CN102267095A

CN102267095A - Method for monitoring and dressing grinding wheel on line

Info

Publication number: CN102267095A
Application number: CN 201110247142
Authority: CN
Inventors: 王洪; 许世雄; 许君; 王东昱; 戴瑜兴
Original assignee: HUNAN YUHUAN TONGXIN CNC MACHINE TOOL CO Ltd
Current assignee: HUNAN YUHUAN TONGXIN CNC MACHINE TOOL CO Ltd
Priority date: 2011-08-26
Filing date: 2011-08-26
Publication date: 2011-12-07
Anticipated expiration: 2031-08-26
Also published as: CN102267095B

Abstract

The invention discloses a method for monitoring and dressing a grinding wheel on line. The method comprises the following steps of: detecting acoustic emission (AE) digital signals which are sent by the grinding wheel in a grinding process and processed by using the conventional grinding wheel dressing system; performing fast Fourier transformation on N AE digital signals a(n) in the previous step to obtain different spectral components of N/2 real parts AR(k) and imaginary parts AI(k); and calculating by formulae (1, 2 and 3) to obtain a root-mean-square value (Y), total energy (E), and a center frequency (fc) of a frequency domain, which serve as input samples of a back propagation (BP) neural network, of the AE digital signals. In the method, the root-mean-square value (Y), the total energy (E) and the center frequency (fc) of the frequency domain of the AE digital signals serve as feature extraction values for the first time and then serve as input values of the BP neural network; and the starting time and ending time for dressing of the grinding wheel are accurately predicted by training and test of the BP neural network, so an automatic dressing function of the grinding wheel is realized. The method is high in prediction accuracy.

Description

Grinding wheel online monitoring and dressing method

Technical Field

The invention belongs to a grinding machine processing control method, and particularly relates to an online monitoring and trimming method for a grinding wheel.

Background

The dressing of the grinding wheel has great influence on the grinding performance and the grinding effect of the grinding wheel. With the development of intelligent and adaptive control of grinding processing, the accurate time for finishing the passivation of the grinding wheel and the finishing of the grinding wheel can be automatically detected, and the finishing of the grinding wheel can be started and finished in time, so that the grinding quality and the production efficiency are improved. In the past, timing dressing is mostly adopted to avoid workpiece grinding burn, and when the grinding wheel does not reach the working life limit, the grinding wheel is dressed in advance, so that the method is blindly carried out ("grinding quality on-line monitoring method research", Liu Guijie et al, diamond and abrasive grinding tool engineering, 2004.10). Frequent dressing of the grinding wheel not only affects machining efficiency, but also accelerates wear of the grinding wheel, and particularly, the use of Cubic Boron Nitride (CBN) grinding wheels is expensive, increasing production costs. On the contrary, if the finishing period is delayed, the precision and the surface quality of the workpiece are influenced, and waste products are caused. Therefore, the grinding wheel can be accurately and automatically dressed in time, and the method is an important way for improving the grinding processing productivity and ensuring the grinding processing quality.

In recent development and research, Acoustic Emission (AE) signals have been widely used as an information source for grinding control. Many studies have shown that: by monitoring the amplitude change of the AE signals, the sharpness of the grinding wheel can be evaluated and the service life of the grinding wheel can be determined (the neural network-based online monitoring of the grinding wheel state, Liu Gui Jie, etc., the university report of northeast China, 2002.10; the application of STFT in the AE signal feature extraction, Liao Jun, etc., the instrument and meter report, 2008.9; the AE monitoring model of the grinding process, Wu national month, grinding machine and grinding, 1999.1). However, when the material, the processing condition and the processing parameters of the processed workpiece change frequently, the amplitude of the acoustic emission signal also changes violently, and the degree of the grinding wheel passivation cannot be judged only by detecting the amplitude of the acoustic emission signal. Therefore, the FFT spectral analysis method is provided, namely, the amplitude of the high-frequency component generated by machining when the grinding wheel is sharp is large and the spectral component is more, and the amplitude of the high-frequency component is small and the spectral component is less when the grinding wheel is passivated. A band limit is set, when the high-frequency component is less than a certain value, the grinding wheel needs to be dressed, and when the high-frequency component reaches a certain range, the dressing of the grinding wheel is finished, so that the automatic dressing of the grinding wheel is realized.

Disclosure of Invention

The invention aims to provide an online monitoring and trimming method for a grinding wheel.

The invention provides an online monitoring and trimming method of a grinding wheel, which is realized by the following steps:

step 1, detecting an Acoustic Emission (AE) digital signal which is emitted and processed in a grinding process of a grinding wheel by using an existing grinding wheel dressing system;

step 2, obtaining N Acoustic Emission (AE) digital signals a (N) of the step 1 through Fast Fourier Transform (FFT)

Different spectral components of real part ar (k) and imaginary part ai (k);

and 3, calculating by formulas (1), (2) and (3) to obtain a root mean square value Y, a total energy E and a frequency domain center frequency fc of the Acoustic Emission (AE) digital signal as an input sample of the BP neural network:

in the formula: y is the root mean square value of the digital signal of the Acoustic Emission (AE), a (N) is the digital signal of the Acoustic Emission (AE), N is the sampling number of the digital signal of the input Acoustic Emission (AE), k is 0, 1, N, AR (k) is the real part frequency spectrum after FFT calculation, AI (k) is the imaginary part frequency spectrum after FFT calculation, E is the total energy of the digital signal of the Acoustic Emission (AE), f_cThe frequency-domain center frequency of the Acoustic Emission (AE) digital signal, f being the resolution frequency of the EFT spectral analysis, e.g. the Acoustic Emission (AE) signalMaximum frequency of number f_maxSampling frequency of f_sIf N is selected as the sampling point, the resolution frequency is f ═ f_s/N。

Step 4, the BP neural network takes a part of the samples in the step 3 for learning and training, so that the error value generated by the sample training is within a set range, and the weight V of the BP neural network is obtained_ij、W_JkAnd a control value R_j、Q_kInputting the rest samples into the BP neural network matrix to be tested to obtain numerical values in two different ranges in the passivation state and the sharpness state of the grinding wheel, wherein the numerical values are used for predicting the starting time and the ending time of grinding wheel dressing respectively;

step 5, programming the processes in the

steps

2, 3 and 4 by using VB and Mitrix-VB software tools, and calculating a root mean square value Y, signal energy E and central frequency f of a signal_cWeight V of sum BP neural network_ij、W_jkControl value R_j、Q_kAnd the numerical values of the grinding wheel in two different ranges in the passivation state and the sharp state are used for verifying whether the result in the step 4 achieves the expected aim of automatic grinding wheel dressing, and the program is transplanted into a numerical control system of a camshaft grinding machine to realize the automatic grinding wheel dressing function.

The method firstly uses the root mean square value (Y), the total energy (E) and the frequency domain center frequency (f) of an Acoustic Emission (AE) digital signal_c) The method has the advantages that the prediction accuracy is high, the machining precision and the machining efficiency are improved, the service life of the grinding wheel is prolonged, and the production cost of an enterprise is reduced.

The invention is further described with reference to the following figures and detailed description.

Drawings

FIG. 1 is a block diagram of an automatic grinding wheel dressing system for a camshaft grinding machine.

FIG. 2 is a diagram of a BP neural network model.

Fig. 3 is a flowchart of the BP neural network training procedure.

Fig. 4 is a flowchart of a BP neural network test procedure.

FIG. 5 is a parameter input diagram of a BP neural network.

FIG. 6 is a graph of the result of the BP neural network training operation.

FIG. 7 is a graph of the result of the BP neural network training operation.

Detailed Description

Fig. 1 shows a block diagram of a grinding wheel dressing system of a conventional YTMCNC8326 numerically controlled camshaft grinder, which comprises a grinding wheel, an AE sensor, a diamond roller dresser, a workpiece clamping device, a marsops P7WB dynamic balancer, a siemens 840D numerically controlled system, a siemens 611D driver and the like. The AE sensor is arranged at the axle center of the dynamic and static piezoelectric main shaft, the MARPOSS dynamic balancing instrument detects an AE signal, digitalizes and filters the AE signal, and then sends the AE signal to the NCU through the Profibus bus, the MMC user interface (OEM) through the NCU or the MMC user interface (OEM) through the serial port COM. And performing spectrum analysis in an OEM (original equipment manufacturer), and controlling the moment of starting or finishing the grinding wheel dressing according to different spectrums at different stages to realize automatic dressing of the grinding wheel.

The steps and the principle of the method are as follows:

Different spectral components of real part ar (k) and imaginary part ai (k);

principle of FFT algorithm

In digital signal processing, Discrete Fourier Transform (DFT) provides a mathematical method for Fourier Transform operations using a digital computer. The DFT calculation is:

wherein,

referred to as fourier factors.

However, the calculation workload of DFT is very large, and Fast Fourier Transform (FFT) algorithm is an algorithm for reducing the calculation time of DFT developed on the basis of DFT, which greatly improves the calculation efficiency. The present invention is directed to a typical form of FFT (i.e., the kurily-graph-radix algorithm), which is a radix-2 FFT algorithm that takes N to the power M of 2: n is 2^MM is a positive integer, and the algorithm starts by decomposing the N-point DFT operation into two groups

And (3) point DFT operation, wherein the calculation formula is as follows:

wherein W is as defined above; r is 0, 1, …,

2r denotes an even number, and 2r +1 denotes an odd number.

By the method, the operation can be divided into four groups of even numbers and odd numbers in the even numbers and the odd numbers, and the like.

The spectral distributions of the various signals can be obtained by the algorithm described above.

Step 3, using a well-known feature space expression method as a proper 'feature', wherein the root mean square value Y and the frequency domain center frequency f of the Acoustic Emission (AE) digital signal_cAnd total energy E is selected, and the root mean square value Y, the total energy E and the frequency domain center frequency f of the Acoustic Emission (AE) digital signal are obtained through calculation of the formulas (1), (2) and (3)_cAnd as input samples for the BP neural network:

in the formula: y is the root mean square value of the digital signal of the Acoustic Emission (AE), a (N) is the digital signal of the Acoustic Emission (AE), N is the sampling number of the digital signal of the input Acoustic Emission (AE), k is 0, 1, N, AR (k) is the real part frequency spectrum after EET calculation, AI (k) is the imaginary part frequency spectrum after FFT calculationSpectrum, E is the total energy of the Acoustic Emission (AE) digital signal, f_cThe frequency domain center frequency of the Acoustic Emission (AE) digital signal, f is the resolution frequency of FFT spectral analysis, e.g. the maximum frequency of the Acoustic Emission (AE) signal is f_maxSampling frequency of f_sIf N is selected as the sampling point, the resolution frequency is f ═ f_s/N。

establishing a BP neural network model:

(1) topology of neural network

The neural network can establish a relation model for a multi-input and multi-output complex nonlinear system, and is very suitable for replacing human decision behaviors based on experience.

In the application, a three-layer BP neural network is adopted as a grinding wheel state identification model, and as shown in FIG. 2, a general three-layer BP neural network model diagram is provided. In fact, the selection of the number of hidden layer nodes is not regular and the solution is to gradually increase the number of neurons from small to large until higher precision and faster convergence rate are obtained. The self-constructed neural network is selected, a hidden layer node is given to the network firstly, and the error of the hidden layer node is calculated; then adding a hidden layer node; and if so, exiting and determining the number of hidden layer nodes of the network. The result of training according to the method shows that the activation function is a Sigmoid function, the topological structure is a 3-15-1 neural network, and the recognition accuracy is about 98%, so that the topological structure is selected.

(2) Neural network learning algorithm

Assuming a given pattern x_iFor network input, O_kFor network output, d_kFor the target value, j is 0, 1, …, m for each output layer; k is 1, 2, … l; the hidden layers are all i-0, 1, …, n; j is 1, 2, …, m.

d_kTo a target value, O_kTo output a value, E_pIs the sum of the squares of the error values

For the output layer or layers, the number of layers,

can be unfolded as follows:

wherein is

f (x) = \frac{1}{1 + e^{- x}},

An error signal is output.

In the case of a hidden layer or layers,

can be unfolded as follows:

wherein

Latent layer error signal, W_jkOutput layer weight

<math> <mrow> <msup> <mi>f</mi> <mo>′</mo> </msup> <mrow> <mo>(</mo> <msub> <mi>net</mi> <mi>k</mi> </msub> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <msub> <mrow> <mo>&PartialD;</mo> <mi>O</mi> </mrow> <mi>k</mi> </msub> <mrow> <mo>&PartialD;</mo> <msub> <mi>net</mi> <mi>k</mi> </msub> </mrow> </mfrac> <mo>=</mo> <msub> <mi>O</mi> <mi>k</mi> </msub> <mrow> <mo>(</mo> <msub> <mrow> <mn>1</mn> <mo>-</mo> <mi>O</mi> </mrow> <mi>k</mi> </msub> <mo>)</mo> </mrow> </mrow> </math>

<math> <mrow> <msup> <mi>f</mi> <mo>′</mo> </msup> <mrow> <mo>(</mo> <msub> <mi>net</mi> <mi>j</mi> </msub> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <msub> <mrow> <mo>&PartialD;</mo> <mi>y</mi> </mrow> <mi>j</mi> </msub> <mrow> <mo>&PartialD;</mo> <msub> <mi>net</mi> <mi>j</mi> </msub> </mrow> </mfrac> <mo>=</mo> <msub> <mi>y</mi> <mi>j</mi> </msub> <mrow> <mo>(</mo> <msub> <mrow> <mn>1</mn> <mo>-</mo> <mi>y</mi> </mrow> <mi>j</mi> </msub> <mo>)</mo> </mrow> </mrow> </math>

Thus, there are:

wherein: yj is the hidden layer output value.

The weight correction of the output layer is

Aw_jk＝η(d_k-O_k)O_k(1-O_k)y_j (11)

The hidden layer weight correction amount is

In order to improve the network convergence characteristics, the correction expression Δ W (n) is selected

ΔW(n)＝ηδX+αW(n-1)

Wherein, W is a certain layer weight matrix, X is a certain layer input vector, and eta is a learning factor; α is a proportionality constant for adjusting the variation.

(3) VB6.0 programming for realizing FFT spectral analysis and training and testing of BP neural network

The grinding wheel dressing control interface is mainly used for collecting AE signals, performing spectral analysis on the AE signals, drawing an AE signal spectrogram, extracting a characteristic root mean square value Y of the AE signals, the central frequency fc of frequency domain amplitude and the total energy E, and judging the starting time and the ending time of grinding wheel dressing through BP neural network prediction.

The slave balancer receives digital AE signals through a serial port or a Profibus bus, and the digital AE signals are divided into 4 groups: the 1 st group is the digital AE signal that the emery wheel was gathered when sharp grinding work piece, the 2 nd group is the digital AE signal that the emery wheel was gathered when passive grinding work piece, the 3 rd group is the digital AE signal that the passivation emery wheel was gathered when the diamond-impregnated roller was maintained, the 4 th group is the digital AE signal that the diamond-impregnated roller was gathered when the sharp emery wheel was maintained. Each group collects 200 segments, and each segment collects 256 sample points. Respectively carrying out spectrum analysis on the 4 groups of AE signals, calculating the root mean square value Y of the AE signals, the central frequency fc of the frequency domain amplitude and the characteristic value of the total energy E, then training the neural network by taking the characteristic values of the 1 st group and the 2 nd group as the input value of the BP neural network until the error value reaches the allowable range, and identifying the grinding workpiece of the grinding wheel by using the network to obtain the initial time of the grinding wheel needing to be dressed. The characteristic values of the 3 rd and 4 th groups are trained by the same method, and the finishing time of the grinding wheel can be obtained.

Fig. 3 is a flowchart of a training procedure of the BP neural network, and fig. 4 is a flowchart of a testing procedure of the BP neural network.

Step 5, programming the processes in the

steps

Examples of the applications

(1) The AE digital signals are read from the marsposs dynamic balancer through the serial port, and the digital AE signals are divided into 4 groups: the 1 st group is the digital AE signal that the emery wheel was gathered when sharp grinding work piece, the 2 nd group is the digital AE signal that the emery wheel was gathered when passive grinding work piece, the 3 rd group is the digital AE signal that the passivation emery wheel was gathered when the diamond-impregnated roller was maintained, the 4 th group is the digital AE signal that the diamond-impregnated roller was gathered when the sharp emery wheel was maintained. Each group collects 200 segments, and each segment collects 256 sample points.

(2) And (3) carrying out Fast Fourier Transform (FFT) on 4 groups, 200 sections in each group and 256 sampling points in each section through the formula (5) in the step (2), programming the formula (5) in the step (5), and calculating different spectral components of the real parts AR (0) -AR (127) and the imaginary parts AI (0) -AI (127) at each end of the 4 groups, 200 sections in each group.

(3) 4 groups, 200 segments in each group, 256 sampling points in each segment, 4 groups, 200 segments in each group, and different frequency spectrum components of real parts AR (0) -AR (127) and imaginary parts AI (0) -AI (127) at each end respectively. The root mean square value Y, the frequency domain center frequency fc and the total energy E characteristic value of 4 groups of 200 Acoustic Emission (AE) digital signals can be calculated by respectively calculating the formulas (1), (2) and (3) in the step 3 and programming the formulas (1), (2) and (3) in the step 5, wherein the table 1 shows partial calculated values of one group. f. of_s250kHz, N1024, then: f ═ f_s/N＝250000/1024＝244Hz。

TABLE 1 extracted eigenvalues

(4) The 4 groups of characteristic values are divided into two groups respectively to carry out BP neural network training, one group comprises the root mean square value Y, the center frequency fc and the total energy E characteristic value of 200 Acoustic Emission (AE) digital signals when the grinding wheel of the 1 st group grinds the workpiece in a sharp manner, and the root mean square value Y, the frequency domain center frequency fc and the total energy E characteristic value of 200 Acoustic Emission (AE) digital signals when the grinding wheel of the 2 nd group grinds the workpiece in a passive manner. And (4) sending the calculated formula into a formula calculation in the step (4), programming the formula in the step (4) through a step 5, sending the characteristic values of the root mean square value Y, the frequency domain center frequency fc and the total energy E characteristic value of the 160 AE signals when the grinding wheel of the 1 st group in one group grinds the workpiece in a sharp manner and the characteristic values of the root mean square value Y, the frequency domain center frequency fc and the total energy E of the 160 AE signals when the grinding wheel of the 2 nd group grinds the workpiece in a passive manner to a BP neural network, training, setting the ideal output when the grinding wheel is sharp to be 1, setting the ideal output when the grinding wheel is passive to be 0, and then training to enable an error curve to be within an allowable range. Training is carried out by using a neural network interface programmed by VB6.0, a BP neural network parameter input graph is shown in fig. 5, and a BP neural network operation result graph is shown in fig. 6. Note: the AE signal has a root mean square value Y, a frequency domain center frequency value fc × 0.001 and an energy value E × 0.1.

(5) The weight and threshold values trained by the neural network are stored (as shown in FIG. 6, V in a graph showing the operation result of BP neural network training_ji、W_kj、R_j、Q_k) And then taking out the remaining 40 groups of the groups to test the neural network, judging the correctness of the grinding wheel at the time of sharpness or passivation when the grinding wheel processes the workpiece, wherein the operation result is shown as a BP neural network test result chart in fig. 7. Similarly, another group of characteristic values of the root mean square value Y, the frequency domain center frequency fc and the total energy E of 160 AE signals when the passivated grinding wheel is dressed by the diamond roller of the group 3 and a group 4 of characteristic values of the root mean square value Y, the frequency domain center frequency fc and the total energy E of 200 Acoustic Emission (AE) digital signals when the sharp grinding wheel is dressed by the diamond roller of the group 4 are used for training and testing another neural network, and the correctness of whether the grinding wheel is passivated or sharp at the time of dressing is judged. So as to realize the automatic dressing function of the grinding wheel.

The method of the invention independently adopts the acoustic emission AE digital signal output by the dynamic balancing instrument, carries out FFT spectrum analysis on the signal, calculates the root mean square value Y, the frequency domain center frequency fc and the total energy E characteristic value of the Acoustic Emission (AE) digital signal, then sends the signal to the BP neural network for training and testing, obtains the accurate grinding wheel passivation and grinding wheel sharp finishing time, and realizes the grinding wheel on-line finishing function. Compared with the original method, the method has the advantages of intellectualization, high grinding wheel prediction accuracy, improvement of machining precision and machining efficiency, prolongation of the service life of the grinding wheel and reduction of the production cost of enterprises.

Claims

1. An online monitoring and finishing method for a grinding wheel is characterized by comprising the following steps:

step 2, obtaining N Acoustic Emission (AE) digital signals a (N) of the step 1 through Fast Fourier Transform (FFT)A sum of real part AR (k) and imaginary part AI (k)Co-frequency spectral components;

step 3, calculating by formulas (1), (2) and (3) to obtain the root mean square value Y, the total energy E and the frequency domain center frequency f of the Acoustic Emission (AE) digital signal_cAnd as input samples for the BP neural network:

in the formula: y is the root mean square value of the digital signal of the Acoustic Emission (AE), a (N) is the digital signal of the Acoustic Emission (AE), N is the sampling number of the digital signal of the input Acoustic Emission (AE), k is 0, 1, N, AR (k) is the real part frequency spectrum after FFT calculation, AI (k) is the imaginary part frequency spectrum after FFT calculation, E is the total energy of the digital signal of the Acoustic Emission (AE), f_cThe frequency domain center frequency of the Acoustic Emission (AE) digital signal, f is the resolution frequency of FFT spectral analysis, e.g. the maximum frequency of the Acoustic Emission (AE) signal is f_maxSampling frequency of f_sIf N is selected as the sampling point, the resolution frequency is f ═ f_s/N。

step 5, programming the processes in the steps 2, 3 and 4 by using VB and Mitrix-VB software tools, and calculating a root mean square value Y, signal energy E and central frequency f of a signal_cWeight V of sum BP neural network_ij、W_jkControl value R_j、Q_kAnd the numerical values of the grinding wheel in two different ranges in the passivation state and the sharp state are used for verifying whether the result in the step 4 achieves the expected aim of automatic grinding wheel dressing, and the program is transplanted into a numerical control system of a camshaft grinding machine to realize the automatic grinding wheel dressing function.