Digital TV Network Management System Based on WEB

With the development and popularization of computer artificial intelligence technology, more and more intelligent machines are gradually produced. These intelligent machines have brought great convenience to people’s lives. This paper studies the control method of snake robot based on environment adaptability, which mainly explains the construction and stability of multi-modal CPG model. In addition, this paper also studies the trajectory tracking and dynamic obstacle avoidance of mobile robot based on deep learning.


Introduction
With the emergence of intelligent robots, people's lives have become more intelligent. Although traditional robots have brought great convenience to the people, their skills are fixed.Nowadays, researchers combine deep learning with intelligent robots, which not only allows robots to have original fixed functions, but also have the ability to learn independently.This will surely bring extremely high learning efficiency and improve the extremely fast robot working accuracy, and this also omit tedious programming [1].

Construction of multi-modal CPG model
Biological principles of multimodal CPG model. Snake's long and narrow skeleton, soft body and various movements make it highly adaptable to the environment.It can choose different motions in different environments. Researchers have developed a snake-shaped intelligent machine based on the characteristics of snakes. This kind of machine retains the characteristics of snakes well. It can slide, roll sideways and change waveforms.In order to improve the environment of the snake-like robot, it also enables its CPG control model to be displayed in a multi-state form [2]. The biological principle of the multi-modal CPG model is shown in Figure 1.
Equation (1) represents the phase coupling relationship between neurons, equation (2) represents the amplitude of the neural output, and equation (3) is the final output of the neuron [3].
The snake-shaped machine is composed of multiple joints, each single joint is a neuron group, and a neuron group is composed of two neurons, and the shape is shown in Figure 2.

Figure 2. Single joint CPG model
Since the serpentine machine contains a variety of movements, the serpentine machine is controlled by the neuron group of the horizontal joints when performing horizontal movements.It is controlled by the neuron group of the vertical joint during vertical movement. The connection of different movement groups of the serpentine machine uses an appropriate number of nodes to construct a CPG model.This model can produce a variety of output modes by adjusting external excitation, so it is called a multi-modal CPG model [4]. The structure of the multi-modal CPG model is shown in Figure  3.

Stability of the multi-modal CPG model
Whether it is the movement of an animal or the movement of a biomimetic robot, it needs to be controlled by responsive joints. Of course, the most indispensable part of this process is to stabilize the joints.The snake-shaped machine has limbs different from other bionic robots, so this increases the complexity of the CPG model to a certain extent, and most of the relevant researchers only show the most basic stability of the snake-shaped machine. That is, they only proved the stability of a certain joint of the serpentine machine, which is theoretically incomplete.Based on the structural characteristics of the serpentine robot, this paper proves the stability of the multi-modal CPG model when the number of connections is random, which guarantees the theoretically smooth and stable performance of the system [5].

Serpentine robot trajectory tracking and dynamic obstacle avoidance based on deep learning
The system structure of the snake robot trajectory tracking and dynamic obstacle avoidance algorithm based on deep learning is shown in Figure 4 [6].

Data preprocessing
When the snake-shaped robot obtains the motion trajectory and motion obstacle avoidance, it mainly uses the neural network to obtain the image data of the robot's perception. In this way, even if the image data becomes smaller, larger or displaced, the invariance of the position characteristics between the robot and the obstacle can be maintained.When the robot recognizes obstacles, it involves the study of the scene recognition system, as shown in Figure 5 is the organizational structure of the recognition system [7].

Strategy approximator selection
The approximator of the strategy chooses to use a deep convolutional neural network, and the optimization objective function of the model is formula (4) [8].
The update of the weight parameters of the convolutional neural network model uses a small batch data stochastic gradient descent algorithm, as shown in formula (5).
The neural network model is shown in Figure 6.

Model training
When training a neural network, it is assumed that the training data is independent and evenly distributed, and there is a correlation between the data collected from the environment [9,10]. The network model training process is shown in Figure 7.