Real-time detection of tool status in CNC machining

introduction

In CNC machining, the detection of tool status is of great importance, as tool damage not only affects the quality and efficiency of the process, but can also lead to serious machine and personal accidents. The damage of the tool is caused by wear and damage. The wear is the phenomenon of the surface material consumed by the contact and friction between the tool and the workpiece during the machining process; and the damage is caused by chipping, breaking, plastic deformation, etc. Loss of cutting ability, which includes brittle fracture and plastic breakage. Brittle fracture is the chipping, chipping and spalling of the tool under mechanical and impact effects without obvious wear. The plastic breakage is a phenomenon in which the tool is plastically flowed on the surface layer in contact with the workpiece due to high temperature and high pressure during cutting, and the cutting ability is lost. At present, the detection of tools mainly adopts three strategies: manual detection, offline detection and online detection. Manual inspection is the inspection of the state of the tool by the worker during the processing; offline inspection is to specifically test the tool before processing, and predict its life to see if it is qualified for the current processing; online detection is also called real-time detection, that is The tool is detected in real time during the machining process, and the corresponding processing is performed according to the detected result. At present, there are many algorithms for tool detection. Some use the theoretical calculation of the change of the stress on the tool to judge the damage of the tool. Some use the method of time series analysis to detect the tool, and some use the neural network technology. The tool is tested, and the other uses the wavelet transform theory and neural network technology to detect the tool, but they are mainly discussed theoretically. Considering that the plastic damage of the tool in NC machining is relatively rare, and the safety relationship between the wear and the CNC is not very large and it can be processed by offline detection. This paper takes the ball-end tool commonly used in NC machining as the research object, and it is in the brittle fracture. The real-time detection of brittle fractures is studied, and the occurrence of such fractures will have a serious impact on the quality of the processing and the machine itself. We believe that there are tiny cracks in the tool itself, and use the neural network to establish the load model of the ball-end tool. It is judged by online detection whether the micro-crack will expand under the load condition at this time. If it is possible to expand, we think that the load is Dangerous and reduce the load on the tool by reducing the feed of the tool to ensure the safety of the tool.

1 Real-time detection of tools

(1) The establishment of the ball head tool load model As mentioned above, the load on the tool is related to many factors during NC machining, but considering the characteristics of the ball head tool and the need of real-time machining, this paper only considers the influence of the ball. Several factors, namely the spindle speed, feed rate, depth of cutting, and cutting performance of the machined material, the model of the ball load is F=f(s,v,h,m) (1)
Where: F - load vector; h - depth of cutting;
s - the speed of the spindle; m - the cutting performance of the material.
v - the amount of feed;
Obviously, equation (1) only gives a general relationship between the load and various influencing factors. In order to find the specific expression of the relationship between the load and each influencing factor, the specific size of each factor's influence on the load must be obtained. For this reason, the mathematical method such as differential geometry is used to carry out complex derivation, or the experimental method is used to obtain the influence coefficient of each factor, but the model thus established is difficult to adapt to the changing environment, and the real-time detection effect in numerical control processing is not very good. ideal. This paper uses neural network technology to process the model and use it in real-time detection of tools.

(2) Principle of real-time tool detection

The principle of real-time detection of the tool is to measure the cutting depth and feed amount of the tool in real time and input the load of the neural network controller with the spindle speed and the material type of the machined part. The load input detector is calculated and compared. If the load exceeds the crack propagation load under the fatigue condition of the tool, the feed rate of the tool is reduced, and the decrease of the feed rate is fed back to the input information of the CNC controller, so that the CNC controller performs corresponding control. To make the size of the load change to a safe level. The real-time detection principle of the tool is shown in the figure.

(3) Structure of the neural network The structure of the neural network has a decisive influence on the characteristics of the entire neural network system. The load adaptive control neural network system adopts a three-layer BP structure. According to the above analysis, it is obvious that the input layer has four nodes, and the output layer has three nodes, that is, the size of the load in three directions of xyz. The problem now is to determine the number of nodes in the middle hidden layer. The choice of the number of nodes in the middle hidden layer has a very important influence on the learning and computing characteristics of the network. It is the key to the success or failure of the network structure. If there are too few nodes in the middle hidden layer, The network is difficult to handle complex problems. However, if there are too many nodes in the middle hidden layer, the network learning time will increase sharply, and the network learning may be excessive, which may reduce the network anti-interference ability. At present, there is no perfect theory to guide the selection of the number of nodes in the middle hidden layer, but only the tentative selection and further optimization according to the actual situation. Considering the characteristics of the load adaptive control system, we consider the load to be a continuous function of the feed rate. According to the Kolmogorov theorem (continuous function representation theorem), in order to theoretically accurately simulate the continuous function, if the input of the three-layer neural network The layer is M nodes, and the output layer is N nodes, then the middle layer should be 2M+1 nodes. To this end, we choose the number of nodes in the middle hidden layer should be 2M + 1 = 2 × 4 + 1 = 9 nodes [9]. Therefore, the neural network structure has four nodes in the input layer, nine nodes in the middle layer, and three nodes in the output layer.

(4) Off-line learning of neural networks An important feature of neural networks is that they have the function of learning, that is, the ability to automatically adjust the weight of each node to meet the established target according to the input and output relationship of a certain amount of samples. In the learning process, the selection of the number of samples is very important. If there are too few samples, the performance of the learned network is not good. If the number of samples increases, it will increase the workload of collecting sample data and the learning time of the network. At the same time, since the neural network has better interpolation performance and the extrapolation performance is poor, the sample data must include all modes and consider the influence of possible random noise. For the neural network load adaptive control system, which has four input nodes, according to the above analysis, we take four values ​​given by each node, and use their different combinations as sample input data, so that 256 samples can be obtained. The specific method is to divide the input quantities into four equal parts within the possible variation range, and experimentally measure the load value under each input condition. After getting 256 samples, we use offline learning to get the weight between each connected node, so that the learned neural network establishes the corresponding tool load model to provide conditions for real-time detection of the tool.

(5) Detector design The function of the detector is to detect whether the load applied by the tool exceeds the crack propagation load under stress fatigue conditions. If it is possible to expand, we consider the load to be dangerous and reduce the load on the tool by reducing the feed of the tool to ensure the safety of the tool. To do this, we first establish the mechanical model of the tool. We simplify the tool in the machining to the cantilever beam with the end point, and the force at the end point is the load obtained by the neural network. Thus, according to the relevant theory of material mechanics It can be concluded that the part of the tool with the greatest stress is the joint between the tool and the machine tool, and the stress here can be obtained. Then, according to the relevant theory of fracture mechanics, there is the formula da/dn=f(σ, a, c), where a is the length of the crack, N is the frequency of the stress, σ is the normal stress, and c is the constant related to the material [ 10]. In the above formula, σ, N, and c are known quantities or can be found by data, and the function f and the length a of the crack are to be determined. For f, the measures we take are: assume that the microcracks at the joint with the machine tool are composite cracks of type I, II, and III cracks, and determine the proportion of the three types of cracks according to the magnitude of the normal stress and the shear stress. In this way, the formula can be established according to various specific crack types. As for a, we are based on the average crack length of the tool during its service life. The average length can be detected by non-destructive testing of tools in different life periods.

2 Conclusions This paper proposes a method for real-time detection of ball-end tools in NC machining by neural network. This method can calculate the load of the ball-end tool in the real-time process, and determine whether the load is detected by real-time detection. Exceeding the load level of the crack propagation of the tool under stress fatigue conditions and corresponding treatment. The method simplifies the factors affecting the load reasonably, making the algorithm of the control model very efficient, so it is especially suitable for real-time detection. Although this paper takes the ball-end tool in numerical control as the research object, in fact, the principle of this method can also be used in other machining and other tools, such as electric machining.