image sequence after removing vibration artifacts. Subpixel edge detection employs a subpixel edge localization algorithm based on grayscale moments; The subpixel edge localization algorithm first uses a Gaussian filter to smooth and denoise the image; The initial edge point set is extracted using the Canney edge detection operator; The initial edge points are calculated with sub-pixel precision based on the principle of gray-scale moment invariance.

5. The method for compensating for positioning deviations in tube sheet and pump casing machining based on image correction according to claim 1, characterized in that, The steps for extracting time-series data of processing deformation based on adjacent frame time-series image sequences and constructing a deformation prediction model include: Based on the geometric transformation relationship between adjacent frame images, calculate the deformation increment of the tube sheet flange surface or pump casing sealing surface within the time sampling interval.

6. The method for compensating for positioning deviations in tube sheet and pump casing machining based on image correction according to claim 5, characterized in that, The cumulative deformation increments are arranged in chronological order to construct a time-series data sequence of deformation evolution over time. The deformation prediction model adopts an exponential decay model with a characteristic time constant.

7. The method for compensating for positioning deviations in tube sheet and pump casing machining based on image correction according to claim 5, characterized in that, The deformation prediction model is estimated and updated online using the real-time recursive least squares method; The recursive least squares method employs an adaptive update strategy with a forgetting factor.

8. The method for compensating for positioning deviations in tube sheet and pump casing machining based on image correction according to claim 6, characterized in that, The steps include predicting the deformation amount at future moments based on the deformation prediction model, calculating the compensation amount for the positioning deviation based on the prediction results, and outputting the results: The deformation amount at future moments is predicted based on the deformation prediction model. The predicted deformation amount is compared with the preset fit gap threshold, and the direction and magnitude of the compensation amount are calculated. The compensation amount calculation adopts a predictive control strategy, which predicts several sampling periods in advance based on the known deformation evolution trend at the current moment, and determines the compensation amount at the current moment based on the prediction results.

9. The method for compensating for positioning deviations in tube sheet and pump casing machining based on image correction according to claim 1, characterized in that, After compensation, time-series image sequences are reacquired, and the measured deformation values ​​are compared with the predicted values ​​from the deformation prediction model. The steps for recursively updating the parameters of the deformation prediction model include: After the CNC machining system executes the compensation command, the measurement area of ​​the tube sheet or pump casing is re-acquired.

10. The method for compensating for positioning deviations in tube sheet and pump casing machining based on image correction according to claim 1, characterized in that, The deformation measurement values ​​generated by image acquisition are compared with the predicted values ​​before compensation; The parameters of the deformation prediction model are updated recursively using the new measurement data.

Images