1. A multi-olfactory robot cooperative odor source positioning method is characterized in that: firstly, in an odor source discovery stage, a plurality of olfactory robots are randomly distributed at different positions of a two-dimensional search plane, and then all olfactory robots perform outward expanding spiral motion; once the gas concentrations measured by any three olfactory robots are all larger than a low threshold value, the three olfactory robots enter an odor source tracking stage at the moment and start to do s-shaped segmented sinusoidal motion towards the direction of an odor source; based on a BP neural network and an RFID positioning system, the BP neural network dynamically adjusts a weight by utilizing periodic gradient information to control the motion slope of each period of the olfactory robot; the RFID positioning system dynamically modifies the amplitude and angular frequency of sinusoidal motion in each period according to the gas concentration difference by transmitting the gas concentrations measured by a plurality of olfactory robots; once the gas concentration measured by any olfactory robot is greater than the high threshold, the olfactory robot enters an odor source confirmation stage and performs odor source positioning by using a statistical method.

2. The multi-olfactory robot cooperative odor source locating method according to claim 1, wherein: the method specifically comprises the following steps:

step 1: in an odor source discovery stage, n olfactory robots are respectively marked as robots l, l is 1,2,3, n, and randomly distributed at any initial position of a two-dimensional search plane, and the robots l respectively start to make spiral motion with increasing radius by taking the respective initial positions as original points, as shown in formula (1); and measuring the gas concentration at the location at intervals of Δ t;

wherein t is b × Δ t, b is 1,2,3, and t represents the current measurement time; (x)_l(t),y_l(t)) corresponding coordinates representing the position of the robot l at time t;

step 2: in the stage of odor source discovery, once any three olfactory robots are at time t₁The measured gas concentrations are all larger than the low threshold value gamma, and are marked as Robot according to the sequence satisfying the low threshold value condition₁、Robot₂、Robot₃Then the three olfactory robots enter the odor source tracking immediatelyA stage; the rest olfactory robots stop moving and quit the positioning process of the odor source;

and step 3: in the odor source tracking phase, in order to get close to the odor source quickly, the olfactory Robot Robot_βAt time t₁Stopping the outward expanding spiral motion, and starting the s-shaped segmental sinusoidal motion with beta being 1,2 and 3, and starting the time t₁The direction of the spiral motion of (a) is taken as the initial direction of the sinusoidal motion; setting the period of sinusoidal motion to be 2 pi, the fixed motion time of each period to be T, and the olfactory Robot Robot in the k-th period_βThe sinusoidal motion curve of (2);

in the formula (x)_β′,y_β') Robot for olfaction_βTwo-dimensional coordinates relative to the initial time of each cycle; t' represents the relative time within each cycle; a. the_β(k),ω_β(k) Respectively represent the k cycle olfactory Robot_βAmplitude and angular frequency of sinusoidal motion, where A_β(k),ω_β(k) Keep constant value constant during each cycle while ω_β(k)×x_β′∈[2(k-1)π,2kπ)；

Considering omega_β(k)x_β' 2k pi-pi is the symmetric center point of the kth period, and has certain representativeness to the average gas concentration of the period, and the kth period olfactory Robot Robot is assumed_βAt omega_β(k)×x_βThe gas concentration measured when is 2k pi-pi is respectively PL_β(ii) a According to the RFID positioning algorithm, the gas concentration presents the characteristic of logarithmic form attenuation along with the increase of the distance, and the Robot is a smell Robot₃The measured gas concentration is used as a reference to obtain the k cycle olfactory Robot Robot_βAnd the distance between the odor source, as shown in formula (3);

in the formula (d)₁,d₂,d₃Respectively represent the k cycle olfactory Robot_βActual distance to the odor source; ζ represents a path loss coefficient, and is usually set to 2 to 5;

according to the formula (3), without loss of generality, Robot with a Robot₁And Robot₃The distance ratio between them is calculated as an example, as shown in formula (4);

robot Robot_βThe difference in distance from the odor source will determine the amplitude A in equation (2)_β(k) And angular frequency ω_β(k) A difference of (a); when the olfactory robot is far away from the odor source, the gas concentration is low, sinusoidal motion with large amplitude and small angular frequency is required, and the olfactory robot is quickly close to the odor source and collects gas concentration data in a large range; when the olfactory robot is close to the position of the odor source, the gas concentration is higher, the sinusoidal motion with smaller amplitude and larger angular frequency is required, so that the gas concentration can be collected in a small range near the odor source for the following odor source confirmation;

at the end of the k-th cycle, the olfactory robot will adjust the amplitude A of the k + 1-th cycle according to the measured gas concentration_β(k +1), angular frequency ω_β(k +1) and an initial movement direction; when k is 1, the olfactory robot adjusts according to the gas concentration measured at the moment when the period t' is 0; wherein the amplitude A_β(k +1), angular frequency ω_β(k +1) realizing dynamic adjustment by using a sigmoid activation function, and specifically carrying out the dynamic adjustment according to the following rules: robot for olfaction₃As a reference, the amplitude A of the reference is set₃(k +1) and angular frequency ω₃(k +1) is a fixed value epsilon and rho, and the k +1 th cycle olfactory Robot Robot is obtained₁And Robot₂The amplitude and angular frequency of (a) are as shown in formulas (5) and (6);

and 4, step 4: in the odor source tracking stage, the adjustment of the motion direction of the olfactory robot is carried out according to the following rules: robot with sense of smell₁For example, the olfactory robot in the k cycleMeasuring gas concentration, and using the concentration values as input of BP neural network, so that the number of nodes of BP neural network input layer is 4, respectively_pRepresents, p ═ 1,2,3, 4; the output layer has the kth period omega_β(k)×x_βActual scent source predicted direction angle slope of' 2k piThe corresponding node number is 1, and the node number of the hidden layer is set to beWherein m represents the number of nodes of an input layer, m' represents the number of nodes of an output layer, and a is an integer between 1 and 10;

firstly, considering that the olfactory robot has large average difference of gas concentration measured in different periods, which can cause large influence on the training of BP neural network, the input volume v of the k period is used_p(k) Normalizing to obtain the input V of the BP neural network_p(k) As shown in formula (7);

wherein, min (v)_p(k) Represents an input node v_pMinimum value of (d); max (v)_p(k) Represents an input node v_pMaximum value of (d);

the desired odor source predicted direction angle slope isThe slope is the kth period omega_β(k)×x_βObtaining a loss function E (k) of the kth period as shown in a formula (8) by the slope of an included angle between the position where the 2k pi is located and the connecting line of the odor source;

in the formula (I), the compound is shown in the specification,representing an actual scent source prediction direction angle slope;

then the weight value change quantity delta w of the k-1 th period_i,j(k-1) the weight adjustment value Δ w for the kth cycle is obtained by applying the weight adjustment value to the kth cycle, and not considering that k is 1_i,j(k) As shown in formula (9);

in the formula, η represents a learning rate;respectively representing gradient information of the k-th cycle, w_i,j ^γ(k) Representing the weight value from a node i of a gamma layer to a node j of a next layer in the k-th periodic neural network, wherein gamma-1 is represented as an input layer, and gamma-2 is represented as an implied layer; -0.5 < lambda^-＜0＜λ⁺< 0.5 denotes the adjustment factor;

then, updating the weight value to obtain the weight value w of the (k +1) th period_i,j(k+1)；

w_i,j(k+1)＝w_i,j(k)+Δw_i,j(k) (10)

Finally, according to the actual output value of the BP neural network in the kth period, obtaining the actual sinusoidal motion curve of the kth period by using the sinusoidal motion curve in the rotation matrix correction formula (2), as shown in a formula (11);

in the formula, X_β,Y_βRobot for representing smell_βActual two-dimensional coordinates relative to the initial time of each cycle;

and 5: in the stage of odor source tracking, once the gas concentration measured by a certain olfactory robot is greater than a high threshold value delta, the olfactory robot enters the stage of odor source confirmation and resets the current period to 0, and because the gas concentration at the position of the odor source is the highest, the gas concentration transmitted to the periphery is gradually reduced, the olfactory robot performs inward-contracting spiral motion near the point with the higher gas concentration, and the rest olfactory robots temporarily stop moving; in the process, the olfactory robot in the r-th period is judgedMeasured gas concentration value v_p(r) whether it is greater than the high threshold δ as shown in equation (12);

wherein μ represents a gas concentration value v measured in the r-th cycle_p(r) mean value, σ represents the value of the gas concentration v measured in the r-th cycle_p(r) variance of (r);

if the olfactory robot continuously meets the conditions for three periods, the point can be judged as an odor source, and the positioning of the odor source is finished at the moment; when the olfaction Robot can not satisfy the above conditions, the olfaction Robot Robot at this time_βAnd returning to execute the step 3 and continuing to make the sinusoidal motion.

3. The multi-olfactory robot cooperative odor source locating method according to claim 2, wherein: and a is an integer of 3, and the number of nodes of the hidden layer is defaulted to 5.

Background

The rapid location of the dangerous odor source is of great significance to both human health and environmental safety. Traditional odor source positioning methods track odor sources according to local concentration gradients, such as zigzag search and trilateral positioning search, and because such algorithms lack a priori knowledge, the whole search area is usually traversed, so that the search efficiency of the algorithms is low, and the degree of intelligence is weak.

In recent years, the odor source positioning method based on the artificial neural network is widely applied. For example, research on improving the convergence rate of neural network training by combining a mixed frog-leaping algorithm and research on introducing a radial basis neural network into the active olfactory sensation of a robot are carried out. Although the method improves the global quick search capability to a certain extent, the method lacks information sharing of multiple robots and cannot well realize linkage perception of iterative training results and movement path planning.

Disclosure of Invention

Aiming at the defects of the prior art, the invention constructs a multi-olfactory robot cooperative odor source positioning method. The olfactory robot finds the odor source quickly through spiral motion, optimizes the weight adjustment of the BP neural network to improve the accuracy of tracking the motion slope, quickly tracks the odor source through s-shaped segmented sinusoidal motion, and simultaneously strengthens the information coupling among multiple robots by utilizing the characteristics of the RFID positioning system, so that the expandability and the robustness of the system are improved to a certain degree, and the generalization capability of the model is strengthened while the quick positioning of the odor source is realized.

A multi-olfactory robot cooperative odor source positioning method comprises the following steps: firstly, in an odor source discovery stage, a plurality of olfactory robots are randomly distributed at different positions of a two-dimensional search plane, and then all olfactory robots perform outward expanding spiral motion; once the gas concentrations measured by any three olfactory robots are all larger than a low threshold value, the three olfactory robots enter an odor source tracking stage at the moment and start to do s-shaped segmented sinusoidal motion towards the direction of an odor source; based on a BP neural network and an RFID positioning system, the BP neural network dynamically adjusts a weight by utilizing periodic gradient information to control the motion slope of each period of the olfactory robot; the RFID positioning system dynamically modifies the amplitude and angular frequency of sinusoidal motion in each period according to the gas concentration difference by transmitting the gas concentrations measured by a plurality of olfactory robots; once the gas concentration measured by any olfactory robot is greater than the high threshold, the olfactory robot enters an odor source confirmation stage and performs odor source positioning by using a statistical method.

The invention specifically comprises the following steps:

step 1: in an odor source discovery stage, n olfactory robots are respectively marked as robots l, l is 1,2,3, n, and randomly distributed at any initial position of a two-dimensional search plane, and the robots l respectively start to make spiral motion with increasing radius by taking the respective initial positions as original points, as shown in formula (1); and measuring the gas concentration at the location at intervals of Δ t;

where t is b × Δ t, b is 1,2,3 …, and t represents the current measurement time; (x)_l(t),y_l(t)) corresponding coordinates representing the position of the robot l at time t;

step 2: in the stage of odor source discovery, once any three olfactory robots are at time t₁The measured gas concentrations are all larger than the low threshold value gamma, and are marked as Robot according to the sequence satisfying the low threshold value condition₁、Robot₂、Robot₃Then the three olfactory robots immediately enter an odor source tracking stage; the rest olfactory robots stop moving and quit the positioning process of the odor source;

and step 3: in the odor source tracking phase, in order to get close to the odor source quickly, the olfactory Robot Robot_βAt time t₁Stopping the outward expanding spiral motion, and starting the s-shaped segmental sinusoidal motion with beta being 1,2 and 3, and starting the time t₁The direction of the spiral motion of (a) is taken as the initial direction of the sinusoidal motion; setting the period of sinusoidal motion to be 2 pi, the fixed motion time of each period to be T, and the olfactory Robot Robot in the k-th period_βThe sinusoidal motion curve of (2);

in the formula (x)_β′,y_β') Robot for olfaction_βTwo-dimensional coordinates relative to the initial time of each cycle; t' represents the relative time within each cycle; a. the_β(k),ω_β(k) Is divided intoRobot for distinguishing k-th cycle_βAmplitude and angular frequency of sinusoidal motion, where A_β(k),ω_β(k) Keeping constant value in each period while

ω_β(k)×x_β′∈[2(k-1)π,2kπ)；

Considering omega_β(k)x_β' 2k pi-pi is the symmetric center point of the kth period, and has certain representativeness to the average gas concentration of the period, and the kth period olfactory Robot Robot is assumed_βAt omega_β(k)×x_βThe gas concentration measured when is 2k pi-pi is respectively PL_β(ii) a According to the RFID positioning algorithm, the gas concentration presents the characteristic of logarithmic form attenuation along with the increase of the distance, and the Robot is a smell Robot₃The measured gas concentration is used as a reference to obtain the k cycle olfactory Robot Robot_βAnd the distance between the odor source, as shown in formula (3);

in the formula (d)₁,d₂,d₃Respectively represent the k cycle olfactory Robot_βActual distance to the odor source; ζ represents a path loss coefficient, and is usually set to 2 to 5;

according to the formula (3), without loss of generality, Robot with a Robot₁And Robot₃The distance ratio between them is calculated as an example, as shown in formula (4);

robot Robot_βThe difference in distance from the odor source will determine the amplitude A in equation (2)_β(k) And angular frequency ω_β(k) A difference of (a); when the olfactory robot is far away from the odor source, the gas concentration is low, sinusoidal motion with large amplitude and small angular frequency is required, and the olfactory robot is quickly close to the odor source and collects gas concentration data in a large range; when in useWhen the olfactory robot is close to the position of the odor source, the gas concentration is higher, the sinusoidal motion with smaller amplitude and larger angular frequency is required, so that the gas concentration can be collected in a small range near the odor source for the following odor source confirmation;

at the end of the k-th cycle, the olfactory robot will adjust the amplitude A of the k + 1-th cycle according to the measured gas concentration_β(k +1), angular frequency ω_β(k +1) and an initial movement direction; when k is 1, the olfactory robot adjusts according to the gas concentration measured at the moment when the period t' is 0; wherein the amplitude A_β(k +1), angular frequency ω_β(k +1) realizing dynamic adjustment by using a sigmoid activation function, and specifically carrying out the dynamic adjustment according to the following rules: robot for olfaction₃As a reference, the amplitude A of the reference is set₃(k +1) and angular frequency ω₃(k +1) is a fixed value epsilon and rho, and the k +1 th cycle olfactory Robot Robot is obtained₁And Robot₂The amplitude and angular frequency of (a) are as shown in formulas (5) and (6);

and 4, step 4: in the odor source tracking stage, the adjustment of the motion direction of the olfactory robot is carried out according to the following rules: robot with sense of smell₁For example, the olfactory robot in the k cycleMeasuring gas concentration, and using the concentration values as input of BP neural network, so that the number of nodes of BP neural network input layer is 4, respectively_pRepresents, p ═ 1,2,3, 4; the output layer has the kth period omega_β(k)×x_βActual scent source predicted direction angle slope of' 2k piThe corresponding node number is 1, and the node number of the hidden layer is set to beWherein m represents the number of nodes of the input layer, m' represents the number of nodes of the output layer, a is an integer between 1 and 10, and an integer 3 is selected as default, so that the number of nodes of the hidden layer is 5 as default; considering that the traditional BP neural network always has the problem of low convergence speed, and the optimization of the weight can effectively improve the training speed of the neural network, so that the dynamic adjustment of the weight by utilizing the periodic gradient information can still accelerate the adjustment speed when the variable quantity of the weight is small, and play a role in damping when the weight is changed greatly;

firstly, considering that the olfactory robot has large average difference of gas concentration measured in different periods, which can cause large influence on the training of BP neural network, the input volume v of the k period is used_p(k) Normalizing to obtain the input V of the BP neural network_p(k) As shown in formula (7);

wherein, min (v)_p(k) Represents an input node v_pMinimum value of (d); max (v)_p(k) Represents an input node v_pMaximum value of (d);

the desired odor source predicted direction angle slope isThe slope is the kth period omega_β(k)×x_βObtaining a loss function E (k) of the kth period as shown in a formula (8) by the slope of an included angle between the position where the 2k pi is located and the connecting line of the odor source;

in the formula (I), the compound is shown in the specification,representing an actual scent source prediction direction angle slope;

then the weight value change quantity delta w of the k-1 th period_i,j(k-1) the weight adjustment value Δ w for the kth cycle is obtained by applying the weight adjustment value to the kth cycle, and not considering that k is 1_i,j(k) As shown in formula (9);

in the formula, η represents a learning rate;respectively representing gradient information of the k-th cycle, w_i,j ^γ(k) Representing the weight value from a node i of a gamma layer to a node j of a next layer in the k-th periodic neural network, wherein gamma-1 is represented as an input layer, and gamma-2 is represented as an implied layer; -0.5 < lambda^-＜0＜λ⁺< 0.5 denotes the adjustment factor;

then, updating the weight value to obtain the weight value w of the (k +1) th period_i,j(k+1)；

w_i,j(k+1)＝w_i,j(k)+Δw_i,j(k) (10)

Finally, according to the actual output value of the BP neural network in the kth period, obtaining the actual sinusoidal motion curve of the kth period by using the sinusoidal motion curve in the rotation matrix correction formula (2), as shown in a formula (11);

in the formula, X_β,Y_βRobot for representing smell_βActual two-dimensional coordinates relative to the initial time of each cycle;

and 5: in the odor source tracking stage, once the gas concentration measured by a certain olfactory robot is greater than a high threshold value delta, the olfactory robot followsThe smell source confirming stage is entered, the current period is reset to 0, and the gas concentration at the position of the smell source is the highest, and the gas concentration transmitted to the periphery is gradually reduced, so that the smell robot can do inward-contracting spiral motion near the point with higher gas concentration, and the rest smell robots stop moving temporarily; in the process, the olfactory robot in the r-th period is judgedMeasured gas concentration value v_p(r) whether it is greater than the high threshold δ as shown in equation (12);

wherein μ represents a gas concentration value v measured in the r-th cycle_p(r) mean value, σ represents the value of the gas concentration v measured in the r-th cycle_p(r) variance of (r);

if the olfactory robot continuously meets the conditions for three periods, the point can be judged as an odor source, and the positioning of the odor source is finished at the moment; when the olfaction Robot can not satisfy the above conditions, the olfaction Robot Robot at this time_βAnd returning to execute the step 3 and continuing to make the sinusoidal motion.

Preferably, a is an integer of 3, and the number of nodes of the hidden layer is defaulted to 5.

The invention has the characteristics that:

1. the olfactory robot adopts s-shaped segmented sinusoidal motion, dynamically adjusts the amplitude, angular frequency and direction angle slope of sinusoidal motion in each period, realizes the autonomous search of a complete odor source, and ensures less energy consumption of an olfactory robot system.

2. The back propagation of the traditional BP neural network is improved, the weight change obtained in the k-1 th period is used as prior information, and the prior information is applied to the weight adjustment of the k-th period, so that the flexibility and the efficiency of the search can be greatly improved.

3. The information sharing among the multiple olfactory robots is realized by utilizing the characteristics of the RFID positioning system, the motion trail of the olfactory robots is controlled, and the requirement of quick positioning in practice is met.

Drawings

In order to make the object, technical scheme and beneficial effect of the invention more clear, the invention provides the following drawings for explanation:

fig. 1 is a flowchart of the method for locating odor sources by using multi-olfactory robot cooperation.

Fig. 2 is a schematic view of the spiral motion of the outward expansion during the odor source discovery phase.

FIG. 3 is a schematic diagram of the odor source tracking phase-s-shaped segmented sinusoidal motion.

Fig. 4 is a schematic diagram of the spiral motion of the inward contraction during the scent source confirmation phase.

Detailed Description

As shown in fig. 1, the invention provides a method for positioning an odor source by cooperation of multiple olfactory robots. Firstly, in an odor source discovery stage, a plurality of olfactory robots are randomly distributed at different positions of a two-dimensional search plane, and then all olfactory robots perform outward expanding spiral motion; once the gas concentration measured by any three olfactory robots is greater than a low threshold value, the three olfactory robots enter an odor source tracking stage and start to do s-shaped segmented sinusoidal motion towards the direction of an odor source. Based on a BP neural network and an RFID positioning system, the BP neural network dynamically adjusts a weight by utilizing periodic gradient information to control the motion slope of each period of the olfactory robot; the RFID positioning system dynamically modifies the amplitude and angular frequency of sinusoidal motion in each period according to the gas concentration difference by transmitting the gas concentrations measured by a plurality of olfactory robots. Once the gas concentration measured by any olfactory robot is greater than the high threshold, the olfactory robot enters an odor source confirmation stage and performs odor source positioning by using a statistical method.

The odor source and the sensor configured by the olfactory robot have a corresponding relationship, namely, under the stimulation of the odor source, the sensor can generate a response with certain intensity, wherein the range of the gas concentration value measured by the sensor is assumed to be 256 levels, the range corresponds to 16-system 0X 00-0 XFF, and the low threshold value and the high threshold value of the odor concentration are respectively set to be 0X20 and 0X 80. The patent specifically comprises the following steps:

step 1: as shown in fig. 2, in the odor source discovery phase, n olfactory robots (each identified as a robot l (l ═ 1,2,3.., n)) are randomly distributed at any initial position of the two-dimensional search plane, and the robots l (l ═ 1,2,3.., n) start to make spiral motions with increasing radii by using the respective initial positions as origins, as shown in equation (1). And the gas concentration at that location is measured at intervals at.

In the formula, t ═ b × Δ t (b ═ 1,2, 3.) denotes the current measurement time; (x)_l(t),y_l(t)) represents the corresponding coordinates of the position of the robot l at time t.

Step 2: in the stage of odor source discovery, once any three olfactory robots are at time t₁The measured gas concentrations are all greater than a low threshold γ (γ is set to 0X20 by default), and are labeled as Robot in order of the order in which the low threshold condition is satisfied₁、Robot₂、Robot₃Then the three olfactory robots immediately enter an odor source tracking stage; and the rest olfactory robots stop moving and exit the positioning process of the odor source.

And step 3: in the odor source tracking phase, in order to get close to the odor source quickly, the olfactory Robot Robot_β(β ═ 1,2,3) at time t₁Stopping the outward expanding spiral motion, starting the s-shaped segmental sinusoidal motion (sinusoidal motion for short), and stopping the time t₁As the initial direction of the sinusoidal motion. Setting the cycle of sinusoidal motion to be 2 pi, the fixed motion time of each cycle to be T, and the olfactory Robot Robot in the k-th cycle (k is 1,2, 3.)_βThe sinusoidal motion curve of (β ═ 1,2,3) is shown in formula (2).

In the formula, x_β′,y_β' indicating olfactory Robot Robot_βTwo-dimensional coordinates relative to the initial time of each cycle; t' represents the relative time within each cycle; a. the_β(k),ω_β(k) Respectively represent the k cycle olfactory Robot_βAmplitude and angular frequency of sinusoidal motion, where A_β(k),ω_β(k) Keeping constant value in each period while

ω_β(k)×x_β′∈[2(k-1)π,2kπ)。

Considering omega_β(k)x_β' 2k pi-pi is the symmetric center point of the kth period, and has certain representativeness to the average gas concentration of the period, and the kth period olfactory Robot Robot is assumed_β(β ═ 1,2,3) at ω_β(k)×x_βThe gas concentration measured when is 2k pi-pi is respectively PL_β(β ═ 1,2, 3). According to the RFID positioning algorithm, the gas concentration presents the characteristic of logarithmic form attenuation along with the increase of the distance, and the Robot is a smell Robot₃The measured gas concentration is used as a reference to obtain the k cycle olfactory Robot Robot_βThe relative relationship between (β ═ 1,2,3) and the distance between odor sources is shown in formula (3).

In the formula (d)₁,d₂,d₃Respectively represent the k cycle olfactory Robot_β(β ═ 1,2,3) actual distance to the odor source; ζ represents a path loss coefficient, and is usually set to 2 to 5.

According to the formula (3), without loss of generality, Robot with a Robot₁And Robot₃The distance ratio therebetween is calculated as shown in equation (4), for example.

Robot Robot_β(beta. 1,2,3) to odorThe difference in source distance will determine the amplitude A in equation (2)_β(k) And angular frequency ω_β(k) The difference in (a). When the olfactory robot is far away from the odor source, the gas concentration is low, sinusoidal motion with large amplitude and small angular frequency is required, and the olfactory robot can be quickly close to the odor source and collect gas concentration data in a large range; when the olfactory robot is close to the position of the odor source, the gas concentration is higher, and a sinusoidal motion with smaller amplitude and larger angular frequency is required, so that the gas concentration can be collected in a small range near the odor source for the next odor source confirmation.

At the end of the k-th cycle, the olfactory robot will adjust the amplitude A of the k + 1-th cycle according to the measured gas concentration_β(k +1), angular frequency ω_β(k +1) and an initial movement direction; when k is 1, the olfactory robot adjusts the gas concentration measured at the time when the period t' is 0. Wherein the amplitude A_β(k +1), angular frequency ω_β(k +1) realizing dynamic adjustment by using a sigmoid activation function, and specifically carrying out the dynamic adjustment according to the following rules: robot for olfaction₃As a reference, the amplitude A of the reference is set₃(k +1) and angular frequency ω₃(k +1) is a fixed value epsilon and rho, and the k +1 th cycle olfactory Robot Robot is obtained₁And Robot₂The amplitude and angular frequency of (d) are shown in equations (5) and (6).

And 4, step 4: as shown in fig. 3, in the odor source tracking phase, the adjustment of the moving direction of the olfactory robot is performed according to the following rules: robot with sense of smell₁For example, the olfactory robot in the k cycleMeasuring the gas concentration asThe concentration values are used as the input of a BP neural network, so that the number of nodes of the input layer of the BP neural network is 4, and v is used as_p(p is 1,2,3, 4); the output layer has the kth period omega_β(k)×x_βActual scent source predicted direction angle slope of' 2k piThe corresponding node number is 1, and the node number of the hidden layer is set to beWherein m represents the number of nodes of the input layer, m' represents the number of nodes of the output layer, a is an integer between 1 and 10, and the default is an integer 3, so that the number of nodes of the hidden layer is 5 by default. Considering that the traditional BP neural network always has the problem of low convergence speed, and the optimization of the weight can effectively improve the training speed of the neural network, so that the adjustment speed can be still accelerated when the variation of the weight is small by utilizing the periodic gradient information to dynamically adjust the weight, and the damping effect is realized when the weight is greatly changed.

Firstly, considering that the olfactory robot has large average difference of gas concentration measured in different periods, which can cause large influence on the training of BP neural network, the input volume v of the k period is used_p(k) Normalizing to obtain the input V of the BP neural network_p(k) As shown in formula (7).

Wherein, min (v)_p(k) Represents an input node v_p(p ═ 1,2,3,4) minimum; max (v)_p(k) Represents an input node v_p(p is 1,2,3, 4).

The desired odor source predicted direction angle slope isThe slope is the kth period omega_β(k)×x_β' -2 k piAnd (3) setting the slope of an included angle between the line and the odor source to obtain a loss function E (k) of the kth period, as shown in a formula (8).

In the formula (I), the compound is shown in the specification,representing the actual scent source prediction direction angle slope.

Then the weight value change quantity delta w of the k-1 th period_i,j(k-1) is applied to weight adjustment in the kth cycle (not considered when k is 1), and a weight adjustment value Δ w for the kth cycle is obtained_i,j(k) As shown in formula (9).

In the formula, η represents a learning rate;respectively representing the gradient information of the k-1 th and k-th periods, w_i,j ^γ(k) Representing the weight value from a node i of a gamma layer to a node j of a next layer in the k-th periodic neural network, wherein gamma-1 is represented as an input layer, and gamma-2 is represented as an implied layer; -0.5 < lambda^-＜0＜λ⁺< 0.5 indicates an adjustment coefficient.

Then, updating the weight value to obtain the weight value w of the (k +1) th period_i,j(k+1)。

w_i,j(k+1)＝w_i,j(k)+Δw_i,j(k) (10)

And finally, according to the actual output value of the BP neural network in the kth period, obtaining the actual sinusoidal motion curve of the kth period by using the sinusoidal motion curve in the rotation matrix correction formula (2), as shown in a formula (11).

In the formula, X_β,Y_βRobot for representing smell_β(β ═ 1,2,3) relative to the actual two-dimensional coordinates at the initial instant of each cycle.

And 5: as shown in fig. 4, in the odor source tracking phase, once the gas concentration measured by a certain olfactory robot is greater than a high threshold value δ (default is δ being 0X80), the olfactory robot enters the odor source confirmation phase and resets the current cycle to 0, and since the gas concentration at the position of the odor source is the highest and the gas concentration transmitted to the periphery is gradually decreased, the olfactory robot makes an inward-contracting spiral motion near the point where the gas concentration is high, and the rest olfactory robots stop moving temporarily. In the process, the olfactory robot in the r-th period is judgedMeasured gas concentration value v_p(r) is greater than the high threshold δ (confidence 99.7%), as shown in equation (12).

Wherein μ represents a gas concentration value v measured in the r-th cycle_p(r) mean value, σ represents the value of the gas concentration v measured in the r-th cycle_pThe variance of (r).

If the olfactory robot continuously meets the conditions for three periods, the point can be judged as an odor source, and the positioning of the odor source is finished at the moment; when the olfaction Robot can not satisfy the above conditions, the olfaction Robot Robot at this time_β(β ═ 1,2,3) will return to step 3 and continue sinusoidal motion.