Robot fish bionic control method and system integrating Spiking neural network and CPG

[0044] Example 1

[0045] This embodiment discloses a robotic fish bionic control method fusing Spiking neural network and CPG, imitating the biological motion mechanism to design the robotic fish motion control system, the Spiking neural network is used as the upper controller to process environmental information and generate decision commands; The saturation function enables the CPG model to have an input function; the CPG neuron as a lower-level controller receives spiking excitation signals and outputs control commands. This disclosure proposes the possibility of a robotic fish underwater to perceive the environment and move autonomously.

[0046] Step 1: Spiking neural network modeling

[0047] Neurons are the basic structural components of the brain and are also the most basic unit of the Spiking neural network, which mainly processes impulse signals. This disclosure uses the Izhikevich neuron model as the upper controller of the hierarchical control system, which has sufficient functions to form a CPG network, which is expressed as follows:

[0048]

[0049] Among them, v is the membrane voltage of the neuron, u is the voltage recovery variable, and I is the external input signal.

[0050] The conditions for resetting spike neurons are as follows:

[0051]

[0052] Among them, a, b, c, d in formulas (1) and (2) are constants. The meanings of these four parameters are: a represents the time scale parameter of the recovery variable u; represents the dependence of neurons on membrane voltage; c and d represent the potential reset value after u and v reach a peak value.

[0053] After the peak reaches 30 mv at its apex, the membrane voltage and recovery variables are reset according to equation (2). If v exceeds 30mv, then it is first reset to 30mv so that all spikes have equal amplitude. The model can display all known neuron patterns by changing the parameters a, b, c, and d. +30mv is not a threshold, but a spike of impulse neurons. The threshold of the neuron model is 70mv and 50mv, and it is dynamic. figure 2 The simulation simulates the firing behavior of 20 biological neurons.

[0054] Using unsupervised learning method, an unsupervised algorithm based on Hebb learning rules is proposed. The synaptic connections between neurons are the neural basis of learning and memory. The connection strength of the synapse depends on the degree of stimulation acting on the neurons at both ends of the synapse. This change in the strength of the synaptic connection is called synaptic plasticity. The pulse emitted by the neuron has a continuous effect on the synapse and satisfies the mechanism of Spike Time Dependent Plasticity (STDP). The exact form of STDP varies between different types of synapses. The sequence can be divided into long-term potentiation (LTP), which increases the synaptic weight, and long-term inhibition (LTD). The mathematical model can be expressed as formula (3), and W is used to represent the change in synaptic weight.

[0055]

[0056] Which takes A + =1,A - =-1,τ=8ms, the waveform of STDP is as image 3 Shown. LTP and LDP time are not symmetrical, that is, t> At 0, the long-term synaptic plasticity increases, and the synaptic weight of two consecutive neurons increases; on the contrary, when t When it is less than 0, the synaptic plasticity is inhibited for a long time, and the weight of the synapse connecting the two neurons becomes smaller. This asymmetry can prevent the instability of neuronal signal transmission caused by constant enhancement or constant inhibition.

[0057] Step 2. Establish the dynamic model of robotic fish

[0058] The robotic fish used in the present disclosure has four joints and a pair of pectoral fins, and can perform reciprocating swing motions. The CPG control structure of robotic fish is like Figure 4 Shown.

[0059] The CPG signal is used to drive the swing of each joint, and the phase oscillator model is used, which is expressed as formula (4):

[0060]

[0061] Where θ i And r i Is the oscillator state variable, representing the phase and amplitude respectively; f i And R i Determine the internal frequency and amplitude of the oscillator; τ i Is the normal amount, decide r i Converges to R i The speed; the mutual coupling relationship between oscillators is determined by the weight w ij And phase difference It is decided that i represents the number of CPGs used in the hierarchical control system of the present disclosure, and the value of i is 1,...,n.

[0062] In this embodiment, the robotic fish is controlled by a steering gear, and the action of the steering gear is driven by the output signal of the oscillator; a nonlinear oscillator model is established for a robotic fish with four joints and a pair of pectoral fins to control the robotic fish to swim and swim directly. Turning movement.

[0063] Use the above nonlinear oscillator model as the CPG neuron, the input of the CPG neuron is the excitation, and it is divided into left excitation d l And right incentive d r , Get the oscillator parameter f after saturation function i And A i , Respectively drive the left body and the right body; define the behavior under the phase coupling relationship of the model φ d , On the behavior φ u.

[0064] This disclosure determines the parameters of the robot fish when it swims straight and when it turns, such as the upward and downward phase coupling coefficients, the length of the tail fin, the left and right excitation values ​​of the CPG frequency corresponding to each joint, etc. The control flow chart of the CPG mathematical model is as follows Figure 5 Shown.

[0065] Step 2.1: Define the key parameters of the CPG model and perform robotic fish control simulation

[0066] This disclosure uses MATLAB software to perform simulation on this platform. Set the drive signal d to drive both sides of the robotic fish. The drive signal d increases from 0 to 5. The other parameters of the CPG model are shown in Table 1. The joints and caudal fin neurons of the robotic fish oscillate sequentially, and the frequency and amplitude of the neurons follow As the driving signal increases, the control signal of each joint servo motor is obtained, and each shutdown is controlled by two neurons with a phase difference of 180 degrees.

[0067] Table 1 Parameters of CPG Shenjiangyuan Mathematical Model

[0068] parameter Joint 1 Joint 2 Joint 3 Caudal fin pectoral fins n 2 2 2 2 2 τ 20 20 20 20 20 dl

[0069] Step 3. Saturation function design

[0070] The saturation function is designed as the input function of the CPG lower-level controller. It is divided into left and right parts. After the output excitation signal of the Spiking upper-level controller passes the saturation function, the CPG lower-level controller obtains the oscillator parameters. There are two input signals for driving One of the left and right sides of the bionic robotic fish, the saturation function can be expressed as follows:

[0071]

[0072]

[0073] Where k 1 ,k 2 ,b 1 ,b 2 Represents constant coefficient, d l ,d r Respectively represent the excitation on the left and right sides, d low Indicates the minimum excitation value.

[0074] The saturation function in equation (5) is used for the tail fin and flexible body of the robotic fish, and the saturation function in (6) is used for the pectoral fin of the robotic fish.

[0075] When the excitation stimulus is less than the minimum excitation, the oscillator stops oscillating. With the gradual enhancement of the excitation signal, the oscillation frequency increases and the amplitude gradually increases; the oscillator gradually oscillates. When the excitation signal exceeds the maximum value, the oscillator will not stop and will continue to drive the bionic robotic fish with the highest frequency and highest amplitude.

[0076] Step 4: Spiking neural network drives the basic control flow chart design of the CPG unit

[0077] After the above steps, the modeling and design of the controllers at all levels of the hierarchical control system are completed. The upper controller adopts the Izhikevich neuron model. By adjusting the values ​​of parameters a, b, c, and d, it can simulate the regular burst, internal burst, and Quickly read and release different discharge states, according to this characteristic, select the discharge state with different characteristics as the input of the CPG mathematical model. The basic control flow chart of the Spiking neural network driving the CPG unit is as follows Figure 5 Shown.

[0078] Spiking neural network is used as the upper-level controller of the hierarchical structure of the bionic control system, processing sensors and environmental information, making decisions on environmental changes and generating excitation signals;

[0079] CPG is the lower-level controller of the hierarchical structure of the bionic control system. It receives the excitation signal generated by the Spiking neural network and outputs control commands to drive the motion of each joint of the bionic robotic fish;

[0080] The saturation function is used as the connection interaction between the upper and lower controllers, so that the CPG has an enhanced input interface and can receive the excitation signal output by the upper controller of the Spiking neural network;

[0081] Determine the control process of Spiking neural network to drive CPG neurons: Based on the adopted Izhikevich neuron model, by adjusting key parameters, it simulates the regular burst, internal burst, fast reading and discharge of neurons and other different characteristics of the discharge state to simulate environmental changes, and select different characteristics The discharge state is used as the input of the CPG model. The motion control of the robotic fish is simulated in the matlab environment, and the excitation signal output by the Spiking neural network and the control signal of each joint of the robotic fish are obtained.

[0082] Carry out spiking neural network drive CPG control simulation

[0083] This disclosure uses MATLAB simulation software to set a set of key parameters a, b, c, and d of the Spiking neural network Izhikevich neuron model as external excitation signals to drive the CPG mathematical model of the bionic robotic fish, and simulate the direct swimming and turning state of the bionic robotic fish . The parameter settings are shown in Table 2. Set the state parameters of the upper-level controller Spiking neural network and obtain the output excitation signal. Through the spiking neural network, a 0/1 pulse sequence can be obtained, which is input to the lower-level controller CPG as the driving signal d The unit can obtain the output result of the CPG unit after the input pulse sequence driving signal, and generate the control signal of all the joints of the robotic fish when driven by the Spiking neural network.

[0084] Table 2 Spiking neural network parameters

[0085] parameter a b c d I Set value 1.0 1.5 -60 0 -65

[0086] Example 2

[0087] The present disclosure provides a robotic fish bionic control system integrating Spiking neural network and CPG, including:

[0088] The CPG model building module is used to model the dynamics of a robotic fish with four joints and pectoral fins, and use the nonlinear oscillator model as the CPG neuron to determine the input excitation on the left and right sides, the downward and upward phase coupling coefficients, and the upward and downward coupling Coefficient weight, corresponding to the CPG frequency of each joint;

[0089] The Spiking neural network model building module is used to determine the Izhikevich neuron model, set various parameters to simulate different discharge states, use an unsupervised algorithm based on Hebb learning rules to conduct neural network training for different discharge states, and send the trained data To CPG neuron, as the input signal of CPG neuron, drive CPG to output bionic robotic fish control signal;

[0090] Spiking neural network and CPG hierarchical control model establishment module, which is used for piking neural network model as the upper-level controller, CPG model as the lower-level controller, design saturation function to connect the upper and lower-level controllers, CPG model receives Spiking neural network through saturation function The excitation signal generated by the model and the output control signal drive the movement of each joint of the bionic robotic fish.

[0091] The saturation function is expressed as follows:

[0092]

[0093]

[0094] Where k 1 ,k 2 ,b 1 ,b 2 Represents constant coefficient, d l ,d r Indicates the excitation on the left and right sides respectively, d low Indicates the minimum excitation value.