www.melanexjou.fora.pl

[loiwtjsf]

,[link widoczny dla zalogowanych]
Recurrent Network Model of Dynamic Compensation


】 (14) H () calculated by the following formula: 1H () =- Σ (?,[link widoczny dla zalogowanych],)(?,[link widoczny dla zalogowanych],)( 15) RPE algorithm based on the above principle is given by b: e (t) = Y (t) A tt (t) (16a) p (?)={ p (? a 1) a p (? a 1 )(?)[( t) l + (?) p (t-1) ( t)] a (?) p (t A 1)} (16b) 24 No. 1 Tian Social Equality: Recurrent Network Model on Dynamic Compensation of 5lW (t): W (t A 1) + P (t) (t) ? (t) (16c) a large increase. Where,[link widoczny dla zalogowanych], P (t) as the matrix in the Division I,[link widoczny dla zalogowanych], the initial duty] to take P (0) = 10 \?) <1, in order to achieve self-adaptive algorithm; and when the t-∞, taking (t) I 1. (T) is calculated as ?)( 17a) woa-Ad: 0 (?) (17b) aW a \) / 0, Q (t) = a (t) / aW;, and meet 『(?)=(? a 1) + Pj (Bu 1), P (0) = 0 (18a) 【Q (t) =,. (T A 1) + Q (t A 1), Q (0) = 0 (18b) 3 calculated by the above analysis, we can conclude recursive network dynamic compensation of the sensor, follow these steps: 1. According to the characteristics of the sensor, the sensor selected the desired characteristics (reference model); selected middle layer recurrent network nodes; selected on the initial value. 2. Information about the test data, the use of RPE algorithm to train recurrent network model, until the weights converge. 3. The recurrent network model of access to the sensor, the sensor dynamic compensation. Example: The following sensor dynamic compensation, the transfer function expressed as (s) = 2 × 106 / (s +200 s +10), requires the compensated sensor output characteristic after the meeting (s) = 25 × 10 / ( s +7 × 10. s +25 × 10) (reference model). In the simulation, the time domain waveform superimposed on a mean of 0 and standard deviation of 0.1 random noise of normal distribution. Figure 4 shows the before and after compensation to compensate the step response results. Recurrent network nodes in the middle layer taken q = 4. From the results, the compensation of the dynamic characteristics and H (s) of the dynamic characteristics of good agreement. Figure 5 shows the result of a force sensor compensation. Where 1 is the step response curve, using system identification method modeling, model order for the 7-order results can be better. The model order higher, to the dynamic compensation brought certain difficulties. Take q = 8 recursive compensation network shown in Figure 2 shows the curve after compensation to reach steady state in less than 5ms, the corresponding dynamic performance indexes have been very 4 Conclusion 1. Recurrent network model are discussed in the application of dynamic compensation of sensors, from the test results, the sensor through the compensation required to meet the dynamic characteristics, indicating that the method is effective. 2. From the above analysis, the recurrent network dynamic mapping ability of the model itself, its structure only with the input layer and the number of nodes on the middle layer, simplifying the structure of the dynamic compensator design, which have high dynamic compensation of the sensor the more important significance. 3. In order to ensure proper compensation for the dynamic characteristics of the sensor, the selected training data must be typical. 4. Recurrent network model to increase number of nodes in the hidden layer, help to improve the accuracy of dynamic compensation, but will increase the network training time. r / ms; Figure 4, before and after compensation compensation step response tfms; Figure 5 Force sensor dynamic compensation results 【References] [1] Lin Mingbang, Zhao Honglin. Mechanical measurement [M]. Beijing: China Machine Press. 1992. [2] OF SCIENTIFIC Chen Bao, ZHANG Chong. Automatic detection of common technology and instrumentation in the [M]. Beijing: Tsinghua University Press, 2000. [3] LjungL, SoderstromT. TheoryandPracticeofRecursiveIdentificationlMJ. 1ondon: TheMITPress, 1983. [4] KuCC, LeeKY. DiagonalRecurrentNeuralNetworkforDynamicSystemsControl [J]. IEEETransOnNN. 1995,6 (1) :144-155.

相关的主题文章：


[link widoczny dla zalogowanych]

[link widoczny dla zalogowanych]

[link widoczny dla zalogowanych]