CN116346636A - Neural network sliding mode control-based network control system time delay compensation method - Google Patents
Neural network sliding mode control-based network control system time delay compensation method Download PDFInfo
- Publication number
- CN116346636A CN116346636A CN202310005861.8A CN202310005861A CN116346636A CN 116346636 A CN116346636 A CN 116346636A CN 202310005861 A CN202310005861 A CN 202310005861A CN 116346636 A CN116346636 A CN 116346636A
- Authority
- CN
- China
- Prior art keywords
- network
- sliding mode
- time delay
- optimal
- value
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
- H04L41/142—Network analysis or design using statistical or mathematical methods
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
- H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
- H04L41/147—Network analysis or design for predicting network behaviour
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/16—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks using machine learning or artificial intelligence
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/02—Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]
Abstract
The invention relates to a network control system time delay compensation method based on neural network sliding mode control, which comprises the following steps: establishing an echo state network; updating the echo state network training; introducing a mucor algorithm to optimize an echo state network model, and forming a combined prediction model prediction network delay value of the echo state network optimized based on the mucor algorithm; and designing a sliding mode function according to the network time delay value to obtain a state equation when the system is positioned on a sliding mode surface, and solving the control quantity of the system. The invention solves the control problem of the linear system with input delay, effectively improves the prediction precision, outputs the future control quantity by combining the predicted result with the sliding mode control algorithm, compensates the time-varying delay of the network control system and improves the tracking capability of the network control system signal.
Description
Technical Field
The invention belongs to the technical field of network control systems, and particularly relates to a time delay compensation method of a network control system based on SMA (slime mold algorithm, SMA) -ESN (echo state network, ESN) neural network sliding mode control in a network environment.
Background
The network control system (networked control system, NCS) NCS is a novel control system which is developed after integrating network communication technology, computer information technology and automatic control technology, and has the advantages of good real-time performance, easy expansion and maintenance, high reliability and the like. In the design work of a network control system, a lot of information sources are needed, and a network communication line is used in the process of data transmission, so that certain network delay is generated in the process of information transmission, the delay can cause the loss and the timing disorder of data packets in the transmitted data, and the problems can bring serious obstruction to the use and popularization of the network control system.
The sliding mode variable structure control (sliding mode control, SMC) is robust to system model structural uncertainty and external disturbances, which makes SMC an ideal choice for NCS design. NIZARA et al propose using predictive sliding mode variable structure control (prediction sliding mode control, PSMC) for time delay systems, and for linear systems with state delay, an enhanced predictive discrete time sliding mode control with sliding function is proposed, which shows that PSMC has a better effect on systems with time lags, but in some practical applications, due to the problem of input delay, system buffeting occurs when SMC method is implemented, and in addition, the buffeting problem is aggravated by data packet loss and wrong sequence.
In summary, the existing algorithm has comprehensive researches on the linear system with state delay, and relatively less researches on input delay, so that the method can effectively compensate the input delay problem in the system.
Disclosure of Invention
The invention provides a neural network sliding mode control-based Network Control System (NCS) time delay compensation method, which aims to solve the control problem of a linear system with input delay so as to achieve the purposes of accurately predicting and effectively compensating the time delay of the network control system and having good development prospect.
In order to achieve the above purpose, the technical scheme of the invention is as follows:
a network control system time delay compensation method based on neural network sliding mode control comprises the following steps:
step one, an echo state network is established;
step two, echo state network training and updating;
step three, introducing a mucor algorithm to optimize an echo state network model, and forming a combined prediction model prediction network delay value tau (t) of the echo state network optimized based on the mucor algorithm;
and fourthly, designing a sliding mode function according to the network delay value tau (t), obtaining a state equation when the system is positioned on a sliding mode surface, and solving the control quantity of the system.
Furthermore, in the first step, an input unit u (t) of the echo state network is time delay data of the past moment, and an input layer is provided with K nodes; the reserve pool x (t) is in a current moment state, and N nerve nodes are arranged in the reserve pool; the output unit y (t) is a network delay value output at the next moment, and the network has L output nodes; the states of the input unit, the reserve tank and the output unit are as follows:
u(t)＝[u 1 (t),u 2 (t),...u K (t)] T
x(t)＝[x 1 (t),x 2 (t),...x N (t)] T
y(t)＝[y 1 (t),y 2 (t),...y L (t)] T 。
further, in the step one, the output state equation of the echo state network is:
y(t+1)＝f out ·(W out ·(u(t+1),x(t+1))，
where fout is the output layer neuron activation function, W out The connection weight of the reserve tank to the output layer is the output of the next moment of y (t+1), the control quantity of the next moment of u (t+1) and the input of the next moment of x (t+1).
Further, the optimization process of the echo state network optimized based on the mucosae algorithm in the third step is as follows:
(1) Initializing parameters related to a colistin algorithm, initializing a population quantity parameter N, the maximum iteration times maxT and a parameter Z in a population position updating formula;
(2) Calculating the fitness value of each population by applying a fitness function;
(3) Applying a mucositis individual position update model to the population position, the optimal fitness value and the optimal population position information X b Updating; if the situation that the global optimal fitness value is higher than the optimal fitness value occurs in the iteration process, replacing the global optimal fitness value with the fitness value obtained by iteration, and updating the optimal position information of the population to the current position information;
(4) After updating the position information, calculating the fitness of each population, and updating the global optimal position of the population;
(5) Judging whether the maximum iteration times are reached, if so, outputting a reserve pool parameter corresponding to the optimal individual position information, establishing an echo state network prediction model optimized based on a myxobacteria algorithm by applying the combination of the two optimal parameters, and then predicting test set data to output y (t), namely outputting a y (t) namely a network delay value tau (t); if the maxT is not reached, repeating the steps (2) - (4) to perform the parameter continuous optimizing operation.
Further, the step of searching for an optimal fitness value in the step (3):
firstly, the approach behavior of the coliform bacteria is mathematically modeled, and the following rules are proposed to simulate the shrinkage mode of the coliform bacteria:
wherein vb is [ -a, a]Random numbers in between, W represents weight coefficient of coliform bacteria, X A (t) and X B (t) is two random individual positions, vc is at [ -1,1 [ -1 ]]Parameters of oscillation and finally approaching zero, X b (t) represents the current fitness-optimal individual position, and X (t) represents the current mucositis individual position;
next, the update formulas of the control parameter p, the parameter vb, the parameter a and the weight coefficient W are as follows:
p＝tanh|S(i)-DF|
wherein i epsilon 1,2, …, n, S (i) represents the fitness value of the ith mucosae, DF is the optimal fitness value obtained currently;
vb＝[-a,a]
SmellIndex＝sort(S)
the condition represents that fitness is ranked in the first half of individuals in the population, other represents the rest individuals, r represents a random number between [0,1], bF represents the best fitness value obtained by the current iteration, wF represents the worst fitness value of the current iteration, T represents the current iteration times, maxT represents the maximum iteration times, smellIndex represents the ordering condition of the fitness sequence, smellIndex, bF, wF is obtained, and finally W is calculated;
finally, in order to find the optimal fitness value, that is, when the evaluation index approaches 0, the performance evaluation of the proposed prediction model is performed by considering the root mean square error evaluation index, and the root mean square error is set as
m represents the total length of the time series, i is expressed as the number of iterations, y k And
the actual value of the time series at time k and the output predicted value of the ESN model are respectively represented.
Further, the location update model of the colistin individual in the step (3) is as follows:
wherein UB and LB are upper and lower bounds, rand is a random number uniformly distributed between 0 and 1, z is a self-defined parameter, the value of vb randomly oscillates between [ -a, a ], the value of vc oscillates between [ -1,1] along with the increase of iteration times, and finally approaches zero, and the calculation formula is as follows:
vc＝[-b,b]
b＝1-T/maxT
the synergy between vb and vc mimics the selection behavior of slime.
Further, the sliding mode function in the fourth step is as follows:
S(t)＝Cx(t)-∫C(A-BK)x(t)dt+∫CB[u(t)-u(t-τ(t))]dt
wherein, C is a constant matrix with proper dimension, and can ensure CB to be a nonsingular matrix, K is a parameter to be designed, t is time, A is an m×m order matrix, B is an m×n order column full order matrix, τ (t) is a network delay value, and u (t) is a control input.
Further, in the fourth step, the system control amount is:
u(t)＝-Kx(t)-(CB) -1 (ηS+εsgnS)
wherein, take eta=1, epsilon=0.005, S is the sliding mode function;
the equation of state when the system is at the slip plane can be described as:
by adopting the technical scheme, the invention can produce the following technical effects:
the prediction model of time-varying time delay in NCS is established, the SMA-ESN algorithm solves the control problem of the linear system with input delay, the prediction precision is effectively improved, the predicted result is used for outputting future control quantity in combination with the sliding mode control algorithm, the time-varying time delay of the network control system is compensated, and the tracking capability of the network control system signal is improved. The invention has the characteristics of simplicity, high stability, simple operation, strong portability, low cost and the like, and can be applied to engineering practice.
Drawings
FIG. 1 is a comparison of time delay predictions for an SMA-ESN neural network prediction algorithm and an ESN prediction algorithm;
FIG. 2 is a diagram of the experimental results of a generic slip-form controller method;
FIG. 3 is a graph of experimental results of the design of the present method;
FIG. 4 shows the control rate obtained by a common slip-form controller;
FIG. 5 is a control rate obtained by the controller designed by the method;
FIG. 6 is an evolution of the slip form face of a conventional slip form controller;
FIG. 7 is an evolution of the slip plane of the controller of the present method;
FIG. 8 is a block diagram of a network control system of the present method;
FIG. 9 is a schematic diagram of an NCS network delay compensation scheme based on neural network sliding mode control of the present invention;
fig. 10 is an echo state network schematic diagram.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.
According to the invention, according to the actual measured original network time delay data, an echo state network (SMA-ESN) algorithm based on a mucosae algorithm is established, the mucosae algorithm (slime mold algorithm, SMA) is set to iterate the reserve pool parameters of the echo state network, and the optimal solution obtained by iteration is assigned to the echo state network. Here SMA updates the population position. And calculating the fitness value, updating the global optimal position, judging whether the end condition is reached, outputting an optimal result if the end condition is reached, and otherwise repeating the steps. After the weight of the echo state network is optimized by the myxobacteria algorithm, training is carried out by utilizing a data set of the network time delay measured before, and then a trained model is used for time delay prediction; the network time delay prediction model predicts and obtains the network time-varying time delay at the next moment according to an SMA-ESN algorithm:
as shown in fig. 9, the invention provides an NCS network delay compensation method based on neural network sliding mode control, which comprises the following steps:
step one, establishing an Echo State Network (ESN);
fig. 10 shows the structure of an echo state network. The input u (t) of the network is time delay data of the previous moment, and the output y (t) is time delay data of the next moment. In the figure, u (t) represents the input of the moment t in the network, the input layer has K nodes, x (t) represents the current moment state of the storage pool, N neural nodes are arranged in the storage pool, y (t) represents the output of the echo state network, and L output nodes are arranged in the network. When the system is provided with K input units, N internal processing units and L output units, the states of the input units, the internal states of the storage pool and the output units are as follows:
u(t)＝[u 1 (t),u 2 (t),...u K (t)] T
x(t)＝[x 1 (t),x 2 (t),...x N (t)] T
y(t)＝[y 1 (t),y 2 (t),...y L (t)] T
the reserve pool in the figure is replaced by an implicit layer structure in a conventional neural network, and the connection weight of an input layer to the reserve pool is W in Is (N.times.K) order matrix, the internal connection weight from the reserve pool to the state of the reserve pool at the next moment is W, is (N.times.N) order matrix, and the connection weight from the reserve pool to the output layer is W out Is an L (K+N+L) order matrix, and there is a connection W from the output layer at the previous moment to the reserve pool at the next moment back Is an (N x L) order matrix, but this connection is not required.
At each moment, u (t) is input, the storage pool is updated with the state update equation:
x(t+1)＝f(W in ·u(t+1)+W·x(t)+W back ·y(t))
in which W is in And W is back Are randomly initialized and fixed at the time of initial network establishment.
u (t+1) is the input at this time, x (t+1) is the pool state at this time, and x (t) is the pool state at the last time. f is the DR internal neuron activation function.
The output state equation for ESN is: y (t+1) =f out ·(W out (u (t+1), x (t+1)), where fout is the output layer neuron activation function, W out Is an output weight matrix.
The final performance of the ESN is determined by the individual parameters of the pool, including: the internal connection weight spectrum radius SR of the reserve tank, the reserve tank scale N, the reserve tank input unit scale IS and the reserve tank sparseness SD.
The reservoir spectrum radius SR is the absolute value of the maximum eigenvalue of the reservoir internal connection weight matrix W. The spectral radius of the echo state network reservoir affects the memory capacity of the echo state network. This value may determine the memory capacity of the reservoir. Therefore, the spectral radius selection of the reservoir is very critical, and it is generally considered that the ESN has a stable echo state property when 0< sr < 1. When the spectrum radius is selected too small, the memory capacity of ESN is poor, the nonlinear processing capacity is weak, but when the spectrum radius is selected too large, the stability of the state of the reserve pool is affected, and the prediction capacity of the echo state network is further affected.
Pool size N, which is the number of neurons in the pool, the pool size choice is related to the number of samples, which has a significant impact on network performance, the larger the pool size, the more accurate the ESN describes a given dynamic system, but which can lead to overfitting problems.
The pool input unit scale IS, which IS a scale factor by which the input signal of the pool needs to be multiplied before being connected to neurons inside the pool, IS a scaling of the input signal. The more nonlinear the object that generally needs to be processed, the greater IS.
Reservoir sparsity SD, which represents the connection between neurons in a reservoir, is not all of which are connected. SD represents the percentage of the total number of interconnected neurons in the reservoir to the total number of neurons N, with greater values being more non-linear approximation capable.
N is the number of interconnected neurons, and N is the total number of neurons.
Step two, echo state network training and updating;
the training process of the ESN is given by the parameters of the ESN network described above.
1. Initializing an ESN network
First initialize W in ESN in 、W、W back And an initial state x (0) of the reserve tank. Wherein W is in ，W back And for the randomly generated connection matrix, W is a randomly generated sparse matrix, and the spectrum radius is smaller than 1. The pool has no initial state prior to network training.
2. Updating a pool state matrix
Inputting network training samples (u (1), yt (1)), … (u (n), yt (n)), and updating the system internal state information according to the state of the storage pool to obtain a formula x (t+1) =f (W) in ·u(t+1)+W·x(t)+W back Y (t)) is updated. At each update, x (t) and y (t) represent the current state of the pool and the outputs of the network x (t+1) and u (t+1) correspond to the next state of the network pool and the input of the next network. In the network learning process, W in 、W、W back All remain unchanged.
3. Collecting state vectors
In the process of updating the state in the reserve tank, the data of the previous stage is selected to be discarded in order to remove the influence of the initial state of the reserve tank, and the data of the previous stage is selected to be removed from the Mth 0 Starting recording of the internal state x (i) of the reservoir and the input samples u (i) at the moment, forming a state vector
At this time, the system output does not perform ESN output equation
y(t+1)＝f out ·(W out The operation of (u (t+1), x (t+1)), at which stage the information in the output nodes is simply an overlay of training data, the corresponding training samples constituting the state vector as an output matrix
1, then the state matrix S is t× (n+k), D is the state vector of dimension t×l.
4. Calculating an output weight matrix Wout
The output connection weight matrix Wout needs to be calculated by the internal state matrix x (n) in the system and the sample data ((u (1), yt (1)), … (u (n), yt (n))). After the output weight matrix Wout is obtained, the actual output y (n) in the network is utilized to approach the expected output yt (n), namely:
i.e. the matrix of desired weights W out The minimum system mean square error E (y, yt) is reached:
Step three, introducing a mucor algorithm to optimize an echo state network model, and forming a combined prediction model prediction delay tau (t) of an echo state network (SMA-ESN) optimized based on the mucor algorithm;
fig. 8 is a block diagram of a network control system, where the sensor and controller, controller and actuator are all interconnected by a network. There is a time delay τ between the sensor and the controller sc (t) while there is a delay τ between the controller and the actuator ca (t). The total input delay in the network is
Here->
In order to compensate the influence of time delay in a network on a system, a mucosae (SMA) -Echo State Network (ESN) predictor is arranged, the designed SMA-ESN neural network trains the network through the past time delay data, then predicts and obtains a time delay value tau (t) at the next moment, and the tau (t) is designed into a corresponding sliding mode controller. Due to the presence of the time delay,the executor can not necessarily receive the control signal in time, a time delay compensator is arranged, or a buffer is added before the executor, the control signal sequence u (t) of the previous sliding mode controller is stored, so that once the executor does not receive the control signal, the control signal can be selected from the previous control sequence according to the time tag and transmitted to the controlled object.
The core of the ESN training process is to use a dynamic reserve tank to replace the hidden layer of the traditional neural network, wherein the performance of the ESN is mainly determined by each parameter of the reserve tank, the optimal parameter of the current reserve tank cannot be directly deduced through a mathematical formula due to the black box characteristic of the reserve tank, and the prediction performance of the ESN is difficult to be optimized by a randomly initialized reserve tank, so that an echo state network model is optimized by introducing a slime algorithm, and a combined prediction model based on the SMA-ESN is formed.
Initializing weights by an echo state network, mapping neurons in a reserve pool of the echo state network into a myxobacteria individual, mapping the weights of the echo state network into positions of the myxobacteria individual, and updating the positions of the myxobacteria individual through a myxobacteria algorithm; and assigning the updated optimal solution to the echo state network to form an echo state network SMA-ESN optimized based on the mucosae algorithm.
Taking four pool parameters (pool scale N, spectrum radius SR, input contraction factor IS, pool sparsity SD) of the echo state network as objects to be optimized, assuming that coordinates of a point in space accord with four optimal parameters of the pool, then continuously carrying out iterative search by using a slime algorithm to enable the positions of slime individuals to approach to optimal value points infinitely, namely the optimal parameters of the pool, carrying out weight training on the optimized parameters back to the echo state network, and obtaining an output weight matrix, thus finally obtaining a prediction result of the echo state network.
The optimization process of the four parameters of the ESN reservoir by using the SMA is as follows:
(1) Initializing parameters related to an SMA algorithm, initializing a population quantity parameter N, the maximum iteration times maxT and a parameter Z in a population position updating formula;
(2) A fitness function is applied to calculate fitness values for each population.
Firstly, the approach behavior of the coliform bacteria is mathematically modeled, and the following rules are proposed to simulate the shrinkage mode of the coliform bacteria:
wherein vb is [ -a, a]The random number between, vc is at [ -1,1]The parameter of oscillation and finally tending to zero, t is the current iteration number, X b (t) represents the position of the individual with the optimal fitness at present, X (t) represents the position of the current coliform individual, X A (t) and X B (t) two random individual positions, W representing the weight coefficient of the slime.
The update formulas of the control parameter p, the parameter vb, the parameter a and the weight coefficient W are as follows:
p＝tanh|S(i)-DF|
wherein i epsilon 1,2, …, n, S (i) represents the fitness value of the ith mucor individual, and DF is the currently obtained optimal fitness value.
vb＝[-a,a]
SmellIndex＝sort(S)
In the formula, conditions represent that fitness is ranked in the first half of individuals in the population, other represents the rest individuals, r represents a random number between [0,1], bF represents the best fitness value obtained by the current iteration, wF represents the worst fitness value of the current iteration, T represents the current iteration, maxT is the maximum iteration number, smellIndex is the ranking condition of the fitness sequence, smellIndex, bF, wF is obtained, and finally W is calculated.
In order to find the optimal fitness value, that is, when the evaluation index approaches 0, the performance evaluation of the proposed prediction model is carried out by taking the root mean square error evaluation index into consideration, and the root mean square error is set as
m represents the total length of the time series, i is expressed as the number of iterations, y k And
the actual value of the time series at time k and the output predicted value of the ESN model are respectively represented.
(3) Applying a mucositis individual position update model to the population position, the optimal fitness value and the optimal population position information X b Updating; if the situation that the global optimal fitness value is higher than the optimal fitness value occurs in the iteration process, the global optimal fitness value is replaced by the fitness value obtained through iteration, and the optimal position information of the population is updated to be the current position information.
The location update model for the slime individuals is as follows:
wherein UB and LB are upper and lower bounds, rand is a random number uniformly distributed between 0 and 1, z is a self-defined parameter (value is 0.03), and X b (t) represents the position coordinate condition corresponding to the highest odor concentration of the food source, X A (t) and X B (t) represents two randomly defined mucositis position coordinate information. vb has a value of [ -a, a]Random oscillation between them, and gradually approaches zero as the number of iterations increases. The value of vc is [ -1,1]And (3) oscillating, and finally tending to zero, wherein the calculation formula is as follows:
vc＝[-b,b]
b＝1-T/maxT
the synergy between vb and vc mimics the selection behavior of slime.
(4) And calculating the fitness and updating the global optimal position of the population.
Using the third step
Carrying out update calculation of parameters a and b with b=1-T/maxT, and then updating global optimal position information;
(5) Judging whether the maximum iteration times are reached, if so, outputting the storage pool parameters corresponding to the optimal individual position information, establishing an SMA-ESN prediction model by applying the optimal parameter combination of the two parameters, and then predicting test set data to output y (t), namely outputting y (t), namely a network delay value tau (t); if maxT is not reached, repeating the step (2)
(4) To perform parameter continuous optimizing operation.
And fourthly, designing a sliding mode function according to the network delay value tau (t), obtaining a state equation when the system is positioned on a sliding mode surface, and solving the control quantity of the system.
The predicted delay tau (t) obtained by the previous step considers a multiple-input multiple-output continuous system. The SMA-ESN neural network designed in the previous step predicts the time delay value at the next moment through the previous time delay data and transmits the time delay value to a sliding mode controller to be designed. Because of the existence of time delay, the executor can not receive the control signal in time, a time delay compensator is arranged, or a buffer is added before the executor, and the control signal sequence u (t) is stored, so that once the executor does not receive the control signal, the control signal can be selected from the previous control sequence according to the time label. The continuous state model that considers going from controller to actuator to object can be described as:
The sliding mode function is designed according to the network delay value tau (y) as follows:
S(t)＝Cx(t)-∫C(A-BK)x(t)dt+∫CB[u(t)-u(t-τ(t))]dt
wherein, C is a constant matrix with proper dimension, and can ensure CB to be a nonsingular matrix, and K is a parameter to be designed.
Then the following exponential approach law is set: order the
From this, it can be seen that
Thus meeting the accessibility condition of the sliding die surface, the following can be obtained:
it can be seen that the sliding mode function eliminates the inclusion of u (t- τ (t)) after differentiation, taking
Can find the equivalent control item as u eq ＝-Kx(t)，
Combining the equivalent control items u obtained above eq And an exponential approach law, the final control law can be obtained as follows:
u(t)＝-Kx(t)-(CB) -1 (eta+epsilon sgnS). The equation of state when the system is at the slip plane can be described as:
it follows that the original continuous state model +.>
The input delay τ (t) in (b) is compensated.
In summary, the invention improves the algorithm of the prediction time delay model, combines the echo state network based on the mucosae algorithm with the sliding mode controller, establishes the prediction model of random time delay in NCS in a very short time, effectively reduces the training time of the prediction model and has good prediction precision, outputs the future control quantity by combining the prediction result with the sliding mode control algorithm, compensates the random time delay of the network control system, and improves the tracking capability of the network control system signal. The invention has the characteristics of simple structure, high stability, simple operation, strong portability, low cost and the like, and can be applied to engineering practice.
In the present invention, we provide an example to illustrate the effectiveness of the proposed control measures. And comparing the obtained result with the design effect of the common sliding mode controller. The continuous state space model parameters of the controlled object are given as follows:
taking out
First, network-controlled delay prediction will be discussed with emphasis. By simulating the system by Matlab, FIG. 1 shows that the SMA-ESN neural network prediction algorithm has higher prediction precision than the ESN prediction algorithm, better control effect and more contribution to prediction delay.
In order to compensate the time delay generated by the data transmission of the system, a neural network sliding mode controller is designed to ensure the stable state of the network control system, fig. 2 is an experimental effect of a common sliding mode controller method, and fig. 3 is an experimental effect designed by the method. The control signal output by the common sliding mode controller is shown in fig. 4, and the control rate u (t) obtained by the controller designed by the method is shown in fig. 5, compared with the control rate u (t), the control rate u (t) is obviously higher, and the control speed is higher. Fig. 6 and 7 respectively depict the evolution of the sliding mode surfaces of a common sliding mode controller and a controller of the method, and fig. 7 shows that the method has higher convergence speed, smaller overshoot and stronger robustness.
To sum up: all variables corresponding to the method are faster in convergence speed than the common sliding mode method, and the method has smaller oscillation and better robustness.
Claims (8)
1. A network control system time delay compensation method based on neural network sliding mode control is characterized in that: the method comprises the following steps:
step one, an echo state network is established;
step two, echo state network training and updating;
step three, introducing a mucor algorithm to optimize an echo state network model, and forming a combined prediction model prediction network delay value tau (t) of the echo state network optimized based on the mucor algorithm;
and fourthly, designing a sliding mode function according to the network delay value tau (t), obtaining a state equation when the system is positioned on a sliding mode surface, and solving the control quantity of the system.
2. The network control system time delay compensation method based on neural network sliding mode control according to claim 1, wherein the method is characterized in that: in the first step, an input unit u (t) of the echo state network is time delay data of the past moment, and an input layer is provided with K nodes; the reserve pool x (t) is in a current moment state, and N nerve nodes are arranged in the reserve pool; the output unit y (t) is a network delay value output at the next moment, and the network has L output nodes; the states of the input unit, the reserve tank and the output unit are as follows:
u(t)＝[u 1 (t),u 2 (t),...u K (t)] T
x(t)＝[x 1 (t),x 2 (t),...x N (t)] T
y(t)＝[y 1 (t),y 2 (t),...y L (t)] T 。
3. the network control system time delay compensation method based on neural network sliding mode control according to claim 1, wherein the method is characterized in that: the output state equation of the echo state network in the first step is:
y(t+1)＝f out ·(W out ·(u(t+1),x(t+1))，
where fout is the output layer neuron activation function, W out The connection weight of the reserve tank to the output layer is the output of the next moment of y (t+1), the control quantity of the next moment of u (t+1) and the input of the next moment of x (t+1).
4. The network control system time delay compensation method based on neural network sliding mode control according to claim 1, wherein the method is characterized in that: in the third step, the optimization process of the echo state network based on the mucosae algorithm optimization is as follows:
(1) Initializing parameters related to a colistin algorithm, initializing a population quantity parameter N, the maximum iteration times maxT and a parameter Z in a population position updating formula;
(2) Calculating the fitness value of each population by applying a fitness function;
(3) Applying a mucositis individual position update model to the population position, the optimal fitness value and the optimal population position information X b Updating; if the situation that the global optimal fitness value is higher than the optimal fitness value occurs in the iteration process, replacing the global optimal fitness value with the fitness value obtained by iteration, and updating the optimal position information of the population to the current position information;
(4) After updating the position information, calculating the fitness of each population, and updating the global optimal position of the population;
(5) Judging whether the maximum iteration times are reached, if so, outputting a reserve pool parameter corresponding to the optimal individual position information, establishing an echo state network prediction model optimized based on a myxobacteria algorithm by applying the combination of the two optimal parameters, and then predicting test set data to output y (t), namely outputting a y (t) namely a network delay value tau (t); if the maxT is not reached, repeating the steps (2) - (4) to perform the parameter continuous optimizing operation.
5. The network control system time delay compensation method based on neural network sliding mode control according to claim 4, wherein the method is characterized in that: a step of searching for an optimal fitness value in step (3):
firstly, the approach behavior of the coliform bacteria is mathematically modeled, and the following rules are proposed to simulate the shrinkage mode of the coliform bacteria:
wherein vb is [ -a, a]Random numbers in between, W represents weight coefficient of coliform bacteria, X A (t) and X B (t) is two random individual positions, vc is at [ -1,1 [ -1 ]]Parameters of oscillation and finally approaching zero, X b (t) represents the current fitness-optimal individual position, and X (t) represents the current mucositis individual position;
next, the update formulas of the control parameter p, the parameter vb, the parameter a and the weight coefficient W are as follows:
p＝tanh|S(i)-DF|
wherein i epsilon 1,2, …, n, S (i) represents the fitness value of the ith mucosae, DF is the optimal fitness value obtained currently;
vb＝[-a,a]
SmellIndex＝sort(S)
the condition represents that fitness is ranked in the first half of individuals in the population, other represents the rest individuals, r represents a random number between [0,1], bF represents the best fitness value obtained by the current iteration, wF represents the worst fitness value of the current iteration, T represents the current iteration times, maxT represents the maximum iteration times, smellIndex represents the ordering condition of the fitness sequence, smellIndex, bF, wF is obtained, and finally W is calculated;
finally, in order to find the optimal fitness value, that is, when the evaluation index approaches 0, the performance evaluation of the proposed prediction model is performed by considering the root mean square error evaluation index, and the root mean square error is set as
m represents the total length of the time series, i is expressed as the number of iterations, y k And
the actual value of the time series at time k and the output predicted value of the ESN model are respectively represented.
6. The network control system time delay compensation method based on neural network sliding mode control according to claim 4, wherein the method is characterized in that: the location update model of the myxobacteria individual in step (3) is as follows:
wherein UB and LB are upper and lower bounds, rand is a random number uniformly distributed between 0 and 1, z is a self-defined parameter, the value of vb randomly oscillates between [ -a, a ], the value of vc oscillates between [ -1,1] along with the increase of iteration times, and finally approaches zero, and the calculation formula is as follows:
vc＝[-b，b]
b＝1-T/maxT
the synergy between vb and vc mimics the selection behavior of slime.
7. The network control system time delay compensation method based on neural network sliding mode control according to claim 1, wherein the method is characterized in that: the sliding mode function in the fourth step is as follows:
S(t)＝Cx(t)-∫C(A-BK)x(t)dt+∫CB[u(t)-u(t-τ(t))]dt
wherein, C is a constant matrix with proper dimension, and can ensure CB to be a nonsingular matrix, K is a parameter to be designed, t is time, A is an m×m order matrix, B is an m×n order column full order matrix, τ (t) is a network delay value, and u (t) is a control input.
8. The network control system time delay compensation method based on neural network sliding mode control according to claim 1, wherein the method is characterized in that: the system control amount in the fourth step is as follows:
u(t)＝-Kx(t)-(CB) -1 (ηS+εsgn S)
wherein, take eta=1, epsilon=0.005, S is the sliding mode function;
the equation of state when the system is at the slip plane can be described as:
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310005861.8A CN116346636A (en) | 2023-01-04 | 2023-01-04 | Neural network sliding mode control-based network control system time delay compensation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310005861.8A CN116346636A (en) | 2023-01-04 | 2023-01-04 | Neural network sliding mode control-based network control system time delay compensation method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116346636A true CN116346636A (en) | 2023-06-27 |
Family
ID=86882898
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310005861.8A Pending CN116346636A (en) | 2023-01-04 | 2023-01-04 | Neural network sliding mode control-based network control system time delay compensation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116346636A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117094074A (en) * | 2023-07-05 | 2023-11-21 | 盐城工学院 | Control method for processing inherent input time delay of intelligent automobile |
-
2023
- 2023-01-04 CN CN202310005861.8A patent/CN116346636A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117094074A (en) * | 2023-07-05 | 2023-11-21 | 盐城工学院 | Control method for processing inherent input time delay of intelligent automobile |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Han et al. | Adaptive computation algorithm for RBF neural network | |
| CN111047085B (en) | Hybrid vehicle working condition prediction method based on meta-learning | |
| CN111158237B (en) | Industrial furnace temperature multi-step prediction control method based on neural network | |
| JP2022544474A (en) | Improved machine learning for technical systems | |
| CN110046408A (en) | Bounded-but-unknown uncertainty optimization method based on BP neural network differential and interval analysis | |
| CN115186097A (en) | Knowledge graph and reinforcement learning based interactive recommendation method | |
| CN112685657B (en) | Conversation social recommendation method based on multi-mode cross fusion graph network | |
| CN113641722A (en) | Long-term time series data prediction method based on variant LSTM | |
| CN116346636A (en) | Neural network sliding mode control-based network control system time delay compensation method | |
| CN114219181A (en) | Wind power probability prediction method based on transfer learning | |
| CN116432037A (en) | Online migration learning method, device, equipment and storage medium | |
| CN108959787B (en) | Thermal deformation prediction method and system of macro-macro dual-drive system considering actual working conditions | |
| Wang et al. | An Improved Grey Prediction Model Based on Matrix Representations of the Optimized Initial Value. | |
| CN113821724A (en) | Graph neural network recommendation method based on time interval enhancement | |
| CN113052373A (en) | Monthly runoff change trend prediction method based on improved ELM model | |
| Abiyev | Fuzzy wavelet neural network for prediction of electricity consumption | |
| CN116663419A (en) | Sensorless equipment fault prediction method based on optimized Elman neural network | |
| CN109740221B (en) | Intelligent industrial design algorithm based on search tree | |
| KR20200105365A (en) | Method for reinforcement learning using virtual environment generated by deep learning | |
| Tian et al. | A new incremental learning modeling method based on multiple models for temperature prediction of molten steel in LF | |
| Xu et al. | Elman neural network for predicting aero optical imaging deviation based on improved slime mould algorithm | |
| CN116054144A (en) | Distribution network reconstruction method, system and storage medium for distributed photovoltaic access | |
| CN115600492A (en) | Laser cutting process design method and system | |
| CN112183814A (en) | Short-term wind speed prediction method | |
| Xuewu et al. | The city taxi quantity prediction via GM-BP model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |