BP (Backpropagation) algorithm is an effective algorithm, its model has become one of the important models of the neural network。 The BP network is a unidirectional propagation multi-layer feed-forward network consisting of an input layer, an output layer, and an implicit layer (which can be one or more layers). The BP neural network for molybdenum wire-cut EDM is the input layer and the input layer is four. Units: workpiece thickness, pulse width, pulse interval, and pulse peak current; output layer has two units: cutting speed and surface roughness. Theorem proves that a three-layer BP network can complete any mapping from N dimension to M dimension. For this problem, use a hidden layer BP network. The choice of the number of hidden units is a very complicated problem. Generally speaking, the number of network hidden units may not be able to train a good network, or the network is not strong, and the samples that have not been learned before may not be identified, and fault tolerance is poor; however, there are too many hidden units.
The learning time is too long and the error is not necessarily the smallest, so there are an optimal number of hidden units. Through trial and error tests, the number of best hidden units was determined to be 9. The constructed BP network topology is shown in Figure 6-10. [图片] FIGURE 6-10 BP network topology during wire cutting The initial weight of the network has a great influence on the convergence of learning. The weights are initialized by the number of random numbers within the range of [− o. 5, + 0.5]. The choice of several network parameters, such as the learning factor, momentum factor, and network error termination value, is crucial to the BP network. . (Related knowledge:molybdenum wire china, molybdenum wire cutter, cnc molybdenum wire cut machine, golden molybdenum wire,) 3. Online learning and trainingIn order to avoid local minimum problems as much as possible and improve the speed of network learning, improved learning rules are adopted. For an input vector Xk, the network output is y k ‘. Expected output is Yk ‘.
The output of node I is O, and the input of k’ node j is neit k.The transfer function chooses asymmetry function [图片] its learning formula is as follows: [图片] The network error evaluation formula uses the squared error sum formula:[图片] Running the neural network, adjusting the parameters, and after repeated experiments, determined the optimal network parameters: learning factor 11 = 0.4, the dynamic address factor a = 0.9, and the network error termination value taken as 0.02.The learning process of the BP network consists of forward propagation and reverse propagation. In the forward propagation process, the input information is gradually processed from the input layer through the hidden layer and transmitted to the output layer. Each layer of neurons only affects the state of the next layer of neurons. If the desired output cannot be obtained at the output layer, it is transferred to the reverse propagation, the error signal is returned along the original connection channel, and the link weights between layers are modified layer by layer until the molybdenum wire-cut EDM network error reaches the accuracy requirement. The specific learning process is shown in Figure 6-11. 4. Wire cutting simulation example based on BP networkThe effect of wire(molybdenum cutting wire or brass wire) cutting process is a multivariable and multi-parameter problem. The process effect is not only related to the pulse power supply parameter, but also affected by many non-electrical factors, such as electrode molybdenum wire speed, working fluid chemical composition, flushing speed, workpiece material, etc. All will have effect on the process effect, and the interaction between the process parameters, making it difficult to establish a mathematical model between the process parameters and process effects, and there is a large error. Because artificial neural network has the functions of self-organization, self-learning, associative memory, etc., and has the characteristics of distribution, parallelism, and high robustness, it is very suitable for the modeling of thousands of complex systems.Using BP network to simulate the complex function relationship between the input and output of the actual process of line cutting, an artificial neural network model of molybdenum wire cutting machining was established. Based on this, a prediction system for the processing effect can be established.
The prediction of the effect of processing technology and the optimization of process parameters are realized. The overall structure is shown in Figure 6-12. [图片] The process effect prediction system consists of two modules: training module and prediction module. During the training phase, input learning samples to the BP network, including process parameters (workpiece thickness, pulse width, pulse interval, and pulse peak current) and process effectiveness(Cutting speed and surface roughness). The BP network normalizes the input and output of the learning sample and transforms each parameter into the (0, I) interval. Initialize the weights and阙 values​​with a random number between (-0.5, 0.5). The network then has teacher learning, the input signal is transmitted from the input layer to the output layer via the hidden layer, and the network output is compared with the teacher signal. If the network error does not meet the accuracy requirements, the user is transferred back to the layer and the layers are modified layer by layer for weights and thresholds, until the BP network convergence, link network parameters with weights, thresholds, and data files for prediction. In order to improve the prediction accuracy of the BP network, additional learning samples can be used to retrain the BP network. Read out network parameters, link weights and thresholds from the data file, and use the built B P network to predict the effect of the process. For the processing conditions set by thousands, the user inputs the processing parameters, and the prediction system can give the corresponding processing effects.
Related knowledge: molybdenum wire for edm,guangming molybdenum wire,molybdenum wire price.Process tests were performed on a DK 772 5 high speed molybdenum wire cutting machine to obtain a sample set of network learning. The input and output parameters are normalized so that the range of each parameter is between ( 0 , 1 ), and this is used as the input and output of the BP neural network for learning and training. The results show that the BP neural network with 4 – 9 – 2 structures can very well approximate the actual process of molybdenum wire cutting. The network error of BP model quickly approaches 1000 in the training process, and the final learning accuracy of the network reaches 0. 013.Table 6-4 shows the cutting speed and surface roughness prediction results ofmolybdenum wire -cutting iron-based alloys with BP networks. The average prediction error of the cutting speed and surface roughness of the BP network is 3.2% and 3.4%, respectively, and the accuracy predicted by the BP network is relatively high.From the wire cutting test data and simulation prediction results listed in Table 6-4, it can be seen that the prediction using the B P network. The effectiveness of the arts results is very close to the results of the process tests. It is feasible to establish a process effect prediction system using the BP network. The prediction accuracy is high and can reflect the actual process of line cutting and machining.
