CN106627279A - Vehicle and collision energy absorption control device and method thereof - Google Patents

Abstract

The invention relates to a vehicle and a collision energy absorption control device and method thereof. The vehicle collision energy absorption control device comprises a seat body, a damping assembly, an elastic sealing connecting component, a box body, a collision information collecting unit, a passenger information collecting unit, a seat information collecting unit and an ECU, wherein the damping assembly comprises damping plates perpendicularly and fixedly connected to the bottom of the seat body, and damping holes; the box body internally filled with magnetorheological fluid is connected with the damping assembly in a sealed mode through the elastic sealing connecting component; the damping assembly is suspended in the magnetorheological fluid and can be kept separated from the inner surface of the box body; the collision information collecting unit is used for collecting vehicle collision information; the passenger information collecting unit is used for collecting passenger information; the seat information collecting unit is used for collecting seat information; and the ECU receives vehicle collision information, passenger information and seat information, the current control amount acting on the magnetorheological fluid is calculated according to the input information, and thus the damping characteristic of the magnetorheological fluid flowing through the damping holes is controlled. The control device and method can limit displacement of the seat body and well absorb collision energy when collision of the vehicle occurs, passengers can be protected, and the collision safety is improved.

Description

A kind of vehicle and its collision energy-absorbing control device, control method

Technical field

The present invention relates to technical field of vehicle, more particularly to a kind of vehicle and its collision energy-absorbing control device, controlling party Method.

Background technology

The safety issue of automobile is directly related with the life security of people, therefore receives much concern, for example：Safety chair seats of car Security performance.The design of seat can only be directed at present front hit, a kind of crash type in side crash and rear-end collision, and not There is effective energy absorption device to reduce the energy of collision.Prior art prevents from hitting before automobile under operating mode using novel mechanical structures Seat dive and reach, similar device can enter row constraint during collision to the position of occupant, so as to serve Reasonable protective action, but seat still lacks effective energy absorption device, and the energy of collision is still very big.At this stage can Constrain the position of occupant and with energy-absorbing effect, and it is few to be directed to the general seat unit of different crash types.

On the other hand, in terms of concentrating on the comfortableness for improving seat substantially to the active control of seat at present, and in seat Security aspect carries out the invention of active control and device is less, so if seat peace can be improved on the basis of active control Quan Xing, it will make occupant be protected by more comprehensively.

Thus, it is desirable to have a kind of technical scheme come overcome or at least mitigate prior art drawbacks described above at least one It is individual.

The content of the invention

It is an object of the invention to provide a kind of vehicle collision energy-absorbing control device and its method are overcoming or at least mitigate At least one of drawbacks described above of prior art.

For achieving the above object, the present invention provides a kind of vehicle collision energy-absorbing control device, the vehicle collision energy-absorbing control Device processed includes that seat body, damper assembly, elastic packing connecting elements, casing, collision information collecting unit, occupant information are adopted Collection unit, seat information acquisition unit and ECU, wherein：The damper assembly includes being fixedly connected on the seat in a vertical manner The damping sheet of chair body bottom portion and the damping hole through the damping sheet；The casing that enclosed inside has magnetic flow liquid passes through Elastic packing connecting elements is tightly connected with the damper assembly；The damper assembly is suspended in the magnetic flow liquid and can Remain with the inner surface of the casing and separate state；The collision information collecting unit is used for collection vehicle collision information；Institute Occupant information collecting unit is stated for gathering occupant information；The seat information acquisition unit is used to gather the seat body phase For the velocity information of vehicle floor；The ECU is used to receive the vehicle collision information, occupant information and the seat body Relative to the velocity information of vehicle floor, and the electric current control for being applied to the magnetic flow liquid is calculated according to each described information of input Amount processed, to control the damping characteristic when magnetic flow liquid flows through the damping hole.

Further, the ECU includes：First BP neural network model, it is used to choose vehicle and touch in the way of emulating Hit information and occupant information be the first input layer input vector, choose magnetorheological fluid damp power and be the first output layer output vector, Select the sample data of the first input layer input vector and the first output layer output vector and instructed using the sample data Practice study, to build the first corresponding relation between the damping force that vehicle collision information, occupant information and magnetic flow liquid are produced；With And for receiving actual vehicle collision information and the occupant information collecting unit that the collision information collecting unit is collected The actual occupant information for collecting, and according to first corresponding relation, damping force value is expected in output；Second BP neural network mould Type, its damping force and seat body for being used to experimentally choose magnetic flow liquid generation are believed relative to the speed of vehicle floor Cease for the second input layer input vector, choose magnetic flow liquid input current for the second output layer output vector, selection described second The sample data of input layer input vector and the second output layer output vector and using sample data training study, to build Magnetic flow liquid produce damping force, seat body relative to vehicle floor velocity information and magnetic flow liquid input current second Corresponding relation, and for receiving actual chair body that the seat information acquisition unit collects relative to vehicle floor Velocity information and the expectation damping force value of the first BP neural network model output, and according to second corresponding relation, it is defeated Go out to be applied to the controlled quentity controlled variable of magnetic flow liquid input current.

Further, the of the first corresponding relation of the first BP neural network model and the second BP neural network model Two corresponding relations are all obtained according to below scheme：

The first step, definition：Input layer has n neuron, and hidden layer has p neuron, and output layer has q neuron, n, p Positive integer is with q；

X=(x₁,x₂,...,x_n) be the input layer input vector；

Hi=(hi₁,hi₂,…,hi_p) be the hidden layer input vector；

Ho=(ho₁,ho₂,…,ho_p) it is the hidden layer output vector；

Yi=(yi₁,yi₂,…,yi_q) be the output layer input vector；

Yo=(yo₁,yo₂,…,yo_q) it is the output layer output vector；

d_o=(d₁,d₂,…,d_q) it is desired output vector；

b_hFor the threshold value of each neuron of hidden layer

b_oFor the threshold value of each neuron of output layer

w_ihFor the connection weight of the input layer and the hidden layer；

w_hoFor the connection weight of the hidden layer and the output layer；

μ₁、μ₂For learning rate, wherein 0<μ₁<1,0<μ₂<1

F (x) is activation primitive,

E is error function,

K=1,2 ... m is the number of the sample data；

Second step, the initialization of network and sample are chosen：

Using random function to each connection weight assign one it is interval (- 1,1) in random number, set error function e, give Determine computational accuracy ε and maximum study number of times M；

Choose k-th input sample and desired output：

X (k)=(x₁(k),x₂(k),...,x_n(k))

d_o(k)=(d₁(k),d₂(k),...,d_q(k))

3rd step, calculates the input and output of each layer neuron：

ho_h(k)=f (hi_h(k)) h=1,2 ..., p

yo_o(k)=f (yi_o(k)) o=1,2 ... q

4th step, using network desired output and reality output, each neuron of calculation error function pair output layer it is inclined Derivative：

Wherein：

5th step, the partial derivative of each neuron of calculation error function pair input layer：

Wherein：

6th step, using the δ of each neuron of output layer_oK the output of () and each neuron of hidden layer is correcting connection weight w_oh(k)：

7th step, using the δ of each neuron of hidden layer_hThe Introduced Malaria connection weight of (k) and each neuron of input layer：

8th step, calculates global error：

9th step, judges whether network error meets requirement, when error reaches default precision or study number of times more than setting Maximum times, then terminate algorithm；Otherwise, next learning sample and corresponding desired output are chosen, backs into next round Study.

Further, the quantity of the damping sheet in the damper assembly is three pieces, is configured to intersect and mutually hangs down Directly.

Further, the damper assembly also includes connecting plate, and the lower surface of the connecting plate connects each piece of damping The top of plate, upper surface connects the seat body, and the elastic packing connecting elements is arranged outside the connecting plate.

Further, the elastic packing connecting elements is annular in shape, and inner ring is set in outside the damper assembly；The elasticity It is tightly connected component to have for the embedded circumferential slot of the open top of the casing, to be tightly connected the casing and the damping Component.

Further, four sides of the casing are respectively provided with the fixed magnetic pole plate for being wound with magnet exciting coil, the fixation Pole plate is electrically connected by current amplifier with the ECU, and the current control amount of the ECU outputs is put successively via the electric current Big device and fixed magnetic pole plate, are applied to the magnetic flow liquid, to control the damping spy that the magnetic flow liquid flows through the damping hole Property.

The present invention also provides a kind of vehicle, and the vehicle includes vehicle collision energy-absorbing control device as above.

The present invention also provides a kind of vehicle collision energy-absorbing control method, and the vehicle collision energy-absorbing control method includes：

Step 1, Real-time Collection vehicle collision information, occupant information and seat body are believed relative to the speed of vehicle floor Breath；Step 2, builds two BP neural network models, and it is specifically included：Step 21, builds the first BP neural network model, with imitative Genuine mode chooses vehicle collision information and occupant information is the first input layer input vector, chooses magnetorheological fluid damp power for the One output layer output vector, the sample data for selecting the first input layer input vector and the first output layer output vector and Using the sample data training study, to build vehicle collision information, occupant information and magnetic in the first BP neural network model The first corresponding relation between the damping force that rheology liquid is produced；Step 22, builds the second BP neural network model, with the side tested It is the input of the second input layer relative to the velocity information of vehicle floor that formula chooses the damping force and seat body of magnetic flow liquid generation Vector, choose magnetic flow liquid input current and be the second output layer output vector, select the second input layer input vector and the The sample data of two output layer output vectors and using sample data training study, with the second BP neural network model Build the velocity information and magnetic flow liquid input current of damping force, seat body that magnetic flow liquid produces relative to vehicle floor Second corresponding relation；Step 3, actual vehicle collision information and reality that the first BP neural network model receiving step 1 is collected Occupant information, and according to the first corresponding relation built in it, damping force value is expected in output；Step 4, the second BP neural network Actual chair body that model receiving step 1 is collected relative to vehicle floor velocity information, and according to building in it Second corresponding relation, controlled quentity controlled variable of the output action to magnetic flow liquid input current；Step 5, the current control obtained using step 4 Amount control magnetic flow liquid flows through damping characteristic during damping hole, to control displacement and the energy absorption of the seat body.

Further, step 21 and step 22 are all obtained according to below scheme：

The first step, definition：Input layer has n neuron, and hidden layer has p neuron, and output layer has q neuron, n, p Positive integer is with q；

X=(x₁,x₂,...,x_n) be the input layer input vector；

Hi=(hi₁,hi₂,…,hi_p) be the hidden layer input vector；

Ho=(ho₁,ho₂,…,ho_p) it is the hidden layer output vector；

Yi=(yi₁,yi₂,…,yi_q) be the output layer input vector；

Yo=(yo₁,yo₂,…,yo_q) it is the output layer output vector；

d_o=(d₁,d₂,…,d_q) it is desired output vector；

b_hFor the threshold value of each neuron of hidden layer

b_oFor the threshold value of each neuron of output layer

w_ihFor the connection weight of the input layer and the hidden layer；

w_hoFor the connection weight of the hidden layer and the output layer；

μ₁、μ₂For learning rate, wherein 0<μ₁<1,0<μ₂<1

F (x) is activation primitive,

E is error function,

K=1,2 ... m is the number of the sample data；

Second step, the initialization of network and sample are chosen：

Using random function to each connection weight assign one it is interval (- 1,1) in random number, set error function e, give Determine computational accuracy ε and maximum study number of times M；

Choose k-th input sample and desired output：

X (k)=(x₁(k),x₂(k),...,x_n(k))

d_o(k)=(d₁(k),d₂(k),...,d_q(k))

3rd step, calculates the input and output of each layer neuron：

ho_h(k)=f (hi_h(k)) h=1,2 ..., p

yo_o(k)=f (yi_o(k)) o=1,2 ... q

4th step, using network desired output and reality output, each neuron of calculation error function pair output layer it is inclined Derivative：

Wherein：

5th step, the partial derivative of each neuron of calculation error function pair input layer：

Wherein：

6th step, using the δ of each neuron of output layer_oK the output of () and each neuron of hidden layer is correcting connection weight w_oh(k)：

7th step, using the δ of each neuron of hidden layer_hThe Introduced Malaria connection weight of (k) and each neuron of input layer：

8th step, calculates global error：

9th step, judges whether network error meets requirement, when error reaches default precision or study number of times more than setting Maximum times, then terminate algorithm；Otherwise, next learning sample and corresponding desired output are chosen, backs into next round Study.

When vehicle collision accident occurs, the present invention can obtain vehicle collision information using collision information collecting unit, Such as：The information such as crash type (central collision, side impact and knock into the back) and crash severity, is obtained using occupant information collecting unit and is taken advantage of Member's information, such as：The information such as the build and sitting posture of occupant, seat body is obtained for vehicle ground using seat information acquisition unit The velocity information of plate, and corresponding current control amount is provided based on these information and corresponding control strategy, by changing magnetic Field intensity, makes the magnetic-particle in magnetic flow liquid to change disorderly and unsystematic arrangement mode originally in magnetic field, is changed into according to certain Direction ordered arrangement, changes the flow direction of magnetic flow liquid, and the viscosity for controlling magnetic flow liquid is appropriate value, changes chiasma type damping sheet The complexity moved in magnetic rheological liquid, makes seat body produce acceptable displacement, constrains the position of human body, and it is right to improve The protection effect of occupant.Further, since the generation of acceptable displacements, sticking magnetic flow liquid can produce flow at high speed, its stream Heat is produced during damping hole on Jing damping sheets, the process can significantly consume collision energy, mitigate the injury to human body, greatly The crashworthiness of automobile is improve greatly, reaches the effect for significantly improving crashworthiness.

Description of the drawings

Fig. 1 is the structural representation according to the preferred embodiment of vehicle collision energy-absorbing control device one provided by the present invention.

Fig. 2 is the structural representation of the seat in Fig. 1.

Fig. 3 is the structural representation of the elastic packing connector in Fig. 1.

Fig. 4 is the control flow chart of the vehicle collision energy-absorbing control device in Fig. 1.

Specific embodiment

In the accompanying drawings, same or similar element is represented or with same or like function using same or similar label Element.Embodiments of the invention are described in detail below in conjunction with the accompanying drawings.

In describing the invention, term " " center ", " longitudinal direction ", " horizontal ", "front", "rear", "left", "right", " vertical ", The orientation or position relationship of the instruction such as " level ", " top ", " bottom " " interior ", " outward " is to be closed based on orientation shown in the drawings or position System, is for only for ease of the description present invention and simplifies description, rather than indicates or imply that the device or element of indication must have Specific orientation, with specific azimuth configuration and operation, therefore it is not intended that limiting the scope of the invention.

As shown in Figure 1 to Figure 3, the vehicle collision energy-absorbing control device that the present embodiment is provided includes seat body 1, damping Component 2, elastic packing connecting elements 3, casing 4, collision information collecting unit 5, occupant information collecting unit 7, seat information are adopted Collection unit 8 and ECU (Electronic Control Unit, vehicle-mounted computer) 9, wherein：

Damper assembly 2 includes damping sheet 21 and connecting plate 23, wherein, damping sheet 21 is fixedly connected in a vertical manner seat The bottom of body 1, also, it is covered with the damping hole 22 having through damping sheet 21 on damping sheet 21.The lower surface connection of connecting plate 23 is each The top of block damping sheet 21, upper surface is fixedly connected seat body 1 by connector 11.

Elastic packing connecting elements 3 belongs to flexible member, and when its a direction produces vibration, it can play similar bullet The effect of spring, buffer shock-absorbing is carried out to seat, and can limit seat being unlikely to excessive in the displacement of three-dimensional.

The bottom of casing 4 is fixedly connected on floor 12.The enclosed inside of casing 4 has magnetic flow liquid 6, and by elastic packing Connecting elements 3 is tightly connected with damper assembly 2.Damper assembly 2 is placed in magnetic flow liquid 6, is connected using buoyancy and elastic packing The pulling force of component 3 makes damper assembly suspend, and damper assembly 2 can be suspended in magnetic flow liquid 6, with the inner surface of casing 4 all the time Remain and separate state.

Collision information collecting unit 5 is used for the actual collision information of collection vehicle, and the actual collision information of vehicle is mainly wrapped Include two parts, a part is crash type (central collision, side impact and knock into the back), and another part is crash severity.

Preferably, collision information collecting unit 5 can adopt acceleration transducer, acceleration transducer to be used for collection vehicle Acceleration information during collision, the acceleration information includes the direction of acceleration and size.Acceleration transducer is arranged on vehicle body Front portion, both sides and afterbody, the acceleration transducer of anterior and afterbody can detect the acceleration information in longitudinal direction of car, both sides Acceleration transducer can detect acceleration information in lateral direction of car.Above-mentioned acceleration transducer can be the peace on vehicle The sensor that full gas-bag system has in itself, without the need for setting up in addition.

When vehicle crashes, the acceleration information of collision acceleration sensor collection is input to ECU9, by ECU9 judges the actual collision information of vehicle according to acceleration information.The decision method of ECU9 can be said by examples below It is bright：

When vehicle frontal collision hits generation, corresponding acceleration transducer detects the direction of acceleration and size, and should ECU9 is transferred to after acceleration information process, can be with write-in program in ECU9, on the one hand the absolute value of the acceleration to being input into is big It is little to be analyzed, provide crash severity rank；On the other hand the direction of acceleration is analyzed, learns acceleration direction It is towards rear view of vehicle, then can determine that for crash type be central collision.Side impact is similar with the determination methods for knocking into the back, and here is not another One repeats.

Occupant information collecting unit 7 is used to gather occupant information, and occupant information mainly includes build and sitting posture of occupant etc. Information.

Preferably, occupant information collecting unit 7 can adopt imageing sensor and picture processing chip, wherein：Image is passed Sensor is arranged in occupant room, such as the side of room mirror, for carrying out IMAQ to in-car occupant.Image procossing Chip connects imageing sensor, for receiving the occupant image information that imageing sensor is collected, and the occupant for collecting is schemed As information conveyance is to picture processing chip, picture processing chip recognizes human body contour outline based on the characteristic of image border gray scale mutation, The body-shape information of occupant is obtained, ECU9 is input to.When vehicle collides, imageing sensor carries out again IMAQ, should The image information Jing picture processing chip for collecting is processed, and obtains the sitting posture information of human body, is input to ECU9.By accurately acquisition The build and sitting posture information of occupant, effectively can protect occupant.

Seat information acquisition unit 8 is used to gather seat information, seat information mainly include seat body 1 and floor 12 it Between relative displacement information.Preferably, seat information acquisition unit 8 can adopt displacement transducer, displacement transducer to be arranged on Between seat body 1 and floor 12, to gather the relative displacement information between acquisition seat body 1 and floor 12, the relative position Information input is moved to ECU9, is differentiated by ECU9 after processing, obtain the relative velocity between seat body 1 and floor 12.

ECU9 is used to receive the vehicle collision information, occupant information and seat information, and according to the vehicle of input Collision information, occupant information and seat information calculate the current control amount for being applied to magnetic flow liquid 6, are flowed with controlling magnetic flow liquid 6 Damping characteristic during Jing damping holes 22, to change complexity when magnetic flow liquid 6 flows through damping hole 22.

Obviously, damper assembly 2 is suspended in magnetic flow liquid 6, and this is in all the time a suspension equivalent to seat body 1 State, with the change of the motion operating mode of vehicle, seat body 1 can produce corresponding motion, the damping group being connected with seat body 1 The motion of seat body 1 is passed to magnetic flow liquid 6 by part 2, makes magnetic flow liquid 6 flow through resistance towards the direction corresponding with the direction of motion Buddhist nun hole 22.What those skilled in the art could be aware that is：The viscosity of magnetic flow liquid 6 can directly influence the flowing of magnetic flow liquid 6 Property, and then have influence on complexity when magnetic flow liquid 6 flows through damping hole 22.

When vehicle collision accident occurs, the present embodiment can obtain vehicle collision letter using collision information collecting unit Breath, such as：The information such as crash type (central collision, side impact and knock into the back) and crash severity, is obtained using occupant information collecting unit Occupant information is taken, such as：The information such as the build and sitting posture of occupant, seat body is obtained for car using seat information acquisition unit The velocity information on floor, and corresponding current control amount is provided based on each information and corresponding control strategy that collect, By changing magnetic field intensity, make the magnetic-particle in magnetic flow liquid to change disorderly and unsystematic arrangement mode originally in magnetic field, become It is the flow direction for changing magnetic flow liquid according to certain direction ordered arrangement, the viscosity for controlling magnetic flow liquid is appropriate value, changes and hands over The complexity that forked type damping sheet is moved in magnetic rheological liquid, makes seat body produce acceptable displacement, constrains human body Position, effectively prevents human body dive, improves the protection effect to occupant.

At the same time, due to the generation of acceptable displacements, sticking magnetic flow liquid can produce flow at high speed, and it flows through resistance Heat is produced during damping hole on Buddhist nun's plate, the process can significantly consume collision energy, mitigate the injury to human body, carried significantly The high crashworthiness of automobile, reaches the effect for significantly improving crashworthiness.Similarly, side impact, the accident such as knock into the back are sent out When raw, the seat unit can produce acceptable displacement, and then effectively absorb the energy for colliding, and automobile is improved comprehensively Crashworthiness.

In one embodiment, it is to meet safety requirements, the present embodiment is using BP neural network algorithm to magnetic flow liquid formula Safety seat is controlled, and BP neural network algorithm can learn and store substantial amounts of input and output mode mapping relations, and need not The math equation of this mapping relations of description is disclosed in advance.The learning rules of BP neural network algorithm are to use steepest descent method, The weights of each layer neural unit are constantly adjusted by reverse propagated error, makes the error sum of squares of network minimum.BP nerve nets Network model topology structure includes input layer, hidden layer and output layer, and each layer is made up of a number of neuron node.Respectively Cross reaction between layer neuron is described by activation primitive.It is specific as follows：

ECU9 includes the first BP neural network model and the second BP neural network model, wherein：

First BP neural network model is used to choose vehicle collision information in the way of emulating and occupant information is defeated as first Enter a layer input vector, choose magnetorheological fluid damp power be the first output layer output vector, select first input layer be input into Amount is believed with the sample data of the first output layer output vector and using sample data training study with building vehicle collision The first corresponding relation between the damping force that breath, occupant information and magnetic flow liquid are produced；And for Receiving collision information gathering The actual occupant information that the actual vehicle collision information and occupant information collecting unit 7 that unit 5 is collected is collected, and according to institute The first corresponding relation is stated, damping force value is expected in output.

Second BP neural network model is used to experimentally choose damping force and the seat body that magnetic flow liquid is produced It is the second input layer input vector, chooses magnetic flow liquid input current for the second output layer relative to the velocity information of vehicle floor Output vector, the sample data for selecting the second input layer input vector and the second output layer output vector and using the sample Notebook data training study, with build magnetic flow liquid generation damping force, seat body relative to vehicle floor velocity information and Second corresponding relation of magnetic flow liquid input current, and for receiving the actual chair that seat information acquisition unit 8 is collected Information and the expectation damping force value of the first BP neural network model output, and according to second corresponding relation, output is made Use the controlled quentity controlled variable of magnetic flow liquid input current.

In one embodiment, the first corresponding relation of the first BP neural network model and the second BP neural network mould Second corresponding relation of type is all obtained according to below scheme：

The first step, definition：Input layer has n neuron, and hidden layer has p neuron, and output layer has q neuron, n, p Positive integer is with q；

X=(x₁,x₂,...,x_n) be the input layer input vector；

Hi=(hi₁,hi₂,…,hi_p) be the hidden layer input vector；

Ho=(ho₁,ho₂,…,ho_p) it is the hidden layer output vector；

Yi=(yi₁,yi₂,…,yi_q) be the output layer input vector；

Yo=(yo₁,yo₂,…,yo_q) it is the output layer output vector；

d_o=(d_1,d₂,…,d_q) it is desired output vector；

b_hFor the threshold value of each neuron of hidden layer

b_oFor the threshold value of each neuron of output layer

w_ihFor the connection weight of the input layer and the hidden layer；

w_hoFor the connection weight of the hidden layer and the output layer；

μ₁、μ₂For learning rate, wherein 0<μ₁<1,0<μ₂<1

F (x) is activation primitive,

E is error function,

K=1,2 ... m is the number of the sample data；

Second step, the initialization of network and sample are chosen：

Using random function to each connection weight assign one it is interval (- 1,1) in random number, set error function e, give Determine computational accuracy ε and maximum study number of times M；

Choose k-th input sample and desired output：

X (k)=(x₁(k),x₂(k),...,x_n(k))

d_o(k)=(d₁(k),d₂(k),...,d_q(k))

3rd step, calculates the input and output of each layer neuron：

ho_h(k)=f (hi_h(k)) h=1,2 ..., p

yo_o(k)=f (yi_o(k)) o=1,2 ... q

4th step, using network desired output and reality output, each neuron of calculation error function pair output layer it is inclined Derivative：

Wherein：

5th step, the partial derivative of each neuron of calculation error function pair input layer：

Wherein：

6th step, using the δ of each neuron of output layer_oK the output of () and each neuron of hidden layer is correcting connection weight w_oh(k)：

7th step, using the δ of each neuron of hidden layer_hThe Introduced Malaria connection weight of (k) and each neuron of input layer：

8th step, calculates global error：

9th step, judges whether network error meets requirement, when error reaches default precision or study number of times more than setting Maximum times, then terminate algorithm；Otherwise, next learning sample and corresponding desired output are chosen, backs into next round Study.

In one embodiment, the quantity of the damping sheet 21 in damper assembly 2 is three pieces, is configured to intersect and mutually hangs down Directly.Damping sheet 21 is formed and is similar to that in criss-cross construction, therefore, when seat body 1 moves up and down：Horizontally disposed Damping sheet 21 serves the effect of damper.Seat body 1 or so, when moving forward and backward：The two pieces of damping sheets 21 vertically arranged To the effect of damper.

The present embodiment utilizes the orthogonal damping sheet of three dimensions, and is combined with BP neural network algorithm to magnetorheological Liquid formula safety seat is controlled, the damping change control being capable of achieving in space in all directions, in the face of different vehicle collisions When, ECU9 can be calculated after vehicle collision information, occupant information and seat information is received according to each described information of input The current control amount of magnetic flow liquid 6 is applied to, so as to change magnetic field intensity, and then different damping forces is produced, is realized in three-dimensional Space omnidirectional constrains the displacement of seat body, and then the position of constraint human body, effectively prevents human body dive, improves to occupant Protection effect.

As shown in figure 3, elastic packing connecting elements 3 is made up of elastomeric material, and annularly, inner ring is set in damper assembly Outside 2 connecting plate 23.Elastic packing connecting elements 3 has for the embedded circumferential slot 31 of the open top of casing 4, to be tightly connected Casing 4 and damper assembly 2.Elastic packing connecting elements 3 also acts as the effect similar to spring, for producing to damper assembly 2 Pulling force is allowed to and the gravitational equilibrium of damper assembly 2, adds the buoyancy of magnetic flow liquid 6, is allowed to be suspended in magnetic flow liquid 6.

In one embodiment, the vehicle collision energy-absorbing control device is also put including D and D/A converter 10, signal Big modulate circuit 13 and current amplifier 14, wherein：

D and D/A converter 10 is used for the acceleration information, relative between seat body 1 and floor 12 during collision Displacement information changes into data signal by analog signal and the current control amount for ECU9 to be exported is converted by data signal Into analog signal, the analog signal is applied to magnetic flow liquid 6, controls the damping characteristic of magnetic flow liquid 6.

Signal amplifying and conditioning circuit 13 is connected electrically between acceleration transducer and ECU9 and displacement transducer and ECU9 Between, for the relative displacement information between the acceleration information for collecting and seat body 1 and floor 12 to be amplified, whole After stream and filtering, then ECU9 is conveyed to after the process of D and D/A converter 10.

Current amplifier 14 is connected electrically between the D and D/A converter 10 and fixed magnetic pole plate 15, for inciting somebody to action After the current control amount of ECU9 outputs is amplified, the fixed magnetic pole plate 15 is conveyed to.

Four sides of casing 4 are respectively provided with the fixed magnetic pole plate 15 for being wound with magnet exciting coil, and fixed magnetic pole plate 15 is by electricity Stream amplifier 14 is electrically connected with D and D/A converter 10, and the current control amount of ECU9 outputs is successively via modulus/digital-to-analogue conversion Device 10, current amplifier 14 and fixed magnetic pole plate 15, are applied to the magnetic flow liquid 6.

Below by taking automobile front collision as an example, the course of work of the present invention is described in detail：

After occupant gets on the bus, imageing sensor carries out IMAQ to human body, and occupant image information is passed at image Reason chip, picture processing chip obtains buman body type information, people based on the characteristic identification human body contour outline of image border gray scale mutation Body body-shape information is transferred to ECU.When collision accident occurs, imageing sensor can carry out IMAQ to occupant again, should Image information obtains the sitting posture information of human body Jing after picture processing chip process, and is transferred to ECU.At the same time, installed in car The collision acceleration sensor of the front part sides, middle part and afterbody of body, acceleration signal Jing signals during collection collision amplify to be adjusted Reason circuit is carried out passing to after signal transacting in analog-digital converter and carries out data conversion, at the signal input after conversion to ECU Reason, according to the accekeration size after process the order of severity of accident is judged, is sent out according to the walking direction accident of crash acceleration Raw type is front collision.

ECU is received after build, sitting posture, crash severity and crash type this four input quantities, the BP in ECU The expectation damping force that neural network model is obtained.

While collision occurs, displacement transducer obtains relative displacement information of the seat body relative to vehicle floor, Jing after signal amplifying and conditioning circuit carries out signal transacting, be transferred in analog-digital converter carries out modulus turn to the relative displacement information Change, the signal input after analog-to-digital conversion to ECU, to be differentiated by ECU and obtain seat body relative between vehicle floor after processing Relative velocity.

Expectation damping force that first BP neural network model is obtained and seat body are relative to relative between vehicle floor Speed is input among the second BP neural network model as input vector, output current controlled quentity controlled variable.Current control amount ratio contracts Output after little.

Current control amount after scale smaller is input to current amplifier and enters Jing after carrying out digital-to-analogue conversion via weighted-voltage D/A converter Row amplifies, and afterwards electric current is input in the magnet exciting coil in fixed magnetic pole plate, produces magnetic field, makes the magnetic-particle in magnetic flow liquid It is changed into ordered arrangement from disorderly and unsystematic arrangement, the viscosity for controlling magnetic rheological liquid is appropriate value, changes chiasma type damping sheet and exist The complexity moved in magnetic rheological liquid, makes seat body produce acceptable displacement, constrains the position of human body, effectively anti- Only human body dive, improves the protection effect of occupant.

Due to the generation of acceptable displacements, sticking magnetic flow liquid can produce flow at high speed, when it flows through damping plate hole Producing heat can significantly consume collision energy, mitigate the injury to human body, substantially increase the crashworthiness of automobile, reach To the effect for significantly improving crashworthiness.Similar, when side impact, the accident such as knock into the back occur, the seat unit can be produced Acceptable displacement, and then the energy of collision is effectively absorbed, the crashworthiness of automobile is improved comprehensively.

The present invention also provides a kind of vehicle, and the vehicle includes vehicle collision energy-absorbing control as described above described in each embodiment Device.

The present invention also provides a kind of vehicle collision energy-absorbing control method, and the vehicle collision energy-absorbing control method includes：

Step 1, Real-time Collection vehicle collision information, occupant information and seat body are believed relative to the speed of vehicle floor Breath；

Step 2, builds two BP neural network models, and it is specifically included：

Step 21, builds the first BP neural network model, and in the way of emulating vehicle collision information and occupant information are chosen For the first input layer input vector, magnetorheological fluid damp power is chosen for the first output layer output vector, selection first input The sample data of layer input vector and the first output layer output vector and using sample data training study, with a BP Vehicle collision information, the between occupant information and the damping force that magnetic flow liquid is produced first corresponding pass are built in neural network model System；

Step 22, build the second BP neural network model, experimentally choose magnetic flow liquid produce damping force and Seat body is the second input layer input vector, chooses magnetic flow liquid input current for the relative to the velocity information of vehicle floor Two output layer output vectors, the sample data for selecting the second input layer input vector and the second output layer output vector and Using the sample data training study, to build damping force, seat that magnetic flow liquid is produced in the second BP neural network model Body is relative to the velocity information of vehicle floor and the second corresponding relation of magnetic flow liquid input current；

Step 3, the actual vehicle collision information that the first BP neural network model receiving step 1 is collected and actual occupant believe Breath, and according to the first corresponding relation built in it, damping force value is expected in output；

Step 4, the actual chair body that the second BP neural network model receiving step 1 is collected is relative to vehicle floor Velocity information, and according to the second corresponding relation built in it, controlled quentity controlled variable of the output action to magnetic flow liquid input current；

Step 5, the current control amount control magnetic flow liquid obtained using step 4 flows through damping characteristic during damping hole, with Control displacement and the energy absorption of the seat body.

In one embodiment, step 21 and step 22 are all obtained according to below scheme：

The first step, definition：Input layer has n neuron, and hidden layer has p neuron, and output layer has q neuron, n, p Positive integer is with q；

X=(x₁,x₂,...,x_n) be the input layer input vector；

Hi=(hi₁,hi₂,…,hi_p) be the hidden layer input vector；

Ho=(ho₁,ho₂,…,ho_p) it is the hidden layer output vector；

Yi=(yi₁,yi₂,…,yi_q) be the output layer input vector；

Yo=(yo₁,yo₂,…,yo_q) it is the output layer output vector；

d_o=(d₁,d₂,…,d_q) it is desired output vector；

b_hFor the threshold value of each neuron of hidden layer

b_oFor the threshold value of each neuron of output layer

w_ihFor the connection weight of the input layer and the hidden layer；

w_hoFor the connection weight of the hidden layer and the output layer；

μ₁、μ₂For learning rate, wherein 0<μ₁<1,0<μ₂<1

F (x) is activation primitive,

E is error function,

K=1,2 ... m is the number of the sample data；

Second step, the initialization of network and sample are chosen：

Using random function to each connection weight assign one it is interval (- 1,1) in random number, set error function e, give Determine computational accuracy ε and maximum study number of times M；

Choose k-th input sample and desired output：

X (k)=(x₁(k),x₂(k),...,x_n(k))

d_o(k)=(d₁(k),d₂(k),...,d_q(k))

3rd step, calculates the input and output of each layer neuron：

ho_h(k)=f (hi_h(k)) h=1,2 ..., p

yo_o(k)=f (yi_o(k)) o=1,2 ... q

4th step, using network desired output and reality output, each neuron of calculation error function pair output layer it is inclined Derivative：

Wherein：

5th step, the partial derivative of each neuron of calculation error function pair input layer：

Wherein：

6th step, using the δ of each neuron of output layer_oK the output of () and each neuron of hidden layer is correcting connection weight w_oh(k)：

7th step, using the δ of each neuron of hidden layer_hThe Introduced Malaria connection weight of (k) and each neuron of input layer：

8th step, calculates global error：

9th step, judges whether network error meets requirement, when error reaches default precision or study number of times more than setting Maximum times, then terminate algorithm；Otherwise, next learning sample and corresponding desired output are chosen, backs into next round Study.

It is last it is to be noted that：Above example only to illustrate technical scheme, rather than a limitation.This The those of ordinary skill in field should be understood：Technical scheme described in foregoing embodiments can be modified, or it is right Which part technical characteristic carries out equivalent；These modifications are replaced, and the essence for not making appropriate technical solution departs from this Invent the spirit and scope of each embodiment technical scheme.

Claims (10)

1. a kind of vehicle collision energy-absorbing control device, it is characterised in that close including seat body (1), damper assembly (2), elasticity Envelope connecting elements (3), casing (4), collision information collecting unit (5), occupant information collecting unit (7), seat information gathering list First (8) and ECU (9), wherein：The damper assembly (2) is including being fixedly connected in a vertical manner the seat body (1) bottom Damping sheet (21) and the damping hole (22) through the damping sheet (21)；Enclosed inside has the case of magnetic flow liquid (6) Body (4) is tightly connected by elastic packing connecting elements (3) with the damper assembly (2)；The damper assembly (2) is suspended in institute State in magnetic flow liquid (6) and can remain with the inner surface of the casing (4) and separate state；The collision information collecting unit (5) for collection vehicle collision information；The occupant information collecting unit (7) is for gathering occupant information；The seat information Collecting unit (8) is for gathering the seat information；The ECU (9) for receive the vehicle collision information, occupant information and The seat information, and the current control amount for being applied to the magnetic flow liquid (6) is calculated according to each described information of input, to control Make the damping characteristic when magnetic flow liquid (6) flows through the damping hole (22).

2. vehicle collision energy-absorbing control device as claimed in claim 1, it is characterised in that the ECU (9) includes：

First BP neural network model, it is used to be chosen in the way of emulating vehicle collision information and occupant information as the first input Layer input vector, selection magnetorheological fluid damp power are the first output layer output vector, select the first input layer input vector With the sample data of the first output layer output vector and using sample data training study, with build vehicle collision information, The first corresponding relation between the damping force that occupant information and magnetic flow liquid are produced；And for receiving the collision information collection The actual occupant information that the actual vehicle collision information and the occupant information collecting unit (7) that unit (5) is collected is collected, And according to first corresponding relation, damping force value is expected in output；

Second BP neural network model, it is used to experimentally choose the damping force and seat body phase of magnetic flow liquid generation For the velocity information of vehicle floor be the second input layer input vector, to choose magnetic flow liquid input current be that the second output layer is defeated Outgoing vector, the sample data for selecting the second input layer input vector and the second output layer output vector and using the sample Data training study, with build magnetic flow liquid generation damping force, seat body relative to vehicle floor velocity information and magnetic Second corresponding relation of rheology liquid input current, and for receiving the reality that the seat information acquisition unit (8) collects Seat information and the expectation damping force value of the first BP neural network model output, and according to second corresponding relation, it is defeated Go out to be applied to the controlled quentity controlled variable of magnetic flow liquid input current.

3. vehicle collision energy-absorbing control device as claimed in claim 2, it is characterised in that the first BP neural network model The first corresponding relation and the second BP neural network model the second corresponding relation all according to below scheme obtain：

The first step, definition：Input layer has n neuron, and hidden layer has p neuron, and output layer has q neuron, n, p and q It is positive integer；

X=(x₁,x₂,...,x_n) be the input layer input vector；

Hi=(hi₁,hi₂,…,hi_p) be the hidden layer input vector；

Ho=(ho₁,ho₂,…,ho_p) it is the hidden layer output vector；

Yi=(yi₁,yi₂,…,yi_q) be the output layer input vector；

Yo=(yo₁,yo₂,…,yo_q) it is the output layer output vector；

d_o=(d₁,d₂,…,d_q) it is desired output vector；

b_hFor the threshold value of each neuron of hidden layer

b_oFor the threshold value of each neuron of output layer

w_ihFor the connection weight of the input layer and the hidden layer；

w_hoFor the connection weight of the hidden layer and the output layer；

μ₁、μ₂For learning rate, wherein 0<μ₁<1,0<μ₂<1

F (x) is activation primitive,

E is error function,

K=1,2 ... m is the number of the sample data；

Second step, the initialization of network and sample are chosen：

Using random function to each connection weight assign one it is interval (- 1,1) in random number, set error function e, give meter Calculate precision ε and maximum study number of times M；

Choose k-th input sample and desired output：

X (k)=(x₁(k),x₂(k),...,x_n(k))

d_o(k)=(d₁(k),d₂(k),...,d_q(k))

3rd step, calculates the input and output of each layer neuron：

${\mathrm{hi}}_h\left(k\right)=\underset{i=0}{\overset{n}{\mathrm{\&Sigma;}}}{w}_{ih}{x}_i\left(k\right)-b_h,h=1,2,...,p$

ho_h(k)=f (hi_h(k)) h=1,2 ..., p

${\mathrm{yi}}_o\left(k\right)=\underset{h=0}{\overset{p}{\mathrm{\&Sigma;}}}{w}_{ho}{\mathrm{ho}}_h\left(k\right)-b_o,o=1,2,...q$

yo_o(k)=f (yi_o(k)) o=1,2 ... q

4th step, using network desired output and reality output, the partial derivative of each neuron of calculation error function pair output layer：

$\frac{\&part;e}{\&part;w_{ho}}=\frac{\&part;e}{\&part;{\mathrm{yi}}_o}\frac{\&part;{\mathrm{yi}}_o}{\&part;w_{ho}}=-{\mathrm{\&delta;}}_o\left(k\right){\mathrm{ho}}_h\left(k\right)$

Wherein：

${\mathrm{\&delta;}}_o\left(k\right)=-\frac{\&part;e}{\&part;{\mathrm{yi}}_o}=-\frac{\&part;\left(\frac{1}{2}\underset{o=1}{\overset{q}{\&Sigma;}}{\left(d_o\left(k\right)-{\mathrm{yo}}_o\left(k\right)\right)}^2\right)}{\&part;{\mathrm{yi}}_o}=\left(d_o\right(k)-{\mathrm{yo}}_o(k\left)\right)f^{\&prime;}\left({\mathrm{yi}}_o\right(k\left)\right)$

${\mathrm{ho}}_h\left(k\right)=\frac{\&part;{\mathrm{yi}}_o\left(k\right)}{\&part;w_{ho}}=\frac{\&part;\left(\underset{h}{\overset{p}{\mathrm{\&Sigma;}}}{w}_{ho}{\mathrm{ho}}_h\right(k\left)\right)}{\&part;w_{ho}}$

5th step, the partial derivative of each neuron of calculation error function pair input layer：

$\frac{\&part;e}{\&part;w_{ih}}=\frac{\&part;e}{\&part;{\mathrm{hi}}_h\left(k\right)}\frac{\&part;{\mathrm{hi}}_h\left(k\right)}{\&part;w_{ih}}=-x_i\left(k\right){\mathrm{\&delta;}}_h\left(k\right)$

Wherein：

$x_i\left(k\right)=\frac{\&part;{\mathrm{hi}}_h\left(k\right)}{\&part;w_{ih}}=\frac{\&part;\left(\underset{i=0}{\overset{n}{\mathrm{\&Sigma;}}}{w}_{ih}{x}_i\right(k\left)\right)}{\&part;w_{ih}}$

${\mathrm{\&delta;}}_h\left(k\right)=-\frac{\&part;e}{\&part;{\mathrm{hi}}_h\left(k\right)}=-\frac{\&part;\left(\frac{1}{2}\underset{o=1}{\overset{q}{\mathrm{\&Sigma;}}}{\left(d_o\left(k\right)-{\mathrm{yo}}_o\left(k\right)\right)}^2\right)}{\&part;{\mathrm{ho}}_h\left(k\right)}\frac{\&part;{\mathrm{ho}}_h\left(k\right)}{\&part;{\mathrm{hi}}_h\left(k\right)}=\left(\underset{o=1}{\overset{q}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_o\left(k\right)w_{ho}\right)f^{\&prime;}\left({\mathrm{hi}}_h\left(k\right)\right)$

6th step, using the δ of each neuron of output layer_oK the output of () and each neuron of hidden layer is correcting connection weight w_oh (k)：

${\mathrm{\&Delta;w}}_{ho}\left(k\right)=-{\mathrm{\&mu;}}_1\frac{\&part;e}{\&part;w_{ho}}={\mathrm{\&mu;}}_1{\mathrm{\&delta;}}_o\left(k\right){\mathrm{ho}}_h\left(k\right)$

$w_{ho}^{N+1}=w_{ho}^N+{\mathrm{\&mu;}}_1{\mathrm{\&delta;}}_o\left(k\right){\mathrm{ho}}_h\left(k\right)$

7th step, using the δ of each neuron of hidden layer_hThe Introduced Malaria connection weight of (k) and each neuron of input layer：

${\mathrm{\&Delta;w}}_{ih}\left(k\right)=-{\mathrm{\&mu;}}_2\frac{\&part;e}{\&part;w_{ih}}={\mathrm{\&mu;}}_2{\mathrm{\&delta;}}_h\left(k\right)x_i\left(k\right)$

$w_{ih}^{N+1}=w_{ih}^N+{\mathrm{\&mu;}}_2{\mathrm{\&delta;}}_h\left(k\right)x_i\left(k\right)$

8th step, calculates global error：

$E=\frac{1}{2}\underset{k=1}{\overset{m}{\&Sigma;}}\underset{o=1}{\overset{q}{\&Sigma;}}{\left(d_o\right(k)-y_o(k\left)\right)}^2$

9th step, judges whether network error meets requirement, when error reaches default precision or study number of times more than setting most Big number of times, then terminate algorithm；Otherwise, next learning sample and corresponding desired output are chosen, backs into next round study.

4. vehicle collision energy-absorbing control device as claimed any one in claims 1 to 3, it is characterised in that the damping group The quantity of the damping sheet (21) in part (2) is three pieces, is configured to intersect and is mutually perpendicular to.

5. vehicle collision energy-absorbing control device as claimed in claim 4, it is characterised in that the damper assembly (2) also includes Connecting plate (23), the lower surface of the connecting plate (23) connects the top of each piece of damping sheet (21), and upper surface connection is described The elastic packing connecting elements (3) is arranged outside seat body (1), the connecting plate (23).

6. vehicle collision energy-absorbing control device as claimed in claim 5, it is characterised in that the elastic packing connecting elements (3) annular in shape, inner ring is set in the damper assembly (2) outward；The elastic packing connecting elements (3) is with for the casing (4) the embedded circumferential slot (31) of open top, to be tightly connected the casing (4) and the damper assembly (2).

7. vehicle collision energy-absorbing control device as claimed in claim 6, it is characterised in that four sides of the casing (4) The fixed magnetic pole plate (15) for being wound with magnet exciting coil is respectively provided with, the fixed magnetic pole plate (15) is by current amplifier (9) and institute ECU (9) electrical connections are stated, the current control amount of ECU (9) output is successively via the current amplifier (9) and fixed magnetic pole Plate (15), is applied to the magnetic flow liquid (6), to control the damping spy that the magnetic flow liquid (6) flows through the damping hole (22) Property.

8. a kind of vehicle, it is characterised in that include the vehicle collision energy-absorbing control dress as any one of claim 1 to 7 Put.

9. a kind of vehicle collision energy-absorbing control method, it is characterised in that include：

Step 1, the velocity information of Real-time Collection vehicle collision information, occupant information and seat body relative to vehicle floor；

Step 2, builds two BP neural network models, and it is specifically included：

Step 21, builds the first BP neural network model, and vehicle collision information and occupant information are chosen in the way of emulating as the One input layer input vector, selection magnetorheological fluid damp power are the first output layer output vector, select first input layer defeated The sample data of incoming vector and the first output layer output vector and using sample data training study, with neural in a BP The first corresponding relation between the damping force that vehicle collision information, occupant information and magnetic flow liquid are produced is built in network model；

Step 22, builds the second BP neural network model, experimentally chooses damping force and seat that magnetic flow liquid is produced Body relative to the velocity information of vehicle floor be the second input layer input vector, to choose magnetic flow liquid input current be second defeated Go out a layer output vector, select sample data and the utilization of the second input layer input vector and the second output layer output vector The sample data training study, to build damping force, seat body that magnetic flow liquid is produced in the second BP neural network model Second corresponding relation of velocity information and magnetic flow liquid input current relative to vehicle floor；

Step 3, the actual vehicle collision information that the first BP neural network model receiving step 1 is collected and actual occupant information, And according to the first corresponding relation built in it, damping force value is expected in output；

Step 4, speed of the actual chair body that the second BP neural network model receiving step 1 is collected relative to vehicle floor Information, and according to the second corresponding relation built in it, controlled quentity controlled variable of the output action to magnetic flow liquid input current；

Step 5, the current control amount control magnetic flow liquid obtained using step 4 flows through damping characteristic during damping hole, to control The displacement of the seat body and energy absorption.

10. vehicle collision energy-absorbing control method as claimed in claim 9, it is characterised in that step 21 and step 22 all according to Below scheme is obtained：

The first step, definition：Input layer has n neuron, and hidden layer has p neuron, and output layer has q neuron, n, p and q It is positive integer；

X=(x₁,x₂,...,x_n) be the input layer input vector；

Hi=(hi₁,hi₂,…,hi_p) be the hidden layer input vector；

Ho=(ho₁,ho₂,…,ho_p) it is the hidden layer output vector；

Yi=(yi₁,yi₂,…,yi_q) be the output layer input vector；

Yo=(yo₁,yo₂,…,yo_q) it is the output layer output vector；

d_o=(d₁,d₂,…,d_q) it is desired output vector；

b_hFor the threshold value of each neuron of hidden layer

b_oFor the threshold value of each neuron of output layer

w_ihFor the connection weight of the input layer and the hidden layer；

w_hoFor the connection weight of the hidden layer and the output layer；

μ₁、μ₂For learning rate, wherein 0<μ₁<1,0<μ₂<1

F (x) is activation primitive,

E is error function,

K=1,2 ... m is the number of the sample data；

Second step, the initialization of network and sample are chosen：

Using random function to each connection weight assign one it is interval (- 1,1) in random number, set error function e, give meter Calculate precision ε and maximum study number of times M；

Choose k-th input sample and desired output：

X (k)=(x₁(k),x₂(k),...,x_n(k))

d_o(k)=(d₁(k),d₂(k),...,d_q(k))

3rd step, calculates the input and output of each layer neuron：

${\mathrm{hi}}_h\left(k\right)=\underset{i=0}{\overset{n}{\mathrm{\&Sigma;}}}{w}_{ih}{x}_i\left(k\right)-b_h,h=1,2,...,p$

ho_h(k)=f (hi_h(k)) h=1,2 ..., p

${\mathrm{yi}}_o\left(k\right)=\underset{h=0}{\overset{p}{\mathrm{\&Sigma;}}}{w}_{ho}{\mathrm{ho}}_h\left(k\right)-b_o,o=1,2,...q$

yo_o(k)=f (yi_o(k)) o=1,2 ... q

4th step, using network desired output and reality output, the partial derivative of each neuron of calculation error function pair output layer：

$\frac{\&part;e}{\&part;w_{ho}}=\frac{\&part;e}{\&part;{\mathrm{yi}}_o}\frac{\&part;{\mathrm{yi}}_o}{\&part;w_{ho}}=-{\mathrm{\&delta;}}_o\left(k\right){\mathrm{ho}}_h\left(k\right)$

Wherein：

${\mathrm{\&delta;}}_o\left(k\right)=-\frac{\&part;e}{\&part;{\mathrm{yi}}_o}=-\frac{\&part;\left(\frac{1}{2}\underset{o=1}{\overset{q}{\&Sigma;}}{\left(d_o\left(k\right)-{\mathrm{yo}}_o\left(k\right)\right)}^2\right)}{\&part;{\mathrm{yi}}_o}=\left(d_o\right(k)-{\mathrm{yo}}_o(k\left)\right)f^{\&prime;}\left({\mathrm{yi}}_o\right(k\left)\right)$

${\mathrm{ho}}_h\left(k\right)=\frac{\&part;{\mathrm{yi}}_o\left(k\right)}{\&part;w_{ho}}=\frac{\&part;\left(\underset{h}{\overset{p}{\mathrm{\&Sigma;}}}{w}_{ho}{\mathrm{ho}}_h\right(k\left)\right)}{\&part;w_{ho}}$

5th step, the partial derivative of each neuron of calculation error function pair input layer：

$\frac{\&part;e}{\&part;w_{ih}}=\frac{\&part;e}{\&part;{\mathrm{hi}}_h\left(k\right)}\frac{\&part;{\mathrm{hi}}_h\left(k\right)}{\&part;w_{ih}}=-x_i\left(k\right){\mathrm{\&delta;}}_h\left(k\right)$

Wherein：

$x_i\left(k\right)=\frac{\&part;{\mathrm{hi}}_h\left(k\right)}{\&part;w_{ih}}=\frac{\&part;\left(\underset{i=0}{\overset{n}{\mathrm{\&Sigma;}}}{w}_{ih}{x}_i\right(k\left)\right)}{\&part;w_{ih}}$

${\mathrm{\&delta;}}_h\left(k\right)=-\frac{\&part;e}{\&part;{\mathrm{hi}}_h\left(k\right)}=-\frac{\&part;\left(\frac{1}{2}\underset{o=1}{\overset{q}{\mathrm{\&Sigma;}}}{\left(d_o\left(k\right)-{\mathrm{yo}}_o\left(k\right)\right)}^2\right)}{\&part;{\mathrm{ho}}_h\left(k\right)}\frac{\&part;{\mathrm{ho}}_h\left(k\right)}{\&part;{\mathrm{hi}}_h\left(k\right)}=\left(\underset{o=1}{\overset{q}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_o\left(k\right)w_{ho}\right)f^{\&prime;}\left({\mathrm{hi}}_h\left(k\right)\right)$

6th step, using the δ of each neuron of output layer_oK the output of () and each neuron of hidden layer is correcting connection weight w_oh (k)：

${\mathrm{\&Delta;w}}_{ho}\left(k\right)=-{\mathrm{\&mu;}}_1\frac{\&part;e}{\&part;w_{ho}}={\mathrm{\&mu;}}_1{\mathrm{\&delta;}}_o\left(k\right){\mathrm{ho}}_h\left(k\right)$

$w_{ho}^{N+1}=w_{ho}^N+{\mathrm{\&mu;}}_1{\mathrm{\&delta;}}_o\left(k\right){\mathrm{ho}}_h\left(k\right)$

7th step, using the δ of each neuron of hidden layer_hThe Introduced Malaria connection weight of (k) and each neuron of input layer：

${\mathrm{\&Delta;w}}_{ih}\left(k\right)=-{\mathrm{\&mu;}}_2\frac{\&part;e}{\&part;w_{ih}}={\mathrm{\&mu;}}_2{\mathrm{\&delta;}}_h\left(k\right)x_i\left(k\right)$

$w_{ih}^{N+1}=w_{ih}^N+{\mathrm{\&mu;}}_2{\mathrm{\&delta;}}_h\left(k\right)x_i\left(k\right)$

8th step, calculates global error：

$E=\frac{1}{2}\underset{k=1}{\overset{m}{\&Sigma;}}\underset{o=1}{\overset{q}{\&Sigma;}}{\left(d_o\right(k)-y_o(k\left)\right)}^2$

9th step, judges whether network error meets requirement, when error reaches default precision or study number of times more than setting most Big number of times, then terminate algorithm；Otherwise, next learning sample and corresponding desired output are chosen, backs into next round study.