lojmikak
Dołączył: 30 Sie 2010
Posty: 1597
Przeczytał: 0 tematów
Ostrzeżeń: 0/5
Skąd: olvovw
|
 Wysłany: Sob 14:45, 04 Gru 2010 Temat postu: ugg stivali Neural network-based quality control c |
 |
|
Neural network-based quality control chart pattern recognition research
Cover a wide surface. Samples covering a wider pattern, neural network pattern recognition of this stronger. (3) difficult to identify the model number of samples should be more, such as circular pattern difficult to identify than other models. (4) samples of each mode of study should be cross training, otherwise easily lead to algorithm does not converge. (5) In order to overcome the divergence of the algorithm, the first component of random interference pattern less obvious characteristics of the samples for training the neural network,[link widoczny dla zalogowanych], neural network training to be finished, then add some random samples of large component along with the original sample continue to train the neural network. Fourth, pretreatment observations observation data preprocessing of neural network training time and convergence of great influence. Systematic observations x (t) = + d (t) +22- Chen Equality: the quality of neural network based control chart pattern recognition of r (t). Where, u is the system quality characteristics under normal operation the average value, d (t) factors as a result of the quality of some unusual features of the deviation; r (t) solid ground accident caused by random interference pigment composition,[link widoczny dla zalogowanych], the general r (t) zero mean Gaussian white noise sequence variance. . Control chart pattern recognition is to identify the d (t) the variation, x (O can be measured for the known quantity,[link widoczny dla zalogowanych], the system works when d (t) = 0. I (t) ~ x (0 - ~ A a fall. AA> 0i (t) as the neural network input signal, the system works for a 3A ≤ i (t) ≤ 3A probability of 0.9973. A's value is too small, smaller neural network signal to noise ratio abnormal conditions i (t) is not sensitive to the changes, will reduce the network's ability to identify the control charts; A's value is too large, neural network signal is too sensitive to interference. The simulation tests, taking A = 1. V. Simulation 1 . the training sample selection, neural network input eight points for the l0, the output nodes is 6 (number of models for type 6),[link widoczny dla zalogowanych], hidden nodes is 10. r (t) zero mean Gaussian white noise sequence,[link widoczny dla zalogowanych], t A 1,2, ..., l0, a variance of a 1. (1) Normal mode: 20 sets of samples. dn) 10 (2) gradually increased mode: 24 sets of samples. d (t) a (BU to) ・ d ・ A ・., where tt. as the initial starting point, q is the rate of increase. (3) decreased gradually mode: 24 sets of samples d (t) a (a BU a to) ・. ・ A ・, where: d is rate of decline. (4) a sudden increase in mode: 38 sets of samples. d (t) day A ・ a ・】 (t-to), where: t. for the sudden increase in points, B for the rise. When t <to time, 1 (t-to) = 0, when t ≥ to time, 1 (t-to) I 1. (5) a sudden rise mode: 36 sets of samples. d (t) A ・ ・ 1 a day (t-to), where tt. as a sudden drop in points, B for the decline. (6) cyclic changes in mode: 40 sets of samples. d (t) a k ・ ・ sin [2a (t - t.) / T3 , where k is the amplitude gain, T for the cycle, to the initial starting point. r (t) by the random number generator, each sample for different sequence of random numbers. i (t) input, respectively output neural network one layer of 10 input nodes, output nodes for the model type, such as the normal mode is 100,000, and gradually increase mode is 010,000.2. l training samples of the training step by step training samples of each model cross-arranged method of batch training, the first Training mode features samples of the more obvious (for example, a gradual increase in mode. the larger pattern samples), including normal mode l0 samples, samples gradually increased model l2, l2 samples gradually decreased model, a sudden increase in model l8 samples l8 sudden drop model samples, cyclic changes in patterns of 20 samples, a total of 90 samples. inertial coefficient n of 0.3, learning to take 13 steps of 0.4, the number of training I train for the 1000 training samples and then the whole (180) with training, training times for 2000 times. 3. The output value is selected in the training that, if the output value takes 0 and l, the training does not converge. If the output value of 0.01 and obtain 0.9g, training convergence However, a longer training time. If the output value obtained 0.05 and 0.95, the training convergence and training time is shorter. The original solid is the role of the output layer function fix) of a 1 / (1 + e a) of the output value the range (0,】), in the vicinity of 0 and 1, the function of the change rate is small change for a 23 to the input
相关的主题文章：
[link widoczny dla zalogowanych]
[link widoczny dla zalogowanych]
[link widoczny dla zalogowanych]
Post został pochwalony 0 razy
|
|
