Sort
Technical Communicationpreface
The main methods of monitoring tool wear and damage are cutting force method, machine power method, vibration and acoustic emission method, etc. Cutting force method is not easy to be accepted by users because of the need to change machine structure when installing force measuring device.
Compared with the cutting force method, the machine tool power method has the advantages of simple signal measurement, avoiding the interference of chip, oil, vibration and so on in the cutting environment, and easy installation of monitoring device.
Research shows that when the tool is broken, the cutting force is increased by squeezing the broken piece between the tool and the workpiece. Then the cutting force decreases to zero as the piece falls off. When the tool touches the workpiece again, the cutting force suddenly increases again. However, the author found in the test process that the motor power, as an indirect reflection of cutting force, has more variations when the tool is damaged, which reflects the randomness of the tool damage.
Acoustic Emission (AE) signals are elastic stress waves generated by the sudden release of elastic deformation energy inside the material. They directly come from the cutting area and are characterized by high frequency, sensitivity and fast response, which is very suitable for tool damage monitoring.
However, there are many AE sources in the cutting process, such as material rupture, tool damage, chip breaking, chip and workpiece impact, so how to effectively deal with AE signals, extract the characteristic quantity sensitive to tool damage from a large number of AE signals is a key problem.
How to effectively process the collected signals, extract the characteristic quantity reflecting the tool state, and fuse the multi-sensor information to correctly and effectively identify the tool wear and damage state has not been well solved.
There are many model decision methods for multi-sensor information fusion, such as pattern recognition, fuzzy clustering, neural network and fuzzy neural network.
1 net weight grey prediction model
Grey System theory is a new theory of both soft and hard science, which was founded by the famous scholar Professor Deng Jurong in the early 1980s.
In this theory, the system whose information is completely clear is defined as white system, the system whose information is completely unclear is defined as black system, and the system whose information is partly clear and partly unclear is defined as gray system.
For the tool condition monitoring system, the influence of tool wear back cutting tool amount, feed, spindle speed and other information is known, but in the actual cutting process, there are a variety of abnormal factors, such as uneven tool material, workpiece material, blank processing allowance and other unforeseeable factors, in line with the characteristics of the grey system. So the prediction theory and method of grey system can be used to analyze the development and change law of tool wear state.
Different from traditional data series processing methods with random changes, grey system theory regards all random variables as grey variables that change within a certain range. Instead of conducting large sample analysis and research on grey variables from the perspective of statistical rules, data processing methods (data generation and reduction) are used. The chaotic original data is sorted into generated data with strong regularity for study, that is, what grey system theory establishes is not the original data model, but the generated data model.
Through the processing of the original data and the establishment of the grey model, we can find and master the law of the development of the system, and make scientific quantitative prediction of the future state of the system. Grey system theory prediction model GM(1,1) model, the basic method of establishment is shown in the literature.
2 net weight Improvement of grey prediction model
In order to improve the prediction accuracy, the grey prediction model should be modified, mainly to improve the processing of the original data and the prediction model itself.
2.1 Smooth processing of original data
The smoother the original data series, the higher the prediction accuracy.
There are many methods to improve the smoothness of the original data series: it can be weighted. In the time series, different weights are given to the data whose reliability changes in direct proportion to time. Some function (square root function, logarithm function, power exponent function, etc.) transformation can also be used to smooth the transformation of the original data, and an intermediate value can also be inserted between two collection points. All these methods can improve the prediction accuracy to a certain extent.
In order to simplify data analysis and processing, this paper adopts the last method, that is, inserting an intermediate value between two collection points to improve the accuracy of the grey prediction model.
2.2 Equal-dimension new information model
With the passage of time, new detection data are constantly added, and the data series used for prediction will be constantly enlarged. The GM(1,1) model built with all the data is called holographic model.
However, as a matter of fact, new data bring new disturbances and driving factors into the system, so that the system is affected by the latest information and new trends occur. Accordingly, the information significance of the old data will decrease over time. In addition, the addition of new data also makes the dimension of information data sequence increase, the amount of computation is also increasing, which is not convenient for computer processing.
Therefore, the GM(1,1) model should be treated with equal-dimensional new information. That is, every time a new piece of information is added, the oldest piece of information will be removed. In this way, not only the dimension of the original series used for prediction remains fixed, but also the optimal amount of information is guaranteed, thus reducing the gray plane and improving the prediction accuracy.
2.3 Application of improved grey desert type in tool condition monitoring
The vibration signals collected in the process of machining contain very rich working condition information. How to extract the characteristic information is the key to the realization of tool wear and damage monitoring.
Wavelet technology is the most influential method of signal processing at present. Wavelet transform is a mathematical tool for multi-resolution analysis of abrupt signals and non-stationary signals. Its main advantages are: linear transformation, no distortion; The signal can be analyzed locally in time domain and frequency domain simultaneously. Mainly applicable to broadband signal processing and localization analysis (short data analysis). These advantages make it the most suitable for the state number analysis of machining process, and also suitable for the analysis of the tool wear state signal.
Grey prediction theory can use at least 4 original data to establish a grey prediction model conforming to the rule in a certain period, which can solve the problems of less historical data, poor sequence integrity and low reliability. Moreover, the irregular original data can be formed into a series of strong rules, which is easy to operate, high precision and easy to check.
The modeling process is as follows:
(1) Smooth the original data and insert a value between any two collection points, which is the average value of the left and right values in the sequence. Then the original data sequence after inserting the intermediate value is:
X(0)=[0,0.55,1.1,1.2,1.3,1.4,1.5,1.55,1.61,1.67,1.73,2.02,2.3,2.95,3.6]
(2) The grey prediction model established after summing x(0) is as follows:
(3) The sequence can be obtained by calculating Formula (1)
The predicted data is generated after a single reduction and the insertion values are removed, as shown in Table 2.
As can be seen from the prediction curve in the figure, the tool wear process is divided into three distinct stages: initial wear stage, normal wear stage and rapid wear stage.
At the early wear stage (section 0A curve in the figure), the growth of tool wear is fast, mainly because the new grinding tool after the surface of the roughness and micro-cracks, oxidation or shedding layer defects, and the cutting edge is sharp, after the tool surface and machining surface contact area is small, compressive stress is concentrated, soon after the tool surface grinding out a narrow surface. As the pressure is reduced, the rate of wear is stabilized and normal wear is entered.
After the initial wear, the rough surface of the tool has been worn, and the growth of the tool wear value is relatively slow and roughly linear with the cutting time, which is called the normal wear stage (AB curve in the figure). This stage time is longer, is the effective working time of the tool.
The curve of the normal wear phase is close to an upward sloping straight line, the slope of which represents the intensity of the wear. This time should be through the program to compensate the amount of tool wear.
When the width of the wear zone reaches a certain value, the roughness of the machining surface increases, the cutting force and the cutting temperature increase sharply, and the tool wear intensity increases sharply. With the continuation of the cutting time, the tool wear value increases quickly, and the cutting ability will be lost soon. This stage is the sharp wear stage (the curve after point B in the figure).
In the stage of rapid wear, we need to determine the tool is in serious wear through a large number of experimental data, and with the help of wavelet signal singular value detection method, to get the tool wear threshold.
In order to use the tool reasonably, ensure the quality of processing, should avoid making the tool wear into this stage, before this stage, it is necessary to change the tool or replace the blade in time.
After the end of the prediction process, when the time reaches the next acquisition interval, the monitoring system carries out signal acquisition and processing again, and calculates the next tool wear characteristic value.
After adding a new eigenvalue data, the tool state should be re-identified to predict a new wear time. If the newly added data is smaller than the previous value, it belongs to the error value and will not be identified and predicted this time. Such detection, status recognition, and prediction are repeated until the tool is severely worn or damaged, and the tool monitoring is stopped.
► Recognition algorithm/adaptive resonance neural network
Art-2, an Adaptive Resonance neural network based on Adaptive Resonance Theory (ART), is an unsupervised self-organizing neural network. The network can automatically classify and train the input samples of unknown categories according to their inherent characteristics.
When the cutting conditions change, the ART network can automatically generate new categories according to the sample characteristics to adapt to the changes in the machining process. And the network training time is short, there is no local minimal problem, can achieve online training and recognition.
For the specific algorithm, please refer to Automatic Recognition of Tool wear and Damage State in Turning Process and Tool Wear Recognition Based on Fuzzy Neural Network.
4► Power signal/delay variance method
The test shows that when the tool is damaged, the amplitude of power signal in time domain is uncertain, so it is not reliable to use the time-domain threshold method to monitor the tool damage. In this paper, a new data processing method, time delay variance method, is proposed to extract the characteristic quantity of tool damage.
When the tool is cutting normally, the state of the tool can be considered constant for a short period of time, and the signal is stable. After tool damage, the signal mean usually changes and the variance tends to increase.
In general, the mean and variance of the signal are updated in real time according to the sampling points. The mean value of the delay variance method is the mean value of the power signal before a period of time. The advantage of this treatment is that although the variance of the signal variance after the tool is damaged tends to increase, it is relatively small, and the variance of the mean value of the power signal when the power signal is damaged to the normal cutting can reflect the change of the tool state more clearly. So that's the basic idea of the delay variance method. The specific processing algorithm is too professional, not expanded here, interested readers please refer to the original.
5►AE signal/time-frequency analysis
There are many sources of AE signals in the cutting process, including tool damage, microcrack expansion in the tool, chip breaking, impact between the chip and the workpiece, etc. Only in the time domain or frequency domain, it is impossible to distinguish the AE signals sent by tool damage from other AE sources. Therefore, it is more reasonable to analyze the variation of AE signal from both time domain and frequency domain.
The specific processing algorithm is too professional, so it will not be expanded here. Interested readers please refer to the literature "Research and Application of improved Grey Model in tool condition monitoring".