logoxgvb
Joined: 30 Aug 2010
Posts: 1580
Read: 0 topics
Warns: 0/5
Location: hdfttv
|
 Posted: Sat 9:56, 12 Mar 2011 Post subject: BP network in sandy soil liquefaction evaluation |
|
|
BP network in sandy soil liquefaction evaluation of vibration
And thresholds to minimize the error signal (Figure 2). The back propagation learning algorithm network diagram of Figure 1 2P Chen Kui, etc.: BP network in the sandy soil liquefaction evaluation of vibration assumes a total of 47? / T layer (not including input layer), layer number of nodes is n , on behalf of, the output layer node, then: a lJ a: ¨ A z layer nodes, the weight vector,. =, f (·) is the response function, the general sigmoid function: f () = 1 / (1 + exp (a)). for a given training sample (x, y). adjust the network weights, so that the following criterion function (LMS) Minimum: --- E (w) = a iy-P = {Σ (I). a 1 = Yi for the network's output. According to the gradient descent method to strike E (w) gradient to correct weight, weight vector w (For the output layer) l <'+ l8: = Σw sample, here algorithms have been repeatedly revised weights, the network's actual output can be close to the desired output. 2BP BP algorithm analysis and optimization algorithm is a simple static optimization of the steepest descent algorithm, the main drawback in the following areas: (1) using the minimum residual sum of squares as the objective function value will be too great prominence to the impact. (2) Since the gradient based search method, you need a good definition of differentiable surfaces. difficult to resolve with the local minimum The deceptive complexity of the problem, especially when the general strike of non-convex objective function on the global optimal solution. (3) learning convergence speed is very slow, usually takes longer to converge when the castration. There are many optimization algorithm can be used BP network improvements such as: simulated annealing, simplex method, Tabusearch algorithm and genetic algorithm. Based on the above analysis, this paper gave some of BP algorithm optimization, gradient descent method based on the momentum method and adaptive learning rate algorithm to modify the network weights. adjusted (1) as follows: taste DH】 (2) (3) (4) wl factor, 0 ≤ <1. momentum French (2)) can reduce the network for the local details of the sensitivity of the error surface, effectively inhibit the network caught in local minimum, momentum is essentially the equivalent of drag force, which reduces the learning process oscillating trend, thereby improving the convergence. standard BP algorithm is slow convergence is an important cause of the inappropriate choice of learning rate n, is too small the convergence is too slow, too, there may be amendments to overshoot, leading to oscillations and even divergence. Equation (3) (4) that when the two iterations of the gradient in the same direction, indicating that the rate of decline is too slow, then adjust the learning rate to a double. When two consecutive iterative gradient in the opposite direction, it indicates that the rate of decline too fast, halving the learning rate n. 3 sand liquefaction of the BP network evaluation taking into account the characteristics of soil liquefaction, the author uses three layers BP network programming based on the above algorithm to evaluate the liquefaction of sand. nodes in hidden layer is currently selected There is no theoretical guidance, if you choose too small, the nonlinear mapping function of the network and fault tolerance of difference; choice too, forks make learning time increases. learning error is not necessarily the best. usually by examining the model to evaluate the ability to determine . After several operations and to compare experimental data with model calculations obtained through the identification error between the value of the square and after the selection of 6, respectively tangent sigmoid transfer function function and pure linear function. selected as the sum of squared errors of 0.005, learning initial rate of a 0.01, momentum constant chosen as 0.9. Many factors affect the soil liquefaction. can be divided into Soil conditions, burial conditions, dynamic load conditions of the three major aspects. taking into account the simplicity of the test target , easy and representative, select sand water table (/ m), the critical depth (/ m), average standard penetration blow count (N63.5 / hit), the largest ground acceleration of g / (cm · s), vibration duration t / s as the evaluation index. Data from [6], 34 training samples, 20 of which liquefied, 14 not liquefied, the test sample 5. on the numerical treatment of liquefaction is divided into two criteria: liquefaction to take 1, take no liquid 0. Table 1 to Table 4 for data and computing results. 4 Conclusion (1) to compare the Table 1, Table 2, the actual data and evaluation data can
相关的主题文章：
[link widoczny dla zalogowanych]
[link widoczny dla zalogowanych]
[link widoczny dla zalogowanych]
The post has been approved 0 times
|
|
