《分析电阻点焊过程的决策和控制依据.docx》由会员分享，可在线阅读，更多相关《分析电阻点焊过程的决策和控制依据.docx（14页珍藏版）》请在三一办公上搜索。
1、分析电阻点焊过程的决策和控制依据 分析电阻点焊过程的决策和控制依据 人工神经网络是用物理模型模拟生物神经网络的基本功能和结构，可以在未知被控对象和业务模型情况下达到学习的目的。建立神经网络是利用神经网络高度并行的信息处理能力，较强的非线性映射能力及自适应学习能力，同时为消除复杂系统的制约因素提供了手段。人工神经网络在足够多的样本数据的基础上，可以很好地比较任意复杂的非线性函数。另外，神经网络的并行结构可用硬件实现的方法进行开发。目前应用最成熟最广泛的一种神经网络是前馈多层神经网络，通常称为BP神经网络。 Artificial neural network is to use physical
2、model simulation of biological neural network basic function and structure, can be in the case of unknown and business model of the object to achieve the objective of the study. Establish neural network is a highly parallel neural network information processing ability, strong ability of nonlinear m
3、apping and adaptive learning ability, at the same time to eliminate the restrictive factors of complex systems offers a means. Artificial neural network on the basis of enough sample data, is a good way to compare any complex nonlinear function. In addition, the parallel structure of the neural netw
4、ork hardware implementation methods are available for development. One of the most mature and most widely used neural network is a feedforward multilayer neural network (BP), commonly referred to as the BP neural network. 神经网络方法的基本思想是：神经网络模型的网络输入与神经网络输出的数学关系用以表示系统的结构参数与系统动态参数之间的复杂的物理关系，即训练。我们发现利用经过训
5、练的模型进行权值和阈值的再修改和优化时，其计算速度要大大快于基于其他优化计算的速度。 Neural network method the basic idea is: input of neural network model with neural network is applied to the structure of the system output mathematical relationships between the parameters and dynamic parameters of the complex physical relationship, namely
6、 the training. We found that using the trained model weights and the threshold value will be modified and optimized (call it), the calculation speed is much faster than the other optimization based on the speed of calculation. BP神经网络一般由大量的非线性处理单元神经元连接组成的。具有大规模并行处理信息能力和极强的的容错性。每个神经元有一个单一的输出，但可以把这个输出量
7、与下一层的多个神经元相连，每个连接通路对应一个连接权系数。根据功能可以把神经网络分为输入层，隐含层，输出层三个部分。设每层输入为ui(q)输出为vi(q)。同时，给定了P组输入和输出样本 ，dp。 BP neural network is typically composed of a nonlinear processing unit, consisting of neural connections. Has the massively parallel processing information ability and strong fault tolerance. Each neur
8、on has a single output, but can reduce the output connected to more than the next layer of neurons, each connection path corresponds to a connection weights. Neural networks according to their functions can be divided into input layer, hidden layer (or layers), the output layer three parts. Set inpu
9、t on each floor for UI output for vi (q) (q). At the same time, given the P set the input and output sample, dp (P = 200). (6) 该网络实质上是对任意非线性映射关系的一种逼近，由于采用的是全局逼近的方法，因而BP网络具有较好的泛化的能力。 The network in essence is a kind of approximation to arbitrary nonlinear mapping relationship, the adoption of is glob
10、al approximation method, BP network has good generalization ability. 我们主要是利用神经网络的非线性自适应能力，将它用于消音锯片的电阻点焊过程。训练过程是：通过点焊实验获得目标函数与各影响因素间的离散关系，用神经网络的隐式来表达输入输出的函数关系，即将实验数据作为样 本输入网络进行训练，建立输入输出之间的非线性映射关系，并将知识信息储存在连接权上，从而利用网络的记忆功能形成一个函数。不断地迭代可以达到sse最小。 We mainly use nonlinear adaptive ability of neural networ
11、k, and use it for sound attenuation resistance spot welding process of saw blade. Training process is: through the spot welding experiments to obtain the objective function and discrete relationship between various influencing factors, using neural network implicitly to express the function between
12、input and output, the experiment data as sample input network training, establishing the nonlinear mapping relationship between input and output, and information stored in the knowledge connection power, so as to make use of the network memory formed a function. Constant iteration to sse (minimum er
13、ror sum of squares). 我们这次做的消音金刚石锯片电焊机，通过实验发现可以通过采用双隐层BP神经网络就可以很好的反应输入 输出参数的非线性关系。输入神经元为3，分别对应3个电阻点焊工艺参数。输出神经元为1，对应焊接质量指标参数。设第1隐含层神经元取为s1，第2隐含层神经元取为s2。输入层和隐含层以及隐层之间的激活函数都选取Log-Sigmoid型函数，输出层的激活函数选取Pureline型函数。 This time we do the sound attenuation of diamond saw blade welding machine, through the exp
14、eriment discovered by double hidden layer BP neural network can be very good response to input and output parameters of the nonlinear relationship. Input neurons is 3, corresponding to three resistance spot welding process parameters. Output neurons is 1, corresponding welding parameters of quality
15、indicators. Set off for hidden layer neurons 1 s1, 2 off for s2 hidden layer neurons. Between input layer and hidden layer and hidden layer activation function is to select the Log type - Sigmoid function, output layer activation function select Pureline type function. 2点焊样本的选取 2 spot welding sample
16、 selection 影响点焊质量的参数有很多，我们选取点焊时的控制参数，即点焊时间，电极力和焊接电流，在固定式点焊机上进行实验。选用钢种为50Mn2V，600m的消音型薄型圆锯片基体为进行实验。酱油对需要优化的参数为点焊时间，电极力和焊接电流3个参数进行的训练。最后的结果为焊接质量，通常以锯片的抗拉剪载荷为指标。 Parameters affecting the quality of spot welding has a lot of, we select spot welding of the control parameters, the welding time, vigorously
17、 and welding current, electricity experiment in stationary spot welding machine. Choose steel 50 mn2v, 600 m deadened the noise of the thin circular sawblade matrix for the test. To optimize the parameters for the spot welding time, strongly and welding current in three parameters of training. For t
18、he final result of welding quality, usually for saw blade tensile shear load indicator. 建立BP神经网络时，选择样本非常重要。样本的选取关系到所建立的网络模型能否正确反映所选点焊参数和输出之间的关系。利用插值法，将输入变量在较理想的区间均匀分布取值，如果有m个输入量，每个输入量均匀取n个值, 则根据排列组合有nm个样本。对应于本例，有3个输入量，每个变量有5个水平数，这样训练样本的数目就为53=125个。 Establish the BP neural network, choose samples is
19、very important. Sample selection is related to the established network model can correctly reflect the selected welding parameters and the relationship between the output. Interpolation method is used to transform input variables in a relatively ideal uniform distribution of interval values, if ther
20、e is a m input, each uniform input of n value (that is, each input has m level number), then according to the permutation and combination with nm samples. Corresponds to this case, there are three input variables, each variable has 5 levels, so that the number of training samples for 53 = 125. 我们的实验
21、，是以工人的经验为参考依据，发现点焊时间范围为28s，电极力范围为5003000N，点焊电流范围为520kA时，焊接质量比较好。我们先取点焊电流，电极力为定量，在合理的范围内不断改变点焊时间，得到抗拉剪载荷。如此，可以得到不同点焊电流和电极力的抗拉剪载荷。根据点焊数据的发布情况，我们共选用200组数据。部分测试数据如表1： Our experiment, based on workers experience for reference, find spot welding time range is 2 8 s, electric range is 500 3000 n vigorously
22、, spot welding current range is 5 20 ka, the welding quality is better. First we collect the welding current, electricity to quantitative, within a reasonable range changing spot welding time, get the tensile shear load. So, you can get different spot welding current and electricity to the tensile s
23、hear load. According to the welding data released, we choose 200 set of data. Some test data such as table 1: 神经网络建模的关键是训练，而训练时随着输入参数个数的增加样本的排列组合数也急剧增加，这就给神经网络建模带来了很大的工作量，甚至于无法达到训练目的。 Neural network modeling is the key to training, the training sample with the increase of number of input parameters
24、on the number of permutation and combination also increased dramatically, the neural network modeling to bring very great effort, even unable to achieve training purpose. 3神经网络 Three neural networks 我们用200组训练样本对进行神经网络训练，以err_goal=0.01为目标。调用Matlab神经网络工具箱中的函数编程计算，实现对网络的训练，训练完成后便得到一个网络模型。 We use 200 pa
25、irs of training samples for neural network training, aiming at err_goal = 0.01. Matlab neural network toolbox function called programming calculation, and realize the network training, the training is completed after get a network model. 程序 program x1=2.1 2.5 3 3.5 4； %点焊时间输入，取200组 X1 = 2.1 2.1 2.5
26、3 4.) ; % spot welding time input, take 200 groups x2=1.3 1.5 1.9 2.1 2.3；%电极力输入，取200组 X2 = 1.3 1.5 1.9 2.1 1.5. ; % to input, take 200 groups x3=9 10 11 12 13;%点焊电流输入，取200组 The x3 = / 9 10 11 12 13. ; % spot welding current input, take 200 groups y=2756 3167 3895 3264 2877； %输出量，取200组 Y = 2756, 316
27、7, 3895, 3264. 3167. ; % output, take 200 groups net=newff(1 10;0.5 3;5 20,10 10 1,tansigtansigpurelin); Net = newff (20 1 to 10, 0.5, 3, 5, 10 10 1 and tansig tansig purelin); %初始化网络 % initialization of the network net.trainParam.goal = 0.01;%设定目标值 .net. TrainParam. Goal = 0.01; % set target value
28、net=train(net,x1;x2;x3,y);%训练网络 Net = train (.net, (x1, x2, x3), y); % training network figure； %画出图像 Figure; Draw the image % 选取不同的s1,s2，经过不断的神经网络训练，发现当s1=8,s2=6时，神经网络可以达到要求。工具箱示意图如下图1。 Select different s1, s2, after continuous neural network training, found that when s1 = 8, s2 = 6, the neural net
29、work can meet the requirements. Toolbox diagram shown in figure 1 below. 图 1工具箱示意图 Figure 1 kit 工具箱示意图非常清晰地表示了本实验的神经网络的输入，输出以及训练的过程。 Toolbox diagram very clearly according to the input of neural network in this study, the output and the process of training. 神经网络的训练结果，如图2所示: Neural network of trainin
30、g results, as shown in figure 2: 图2神经网络的学习过程 Figure 2 of the neural network learning process 图中可以看出双层网络训练的sse在训练100次时，已经接近0.0001，效果较理想。 Figure you can see in the two-layer network training of sse in training 100 times, is close to 0.0001, the effect is ideal. 为了验证经过训练的网络模型的泛化能力，在输入变量所允许的区域内又另选多个样本进行
31、了计算。发现：利用BP神经网络模型计算的测试输出与期望输出值相符，误差小于2。 In order to verify the trained network model generalization ability, in the input variables are allowed by the area and choose multiple samples were calculated. Found that the BP neural network model is used to calculate the test output and the desired output,
32、 error is less than 2%. 在已经训练好的网络中找出最大值： In has been trained in the network to find maximum: for i=2:10 %点焊时间选择 For I = 2:10 % spot welding time to choose for j=0.5:0.1:3%电极力选择 For j = 0.5, 0.5, 3% power to choose fork=5:0.1:20%点焊电流选择 Fork = 5-0. 1:20 % spot welding current selection a=sim(net,i,j,k
33、)；%仿真 A = sim (.net, I, j, k); % the simulation ifan %比较仿真结果与最大值，取最大值n=a； Ifan % to compare simulation results with the maximum value, maximum value of n = a; i(1)=i；%最大值的时间 I (1) = I; A maximum % of time j(1)=j；%最大值的电极力 J (1) = j; % to maximum value of electricity k(1)=k； %最大值的电流 K = k (1); % of ma
34、ximum current end The end end The end end The end end The end 将i,j,k以及n输出，n为最大值。得到点焊时间为3.4s，电极力为12.7kN，点焊电流为11.8kA，此时的抗剪拉剪载荷为4381N，酿酒设备为训练结果的最大值。将点焊时间为3.4s，电极力为12.7kN，点焊电流为11.8kA在点焊机上进行实验，得到结果为4297N。并且通过与实际的结果相比较，发现误差也在2%以内。 Will I (1), j (1), k (1) and n output, n for maximum. Spot welding time is
35、3.4 s, electricity to 12.7 kN, spot welding current is 11.8 kA, shear tensile shear load is 4381 n at this time, for a maximum of training results. Will spot welding time is 3.4 s, electricity to 12.7 kN, 11.8 kA for spot welding current in spot welding machine for experiments, the results of 4297 n
36、. And by comparing with the actual results, found that the error is within 2%. 4结论 4 conclusion 1）本文采用了插值法作为选取BP神经网络训练样本的方法。并且在数据变化剧烈的地方多选取了75组数据，这样可以得到较高精度的网络模型，使点焊模型的可行性。 1) this paper USES interpolation method as a way to select training samples of BP neural network. And where data changes drasti
37、cally from 75 groups of data, so that we can get a high precision of network model, and make the spot welding the feasibility of the model. 2）基于此方法建立了三个点焊参数的BP神经网络模型，而且所建的BP模型具有较高的精度，可以很好的描述了这三个点焊参数与点焊质量的映射关系。 2) three spot welding parameters is set up based on the method of BP neural network model,
38、 and built the BP model has higher accuracy and can well describe the three spot welding parameters and welding quality of the mapping relationship. 3）由于神经网络模型将系统结构参数与传统动态特性参数之间的物理关系，反映为神经网络模型的网络输入与网络输出的数学关系，因此，在神经网络模型上进行结构修正与优化比在其他模型上更直接，简单与高效。 3) due to the structural parameters of neural network
39、model of the system and traditional physical relations between the dynamic characteristic parameters, reflect the neural network model for network input and output mathematical relationship, therefore, the neural network model structure modification and optimization on direct more than on other model, simple and efficient.
