1 Overview The goal of machining is to pursue the best combination of machining accuracy, cost, and efficiency. To achieve this goal, one of the key technologies that are urgently needed for research and development is the on-line precision measurement technology, especially under the conditions of multi-species and small-batch production. The significance of on-line measurement technology is particularly significant, because on-line measurement is an important part of the integrated process and measurement technology and an important means of ensuring the quality of parts and increasing productivity. It has long been recognized abroad that the importance of on-line measurement technology has been a lot of research, and in the production of a large number of applications. 2 Error Source Analysis Affecting Online Measurement Accuracy The purpose of on-line measurement is to check whether the precision index of the machined part meets the requirements. If it meets the requirements, remove the workpiece. Otherwise, perform the necessary compensation processing until the workpiece processing precision is qualified. We know that we must accurately measure the machining accuracy of the part, and the measurement equipment. The accuracy must be one order of magnitude higher than the accuracy of the part being measured. In ultra-precision machining, the processing environment and the online measurement environment are not much different. To ensure the accuracy of on-line measurement, it can only be achieved through error compensation. That is to say, the online measurement of parts that do not compensate for machining by error compensation can ensure the accuracy of the measurement (error compensation can increase the machining accuracy of the part by an order of magnitude), and measuring the compensated parts by error compensation cannot satisfy the principle of 10 times. However, the accuracy of on-line measurement of the lathe after error compensation is sufficiently high is still meaningful. Lathe 21 Error Source List 3 Error Source Identification and Modeling The lathe mainly processes cylindrical parts, end faces, taper faces, and spherical parts. Without considering the spindle rotation error, only the X-direction error compensation is required to measure the cylindrical surface cylindricity error on-line; the on-line measurement of the end face is required. Z-direction error compensation; Z-axis and X-direction error compensation are required at the same time when measuring surfaces such as taper surfaces and spherical surfaces. The error compensation model must be two-dimensional. Of course, three capacitive sensors can also compensate for the influence of the spindle rotation error on the measurement accuracy. There are usually two methods for identifying the error compensation amount. First, the error sources are identified off-line first, and certain synthesis rules (such as homogeneous coordinate transformation) are adopted. ) Obtain the error compensation amount at each point of the machining space of the machine tool; Second, obtain the error compensation amount by measuring the machined surface of the part. The first method takes time and the assumptions in the modeling process affect the modeling accuracy. The second method can only compensate the measurement. For specific parts, the compensation area cannot be expanded to the entire processing area. This paper combines the two conditions to identify the error compensation amount, which ensures accuracy and saves time. 3.1 Error source identification 1 X-direction error compensation amount identification: δX=δX(x)+δX(z)+αZФ.z+β(z).WZ+α(z).Wh+β(x).TZ (1) Where z is the distance from the origin of the Z-slide; Δ-column X=δX(z)+αZФ.z+β(z).WZ+α(z).Wh (2) Figure 1 Cylindrical bus error measurement In this way, four error sources can be identified. It can be known from equations (1) and (2) that there are δx(x) and β(x) needs to be identified in δX, and β(x) needs to be identified offline by a laser measuring device, δx. (x) is guaranteed by the lathe full closed loop system. In this way, the X-direction error compensation amount can be obtained when measuring, that is: Δx=f(x,z)=δ Column X+β(x).TZ (3) 2 Identification of Z-direction error compensation: Δz=δz(x)+βxФ.x+α(x).Th+β(x).TX+β(z).WX+δz(z) (4) Where x is the distance from the origin in the X direction (center of the face); When measuring the end face of the part (the y-slide is stationary), there are four sources of errors affecting the measurement accuracy of the end face, namely βxФ, δz(x), α(x), and β(x), expressed by the formula: The δ plane X=δ(x)+βxФ.x+α(x).Th+β(x).Tx (5) In order to identify these 4 sources of error, first turn a 200mm diameter (appropriate diameter increase) end face, assuming that the end faces are symmetrical in the radial direction [3]. Mount the prism on the back of the tool post and at the same height as the tool and keep it constant. The tool is along the Z-direction collinear line [4]. At the same time, the tool is replaced with an inductance sensor. The laser is used to measure the straightness of the X slide plate several times. The least squares method is used to fit the angle between the direction of the laser and the X slide motion. The center is (based on this point) the starting point is measured along the feed direction: δ=δ pass-δ straight In the formula δ pass - sensor readings; Δz=f(x,z)=δ face Z+β(z).TX (6) 3.2 Error Source Modeling Taking into account that as long as the number of hidden nodes in the BP neural network is enough, any arbitrary nonlinear function can be fitted with any accuracy, and the interpolation accuracy is very high. In this paper, a neural network is used to establish the error compensation model. John C. Ziegert first applied this method, but failed to solve the training sample problem [5]. This article uses the MATLAB5.1 neural network toolbox to build an online measurement error compensation model, which is very convenient. All network topology structures use the BP network based on the Levenberg-Marquardt optimization algorithm. The hidden layer uses the "tansig" function and the output layer uses the "purelin" function. Fig. 2 Topological diagram of single input and single output neural network 2 Z-direction error compensation model: δ(xi,zj)=δ2X(xi)+δ1X(zj)xi,zj=0,2,4,...,100(mm) The same training sample with the Z-direction error compensation amount is: δZ(xi,zj)=δ1Z(xi)+δ1Z(zj)xi,zj=0,2,4,...,100(mm) From (7) and (8), the neural network training sample in the lathe machining space is [(δX(xi,zj), δZ(xi,zj)), (xi,zj)]xi, zj=0,2 ,4,...,100(mm)i,j=0,1,2,...,50; δX=δX(xi,zj)−δX(xi-1,zj-1) (9) δZ = δZ (xi, zj) - δZ (xi-1, zj-1) (10) Fig. 3 Topology diagram of double-input, double-outlet neural network 4 Online Measurement Experiments and Results Analysis In order to prove the correctness of the established model, after the error compensation model was established, online measurement was performed on the two machined parts by replacing the tool with an inductor with a measuring accuracy of 0.01 μm: Ï=-(θ+Ï€)/2Ï€-π≤θ≤-97Ï€ In order to calculate the flatness conveniently, the polar coordinate equation is converted into the rectangular coordinate parameter equation: Where θ=2Ï€t 0≤t≤48s 5 Conclusion The measurement results show that the error compensation model based on neural network proposed in this paper can improve the online measurement accuracy, the modeling is simple and accurate, and it is easy to use in practice. A good training sample can cover the entire machining space of the machine tool, in order to realize the online measurement automation, The measuring sensor or probe can be stored on the tool holder. For example, a turning tool holder can be used on the lathe. The tool is close to the workpiece during machining. When measuring, the sensor is close to the workpiece. This conversion process can be performed automatically. It is the goal of manufacturing technology to automatically ensure that the machining accuracy of parts is completely guaranteed by the machine tool itself, that is, the integration of machining and measurement technologies. It is not necessary to question whether on-line measurement technology is an important means for quality assurance and productivity improvement of parts, and plays a very important role in machining. Of course, there are still many problems to be solved in the wide application of online measurement technology, such as the development of high-precision sensors, measurement strategies, and optimization of data processing strategies. With the resolution of these problems, online measurement technologies will have brighter application prospects. . SHAOXING CARESO PLASTIC CO., LTD. , https://www.zjcareso.com
The on-line measurement of part machining accuracy can be divided into two cases. One is to directly measure the machining surface of the workpiece during the machining process. Once the machining process is over, the required accuracy index [1] can be obtained. This is the most ideal situation for on-line measurement. Second, after the machining process is finished, the workpiece is still installed on the machine tool, and the workpiece is measured with a reasonable measuring instrument [2]. In ultra-precision machining, the effect of thermal deformation on machining accuracy cannot be ignored. Therefore, constant temperature oil pouring or cooling of the cutting fluid is a must during the machining process. In the case of a high rotational speed of the cooling fluid and the workpiece, the measurement accuracy is achieved. The 0.01μm sensor is not yet available. Therefore, in ultra-precision machining, the accuracy of part machining is mainly measured by the traditional off-line measurement method. The cost of off-line measurement equals or exceeds the processing cost of the part in many cases.
Based on the above reasons, this paper studies the second situation to realize the on-line measurement of parts, and its essence is to use the lathe as a coordinate measuring machine. Because the precision of the movement of the moving parts of the submicron ultra-precision lathes developed is very high, even higher than many measuring instruments and measuring machines, if the machine tool and the appropriate measuring instruments are organically joined, the parts can be processed. Accuracy of online measurement, so that the machine can be used for processing, but also for measurement purposes, to expand the scope of application of the machine tool, but also to solve the measurement problem of the parts [3]. Nowadays, the trend of machining quality assurance is to ensure that the quality assurance is closer to the processing process by replacing offline measurement and statistical quality control with on-line measurement, and that it is a qualified product to ensure that the parts are unloaded from the processing equipment. Of course, this requires a prerequisite for on-line measurement. Efficiency and accuracy must be guaranteed so that comprehensive decisions and necessary compensation can be achieved with minimal time delay. Therefore, it is of great practical significance to study the on-line measurement technology of the part processing accuracy.
Of course, the lathe is used as a coordinate measuring machine, and the on-line measurement accuracy is also affected by the accuracy of the measuring sensor and the measurement strategy and data processing strategy. Many advanced technologies were used in the design and manufacture of the lathe (with a T-shaped layout) to reduce or eliminate the influence of thermal deformation errors on the lathe motion accuracy. For example, the lathe uses an air static pressure spindle and uses a white dense jade as a spindle and bearing. Materials; lathe slide plate adopts air static pressure guide; granite material is used for lathe spindle, slide plate, bed and guide rail; temperature between processing is controlled at 20±0.1°C, etc., so the error source affecting measurement accuracy is mainly the machine tool. Geometric errors, a total of 21 items, ie 6 errors per moving part and 3 mutual position errors between 3 axes. The 21 errors are shown in the table below. Accurate and rapid identification of error sources is the basis for achieving high-precision on-line measurements, taking into account that the movements of the measurement process and the machining process are very similar (tools are replaced by sensors or measuring probes, of course, the error compensation models of the two are similar) Error sources in non-error-sensitive directions have negligible influence on measurement accuracy, ie δ(x), γ(x), δY(z), γ(z), δY(Ф) and α(Ф) are not considered separately. The influence of the measurement accuracy, plus βXZ does not affect the measurement accuracy (included by αZФ and βXФ). In the machine tool qualification process, the spindle rotation accuracy is measured. The measured result is the spindle radial runout error δX(Ф) and Axial turbulence error δZ (Ф) is less than 0.05μm, which is smaller than the straightness error of slideway (δZ(x) ≤ 0.18μm/100mm, δX(z) ≤ 0.20μm/100mm), and the yaw error of the main shaft is β. (Ф) is also very small, so when measuring the part machining accuracy online, the influence of the spindle rotation error on the measurement accuracy is not considered separately. Of course, if high-precision measurement of large-size parts (large-size mirrors) is to be performed on-line, then the spindle's rotary error (eg, β(Ф)) must be considered.
There are six sources of error affecting the X-direction measurement accuracy: δX(z), αZФ, β(z), α(z), δX(x), and β(x). The X-direction error compensation amount is expressed as:
WZ - the distance of the measuring point from the suction cup in the Z direction;
TZ - the distance from the Z-sensor to the center of gravity of the moving part of the X-slide;
Wh - the distance from the center of gravity of the measuring point on the workpiece in the vertical direction to the center of gravity of the Z-slide (the headstock is mounted on the Z-slide to become one);
In order to identify the source of error, the cylindrical surface is turned in the Z direction to the 100mm machining area (X slide is stationary), and the error of the workpiece bus is measured in the direction of the tool relative to the sensor, and the error compensation amount in the X direction when the cylindrical surface is measured can be obtained. Column X, as shown in Fig. 1, has 4 items that affect the accuracy of the workpiece bus: δX(z), αZФ, β(z), and α(z). The error compensation amount is expressed as: 
The error components affecting the Z-direction measurement accuracy are 6 items: δz(x), β(z), βxФ, α(x), β(x), and δz(z). The amount of error compensation is expressed as:
WX - the distance of the measuring point on the workpiece from the center of the suction cup in the X direction;
Th—the distance of the sensor from the center of gravity of the moving part of the X-slide;
TX - the distance from the X sensor to the center of gravity of the X-slide moving part;
δ straight - X slide motion error.
Fitting a straight line to the δ-sequence gives 2βxФ. At the same time, the shape error δ line (x) of the face radial bus bar is obtained. Thus, the error compensation amount from the center of the workpiece to the measurement facet is obtained. δ surface Z=δ line (x)+x .βxФ, in order to overcome the influence of random errors, it is necessary to turn multiple measurements several times. It is known from Equations (4) and (5) that there are β(z) and δz(z) in the Z-direction error source, β( z) Offline measurement and identification with laser measuring equipment, δz(z) is ensured by the lathe full-closed-loop system, and the Z-direction error compensation amount when measuring the end face is:
1 X-direction error compensation modelling:
Take the z position of the slide plate as the input of the network and the δ column X as the output, and use the neural network A (structure shown in Fig. 2) to fit the function δ1X=δ column X=f(z); Position x as the input of the network, β(z).TZ as the output, and neural network B the fitted function δ2X=β(z). TZ=f(x), where δ1X=δ column X=f(z) is Error compensation model for measuring the cylindricity of the part and straightness of the busbar; 
Take X position x as the network input and δ surface Z as output, and directly use neural network C to fit the function δ1Z=δ plane Z=f(x); position z of the Z slide is used as the input of the network. β(z).TX is used as the output, and the neural network D-fitting function δ2Z=β(z).TX=f(z), δ1Z=δ plane Z=f(x) is the error compensation when measuring the planeness of the end face on line model.
In the actual processing, the frequently used area is 100mmmm×100mm. Since the pitch of the driving screw is 5mm, the straightness error clearly shows the component with the pitch as the cycle. In order to accurately model, the processing interval is divided equally. 51 × 51 points (ensuring sufficient density of sampling points), that is, 2601 training samples are required. The training sample calculation method is as follows.
X-axis error compensation amount training sample calculation:
When the δ-column X is measured, sampling points (δ-column X(zj), zj) zj=0, 2, 4, ..., 100 (mm) are obtained; j=0, 1,..., 50 is within 0 to 100 mm. A total of 51 points were sampled; in the measurement of β(x).Tz, the sampling points (δ2X(xi),xi)xi=0,2,4,...,100(mm);i=0,1,..., 50 is to sample 51 points within 0-100mm, that is, the training sample of the X-direction error compensation amount is:
i,j=0,1,2,...,50; (7)
i,j=0,1,2,...,50; (8)
In this way, (xi, zj) is the network input, and (δX(xi,zj), δZ(xi,zj)) is the output. A double-input dual-output neural network can be used to fit the on-line measurement error compensation amount in the lathe machining space. The model (excluding the influence of δX(x), δZ(z)) and its topological structure are shown in Fig. 3. The above modeling process knows that when measuring a surface such as a spherical surface, the amount of error compensation in the X, Z direction at the point (xi, zj) is: 
1 Turn a cylindrical surface (busbar is 100mm), measure the straightness of the busbar of the workpiece online, the straightness error of the busbar is 0.26μm (comparing the straightness after processing is 0.10μm); on the condition that the spindle is mounted on the codewheel, measure the workpiece cylinder online The degree of error is 0.38μm (0.21μm after compensation), and the result is 0.40μm on a TAYLOR cylindrimeter;
2 Turn an end face (diameter 100mm) to measure the flatness error of the workpiece end face on-line. The measurement result is that the non-compensated part flatness error is 0.80μm, the off-line measurement result is 0.78μm, and the compensation part flatness error is 0.12μm. The plane flatness measurement principle is: the spindle speed range of the CNC lathe is 48-3000r/min stepless speed regulation, the spindle speed is reduced to 60r/min; the tool is replaced with a capacitance sensor, and the voltage output is through the MC-1249 optical isolator The entrance interface board is sampled by the computer and measured from the center of the workpiece. After the sensor is centered, the sensor value is used as the zero point. The trajectory of the sensor relative to the workpiece is an Archimedean spiral. The polar coordinate equation is: 
From the above equation, it can be seen that: when the spindle rotates one revolution, the sensor moves radially in the opposite direction to the machining by 1 mm (X slide moving speed range from 0 to 200 mm/min), and the workpiece radius is 50 mm. Considering the probe diameter, the workpiece is sampled. For 48 laps, the sampling interval is 0.1 s, ie, 10 points per lap, sampling is stopped when 481 points are collected, and the deviation distribution of each point after error compensation is zi=f(xi, yi), which can be obtained by least square method. To obtain the flatness error, the sampling interval can be larger in the peripheral area of ​​the workpiece. Of course, according to actual needs, the speed of the spindle and the speed of the carriage can be adjusted to obtain a reasonable sampling point distribution.
It is known from the above measurement results that the on-line measurement accuracy is satisfactory after error compensation. If the spindle rotation error is simultaneously compensated, on-line measurement accuracy will be higher. Because there is no high-precision measurement probe, it cannot measure the surface such as spherical surface online. If there is a high-precision measurement probe and corresponding data processing software, the lathe can be used as a high-precision coordinate measuring machine.