Abstract: The key to optimal adaptive control of machining is the establishment of an adaptive machining model and the formulation of a real-time optimization strategy. This paper proposes to use artificial neural network method to establish the process model, and use genetic algorithm to achieve online optimization. Based on the above algorithm, a self-adaptive optimization system for the processing parameters of the plane milling is constructed, which can make the processing system always obtain the maximum material removal rate without violating the processing constraints.
Keywords : plane milling; adaptive optimization; artificial neural network; genetic algorithm
Classification number: TH391.9 Document code: A
Chapter number: 1001-2265 (2000) 02-0005-04
Adaptive optimization of machining parameters for face milling
Based on artificial neural network and genetic algorithms
Ni Qimin Lin Jianping Li Congxin Ruan xueyu
Abstract: The keys to adaptive control with optimization of machining operations are modeling of cutting process and of of real-time optimal strategy.In this paper, we propose to moldel the machining process using artificial neural network and realize on-line optimization using genetic algorithms .A system for adaptive optimization of cutting parameters such as feed rate and spindle speed in a face milling operation for maximizing the material removal rate without violating machining constraints is set up.
Key words: face milling;adaptive optimization;artificial neural network;genetic algorithms
The choice of metal cutting processing parameters has an important influence on processing productivity and economic efficiency. In conventional CNC machining systems, the machining parameters are determined prior to machining based on the programmer's experience, a processing manual or off-line optimization. However, in the actual processing, the processing conditions are constantly changing. The processing parameters that are specified in advance during programming cannot always maintain the optimum processing target. One way to solve this problem is to provide the CNC machining system with an optimal control system that can optimize the machining parameters in real time based on changes in machining conditions. Such a system is called an adaptive control system and can be divided into Constraint Adaptive Control (ACC) and Optimal Adaptive Control (ACO) .
Constraint adaptive control actually maximizes productivity by maintaining the cutting force or machining power at a limit value. Its main disadvantage is the lack of feedback on the quality of the workpiece. Only processing parameters obtained from a single constraint may be violated by other constraints. And become infeasible .
Although people have conducted a lot of research on optimal adaptive control [3-7], few systems have reached practical use. The main difficulty lies in the establishment of the process model and the formulation of real-time optimization strategies. For these two problems, this paper proposes an adaptive optimization method for face milling based on artificial neural network modeling and genetic algorithm optimization.
2 Artificial neural network modeling during processing
Proper and reasonable processing model is the premise of processing optimization. The machining model should be able to accurately reflect the relationship between machining parameters (cutting speed, feed rate, depth of cut, etc.) and machining output (cutting force, cutting power, surface quality, etc.) over a wide range of operating conditions. In addition, the process model must also have the ability to adapt to changes in the processing environment, especially tool wear. Due to the complexity and non-linear characteristics of the machining process, it is very difficult to establish an analytical process model accurately. Recent studies have shown [8,9] that artificial neural networks with adaptive capabilities and nonlinear processing capabilities, especially multilayer perceptron networks, are suitable for modeling of manufacturing processes and are generally superior to other modeling techniques.
Based on the above analysis, this paper uses artificial neural network to map the relationship between process input variables and process output variables. In order to allow the process model to adapt to changes in tool wear during machining, process input variables include tool wear length variables in addition to machining parameters. In use, it is necessary to perform on-line measurement of tool wear during machining.
2.1 Neural Network Measurement of Tool Wear <br><br><br><br><br> Tool wear conditions have an important influence on cutting forces and workpiece surface quality. Therefore, in order to accurately predict cutting force and machining quality, the tool wear length must be measured first.
There are direct methods and indirect methods for measuring tool wear conditions. Direct methods can be divided into contact measurement methods and non-contact measurement methods. Contact measurement methods can generally only be used for postmortem measurements and it is difficult to achieve online estimates. Non-contact direct measurement methods mostly use laser and image processing techniques, and due to limitations in processing speed or accuracy and relatively high cost, it is also difficult to use for real-time measurement. At present, most of the actual methods used are indirect measurement methods such as cutting force detection method, vibration detection method, AE method, and so on. The research shows [10,11] that because of the complexity of the machining process, the wear status of the tool is only estimated based on the detection information of a single sensor, and high false positives and false positives occur. Therefore, multi-sensor information fusion methods are generally used for detection. , but the amount of calculation is large.
In this paper, based on the large-scale parallel computing capability and information fusion ability of neural network, an artificial neural network measurement method for tool wear length is proposed. The network type is selected as a multi-layer perceptron, and the fused information (ie, network input) includes four characteristic parameters of cutting speed, feed rate, and cutting force spectrum. There is only one network output node, which is the estimated value of tool wear length. The number of hidden layers of the multilayer perceptron and the number of nodes in each hidden layer have a great influence on the performance of the network. Try a variety of network structures, compare their learning speed and accuracy, and finally determine the network structure is 6-4-1.
2.2 Process predictors <br> Process predictor neural networks predict process output variables based on estimated tool wear lengths and machining parameters. The network type is also selected as a multilayer sensor. There are 5 input nodes, namely the tool wear length, cutting speed, feed rate, depth of cut and cut width; there are 3 output nodes, corresponding to cutting force F, cutting power P and surface roughness Ra. The hidden layer structure is determined by an attempt method and uses two hidden layers. The number of hidden layer nodes is 15 and 7 respectively.
2.3 Training Algorithms The above two neural networks must be trained offline before use. The most commonly used training algorithm for multi-layer perceptron neural network is BP algorithm, but this algorithm has the disadvantage of being easily trapped in local minimum. The simulated annealing algorithm  can jump out of the local optimal trap and find the global (or approximate) optimal solution. Combining the two algorithms can give full play to the advantages of both. The training steps are as follows:
1 randomly generated initial state of the network parameters R, so that T = T0
2Generate R's next candidate state R' according to BP algorithm
3 Let E = [Ep(R')-Ep(R)]/Ep(R), Ep be the sum of squared errors of the training samples. 4 If Eâ‰¤0, let R=R'; otherwise, use the probability EXP(- E/KBT) Accept R=R'
5 repeated 2 ~ 4M times 6 lower temperature T
7 Repeat 2 to 6 until the temperature is very low or precision has been reached. The trained neural network can be used for online optimization.
3 Genetic Algorithms for Process Optimization
3.1 Processing Optimization Model <br> From the viewpoint of improving the processing efficiency, the optimization goal is to maximize the material removal rate. For flat milling, the material removal rate MRR can be expressed as the product of feed rate f, depth of cut ap, and width aw. Considering that the on-line adjustment of the depth of cut and the width of the cut in the actual machining is difficult, the on-line control parameters are selected as the easily controllable spindle speed n and the feed rate f (the depth of cut, the cut width can be preliminary based on the workpiece geometry and the machining process during off-line optimization. determine). The processing constraints mainly include the performance limitations of the machine tool and the tool itself and the quality requirements of the workpiece being processed, including the limits of the speed and feed rate, the maximum cutting force constraints, the maximum cutting power constraints, and the workpiece surface roughness constraints. In summary, the optimization model is as follows:
Optimization goal: maxMRR=fapaw
Design Variable: X=[n f]T
Input constraints nmin<n<nmax, fmin<f<fmax
Output constraints F<Fmax, P<Pmax, Ra<Ramax
This is a constrained optimization problem that can be transformed into an unconstrained optimization problem using the penalty function method:
In the formula, R1 and R2 are the penalty factors that are imposed when the input and output constraints are violated. The <> function is defined as:
3.2 Real-time optimization algorithm <br> The efficient and reliable optimization algorithm is a necessary condition for realizing on-line optimization. Inspired genetic algorithm derived from biological evolution  uses heuristic search technology to find the optimal solution. Compared with traditional optimization methods, it has good robustness, high search efficiency, less restriction on objective functions, and easy adoption of parallel machines. Because of the advantages of parallel high-speed operations, it is suitable for on-line solution of the above-mentioned optimization model. The main process of the algorithm is as follows:
A code uses a certain length of binary code to represent various values â€‹â€‹of an argument, and concatenates the binary codes of the respective variables into a string to obtain a binary code string called a chromosome. The length of the chromosome depends on the binary number used in the numerical control system for the rpm and feedrate reset. For example, if the rotation reset bit number is 3 and the feedrate reset bit number is 4, the chromosome length is 7.
2 The initial population is generated by the computer in a random way to generate a series of chromosomes, each chromosome represents an individual, a certain number of individuals constitute the original group.
3 Computation fitness According to the coding rule, the value of the independent variable corresponding to the chromosome of each individual (ni, fi) is substituted into (*) formula, and the value of the objective function is obtained as its fitness.
4 Selection of species Selective stress is selected as suitability, and a pair of individuals are selected at random to serve as parents for breeding offspring.
5 Hybridization Randomly selected parents were crossed.
6 mutations Select individuals in a group with a certain probability. For each individual that has been selected, randomly select a certain digit and flip the digit of the digit.
7 Repeat 3-6 until the group's fitness level becomes stable and no longer rises.
4 Plane Milling Adaptive Optimization System
The key to the optimal adaptive control of machining is the establishment of an adaptive machining model and the formulation of a real-time optimization strategy. In this paper, two artificial neural networks are used to establish the process model of the plane milling process. The online optimization is realized by the genetic algorithm. Based on the above algorithm, an adaptive optimization system for the plane milling shown in the figure below is established.
Plane milling CNC machining adaptive optimization system diagram
The system is mainly composed of three parts: measurement neural network to estimate the tool wear status in real time; and process model neural network to establish the relationship between processing parameters and process output variables under the known tool wear length, in order to optimize the fit of the program. The calculation provides part of the data; the online optimization algorithm guarantees to always provide the CNC system with optimal processing parameters (feedrate, spindle speed) and changes in the machining environment (tool wear, depth of cut due to geometric changes in the workpiece, cut In the case of wide variations, etc., the processing system can still work under optimal conditions, ie, to obtain the largest possible material removal rate without violating the processing constraints.
(1) Artificial neural networks with massive parallel processing capabilities and information fusion capabilities can be used for on-line measurement of tool wear.
(2) The multilayer perceptron network is suitable for modeling of manufacturing processes.
(3) The genetic algorithm with good robustness and high solving efficiency can be used for real-time optimization of the optimal adaptive system.
(4) The plane milling optimization system constructed in this paper can adapt to the changes of the machining environment caused by tool wear and workpiece geometry, but it cannot compensate for the uncertainties such as material non-uniformity. In addition, the process model of this article is obtained through offline training. When the workpiece material or tool changes, it must be retrained. Further research can explore the theory and methods of online adaptive modeling.