An on-line monitoring method for surface micro-topography and roughness in ultra-precision turning

Abstract

The present invention relates to the technical field of ultra-precision cutting processing, and particularly refers to an on-line monitoring method for the surface micro-topography and roughness in ultra-precision turning. This method collects the signal data inside the machine tool and the tool parameters, and performs data processing and result prediction through the geometric simulation model of the machined surface micro-topography, the deep learning model and the prediction result drawing program in the signal analysis and processing module, and draws the simulated roughness distribution map of the machined surface and the three-dimensional micro-topography map of the specified local area. This method does not need to rely on external sensor devices, and only relies on the internal signals generated during the machining process of the machine tool to realize the monitoring of the vibration situation between the tool and the workpiece during the machining process. At the same time, by combining the geometric simulation model and the deep learning model, the monitoring function of the surface micro-topography and roughness during the machining process considering the relative vibration, plastic side flow and elastic recovery effect between the tool and the workpiece is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of ultra-precision cutting machining, and particularly refers to an on-line monitoring method for the generation of surface micro-topography and surface roughness during ultra-precision single-point turning machining. Background Art

[0002] On-line monitoring of the machining process is an important field in industrial production, which involves the use of various sensors, data acquisition systems and data analysis means to monitor the state of the production machining process. Ultra-precision single-point diamond turning technology has a wide range of applications in the fields of machining high-precision and high-surface-quality optical free-form surface components, etc., and is an important technology in modern manufacturing. Since the cutting amount of ultra-precision single-point diamond turning is very small, it is difficult to directly observe the quality of machining during the machining process. Therefore, it is usually only possible to measure the machined surface of the workpiece with a special measuring instrument after machining to evaluate the machining quality, which limits the improvement of machining efficiency and cannot timely reflect the quality of machining. Therefore, during the ultra-precision single-point diamond turning machining process, on-line monitoring of the surface quality of products can ensure product quality, improve production efficiency and reduce production costs, which is of great significance. To achieve this goal, it is necessary to deeply analyze various factors affecting the surface quality during the machining process. Summary of the Invention

[0003] The purpose of this technical solution is to provide an on-line monitoring method for surface micro-topography and roughness in ultra-precision single-point turning machining with high precision and without external sensing equipment.

[0004] The purpose of this technical solution is achieved as follows:

[0005] An on-line monitoring method for surface micro-topography and roughness in ultra-precision turning includes the following steps:

[0006] (a) During the turning machining of the machine tool, obtain the turning machining parameters, the geometric parameters of the cutting tool and the internal signals of the machine tool; the collected internal signals include the reference position signal of the X-axis , the reference position signal of the C-axis and the position error signal of the Z-axis ; the reference position signal is used to determine the position of the collected cutting point in the radial direction, the reference position signal is used to determine the position of the collected cutting point in the circumferential direction, and the position error signal is used to reflect the relative vibration displacement between the cutting tool and the workpiece in the depth-of-cut direction;

[0007] (b) Input the turning parameters, the geometric parameters of the cutting tool, and the internal signals of the machine tool into the geometric simulation program of the machined surface micro-topography in the host computer to achieve the geometric simulation of the machined surface micro-topography;

[0008] (c) Input the turning parameters and the geometric simulation results obtained in step (b) into a pre-trained deep learning model to predict the influence of plastic side flow and elastic recovery on the machined surface micro-topography and obtain the prediction results;

[0009] (d) Use the prediction results obtained in step (c) to calculate and draw the simulated roughness distribution map of the machined surface through the prediction result drawing program in the host computer.

[0010] Preferably, in the step (a), the reference position signal and the reference position signal and the position error signal are collected according to the set sampling duration and sampling frequency . After each completion of duration of sampling, the next duration of sampling is carried out immediately; the sampled signal data is sent to the geometric simulation program of the machined surface micro-topography.

[0011] Preferably, the geometric simulation program of the machined surface micro-topography receives the internal signal data of the machine tool incoming every time in real time and dynamically updates the results of the geometric simulation; the steps for the program to achieve the dynamic geometric simulation of the machined surface micro-topography include:

[0012] (1) According to the internal signal sampling frequency and the spindle rotation frequency of the machine tool , calculate the number of sampling points per revolution ; according to the spindle rotation frequency and the sampling duration , calculate the number of revolutions that the spindle has turned in the previous time; according to the feed rate and the sampling duration , calculate the radial feed distance in the previous time;

[0013] (2) In the previous time, the machined area is an annular area. Discretize the three-dimensional surface of this annular machined area into two-dimensional radial cross-sections, and each radial cross-section contains Cutting sampling points. In the same radial section, the C-axis reference positions of all cutting sampling points are the same, and the X-axis reference positions increase arithmetically; Denote the X-axis reference position when the tool feeds to the workpiece rotation center as , then the radius where the cutting sampling point is located, and the radius is the difference between the current X-axis reference position of the cutting sampling point and the X-axis reference position when the tool feeds to the workpiece rotation center;

[0014] (3) In each radial section, sort all the cutting sampling points in ascending order of the radius to form a set representing the tool tip positions in this radial section,

[0015] ;

[0016] When the tool is a single-point tool with a circular arc edge, its cutting edge profile can be approximately expressed as , where represents the radius of the tool tip arc; According to the set of tool tip positions in the radial section and the cutting edge profile formula, the two-dimensional profile in this radial section can be obtained;

[0017] Discretize the two-dimensional profile into a discrete point set with a radial spacing of ,

[0018] , where the size of the point set;

[0019] The profile height at the discrete points is ,

[0020] ;

[0021] According to the above steps, calculate the coordinates of all discrete points in the radial section to obtain the two-dimensional profile considering only kinematic factors in the radial section;

[0022] (4) According to step (3), calculate the two-dimensional profiles in all radial sections, and thus obtain the microscopic topography geometric simulation results of the annular area machined in the previous time.

[0023] Preferably, in the step (c), the deep learning model used is a deep long short-term memory network, which accepts sequence inputs and outputs sequences of the same length; this model is used to transform the radial profile considering only kinematic factors obtained in step (b) into a radial profile considering plastic side flow and elastic recovery; the length of the input sequence of the network is equal to the number of discrete points of the radial profile , and each step of the input sequence is a four-dimensional vector composed of the profile height value of the discrete point , the depth of cut in turning , the feed rate and the cutting linear velocity corresponding to the discrete point , where ; each step of the output sequence is a scalar value , that is, the profile height value of the discrete point considering plastic side flow and elastic recovery; the network structure is a sequential structure, which is composed of an input layer, 4 bidirectional LSTM layers, a Leaky ReLU activation layer, two fully connected layers and an output layer in sequence; among them, the dimensions of the hidden state and cell state of the 4 bidirectional LSTM layers are both set to 150, and a dropout layer with a probability of 0.2 is added between adjacent bidirectional LSTM layers to alleviate the overfitting of the network; the number of neurons in the two fully connected layers is 300 and 30 respectively, and each fully connected layer also uses Leaky ReLU as the activation function; the steps of the training data collection and training process of the network are as follows:

[0024] (1) For a specific target workpiece material, conduct experiments with several groups of different cutting parameters, and use measuring instruments such as a white light interferometer or a profiler to obtain the radial profile of the machined surface. Extract multiple radial profiles along different radial directions for each machined surface; the cutting parameters include the depth of cut, the feed rate, and the spindle speed;

[0025] (2) Extract the set of tool tip position coordinates from the measured radial profiles; since in the same machining, the elastic recovery amount at the tool tip position is approximately the same and the plastic side flow amount is approximately zero, the relative positions of all tool tip position points remain unchanged before and after the plastic side flow and elastic recovery occur; according to the extracted set of tool tip position coordinates and the theoretical profile of the cutting edge, reverse restore the ideal profile situation before the plastic side flow and elastic recovery occur;

[0026] (3) According to the ideal profile and the corresponding machining parameters obtained in step (2), make an input sequence according to the above input data format requirements, and at the same time, the corresponding measured profile is used as the output sequence to form a training sample of the network; perform the above operations on all measured radial profiles to obtain the training set of the network;

[0027] (4)Train the network. Before training, standardize the training data to accelerate the convergence speed of the network; set the batch size to 20; select the Adam optimizer, and set its parameters as the learning rate of 0.001 and the weight decay rate of 0.001; use MSE as the loss function to describe the difference between the network output and the target value. The network is built and trained using the Pytorch deep learning framework, and the number of training epochs is 100.

[0028] Preferably, in the step (d), the prediction result drawing program in the host computer equally divides the circular area processed in the previous time period into several sub-regions at equal intervals in the radial and circumferential directions respectively. Then, according to the three-dimensional microscopic topography prediction result of the machined surface obtained in the step (c), calculate the surface roughness values of all sub-regions respectively, and then draw the calculation results using a pseudo-color map to obtain a simulated roughness distribution map, which is used to intuitively display the predicted roughness distribution law of different local regions on the entire machined surface; the simulated roughness distribution map is refreshed every second to display the latest machining results in the past seconds, thereby realizing the on-line monitoring of the machining quality during the machining process; the prediction result drawing program can draw the local microscopic topography prediction result within a specified area in the form of a three-dimensional surface diagram and give the calculation result of the roughness index within this area.

[0029] Preferably, the tool used is a single-point diamond tool.

[0030] Preferably, the kinematic factors include the tool nose radius, the feed rate, and the relative vibration.

[0031] The prominent and beneficial technical effects of this technical solution compared with the prior art are:

[0032] The present invention realizes the monitoring of the relative vibration between the tool and the workpiece during ultra-precision single-point turning by online collecting and analyzing the internal signals generated during the machining process of the machine tool. On this basis, a geometric simulation model of the micro-topography of the machined surface considering relative vibration is established. By simplifying the vibration into the superposition of several constant simple harmonic vibrations with different frequencies, it can more accurately reflect the complex and variable vibration conditions during the machining process. Further, this method predicts the influence of plastic side flow and elastic recovery effects on the micro-topography of the machined surface by building and training a deep learning model. Compared with the traditional method of predicting plastic side flow and elastic recovery effects using empirical formulas, this method has higher prediction accuracy and generalization performance, and the network training process is simple. The number of experiments required to collect training data is less than that required by the traditional method to calibrate many coefficients, effectively reducing the implementation cost of the method. The present invention realizes the online monitoring of the micro-topography and roughness of the surface during the ultra-precision single-point turning process by comprehensively considering tool parameters, machining parameters, the relative vibration between the tool and the workpiece during machining, and plastic measurement and elastic recovery effects. The present invention does not require additional sensor equipment, has a low implementation cost, and high production efficiency. Description of the Drawings

[0033] Figure 1 It is a schematic diagram for dividing the cutting sampling points and the radial section.

[0034] Figure 2 It is a schematic diagram for generating and discretizing a two-dimensional profile considering only kinematic factors in the radial section.

[0035] Figure 3 It is a schematic diagram for the influence of plastic side flow and elastic recovery on the micro-topography of the machined surface.

[0036] Figure 4 It is a schematic diagram for the network structure of the deep learning model.

[0037] Figure 5 It is a schematic diagram for obtaining network training data.

[0038] Figure 6 It is a schematic diagram for showing the prediction performance of the network on the validation set.

[0039] Figure 7 It is a schematic diagram for showing a complete online monitoring process. Detailed Description of the Embodiment

[0040] The following further elaborates in detail the specific implementation manners of the technical solution in combination with the appended Figures 1 - 7 drawings.

[0041] For example Figure 1As shown in the figure, taking face turning performed on a three-axis ultra-precision single-point diamond lathe as an example, during the machining process, the motion controller sends command signals to the lathe according to the executed NC program to guide the movement of each feed axis of the machine tool. At the same time, the sensors inside the lathe measure the actual positions of the feed axes and send feedback signals to the motion controller.

[0042] An on-line monitoring method for surface micro-topography and roughness in ultra-precision turning, comprising the following steps:

[0043] Step (a), during the turning machining process of the machine tool, obtain the turning machining parameters, geometric parameters of the cutting tool, and internal signals of the machine tool; the internal signals during the machining process (the reference position signal of the X-axis , the reference position signal of the C-axis and the position error signal of the Z-axis ) are collected. Each set of sampling values constitutes a cutting sampling point, representing the state of the cutting at that moment. Among them, the reference position signal of the X-axis and the reference position signal of the C-axis are respectively used to determine the positions of the collected cutting points in the radial and circumferential directions, and the position error signal of the Z-axis is used to reflect the relative vibration displacement of the tool-workpiece in the depth of cut direction. The C-axis refers to the rotating axis corresponding to the machine tool spindle, which is a general technical term in the mechanical field, and those skilled in the art can understand and clarify the relevant definitions. Establish communication between the motion controller and the upper computer, and transmit the collected internal signal data of the machine tool in the motion controller to the upper computer.

[0044] Reference position signal 、Reference position signal and position error signal are collected according to the set sampling duration and sampling frequency , and after each completion of duration of sampling, the next duration of sampling is carried out immediately. At the same time, the sampling data of the completed duration is sent to the machining surface micro-topography geometric simulation program of the upper computer through the communication program. In this example, is set to 5 seconds, is set to 2000Hz.

[0045] Step (b), in the upper computer, the machining surface micro-topography geometric simulation program can, according to the turning machining parameters, geometric parameters of the cutting tool, and every The internal signal data of the machine tool input by the duration dynamically updates the results of the geometric simulation. The steps for the program to implement the dynamic geometric simulation of the micro-topography of the machining surface include: The program for the dynamic geometric simulation of the micro-topography of the machining surface runs in the MATLAB software.

[0046] (1) According to the internal signal sampling frequency and the spindle rotation frequency , calculate the number of sampling points per revolution . In this example, the spindle speed is 1000 RPM, that is, the spindle rotation frequency ≈ 1000 / 60 Hz, so ; According to the spindle rotation frequency and the sampling duration , calculate the number of revolutions that the spindle has turned in the previous time. In this example, ; According to the feed speed and the sampling duration , calculate the radial feed distance in the previous time. In this example, , so . It should be noted that in the first at the start of cutting and the last at the end of cutting, the actual feed distance is less than the above calculated value because cutting generally starts and ends in the middle of the first and the last , so the actual cutting feed time is insufficient , which needs to be determined according to the time points for judging the start and end of cutting.

[0047] (2) In the previous 5-second time, the machined area is an annular area. The three-dimensional surface of this annular machining area is discretized into 120 two-dimensional radial sections, and each radial section contains 83 or 84 cutting sampling points. In the same radial section, the C-axis reference positions of all cutting sampling points are the same, the X-axis reference positions increase arithmetically (assuming the feed direction is X+), and the Z-axis position errors are different due to vibration. Denote the X-axis reference position when the tool feeds to the workpiece rotation center as . In this example , then the radius where the cutting sampling point is located, that is, the difference between the current X-axis reference position of the cutting sampling point and the X-axis reference position when the tool feeds to the workpiece rotation center.

[0048] (3) As Figure 2 shown, in each radial section, all the cutting sampling points are arranged according to the radius Sort in ascending order to form a set representing the tool tip positions within the radial cross-section ,

[0049] .

[0050] For a single-point tool with an arc-shaped cutting edge, its cutting edge profile can be approximately represented as , where represents the radius of the tool tip arc. In this example , so the cutting edge profile formula is . Based on the set of tool tip positions within the radial cross-section and the cutting edge profile formula, the two-dimensional profile within the radial cross-section can be obtained. For ease of storage and calculation, the two-dimensional profile needs to be discretely represented as a set of discrete points with a radial spacing of ,

[0051] , where the size of the point set , where [] is the rounding symbol. In this example, the discrete spacing , so the size of the point set . Similarly, within the first 5-second time period at the start of cutting and the last 5-second time period at the end of cutting, the size of the point set is less than the above calculated value. At this time, the specific size of the point set is determined according to the time points for judging the start and end of cutting. Assume that the discrete point is located between two adjacent tool tip position points and , that is , then the profile height at this discrete point is ,

[0052] , that is, take the smaller value of the two adjacent cutting edge profiles. In fact, interference may also occur between non-adjacent cutting edge profiles, especially when the vibration amplitude is large, the tool tip radius is large, and the feed per revolution is small. Therefore, when calculating the profile height at discrete points, the coordinate information of multiple adjacent tool tip position points before and after needs to be considered. Taking the previous and next 2 as an example, the profile height calculation formula is then modified to:

[0053] .

[0054] In this example, the number is taken as 5. According to the above steps, calculate the coordinates of all discrete points within the radial cross-section to obtain the two-dimensional profile within the radial cross-section considering only kinematic factors (i.e., tool tip radius, feed speed, and relative vibration).

[0055] (4)According to step (3), calculate the two-dimensional profiles within all 120 radial cross-sections, thereby obtaining the geometric simulation results of the microscopic topography of the annular region machined in the previous 5-second period.

[0056] Step (c): By inputting the turning process parameters and the geometric simulation results obtained in step (b) into a pre-trained deep learning model, predict the influence of plastic side flow and elastic recovery on the microscopic topography of the machined surface, and obtain the prediction results; the prediction results are in the form of coordinate values of a series of discrete points in terms of data. These discrete points together form a three-dimensional surface, that is, the three-dimensional topography prediction result of the machined surface. In fact, the output result of the "geometric simulation program" mentioned in step (b) above is also a series of discrete point coordinate values. The difference is that the result obtained in step (b) only considers geometric factors, while step (c) considers the influence of plastic side flow and elastic recovery on this basis, and this function is realized by the deep learning model in step (c). As Figure 3 shown, in addition to the contribution of the above kinematic factors to the machined surface topography, Figure 3 the plastic side flow shown in (a) in Figure 3 and the elastic recovery shown in (b) in will also change the machined surface topography. Among them, plastic side flow will increase the residual height of the surface, while elastic recovery will decrease the residual height of the surface. In this method, a deep learning model is used to predict the degree of change of plastic side flow and elastic recovery on the surface under specific machining conditions. This network accepts sequence input and the output is a sequence of the same length. Each step of the input sequence is the profile height value of a discrete point of the radial profile when only considering kinematic factors , the depth of cut of turning , the feed rate and the cutting linear velocity corresponding to the discrete point to form a four-dimensional vector, where . Each step of the output sequence is a scalar value , that is, the profile height value of the discrete point after considering plastic side flow and elastic recovery. The constructed deep learning model is a deep long short-term memory (LSTM) network, and the network structure is a sequential structure. As Figure 4 shown, it is composed of an input layer, 4 bidirectional LSTM layers, a Leaky ReLU activation layer, two fully connected layers and an output layer in sequence. Among them, the dimensions of the hidden state and cell state of the 4 bidirectional LSTM layers are both set to 150, and a dropout layer with a probability of 0.2 is added between adjacent bidirectional LSTM layers to alleviate the overfitting of the network; the number of neurons in the two fully connected layers are 300 and 30 respectively, and each fully connected layer also uses Leaky ReLU as the activation function. The steps of the training data collection and training process of the network are as follows:

[0057] (1) For a specific target workpiece material, experiments are carried out with several groups of different cutting parameters (cutting depth, feed rate, spindle speed). Measuring instruments such as white light interferometers or profilometers are used to obtain the radial profiles of the machined surface. Multiple radial profiles are extracted along different radial directions for each machined surface. In this example, the workpiece is a square copper sheet with a size of 10 mm × 10 mm, and the machining area is a circle with a diameter of 9 mm centered on the center of the copper sheet. 15 sets of cutting parameter combinations in Table 1 below are used as the data source for the training set. After each cutting, the three-dimensional topography of the machined surface is measured with a white light interferometer, and the radial profiles are extracted from it. The measurement results of the three-dimensional topography of the machined surface are as Figure 5 shown.

[0058] Table 1. Training set data table

[0059] Serial number Spindle speed (RPM) Feed rate (μm / min) Cutting depth (μm) 1 1000 4 2 2 1000 4 4 3 1000 4 6 4 1000 6 2 5 1000 6 4 6 1000 6 6 7 1000 8 2 8 1000 8 4 9 1000 8 6 10 1000 10 2 11 1000 10 4 12 1000 10 6 13 1000 12 2 14 1000 12 4 15 1000 12 6

[0060] (2) A set of tool tip position coordinates (generally all the minimum points of the measured profile) is extracted from the measured radial profiles. Since in the same machining process, the elastic recovery amount at the tool tip position is approximately the same and the plastic flow measurement amount is approximately zero, the relative positions of all tool tip position points remain unchanged before and after plastic flow measurement and elastic recovery occur. Based on this, the ideal profile situation before plastic flow measurement and elastic recovery can be reversely restored according to the extracted set of tool tip position coordinates and the theoretical profile of the cutting edge, that is, the radial profile result of the machined surface when only considering kinematic factors such as tool parameters, machining parameters, and relative vibration.

[0061] (3) According to the ideal profile obtained in step (2) and the corresponding machining parameters, an input sequence is made according to the above input data format requirements, and the corresponding measured profile is used as the output sequence (label value) to form a training sample of the network. The above operations are performed on all measured radial profiles to obtain the training set of the network.

[0062] (4) The network is trained. Before training, the training data is standardized to accelerate the convergence speed of the network; the batch size (the number of training samples used in a single training) is set to 20; the Adam optimizer is selected as the optimizer, and its parameters are set as the learning rate of 0.001 and the weight decay rate of 0.001; MSE is used as the loss function to describe the difference between the network output and the target value. The network is built and trained using the Pytorch deep learning framework, and the number of training epochs is 100.

[0063] Figure 6 shows the performance of the trained model on the validation set, Figure 6 in which (a) is a sample of Experiment 2, Figure 6In (b), it is a sample of Experiment 7. The figure respectively shows the actually measured radial profile, the ideal radial profile reversely restored according to the above method, and the profile predicted by the deep learning model based on the input of the ideal profile. It can be seen from the results shown in the figure that the trained deep learning model can well predict the influence of plastic side flow and elastic recovery according to the input ideal profile information, so as to approach the actually measured profile morphology.

[0064] In this embodiment, the geometric simulation program of the machined surface micro-topography obtains the prediction result when only geometric factors are considered; based on the above result, the deep learning model considers the influence of plastic deformation and elastic recovery to obtain the final prediction result; the original data of the prediction result obtained in the previous step is a series of data points, and then these data are calculated, analyzed and visualized through the prediction result plotting program.

[0065] In step (d), using the prediction result obtained in step (c), the simulation roughness distribution map of the machined surface is calculated and drawn through the prediction result plotting program in the host computer; the prediction result plotting program in the host computer will use the previous time period (5 seconds in this embodiment) to equally divide the machined annular area into several sub-regions in the radial and circumferential directions respectively. Subsequently, according to the three-dimensional micro-topography prediction result of the machined surface obtained in the above steps, the surface roughness values of all sub-regions are calculated respectively (such as the arithmetic roughness value Sa, the root mean square roughness value Sq, etc.), and then the calculation results are drawn using a pseudo-color map to obtain the simulation roughness distribution map, which is used to intuitively display the predicted roughness distribution law of different local areas on the entire machined surface. The simulation roughness distribution map is refreshed every 5 seconds to display the latest machining results in the past 5 seconds, thereby realizing the online monitoring of the machining quality during the machining process. In addition, the prediction result plotting program can draw the local micro-topography prediction result in a specified area in the form of a three-dimensional surface diagram and give the calculation result of the roughness index in this area, which is convenient for the operator to observe the specific morphology of a local area of interest. Figure 7The application results of the present method in a single-point turning experiment are given. The cutting parameters of the experiment are set as follows: spindle speed 1000 RPM, feed rate 5 mm / min, and depth of cut 5 μm. The left column shows the refreshing process of the simulated roughness distribution map (taking the surface roughness index Sa as an example) during the machining process. The right side lists the comparison results of the simulated predicted three-dimensional topography and the measured three-dimensional topography of three local areas, as well as the corresponding calculated roughness values of the local areas. It can be seen that there is a good correspondence between the simulated predicted topography and the measured topography, and the local prediction results of the roughness also have high accuracy. From the simulated roughness distribution map, it can be seen that although the cutting parameters remain unchanged during the cutting process, the change in the relative vibration between the tool and the workpiece caused by the change in the cutting environment significantly affects the roughness values of different machining areas. This demonstrates the advantage of the online vibration state monitoring adopted by the present method: compared with the traditional method of simplifying the vibration as the superposition of several constant simple harmonic motions, it can more accurately reflect the change of vibration during the machining process. Moreover, the present method also provides the possibility for monitoring abnormal states during the machining process. This is because abnormalities during machining are usually accompanied by abnormal vibrations, resulting in obvious deterioration of the roughness of the machined surface. Therefore, the occurrence of machining abnormalities can be effectively monitored by observing the simulated roughness distribution map. In addition, the prediction results of the micro-topography and roughness also prove that the deep learning model adopted in the present method can effectively predict the influence of plastic side flow and elastic recovery on the surface micro-topography and surface roughness.

[0066] The present invention realizes the monitoring of the relative vibration between the tool and the workpiece during the ultra-precision single-point turning process by online collecting and analyzing the internal signals generated by the machine tool during the machining process. On this basis, a geometric simulation model of the surface micro-topography considering the relative vibration is established. Compared with the traditional method of measuring the relative vibration between the tool and the workpiece before actual machining and simplifying the vibration as the superposition of several constant simple harmonic vibrations with different frequencies, this method can more accurately reflect the complex and variable vibration conditions during the machining process. Further, by building and training a deep learning model, this method predicts the influence of plastic side flow and elastic recovery effects on the surface micro-topography of the machined surface. Compared with the traditional method of predicting plastic side flow and elastic recovery effects using empirical formulas, this method has higher prediction accuracy and generalization performance, and the network training process is simple. The number of experiments required to collect training data is less than the number of experiments required by the traditional method to calibrate many coefficients, effectively reducing the implementation cost of the method. The present invention realizes the online monitoring of the surface micro-topography and roughness during the ultra-precision single-point turning process by comprehensively considering tool parameters, machining parameters, the relative vibration between the tool and the workpiece during machining, and plastic measurement and elastic recovery effects. The present invention does not require additional sensor devices, has a low implementation cost, and high production efficiency.

[0067] The basic principle, main features and advantages of the present technical solution have been shown and described above. Those skilled in the art should understand that the present technical solution is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principle of the present technical solution. Without departing from the spirit and scope of the present technical solution, the present technical solution will have various changes and improvements, and all these changes and improvements fall within the scope of the present technical solution claimed.

Claims

1. An online monitoring method for surface micromorphology and roughness in ultra-finishing turning, characterized in that: The following steps are involved: (a) During the turning process of the machine tool, the turning parameters, the geometric parameters of the tool and the internal signals of the machine tool are obtained; the internal signals collected include the reference position signal x of the X-axis ref , C-axis reference position signal c ref and the Z-axis position error signal z e ; Reference position signal x ref Used to determine the radial position of the collected cutting point, the reference position signal c ref Used to determine the circumferential position of the collected cutting point, the position error signal z e Used to reflect the relative vibration displacement between the tool and the workpiece in the cutting depth direction; (b) inputting turning processing parameters, geometric parameters of the tool and internal signals of the machine tool into a micro-morphology geometry simulation program of the machined surface in the host computer to realize micro-morphology geometry simulation of the machined surface; (c) predicting the effects of plastic lateral flow and elastic recovery on the micromorphology of the machined surface by inputting the turning processing parameters and the geometric simulation results obtained in step (b) into a pre-trained deep learning model, and obtaining the prediction results; (d) using the prediction result obtained in step (c), calculating and drawing a simulated roughness distribution map of the machined surface through a prediction result drawing program in a host computer; In step (a), the reference position signal x ref , reference position signal c ref and the position error signal z e According to the set sampling time ΔT and sampling frequency f s After completing the sampling of each ΔT duration, the next ΔT duration sampling is then performed; The sampled signal data is sent to the micro-morphology geometry simulation program of the machined surface; The machining surface micro-morphology geometry simulation program receives the internal signal data of the machine tool transmitted at intervals of ΔT in real time, and dynamically updates the result of the geometry simulation; the steps of implementing the dynamic machining surface micro-morphology geometry simulation by the program include: (1) According to the internal signal sampling frequency f s And the machine tool spindle rotation frequency f, calculate the number of sampling points per circle According to the spindle rotation frequency f and the sampling time ΔT, the number of revolutions N of the spindle in the last ΔT time is calculated. r =f s ·ΔT; according to the feed speed f r and sampling time ΔT, calculate the radial feed distance L=f within the last ΔT time r ΔT / 60; (2) During the last ΔT time, the tool forms an annular area, and the three-dimensional surface of the annular area is discretized into N c Two-dimensional radial sections, each containing N r cutting sampling points; in the same radial section, the C-axis reference position of all cutting sampling points is the same, and the X-axis reference position increases in equal steps; when the tool is fed to the workpiece rotation center, the X-axis reference position is x o , then the radius of the cutting sampling point r = x o -x ref , the radius r is the difference between the current X-axis reference position of the cutting sampling point and the X-axis reference position when the tool is fed to the rotation center of the workpiece; (3) In each radial section, all cutting sampling points are sorted in order from small to large radius r to form a set representing the tool tip position in the radial section. When the tool is a single-point tool with an arc blade, its blade profile can be approximately expressed as where r n Indicates the radius of the tool tip arc; according to the tool tip position set T in the radial section and the blade profile formula, the two-dimensional profile in the radial section can be obtained; The two-dimensional contour is discretized into a discrete point set P with radial spacing Δl. P={p1(r1,z1),p2(r2,z2),…,p N (r N ,z N )}, where the size of the point set The height of the contour at the discrete point is z k , According to the above steps, the coordinates of all discrete points in the radial section are calculated to obtain a two-dimensional profile in the radial section that only considers kinematic factors; (4) According to step (3), calculate all N c The two-dimensional profile in the radial section is obtained, thereby obtaining the microscopic geometric simulation result of the annular area processed in the last ΔT time; In the step (c), the deep learning model used is a deep long short-term memory network, which accepts sequence input and outputs a sequence of equal length; the model is used to transform the radial profile obtained in step (b) that only considers kinematic factors into a radial profile that considers plastic lateral flow and elastic recovery; the input sequence length of the network is equal to the number of discrete points N of the radial profile, and each step of the input sequence is a profile height value z of the discrete point, a cutting depth a of the turning process, and a linear relationship between the radial profile and the linear relationship. p , feed speed f r A four-dimensional vector consisting of the cutting line velocity v corresponding to the discrete point, where v = 2πfr; each step of the output sequence is a scalar value That is, the height value of the discrete point contour after plastic lateral flow and elastic recovery is considered; the network structure is a sequential structure, which is composed of an input layer, four bidirectional LSTM layers, a Leaky ReLU activation layer, two fully connected layers and an output layer in sequence; among them, the dimensions of the hidden state and cell state of the four bidirectional LSTM layers are set to 150, and a dropout layer with a probability of 0.2 is added between adjacent bidirectional LSTM layers to alleviate the overfitting of the network; the number of neurons in the two fully connected layers is 300 and 30 respectively, and each fully connected layer also uses Leaky ReLU as the activation function; the training data collection and training process steps of the network are as follows: (1) For a specific target workpiece material, several groups of experiments with different cutting parameters are conducted, and the radial profile of the machined surface is obtained using a white light interferometer or a profilometer. Multiple radial profiles are extracted from each machined surface along different radial directions; the cutting parameters include cutting depth, feed speed, and spindle speed; (2) extracting a set of tool tip position coordinates from the measured radial profile; since the elastic recovery amount at the tool tip position is approximately the same and the plastic side flow is approximately zero in the same machining, the relative positions of all tool tip position points remain unchanged before and after the plastic side flow and elastic recovery occur; and reversing the ideal profile before the plastic side flow and elastic recovery occur based on the extracted tool tip position coordinate set and the theoretical profile of the blade; (3) According to the ideal profile and corresponding processing parameters obtained in step (2), an input sequence is prepared according to the input data format requirements of the above-mentioned deep learning model, and the corresponding measured profile is used as an output sequence to constitute a training sample of the network; the above operation is performed on all measured radial profiles to obtain a training set of the network; (4) The network was trained. Before training, the training data was standardized to speed up the convergence of the network. The batch size was set to 20. The Adam optimizer was used as the optimizer, and its parameters were set to a learning rate of 0.001 and a weight decay rate of 0.

001. MSE was used as the loss function to describe the difference between the network output and the target value. The Pytorch deep learning framework was used to build and train the network, and the number of training rounds was 100.

2. The method for online monitoring of surface micromorphology and roughness in superfinishing turning according to claim 1, characterized in that: In the step (d), the prediction result drawing program in the upper computer divides the annular area processed in the previous ΔT time period into a number of sub-areas at equal intervals in the radial and circumferential directions, and then calculates the surface roughness values ​​of all sub-areas according to the three-dimensional micro-morphology prediction results of the processed surface obtained in step (c), and then draws the calculation results using a pseudo-color map to obtain a simulated roughness distribution map, which is used to intuitively display the predicted roughness distribution law of different local areas on the entire processed surface; the simulated roughness distribution map is refreshed every ΔT seconds to display the latest processing results within the past ΔT seconds, thereby realizing online monitoring of the processing quality during the processing process; the prediction result drawing program can draw the local micro-morphology prediction results in the specified area in the form of a three-dimensional surface map, and give the roughness index calculation results in the area.

3. The method for online monitoring of surface micromorphology and roughness in superfinishing turning according to claim 1, characterized in that: The tool used is a single-point diamond tool.

4. The method for online monitoring of surface micromorphology and roughness in superfinishing turning according to claim 1, characterized in that: The kinematic factors include tool nose radius, feed rate and relative vibration.

Images