A large numerical control grinding machine thermal deformation measurement system and method based on machine vision

Abstract

The application discloses a large numerical control grinding machine thermal deformation measurement system and method based on machine vision. The system is composed of an image acquisition module, a data preprocessing module and a data extraction module. The image acquisition module installs a high-resolution industrial camera through a clamp to collect auxiliary mark images of a to-be-measured area of a machine tool. The data preprocessing module combines a long short-term memory network and an adaptive robust Kalman filter to perform noise suppression and dynamic filtering optimization on the images. The data extraction module uses a diffusion model to perform edge detection and converts pixel displacement into micron-level spatial physical displacement through a sub-pixel algorithm to realize accurate quantification of thermal deformation. During measurement, the images are repeatedly collected and processed at a set interval to obtain thermal deformation time series data. The application combines deep learning and filtering technology to significantly improve the accuracy and robustness of measurement under complex working conditions.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of high-end machine tool precision control, and relates to a large numerical control grinding machine thermal deformation measurement system and method based on machine vision. BACKGROUND

[0002] With the increasing requirements of high-end equipment manufacturing industry on machining precision, efficiency and stability, as the core machining equipment, the optimization of the comprehensive performance of numerical control machine tools has become the focus of the industry. During the operation of the machine tool, the thermal deformation problem caused by internal heat sources (such as spindle motor, cutting process) and external environmental temperature fluctuations is one of the main factors leading to the deterioration of machining precision. According to statistics, 60%~70% of the errors in precision machining are caused by thermal deformation of the machine tool, especially on large numerical control grinding machines and other complex structure machine tools, the thermal deformation presents the characteristics of multi-degree-of-freedom coupling, nonlinear time-varying, etc., which puts forward higher requirements on measurement and compensation technology.

[0003] Traditional thermal characteristic analysis focuses on the temperature-deformation relationship of local components or single degree of freedom, and it is difficult to fully reflect the thermal characteristics of the whole machine. In view of the characteristics of large numerical control grinding machine, such as complex structure, wide heat source distribution and variable working condition, combined with simulation analysis results, by means of international standard system and advanced measurement technology, precise and efficient multi-degree-of-freedom thermal deformation measurement and analysis are carried out, multi-degree-of-freedom, precise and efficient data acquisition and analysis in thermal deformation measurement are realized, and the purpose is to provide theoretical support and practical guidance for thermal stability optimization of large numerical control grinding machine.

[0004] At present, the thermal deformation of the machine tool is usually indirectly measured by temperature sensors and laser interferometers. Through the high monochromaticity, high directivity and strong coherence of the laser, the interference phenomenon occurs through the lens and the mirror, and the thermal deformation of the machine tool is calculated by causing the optical path difference at the spindle, and a temperature-displacement model is established with the temperature measurement results. In the work, the temperature data and the laser equipment are combined, and the thermal deformation is predicted through the temperature data, but the above-mentioned method still has the following defects: 1. Low measurement efficiency. Due to the light path characteristics of the laser, only a single degree of freedom of displacement in that direction can be obtained by single measurement. With the development of laser interferometer, the laser interferometer on the market can only measure six degrees of freedom of a single feed axis; 2. Weak robustness. When high-precision measurement is performed, the environmental requirements are relatively high, and the light path is easily affected by air disturbance, temperature fluctuation, vibration, etc., resulting in fluctuation of the light path; 3. High measurement limitation. Considering the installation of the lens and the mirror, the laser interferometer can usually only be used for spindle position deviation measurement; 4. Complex measurement system. A large number of temperature sensors, laser interferometers and data acquisition equipment need to be arranged in the test system, and there are problems of sensor synchronization and data redundancy. SUMMARY

[0005] The purpose of the present application is to provide a large CNC grinding machine thermal deformation measurement system and method based on machine vision to solve the above problems.

[0006] The purpose of the present application can be realized by the following technical solutions: a large CNC grinding machine thermal deformation measurement system based on machine vision, characterized in that it comprises:

[0007] An image acquisition module comprising at least one industrial camera installed at a predetermined distance from a specific part of the machine tool through a clamp, for acquiring image information with auxiliary markers and transmitting to a computer;

[0008] A data preprocessing module in communication with the image acquisition module and integrated with a long short-term memory neural network and an adaptive robust Kalman filter unit for noise suppression and dynamic filtering of the original image;

[0009] A data extraction module in communication with the data preprocessing module, comprising a diffusion model edge detection unit and a sub-pixel edge fitting unit for extracting edge features from the filtered image and converting pixel displacement into spatial micron-level displacement values to realize quantitative measurement of thermal deformation.

[0010] A large CNC grinding machine thermal deformation measurement method based on machine vision, characterized by the following steps:

[0011] S1: Set auxiliary markers on the surface of the measured area of the machine tool, install not less than one industrial camera at a predetermined distance through a clamp, calibrate the industrial camera and set the shooting interval ∆t;

[0012] S2: Start the large CNC grinding machine, acquire the image of the surface of the measured area at the initial time t0, and transmit the image to the computer;

[0013] S3: Use the data preprocessing module to perform noise suppression, time domain filtering and quality optimization on the image data obtained in step S2;

[0014] S4: Process the image output by step S3 through the data extraction module, use the diffusion model for edge detection, and further use the sub-pixel edge fitting algorithm for curve fitting and pixel-level positioning of the auxiliary marker edge points, and combine the camera intrinsic parameters to realize quantitative extraction of thermal deformation;

[0015] S5: Repeat steps S2 to S4 every ∆t time interval to obtain time series data of the thermal deformation of the machine tool.

[0016] In the above-mentioned large CNC grinding machine thermal deformation measurement method based on machine vision, the industrial camera is a line array camera or a plane array camera, and is equipped with a fixed focus industrial lens; when the industrial camera is a plane array camera, it constitutes a binocular measurement system, and the single camera resolution is not less than 12 million pixels.

[0017] In the above-mentioned large CNC grinding machine thermal deformation measurement method based on machine vision, the area to be measured is a linear feed mechanism, a spindle or a worktable.

[0018] In the above-mentioned large CNC grinding machine thermal deformation measurement method based on machine vision, the installation angle of the industrial camera and the axis of the area to be measured should be 30°-45°, the installation angle interval of multiple cameras should be 45°-90°, the interval between the camera and the area to be measured should be 0.3-0.5 meters, and the industrial camera acquisition interval Δt is set to 5-15 seconds.

[0019] In the above-mentioned large CNC grinding machine thermal deformation measurement method based on machine vision, the adaptive robust Kalman filter process in step S3 includes:

[0020] The prediction step estimates the prior state and error covariance based on the state equation;

[0021] Calculate the residual and use the influence function to suppress the disturbance of abnormal values to the state;

[0022] According to the innovation sequence, the observation noise covariance is dynamically adjusted;

[0023] Update the Kalman gain to get the optimal state estimation;

[0024] At the same time, based on the gating mechanism of the long short-term memory neural network, the historical filtering error sequence is analyzed, and the optimized prediction value of the image feature at the current time t is output.

[0025] In the above-mentioned large CNC grinding machine thermal deformation measurement method based on machine vision, the edge detection of the diffusion model in step S4 includes:

[0026] In the forward diffusion process, Gaussian noise is gradually added to the input image data;

[0027] In the inverse diffusion process, the trained denoising model is used for multiple iterations to gradually restore the edge features of the image;

[0028] The pixel-level edge position detected by the sub-pixel edge extraction algorithm is converted into a spatial physical displacement with sub-pixel accuracy;

[0029] When using a binocular camera for image acquisition, feature point matching and fitting are performed on the two views, disparity information is obtained based on the epipolar constraint and binocular stereo matching algorithm, and three-dimensional point cloud data is recovered combined with the camera calibration parameters.

[0030] In the above-mentioned machine vision-based large CNC grinding machine thermal deformation measurement method, the sub-pixel edge extraction algorithm adopts Steger algorithm, Zernike moment algorithm or improved Canny sub-pixel algorithm.

[0031] Compared with the prior art, the machine vision-based large CNC grinding machine thermal deformation measurement system and method has the following advantages:

[0032] 1. It can realize synchronous and efficient measurement of thermal deformation of large CNC grinding machines in multiple regions and multiple degrees of freedom. This multi-dimensional synchronous measurement capability is more suitable for adapting to the thermal deformation characteristics of large CNC grinding machines, which are "multi-region coupling and nonlinear time-varying". It does not need to adjust the light path for multiple measurements and arrange a large number of temperature sensors, which can greatly shorten the measurement period, provide efficient and real-time data support for whole machine thermal state characteristic analysis, and help to quickly master the global thermal deformation law of the grinding machine.

[0033] 2. It can significantly enhance the robustness of the measurement system and provide micron-level visual measurement accuracy in theory. The present application integrates an adaptive robust Kalman filter and a long short-term memory neural network preprocessing module to real-time suppress image distortion caused by environmental interference. At the same time, it combines the iterative denoising of the diffusion model edge detection and the sub-pixel edge detection algorithm to ensure the stability and high precision of the measurement accuracy under complex industrial conditions, avoiding the precision deficiency caused by traditional machine vision. The strong anti-interference ability of this system is of great significance for industrial field application, which can to some extent reduce the measurement deviation caused by environmental fluctuations and ensure the reliability and stability of the thermal deformation data. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is the machine vision-based thermal deformation measurement method of the present application.

[0035] Figure 2 is a schematic diagram of the camera arrangement of the present application.

[0036] In the figure, 1 is a linear array camera; 2 is a surface array camera. DETAILED DESCRIPTION

[0037] The following is a specific embodiment of the present application combined with the drawings, which further describes the technical solutions of the present application, but the present application is not limited to these embodiments.

[0038] As shown in Figure 1 , Figure 2 , the present application mainly consists of an image acquisition module, an image preprocessing module and a data extraction module;

[0039] The working process of the present application mainly includes the following steps:

[0040] S1: Set up an auxiliary mark on the surface of the area to be measured, install no less than one industrial camera at a preset distance through a clamp, calibrate the industrial camera and set the shooting interval Δt;

[0041] S2: Start the large CNC grinding machine, collect the real-time image of the surface of the area to be measured, and transmit the image to the computer through the information transmission system;

[0042] S3: Use the image preprocessing module integrating the adaptive robust Kalman filter algorithm and the long short-term memory neural network to perform physical and time domain filtering and quality optimization on the image data obtained in step S2;

[0043] S4: Process the image information output in step S3 through the data extraction module, execute the diffusion model edge detection algorithm to obtain clearer edge information, further perform curve fitting and pixel-level fine positioning on the edge point position of the auxiliary mark through the sub-pixel edge fitting algorithm (such as Steger sub-pixel edge extraction algorithm, Zernike moment sub-pixel detection, etc.), and realize quantitative extraction of thermal deformation by combining the camera internal parameters, output the edge information to the computer for analysis or for training other models;

[0044] S5: Repeat steps S2 to S4 every interval Δt to obtain the long-term thermal deformation progress.

[0045] The image acquisition module is arranged according to the characteristics of each area of the large CNC grinding machine, and is used to acquire the image of the corresponding area. The module includes an industrial camera, an auxiliary mark, and a corresponding image transmission system, etc. The industrial camera is arranged and calibrated in advance, and receives a timing pulse signal during work to acquire visual information of the corresponding area, providing raw data for subsequent image processing and deformation judgment;

[0046] The data preprocessing module connects the image acquisition module through a data interface, and is used to receive raw image data and output denoised image information. The module is mainly a long short-term memory neural network-adaptive robust Kalman filter unit, which filters and optimizes the noise of the collected raw image, eliminates image distortion caused by large CNC grinding machine vibration and light interference, and provides stable input for subsequent feature extraction. The filtering process mainly includes the following steps:

[0047] S301: According to the state estimation at the last time, the prior state estimation and error covariance at the current time are predicted by using the state equation to evaluate the system trend and uncertainty;

[0048] S302: The influence function is added to the residual error to suppress the residual error;

[0049] S303: The dynamic adjustment covariance stage adjusts the process noise covariance and the observation noise covariance in real time according to the innovation residual size, for improving the adaptability of the filter to time-varying noise and signal-to-noise ratio fluctuations;

[0050] S304: The state update stage calculates the Kalman gain, and updates the state and the covariance by using the residual weight;

[0051] S305: Every ∆t time, the obtained picture is repeatedly subjected to the steps S301-S304, and the picture information before and after processing is recorded;

[0052] S306: Through the gating mechanism of the long short-term memory neural network, the historical filtering error is analyzed according to the recorded picture filtering information of the previous frames, and an optimized prediction value of the tth frame feature is output, to correct the deviation caused by the model assumption and the time domain change of the adaptive robust Kalman filter;

[0053] S307: The steps S301-S306 are repeated, the prediction error in the step S306 is used to fine-tune the parameters of the adaptive robust Kalman filter online by using the incremental learning method, and the data update range is limited through the sliding window mechanism;

[0054] The data extraction module performs secondary processing on the image information output by the data preprocessing module through an algorithm, to obtain clear edge information. The module mainly includes a diffusion model edge detection algorithm unit, which inputs the preprocessing result into the data processing module for model training, extracts features such as displacement and deformation gradient by using the diffusion model edge detection, and further performs curve fitting and pixel-level fine positioning on the edge point position of the auxiliary mark by using a sub-pixel edge fitting algorithm, converts the pixel displacement into physical displacement, and forms a standardized data set. The sub-pixel edge extraction algorithm includes a gray matrix algorithm, a Gaussian curve fitting method, a least squares method, a Steger sub-pixel edge extraction algorithm, a Zernike matrix sub-pixel detection algorithm, and a Canny sub-pixel improved algorithm. When a binocular camera is used for image acquisition, feature point matching and fitting are performed on the double views, disparity information is obtained based on the epipolar constraint and the binocular stereo matching algorithm, and three-dimensional point cloud data is recovered in combination with the camera calibration parameters.

[0055] The diffusion model edge detection process mainly includes two parts. The first part forms “forward diffusion” by gradually adding random noise to the initial data, and finally reaches the data, and the process is as follows:

[0056]

[0057] wherein, is the noise intensity of the tth step, is usually a small positive value, and I is an identity matrix.

[0058] The second part is trained by the image generated in the previous step to remove noise and reconstruct the original data distribution sample through multiple iterations, and the process is as follows:

[0059]

[0060] wherein is a core hyperparameter, and represents the proportion of the image remaining without noise from 0 to the t-th step.

[0061] The following describes an apparatus embodiment of the present application, which can be used to execute the machine vision-based thermal deformation measurement method of the large numerical control grinding machine in the above-mentioned embodiments of the present application, taking a certain numerical control large numerical control hydrostatic grinding machine as an example:

[0062] A1: A certain large numerical control hydrostatic grinding machine is divided into several regions according to parts for the arrangement of machine vision units, as shown in the distribution diagram of industrial cameras in this type of machine tool. Figure 2

[0063] The linear feed mechanism thermal deformation measurement subsystem uses one Dalsa Linea LA-GM-08K08A line array camera to measure (such as guide rails, screws, etc., taking the guide rail as an example here), and pastes black and white chessboard grids on the side of the guide rail as identification aids. This type of line array camera has 8192 pixels, with a pixel size of 7.04μm×7.04μm, and is matched with a Pumec LL4K-M72-35mm line scanning lens. The line array camera is installed parallel to the length direction of the guide rail, and the edge and corner points of the stripe of the chessboard grid are scanned to monitor the deformation of the length direction and the width direction of the guide rail. For some longer guide rail mechanisms, a mechanical sliding table can be used for moving scanning. In the subsequent process, the edge detection algorithm can be used to extract the length of the stripe in each region to obtain the straightness deviation and deformation of the guide rail.

[0064] A plurality of black and white chessboard grids are pasted on the metal surface of the spindle moving area as auxiliary identification (or a spherical target is used as an identification aid at the spindle interface) to enhance the edge extraction accuracy. Two FLIR Blackfly S BFS-U3-120S4M-C industrial cameras are installed about 0.4 meters away from the measured region by an aluminum alloy support through a clamp, and the cameras have a resolution of 12 million pixels (4096×3000) and a pixel size of 3.45. Each camera is matched with a FUJINON HF16SA-1 16mm fixed focus industrial lens (lens distortion rate <0.1%). The binocular camera group is installed at an angle of 30° with the spindle axis, and the two cameras are arranged at an interval of 90° around the axis. After installation, multi-view joint calibration is performed to obtain the intrinsic matrix and distortion parameters. The image acquisition interval is set to be 10 seconds, which is used for dynamic tracking in the subsequent thermal deformation process;

[0065] ​The workbench thermal deformation measurement subsystem is to be cooperatively measured by two industrial cameras, and a standard black and white checkerboard is to be pasted on the surface of the workbench as an auxiliary identification. The specific arrangement scheme is similar to the spindle, and the cameras are installed at an angle of 45° with the length and height directions of the workbench. The binocular camera set is symmetrically arranged at the middle cross section of the workbench, and the three-dimensional coordinates of the corner points of the checkerboard are calculated through a binocular vision model. After installation, multi-view joint calibration is performed to obtain the intrinsic matrix and distortion parameters. The image acquisition interval is set to seconds, which is used for dynamic tracking in the subsequent thermal deformation process. In the subsequent process, the edge detection algorithm can extract the coordinates of each corner point, which is used for the flatness and feeding accuracy of the workbench.

[0066] A2: Start the large-scale numerical control grinding machine according to the preset working condition, record the initial time, and record 1080 groups of data for a single working condition of 180 minutes. The industrial camera automatically acquires images of the spindle and the beam connection area at the set interval, and the images are transmitted to the computer at high speed through the transmission interface. The acquisition process is automatically scheduled and stored by the acquisition control program;

[0067] A3: An image preprocessing module integrating adaptive robust Kalman filtering algorithm and long short-term memory neural network is used to perform physical, time domain filtering and quality optimization on the image data obtained in step S2. The filter module takes the pixel position of the auxiliary marker point in the image as the state variable, constructs a first-order state space model, and dynamically estimates the covariance matrix combined with the observation residual to adjust the Kalman gain in real time. To enhance the tolerance of the system to noise and abnormal values, a Huber robust cost function is introduced into the filter to judge the type of observation error and dynamically switch the residual response strategy;

[0068] A4: The collected data are evenly divided into 18 parts in chronological order, each containing 120 groups of data. For each part of data, the first 60 groups of filtered data are taken as training data, and the last 60 groups of filtered data are taken as label data. Through the learning of the mapping relationship between image features and observation noise, the filtering prediction module is trained, and the key parameters of the prediction module are set as follows: learning rate is 0.0001, iteration number is 500, and loss function is mean square error (MSE);

[0069] A5: The image information output by step S4 is processed through a data extraction module. A diffusion model is used to obtain a high-confidence edge map by simulating noise and denoising process. Further, a Steger sub-pixel edge extraction algorithm is used to curve fit and pixel-level fine positioning of the edge point position of the auxiliary identification. The feature point matching and fitting of the pictures collected by the binocular camera system are combined with the camera intrinsic parameters to convert the pixel offset into spatial micrometer-level displacement value, so as to realize the quantitative extraction of thermal deformation.

[0070] A6: Repeat steps S2 to S5 every interval At to obtain the long-term thermal deformation progress.

[0071] The specific embodiments described herein are merely illustrative of the principles of this application. Numerous modifications or adaptations will be readily apparent to those skilled in the art of this application without departing from the spirit or scope of the application as defined by the following claims.

[0072] Although the terms are used herein, the possibility of using other terms is not excluded. The use of these terms is merely for the convenience of describing and explaining the nature of the application; it is against the spirit of the application to interpret them as any kind of additional limitation.

Claims

1. A machine vision-based thermal deformation measurement system for a large CNC grinding machine, characterized in that, The method comprises the following steps: S1: setting auxiliary marks on the surface of the to-be-measured area of the machine tool, installing at least one industrial camera at a preset distance through a clamp, calibrating the industrial camera, and setting a shooting interval Δt; S2: starting the large numerical control grinding machine, collecting the image of the surface of the to-be-measured area at an initial time t0, and transmitting the image to a computer; S3: using the data preprocessing module to perform noise suppression, time domain filtering, and quality optimization on the image data obtained in step S2; 2. A measuring method using the machine vision-based thermal deformation measuring system for large CNC grinding machines according to claim 1, characterized in that, S4: processing the image output by step S3 through the data extraction module, using a diffusion model to detect edges, and further using a sub-pixel edge fitting algorithm to curve fit and pixel-level locate the edge points of the auxiliary marks, matching and fitting the feature points of the pictures extracted by the binocular or multi-view measurement system, and combining the camera intrinsic parameters to quantitatively extract the thermal deformation amount; S5: repeating steps S2 to S4 every interval Δt to obtain time series data of the thermal deformation of the machine tool. The industrial camera is a linear array camera or a surface array camera, and is equipped with a fixed-focus industrial lens. When the industrial camera is a surface array camera, it constitutes a binocular or multi-view measurement system, and the single camera resolution is not less than 12 million pixels. The to-be-measured area is a linear feed mechanism, a spindle, or a workbench. The installation angle of the industrial camera and the axis of the to-be-measured area should be 30°-45°, the installation angle interval of the multi-camera cooperative cameras should be 45°-90°, the interval between the camera and the to-be-measured area should be 0.3-0.5 meters, and the industrial camera acquisition interval Δt is set to 5-15 seconds. The adaptive robust Kalman filtering process in step S3 includes:

3. The method for measuring thermal deformation of a large CNC grinding machine based on machine vision according to claim 2, characterized in that, The prediction step estimates the prior state and error covariance based on the state equation; 4. The method for measuring thermal deformation of a large CNC grinding machine based on machine vision according to claim 2, characterized in that, Calculate the residual and use the influence function to suppress the disturbance of abnormal values to the state; 5. The method for measuring thermal deformation of a large CNC grinding machine based on machine vision according to claim 2, characterized in that, Dynamically adjust the observation noise covariance according to the innovation sequence; 6. The method for measuring thermal deformation of a large CNC grinding machine based on machine vision according to claim 2, characterized in that, Update the Kalman gain to get the optimal state estimation; At the same time, based on the gating mechanism of the long short-term memory neural network, analyze the historical filtering error sequence, and output the optimized prediction value of the image feature at the current time t. The diffusion model edge detection in step S4 includes: In the forward diffusion process, gradually add Gaussian noise to the input image data; In the inverse diffusion process, use the trained denoising model for multiple iterations to gradually restore the edge features of the image; ​ 7. The method for measuring thermal deformation of a large CNC grinding machine based on machine vision according to claim 2, characterized in that, ​ ​ ​ The detected pixel-level edge position is converted into a spatial physical displacement with sub-pixel accuracy by a sub-pixel edge extraction algorithm. 8.The method of claim 2, wherein, The sub-pixel edge extraction algorithm adopts a Steger algorithm, a Zernike moment algorithm or a modified Canny sub-pixel algorithm.

Images