CN103576548B - Intelligent wiper based on neutral net - Google Patents

Abstract

nullThe invention discloses a kind of intelligent wiper based on neutral net，Including rain sensor and controller，Described rain sensor detects the signal of rainfall and by A/D converter, rainfall signal is transmitted in controller，Described controller controls the swing speed of rain brush，Described sensor includes infrared transmitter、Infrared remote receiver、MCU，It is characterized in that: described infrared transmitter includes infrared transmitting circuit，Described pre-amplification circuit outfan is connected with the input of described Sampling Integral circuit，The sequential level input of upper Sampling Integral circuit with signal output part respectively input with described MCU and differential amplification be connected，The outfan of differential amplifier circuit is connected with the sampling input of MCU and the input port of feedback circuit，The outfan of above-mentioned feedback circuit is connected with the negative feedback input of described infrared transmitting circuit.Due to the fact that employing controller based on neutral net, thus solve conventional PID controllers load parameter and change on a large scale and nonlinear change problem solves the uncertainty of fuzzy control simultaneously.

Description

Intelligent wiper based on neutral net

Technical field

The present invention relates to a kind of Wiper system, be specifically related to a kind of intelligent wiper based on neutral net at automotive field.

Background technology

Rain brush is one of important spare part of automobile, traditional mechanical rain brush, in use can disperse the attention of driver.The important hidden danger of vehicle accident is become at sleety weather.

In rain brush intelligence system, detector is by the size of rain sensor detection rainfall, and then controls the swing speed of rain brush.In intelligent wiper system, it is not necessary to driver intervenes, can automatically keep windshield clear, add traffic safety performance.Above-mentioned intelligent wiper is generally divided into sensor and controller, and controller uses fuzzy control or PID to control technology, and sensor uses numerical model analysis detection technique or single-chip microcomputer detection technique.As Patent No. CN201010581428.1 includes that rain sensor and controller, described controller are fuzzy controller, described rain sensor transmits a signal to fuzzy controller by A/D sampling, and output PWM ripple controls action and the speed of wiper；There is certain uncertainty in above-mentioned fuzzy control, the most traditional pid algorithm cannot overcome change and the nonlinear change on a large scale of load parameter.To this end, propose a kind of controller adapting to above-mentioned rain brush and corresponding sensor control algorithm.

Summary, can design a kind of intelligent wiper for automotive field.

Summary of the invention

The technical problem to be solved is the present situation for prior art, it is provided that a kind of replace traditional PID controller and fuzzy controller, can overcome non-linear and fuzzy uncertainty shortcoming intelligent wiper based on neutral net.

nullThe present invention solves the technical scheme that above-mentioned technical problem used: a kind of intelligent wiper based on neutral net，Including rain sensor and controller，Described rain sensor detects the signal of rainfall and by A/D converter, rainfall signal is transmitted in controller，Described controller controls rain brush speed，Described rain sensor includes infrared transmitter、Infrared remote receiver、MCU，It is characterized in that: described infrared transmitter includes that infrared LED is launched，Described infrared accepter includes pre-amplification circuit、Sampling Integral circuit、Differential amplifier circuit，Feedback circuit，Signal after modulation is sent to described pre-amplification circuit by the transmitting of described infrared LED，Described pre-amplification circuit outfan is connected with described Sampling Integral circuit input end，The sequential level input of described Sampling Integral circuit is connected with described MCU and differential amplifier circuit input respectively with signal output part，The outfan of described differential amplifier circuit is connected with the sampling input of MCU and the input port of feedback circuit，The negative feedback input that the outfan of described feedback circuit is launched with described infrared LED is connected.

Described MCU is provided with two for controlling the sequential level output end of sample frequency and connecting the sequential level input of said two analog switch respectively.

Computational methods are as follows:

The first step: fl transmission

The input value of described hidden layer is the input value weighting sum of all input layers:

$x_j=\underset{i}{\mathrm{\Σ}}{w}_{\mathrm{ij}}{x}_i$

w_ijIt is the weighted value of input layer and output layer, x_iFor output layer raindrop frequency or, x_jInput value for hidden layer；

The output valve of described hidden layer is:

$x_j^{\′}=f\left(x_j\right)=1/(1+e^{-x_j})$

x′_jFor the output valve of described hidden layer,It it is the first numerical function；

To the first described transmission function f (x_j) differentiate: $f^{\′}\left(x_j\right)=\frac{\∂x_j^{\′}}{\∂x_j}=x_j^{\′}(1-x_j^{\′})=f\left(x_j\right)[1-f\left(x_j\right)]$ Described output layer output valve is:

$x_o=\underset{j}{\mathrm{\Σ}}{w}_{\mathrm{jo}}{x}_j^{\′}$

w_joFor the weighted value between described output layer and described hidden layer；x_oThe o output valve for output layer；

Described x_oWith corresponding idea outputError amount be:

$e_o=x_o^o-x_o$

Choosing p sample, the error performance target function of pth sample is:

$E_p=\frac{1}{2}\underset{o=1}{\overset{N}{\mathrm{\Σ}}}{e}_o^2$

Second step: reverse transfer

Described w_joLearning algorithm is as follows:

$\mathrm{\Δ}{w}_{\mathrm{jo}}=-\mathrm{\η}\frac{\∂E_p}{\∂w_{\mathrm{jo}}}=\mathrm{\η}{e}_o\frac{\∂x_o}{\∂w_{\mathrm{jo}}}=\mathrm{\η}{e}_o{x}_j^{\′}$

η is learning rate, η ∈ [0,1],

Δw_joFor the weighting value difference between described output layer and described hidden layer；

Taking the k+1 moment, k is an integer, then+1 w of kth_jo:

w_jo(k+1)=w_jo(k)+Δw_jo

Described w_ijLearning algorithm is as follows:

$\mathrm{\Δ}{w}_{\mathrm{ij}}=-\mathrm{\η}\frac{\∂E_p}{\∂w_{\mathrm{ij}}}=\mathrm{\η}\underset{o=1}{\overset{N}{\mathrm{\Σ}}}{e}_o\frac{\∂x_o}{\∂w_{\mathrm{ij}}}$

Δw_ijWeighting value difference for input layer and output layer；

$\frac{\∂x_o}{\∂w_{\mathrm{ij}}}=\frac{\∂x_o}{\∂x_j^{\′}}\·\frac{\∂x_j^{\′}}{\∂x_j}\·\frac{\∂x_j}{\∂w_{\mathrm{ij}}}=w_{\mathrm{io}}\·\frac{\∂x_j^{\′}}{\∂x_j}\·x_i=w_{\mathrm{jo}}\·x_j^{\′}(1-x_j^{\′})\·x_i$

Take k+1 moment, then kth+1:

w_ij(k+1)=w_ij(k)+Δw_ij

Add factor of momentum α, α ∈ [0,1], then kth+1 is respectively as follows:

w_jo(k+1)=w_jo(k)+Δw_jo+α(w_jo(k)-w_jo(k-1))

w_ij(k+1)=w_ij(k)+Δw_ij+α(w_ij(k)-w_ij(k-1))

3rd step:

Specification error precision e_oFor M, set initial w_ijAnd w_joBeing set to [-1 ,+1], learning rate η is N, and factor of momentum is L, sets frequency of training Q；

Then, substitute into the first and second above-mentioned steps, when after iteration repeatedly, when meeting error precision or training number of repetition: stop iteration；

Compared with prior art, due to the fact that employing controller based on neutral net, thus solve conventional PID controllers load parameter and change on a large scale and nonlinear change problem and the uncertain problem of fuzzy control.

Accompanying drawing explanation

Fig. 1 is the schematic diagram of rain sensor circuit theory of the present invention；

Fig. 2 is the schematic diagram of neural network algorithm of the present invention；

Fig. 3 is rain brush speed MAP of the present invention；

Fig. 4 is that rain brush of the present invention is spaced MAP movement time；

Fig. 5 is the windshield-wiper controller program flow diagram of the present invention；

Fig. 6 is the schematic diagram of the rain brush of the present invention；

Fig. 7 is the mastery routine figure of the rain sensor of the present invention；

Detailed description of the invention

Below in conjunction with accompanying drawing embodiment, the present invention is described in further detail.

A kind of intelligent wiper based on neutral net, include two parts, being respectively rain sensor 21 and controller 22, rain sensor detects the signal of rainfall and by A/D converter, rainfall signal is transmitted in controller 22, and described controller 22 controls rain brush swing speed.

Above-mentioned rain sensor 21 includes infrared LED transmitting 7, infrared accepter, MCU1, windshield 6；

Described infrared remote receiver includes the pre-amplification circuit 4 containing silicon cell 5, Sampling Integral circuit 3, differential amplifier circuit 2, feedback circuit 8, described infrared LED is launched 7 and will be sent to described pre-amplification circuit 4 above-mentioned pre-amplification circuit 4 outfan and be connected with the input of Sampling Integral circuit 3 by signal after modulation, the sequential level input of above-mentioned Sampling Integral circuit 3 with signal output part respectively input with described MCU1 outfan and differential amplifier circuit 2 be connected, the outfan of differential amplifier circuit 2 is connected with the sampling input of MCU1 and the input port of feedback circuit 8, the negative feedback input that the outfan of above-mentioned feedback circuit 8 launches 7 with described infrared LED is connected, in order to the stable voltage pulsation brought by raindrop fluctuation.

Described MCU1 is provided with two for controlling the sequential level output end of sample frequency and connecting the sequential level input of said two analog switch respectively.Above-mentioned MCU1 is by communication chip and in order to control the actuating of relay of controller 22, and then controls windscreen wiper swing speed.

The employing STC51 Strong MCU of above-mentioned single-chip microcomputer, control algolithm, containing A/D sampling module, is write with the form of c language and is stated in STC51 single-chip microcomputer by itself.

The control program employing neural network algorithm of this written:

The rule of neural network: with setting up raindrop size 9, raindrop frequency 10 and rain brush speed 13, corresponding relation between the rain brush action intermittent time 12.Set up BP neutral net, input layer 14: raindrop size 9 and raindrop frequency 10；Hidden layer 15 nodes: 6；Output layer 16: rain brush speed 13 and rain brush action interval time 12.

Specific algorithm rule is as follows: include input layer 14, hidden layer 15 and output layer 16.The unit of input layer passes to hidden layer input value.Hidden layer 15 has two neural units: humidity unit and intensity are first.Hidden layer 15 is most important one layer of neutral net, and it can connect input layer 14 and output layer 16.Output layer 16 has two neural units: output unit and interval are first.

First: information fl transmission process

The first step: fl transmission

The input value of described hidden layer is the input value weighting sum of all input layers:

$x_j=\underset{i}{\mathrm{\Σ}}{w}_{\mathrm{ij}}{x}_i$

w_ijIt is the weighted value of input layer and output layer, x_iFor output layer raindrop frequency size and the size of raindrop, x_jInput value for hidden layer；

The output valve of described hidden layer is:

$x_j^{\′}=f\left(x_j\right)=1/(1+e^{-x_j})$

x′_jFor the output valve of described hidden layer,It it is the first numerical function；

To the first described transmission function f (x_j) differentiate: $f^{\′}\left(x_j\right)=\frac{\∂x_j^{\′}}{\∂x_j}=x_j^{\′}(1-x_j^{\′})=f\left(x_j\right)[1-f\left(x_j\right)]$ Described output layer output valve is:

$x_o=\underset{j}{\mathrm{\Σ}}{w}_{\mathrm{jo}}{x}_j^{\′}$

w_joFor the weighted value between described output layer and described hidden layer；x_oThe o output valve for output layer；

Described x_oWith corresponding idea outputError amount be:

$e_o=x_o^o-x_o$

Choosing p sample, the error performance target function of pth sample is:

$E_p=\frac{1}{2}\underset{o=1}{\overset{N}{\mathrm{\Σ}}}{e}_o^2$

Second step: reverse transfer

Described w_joLearning algorithm is as follows:

$\mathrm{\Δ}{w}_{\mathrm{jo}}=-\mathrm{\η}\frac{\∂E_p}{\∂w_{\mathrm{jo}}}=\mathrm{\η}{e}_o\frac{\∂x_o}{\∂w_{\mathrm{jo}}}=\mathrm{\η}{e}_o{x}_j^{\′}$

η is learning rate, η ∈ [0,1],

Δw_joFor the weighting value difference between described output layer and described hidden layer；

Taking the k+1 moment, k is an integer, then+1 w of kth_jo:

w_jo(k+1)=w_jo(k)+Δw_jo

Described w_ijLearning algorithm is as follows:

$\mathrm{\Δ}{w}_{\mathrm{ij}}=-\mathrm{\η}\frac{\∂E_p}{\∂w_{\mathrm{ij}}}=\mathrm{\η}\underset{o=1}{\overset{N}{\mathrm{\Σ}}}{e}_o\frac{\∂x_o}{\∂w_{\mathrm{ij}}}$

Δw_ijWeighting value difference for input layer and output layer；

$\frac{\∂x_o}{\∂w_{\mathrm{ij}}}=\frac{\∂x_o}{\∂x_j^{\′}}\·\frac{\∂x_j^{\′}}{\∂x_j}\·\frac{\∂x_j}{\∂w_{\mathrm{ij}}}=w_{\mathrm{io}}\·\frac{\∂x_j^{\′}}{\∂x_j}\·x_i=w_{\mathrm{jo}}\·x_j^{\′}(1-x_j^{\′})\·x_i$

Take k+1 moment, then kth+1:

w_ij(k+1)=w_ij(k)+Δw_ij

Add factor of momentum α, α ∈ [0,1], then kth+1 is respectively as follows:

w_jo(k+1)=w_jo(k)+Δw_jo+α(w_jo(k)-w_jo(k-1))

w_ij(k+1)=w_ij(k)+Δw_ij+α(w_ij(k)-w_ij(k-1))

3rd step:

Specification error precision e_oFor M, set initial w_ijAnd w_jo, it being set to [-1 ,+1], learning rate η is N, and factor of momentum is L, sets frequency of training Q；

Then, substitute into the first and second above-mentioned steps, when after iteration repeatedly, when meeting error precision or training number of repetition: stop iteration；

After network struction, according to the input and output in example, BP e-learning also adjusts connection weights and the threshold value of neuron, so that network obtains being currently entered output MAP relation.Form is as follows:

As shown in Figs. 3-4, program in MATLAB, if error precision is M takes 0.01, make initial weight scope from-1 to+1, learning rate N is set and takes 0.50, factor of momentum L is set and takes 0.05, and set frequency of training and take 5000 as Q, then training network is until reaching error precision or training number of repetition.After training, test network also obtains the input-output MAP of network.

As shown in Figure 6: sensor control programming:

Rain sensor function includes sending infrared LED lamp drive pulses and sending corresponding Sampling Integral sequential, differential amplifier circuit output voltage carries out AD sampling and calculates drop size and water droplet frequency, inquiry neural metwork training result table, control to send rain brush control information to rain brush by serial ports.

Program is as follows:

1) provide driving pulse 20KHz and provide Sampling Integral sequential to infrared LED light source, need to use single-chip microcomputer timer internal, 1/4 time in intervalometer timing infrared light supply driving pulse cycle, i.e. 50 microseconds, infrared LED is lighted at the first two timing cycle, latter two timing cycle closes infrared LED, opens Sampling Integral sequential when second timing cycle and opens high level sampling window, opens Sampling Integral sequential when the 3rd timer period and open low level sampling window.

2) in timer interrupt service program, record enters interruption times, and 200 times, i.e. 10 milliseconds, sampling time interval puts 1 to mark.

3) in mastery routine by inquiry sampling time interval to mark, be to sample when 1, sampled value and stable voltage compared, calculate difference absolute value, and add up 100 times, as raindrop size.If accumulated value is more than this Earth noise of sensor circuit, then water droplet detected in being judged to this unit interval.Detect 10 unit interval, calculate and detect that the number of times of water droplet is as raindrop frequency.

4) by raindrop size and raindrop frequency queries by the rain brush speedometer training neutral net to draw and rain brush action interval schedule, draw control rain brush speed and rain brush action interval time, and be sent to Wiper motor controller by serial ports.

As shown in Figure 5: controller circuitry programming:

Windshield-wiper controller includes that receiving the rain brush that sends of rain sensor controls information, controls rain brush speed, time delay corresponding rain brush action interval time.Wiper motor control be by realizing at a high speed with relay, the control of two kinds of duties of low speed, so, by the rain brush speed that will receive with threshold ratio relatively, control the high low-speed run state of Wiper motor.

1) receive rain brush by serial ports and control information.

2) if rain brush speed is more than High Speed Threshold Hight_limit, then the quick shelves of Wiper motor are started.

3) if rain brush speed is more than low velocity threshold Low_limit and less than High Speed Threshold Hight_limit, then the slow shelves of Wiper motor are started.

4) time delay rain brush action interval time, step 1 to 4 is repeated.

The following is the training table of neutral net:

With the rain brush speedometer corresponding to Fig. 3:

2.14,2.38,2.67,3.01,3.41,3.88,4.40,4.96,5.57,6.19},

2.77,3.01,3.30,3.64,4.05,4.51,5.03,5.59,6.20,6.82},

3.64,3.88,4.16,4.51,4.91,5.37,5.89,6.46,7.06,7.68},

4.72,4.96,5.25,5.59,6.00,6.46,6.98,7.55,8.15,8.77},

5.95,6.19,6.48,6.82,7.22,7.68,8.20,8.77,9.37,10.0},

7.17,7.41,7.70,8.04,8.45,8.91,9.43,10,10.60,11.22},

8.26,8.50,8.79,9.13,9.53,10,10.51,11.08,11.68,12.31},

9.12,9.36,9.65,10.00,10.40,10.86,11.38,11.95,12.55,13.17},

9.76,10.00,10.28,10.63,11.03,11.49,12.01,12.58,13.18,13.80},

10.19,10.43,10.72,11.06,11.46,11.93,12.44,13.01,13.61,14.24},

10.47,10.71,11.00,11.34,11.75,12.21,12.73,13.30,13.90,14.52},

10.66,10.89,11.18,11.53,11.93,12.39,12.91,13.48,14.08,14.70},

10.77,11.01,11.30,11.64,12.04,12.50,13.02,13.59,14.19,14.82},

10.84,11.08,11.37,11.71,12.11,12.57,13.09,13.66,14.26,14.89},

10.88,11.12,11.41,11.75,12.16,12.62,13.14,13.70,14.31,14.93},

10.91,11.15,11.43,11.78,12.18,12.64,13.16,13.73,14.33,14.95},

10.92,11.16,11.45,11.79,12.20,12.66,13.18,13.75,14.35,14.97},

10.93,11.17,11.46,11.80,12.21,12.67,13.19,13.76,14.36,14.98},

10.94,11.18,11.47,11.81,12.21,12.68,13.19,13.76,14.36,14.99},

{ 10.94,11.18,11.47,11.81,12.22,12.68,13.20,13.76,14.37,14.99}}

With the rain brush action interval schedule as corresponding to Fig. 4:

1.91,1.89,1.87,1.85,1.82,1.78,1.74,1.68,1.63,1.56},

1.82,1.78,1.74,1.69,1.63,1.56,1.48,1.39,1.30,1.21},

1.63,1.56,1.48,1.39,1.30,1.21,1.10,1.000.89,0.79},

1.30,1.21,1.10,1.000.89,0.79,0.69,0.60,0.51,0.44},

0.89,0.79,0.69,0.60,0.51,0.44,0.37,0.31,0.26,0.21},

0.51,0.44,0.37,0.31,0.26,0.21,0.18,0.14,0.12,0.10},

0.26,0.21,0.18,0.14,0.12,0.10,0.08,0.06,0.05,0.04},

0.12,0.10,0.08,0.06,0.05,0.04,0.03,0.02,0.02,0.02},

0.05,0.04,0.04,0.03,0.02,0.02,0.01,0.01,0.01,0.01},

0.02,0.02,0.01,0.01,0.01,0.01,0.01,0.01,0.00,0.00},

0.01,0.01,0.01,0.00,0.00,0.00,0.00,0.00,0.00,0.00},

0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00},

0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00},

0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00},

0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00},

0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00},

0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00},

0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00},

0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00},

{ 0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00}}.

Claims (1)

1. null1. an intelligent wiper based on neutral net，Including rain sensor (21) and controller (22)，Rainfall signal is also transmitted in controller (22) by A/D converter by the signal of described rain sensor detection rainfall，Described controller (22) controls rain brush speed，Described rain sensor (21) includes infrared transmitter、Infrared remote receiver、MCU(1)，It is characterized in that: described infrared transmitter includes that infrared LED launches (7)，Infrared accepter includes pre-amplification circuit (4)、Sampling Integral circuit (3)、Differential amplifier circuit (2)，Feedback circuit (8)，Described infrared LED is launched (7) and signal after modulation is sent to described pre-amplification circuit (4)，Described pre-amplification circuit (4) outfan is connected with described Sampling Integral circuit (3) input，The sequential level input of described Sampling Integral circuit (3) is connected with described MCU (1) and differential amplifier circuit (2) input respectively with signal output part，The outfan of described differential amplifier circuit (2) is connected with the sampling input of MCU (1) and the input port of feedback circuit (8)，The outfan of described feedback circuit (8) is launched the negative feedback input of (7) and is connected with described infrared LED；Described MCU (1) is provided with two for controlling the sequential level output end of sample frequency and connecting the sequential level input of two analog switches respectively；The control method of described intelligent wiper is, including input layer (14), hidden layer (15) and output layer (16), input layer (14) includes multiple input block, described input layer (14) passes to hidden layer (15) input value, hidden layer (15) includes two neural units, it is respectively humidity unit and intensity unit, described hidden layer (15) connects input layer (14) and output layer (16) respectively, described output layer (16) includes two neural units, it is respectively output unit and interval unit: computational methods are as follows: the first step: fl transmission, the input value of described hidden layer is the input value weighting sum of all input layers:

   $x_j=\underset{i}{\mathrm{\Σ}}{w}_{ij}{x}_i$

   w_ijIt is the weighted value of input layer and output layer, x_iFor output layer raindrop frequency size or the size of raindrop, x_iInput value for hidden layer；

   The output valve of described hidden layer is:

   $x_j^{\′}=f\left(x_j\right)=1/(1+e^{-x_j})$

   x′_jFor the output valve of described hidden layer,It it is the first numerical function；

   To the first described numerical function f (x_j) differentiate:

   Described output layer output valve is:

   w_joFor the weighted value between described output layer and described hidden layer；x_oThe o output valve for output layer；

   Described x_oWith corresponding idea outputError amount be:

   $e^o=x_o^o-x_o$

   Choosing p sample, the error performance target function of pth sample is:

   $E_p=\frac{1}{2}\underset{o=1}{\overset{N}{\mathrm{\Σ}}}{e}^{o^2}$

   e^oThe o output valve and corresponding idea output for output layerError amount；

   Second step: reverse transfer

   Described w_joLearning algorithm is as follows:

   η is learning rate, η ∈ [0,1],

   Δw_joFor the weighting value difference between described output layer and described hidden layer；

   Taking the k+1 moment, k is an integer, then+1 w of kth_jo；

   w_jo(k+1)=w_jo(k)+Δw_jo

   Described w_ijLearning algorithm is as follows:

   ${\mathrm{\Δw}}_{ij}=-\mathrm{\η}\frac{\∂E_p}{\∂w_y}=\mathrm{\η}\underset{o=1}{\overset{N}{\mathrm{\Σ}}}{e}_o\frac{\∂x_o}{\∂w_{ij}}$

   Δw_ijWeighting value difference for input layer and output layer；

   Take k+1 moment, then w_ijValue in kth+1 moment is:

   w_ij(k+1)=w_ij(k)+Δw_ij

   Add factor of momentum α, α ∈ [0,1], then W_joAnd W_ij+ 1 value of kth be respectively as follows:

   w_jo(k+1)=w_jo(k)+Δw_jo+α(w_jo(k)-w_jo(k-1))

   w_ij(k+1)=w_ij(k)+Δw_ij+α(w_ij(k)-w_ij(k-1))

   3rd step:

   Specification error precision e_oFor M, by W_ijAnd W_joInitial value be set to [-1 ,+1], learning rate η is N, and factor of momentum is L, set frequency of training Q；

   Then, substitute into the first and second above-mentioned steps, when after iteration repeatedly, when meeting error precision or training number of repetition: stop iteration.