Wire winding machine tension networked iterative learning control method based on machine learning optimization

Abstract

The application discloses a kind of based on machine learning optimization's winding machine tension networked iterative learning control method, mainly for solving winding machine in tension control process due to network control system output end and control input end data loss caused by tension control system instability and other problems, its main steps include: tension control system modeling;Design has the initial value change winding tension iterative controller;Winding machine tension networked iterative control system design;Improved particle swarm optimization tension networked iterative control system design.The proposed machine learning optimization's winding machine tension networked iterative learning control method aims to handle the initial condition change and data loss in industrial network control system iterative tension control system.At the same time, the results show that the dynamic performance and steady-state performance of the winding machine control system are effectively improved, and the tension control precision is improved.

Description

Technical Field

[0001] This invention belongs to the field of tension control technology for winding machines. It utilizes the idea of ​​machine learning optimization to design a networked iterative learning control scheme for tension control of winding machines based on machine learning optimization. Background Technology

[0002] In recent years, my country's new energy vehicle market has continued to develop, and the competitiveness of new energy vehicles has significantly improved. According to national and industry development trends, the production of compressor motors for new energy vehicles has broad market prospects. The tension control accuracy of winding machines is a major issue affecting the winding of compressor motor windings. However, in actual production, data loss caused by industrial environments and network latency can significantly reduce the performance of the winding machine tension control system. Therefore, a safety compensation strategy for data loss caused by complex industrial environments and network latency must be considered to improve the networked tension control accuracy of winding machines.

[0003] Iterative learning control (ILC) is one of the most commonly used control methods for handling control systems with periodic repetitive operations. This method modifies the current control action or input command based on error information obtained from the previous iteration. Therefore, iterative learning control is widely used in control systems with periodic actions or disturbances. Significant research has also been conducted on iterative learning control in the winding process. Garimella and Srinivasan first proposed using iterative learning control to suppress periodic disturbances caused by friction in tension control. Zhao and CDRahn used a control method combining a PD feedback controller and an ILC feedforward controller to solve the problem of tension control stability in a single-axis winding system. Therefore, in continuous winding systems, where the system operation is periodic or repetitive, iterative learning control has been widely applied in the winding process. Simultaneously, the time-varying radius of the winding material during the winding process gives the tension a certain gradual variation trend. Furthermore, due to the influence of random factors such as the inhomogeneity of the winding material, the tension also exhibits certain random characteristics. Therefore, tension is a complex signal that contains periodic components, a gradual variation trend, and random components. This invention treats the change in the system's spool radius as a problem of changing initial conditions in iterative learning, and proposes a novel fuzzy iterative learning controller for winding machine tension.

[0004] Networked control systems (NCS) have been widely applied in control systems due to their significant advantages in robustness, flexibility, and convenience. The concept of networked control systems was proposed in the 1990s. With the continuous development and integration of computer network and control technologies, NCSs offer advantages such as modular construction, resource sharing within a region, and remote control. Most industrial winding machine factories have implemented networked control. However, during data transmission, physical limitations such as network bandwidth and flow restrictions, as well as time-division multiplexing of communication network information and network instability interference, inevitably lead to data delays and packet loss during network transmission, significantly reducing the performance of networked control systems. In recent years, some successes have been achieved in NCS research using the ILC method to design controllers. Bu Xuhui studied an iterative learning control problem for a stochastic linear system with random data loss modeled by Bernoulli random variables and proposed a solution. Wang and Lim proposed a data-driven networked optimal iterative learning control strategy for a class of discrete linear time-varying systems with single-operation Bernoulli communication delays. The aforementioned literature all focuses on optimizing iterative learning controllers, neglecting the issue of optimizing the system output itself. Data packet loss in networked control systems is essentially a problem of data loss. While improving the controller can increase the convergence speed of the networked control system, in many experimental scenarios, if the packet loss rate is too high, the system remains in a divergent state. Based on this idea, this paper leverages the advantages of the improved particle swarm optimization algorithm in machine learning for fitting and predicting data. Machine learning is combined with networked control systems, treating the fitted predictions from machine learning as lost system output data. This further optimizes the output error under data loss, improving the convergence speed of the networked iterative control system. Summary of the Invention

[0005] This invention aims to improve the accuracy of iterative tension control systems that handle changes in initial conditions and data loss in industrial network control systems, and to reduce the impact of network packet loss on networked tension control systems. It designs an iterative learning controller for initial condition changes, improves the particle swarm optimization algorithm to fit data loss in the networked tension control system, and finally designs a machine learning-based iterative learning control scheme for winding machine tension.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following design scheme:

[0007] S1: Modeling of the tension control system;

[0008] S2: Design an iterative controller for winding tension with varying initial values;

[0009] S3: Design of a networked iterative control system for tension in a winding machine;

[0010] S4: Design of a tension network-based iterative control system optimized by an improved particle swarm optimization algorithm;

[0011] Furthermore, in step S1, the modeling of the tension control system includes:

[0012] According to Hooke's Law, the tension of the winding system can be expressed as:

[0013]

[0014]

[0015]

[0016] Where σ is the surface stress of the winding material, ε is the elastic deformation of the winding material, E is the elastic modulus of the winding material, T is the tension, A is the cross-sectional area of ​​the winding material, L0 is the absolute deformation of the winding material, and L is the length of the winding material deformed by tension in the current state.

[0017] When V1≠V2, the tension change over time t+dt is as follows:

[0018]

[0019]

[0020]

[0021] V=πDn

[0022]

[0023] Where n is the winding speed of the drive motor; D is the diameter of the winding shaft.

[0024] A dynamic model of the speed loop of the tension control system of the winding machine is established, and its equations are as follows:

[0025]

[0026] Where J is the moment of inertia of the 5th winding shaft; K i is the torque constant under unit input voltage; u is the control input; B is the equivalent coefficient combining friction and model; F f (ω1) is the static friction coefficient; △2(t,T,ω1) is an uncertain nonlinearity; ω is the angular velocity of the reel; r1 is the radius of the reel.

[0027] F f(ω1)=a1tanh(c1ω1)+a2[tanh(c2ω1)-tanh(c3ω1)]

[0028] Where a1 and a2 are different friction coefficient levels of the system, c i i = 1, 2, 3 are shape parameters that approximate various friction effects.

[0029]

[0030]

[0031] Furthermore, in step S2, the design of the winding tension iterative controller with initial value variation includes:

[0032] The actual motor rotation angle or COUNT command is obtained from the data fed back by the motor encoder. The change in roll diameter is then calculated using the following equation:

[0033] D un =D u(n-1) -θ u(n-1) A / 2π

[0034] D wn =D w(n-1) -θ w(n-1) A / 2π

[0035] Among them, D un and D wn Let D be the initial diameter of the unwinding spool 1 and the take-up spool 5 before the nth winding. u(n-1) and D w(n-1) Let θ be the initial value of the diameter of the spool before the (n-1)th winding. u(n-1) and θ w(n-1) This is the angular displacement of the two shafts after the (n-1)th winding, calculated using the encoder feedback value.

[0036] The fuzzy subsets of the input and output of the fuzzy controller are defined as {NB, NM, NS, ZP, PS, PM, PB}. Based on the principle of fuzzy adjustment of learning gain, a fuzzy rule base is established, with the control rule in the following form: if E is NM and EC is NM then U is NM. The fuzzy inference method adopts the Mamdani direct inference method, and the centroid method is used to solve the fuzzy problem.

[0037] From the comprehensive error e k The fundamental domain of (t) is [-|e kmin |,|e kmax |], and e k (0) Basic Analects [-|e 0min |,|e 0max |], Overall error change rate and Taking into account the ζ output of the fuzzy controller k γ k The scaling factor k can be calculated from the convergence condition of iterative learning. p k i The design of the iterative learning controller is complete.

[0038] Fuzzy tension iterative controller with initial value variation:

[0039]

[0040] Γ k (t)=Γ k-1 (t)+△ k

[0041] △ k =ξ k ×k p

[0042]

[0043] η k (t)=η k-1 (t)+Θ k

[0044] Θ k =γ k ×k i

[0045] e k (t)=y d (t)-y k (t)

[0046] Among them, u (k) (t) represents the control variable of the system in the k-th iteration, x k (0) represents the initial state of the system during the k-th iteration, y d (t) represents the expected output value of the system at the k-th iteration, y k (t) represents the actual output value of the system at the k-th iteration. The fuzzy controller uses the system output error e as the reference value. k (t) and rate of change of error As input, the adjustment factor ζ is output through fuzzy inference. k In each iteration, the controller extracts the adjustment factor ζ from the output of the fuzzy controller. k Multiply by the scaling factor k p , construct a with Γ k Adjustment matrix with the same number of digits △ k Similarly, the fuzzy controller uses the system input error e k (0) and the rate of change of input error As input, through fuzzy inference factor γ k In each iteration, the controller extracts the adjustment factor γ from the output of the fuzzy controller. k Multiply by the scaling factor k i , construct a with η k Same adjustment matrix Θ k .

[0047] Furthermore, in step S3, the design of the winding machine tension network iterative control system includes:

[0048] The winding system can be simplified into a repetitive SISO nonlinear discrete system as follows:

[0049]

[0050] Where x(t)∈R n ,y(t)∈R m ,u(t)∈R r Let these represent the system's state variables, output variables, and control variables, respectively. For the aforementioned repetitive nonlinear system, this section will provide the following assumptions.

[0051] Assumption 1: The nonlinear functions f(·), b(·), g(·), d(·) satisfy the global Lipschitz condition for the system's state variable x(t), that is, for all t∈[0,N], there exists a finite constant k. f k b k g k d satisfy:

[0052] ||f(x1(t))-f(x2(t))||≤k f ||x1(t)-x2(t)||

[0053] ||b(x1(t))-b(x2(t))||≤k b ||x1(t)-x2(t)||

[0054] ||g(x1(t))-g(x2(t))||≤k g ||x1(t)-x2(t)||

[0055] ||d(x1(t))-d(x2(t))||≤k d ||x1(t)-x2(t)||

[0056] Here, x1(t) and x2(t) are two arbitrary state variables.

[0057] Assumption 2: For a given desired trajectory curve yd (t), where there is a control input u d (t) and state variable x d (t) satisfies the following conditions:

[0058]

[0059] Among them, u d (t) and x d (t) is called the expected input and expected state.

[0060] Assumption 3: The initial conditions of the above nonlinear discrete system satisfy the following equation:

[0061] x (k+1) (0)=x k (0)+ψe k (0)

[0062] Networked iterative control of winding machine tension:

[0063] For nonlinear winding systems that satisfy assumptions 1-3, the aforementioned ILC algorithm is used. When data loss occurs, the control system must satisfy the following equation:

[0064] ||IL(dI+ΦJ)||1<1

[0065] Then we can obtain:

[0066]

[0067] in,

[0068]

[0069]

[0070]

[0071] in

[0072] (1)y k (t) represents the output data of the tension control system;

[0073] (2) The k-th iteration output is the fitted value obtained by machine learning optimization fitting of the output data of the tension control system during the transmission process.

[0074] (3) Indicates by y k (t) and The combined output vector, with the data remaining intact, is represented by y. k (t), the data loss part is taken

[0075]

[0076] (4) This represents the difference between the fitted data and the actual data for the k-th iteration;

[0077] (5) This represents the actual input data for the iterative learning unit;

[0078] (6) This represents the fitted value of the k-th iteration obtained by machine learning optimization fitting of the part lost during the transfer of control variables;

[0079] (7)u k (t) represents the expression by and The output vector is composed of the parts of the vector that have not lost data. Data loss part

[0080]

[0081] (8) This represents the difference between the fitted data and the actual data for the k-th iteration;

[0082] Furthermore, in step S4, the design of the tension networked iterative control system optimized by the improved particle swarm optimization algorithm includes:

[0083]

[0084] The formula shows that the error norm 1 of the networked iterative control system depends on the data loss rate and the data fitting approximation error. On the other hand, it can be shown that, under the same data loss rate, finding a reasonable error fitting approximation strategy can make the system convergence performance better.

[0085] Improved Particle Swarm Optimization Algorithm:

[0086]

[0087] Among them, P i V represents the current position of particle i; i Let P be the current velocity of particle i. a Let P be the optimal position of particle i at the current moment. b The optimal position is defined within the entire particle swarm. c1 and c2 are cognitive coefficients, r1 and r2 are random real numbers belonging to (0,1), and β... i The inertia weight is used to maintain balance in the exploration and development of the search space by the particle swarm. The inertia weight dynamically decreases from 1.0 to near 0 during each generation, and its specific expression is as follows:

[0088]

[0089] Among them, iter max The maximum number of iterations is given by `iter`, which is currently the number of iterations.

[0090] Different types of mutation operators can be used to increase population diversity and help particle swarm optimization (PSO) algorithms escape local minima. The type of mutation operator can be more or less effective, depending on the stage of the optimization process. In this paper, three different types of mutation operators are applied to different stages of the problem to further explore the search space. An adaptive method is proposed to select the appropriate mutation operator for each stage of the problem.

[0091] Cauchy mutation operator:

[0092] V g =V g exp(δ)

[0093] P g =P g +V g δ g

[0094] Among them, P g and V g Position and velocity for the globally optimal example. δ and δ g It is a Cauchy random number with a scaling parameter of 1.

[0095] Gaussian Mutation Operator:

[0096] V g =V g exp(θ)

[0097] P g =P g +V g θ g

[0098] Among them, P g and V g Position and velocity for the globally optimal example. δ and δ g is the Gaussian distribution number with a mean of 0 and a variance of 1.

[0099] Levy mutation operator:

[0100] V g =V gexp(L(α))

[0101] P g =P g +V g L g (α)

[0102] Where, L(α) and L g (α) is a random real number described by the Levy distribution, with the parameter α set to 1.3.

[0103] The proposed method for adaptive mutation utilizes the three mutation operators mentioned above. Initially, the selection ratio is set to 1 / 3 using this ratio, and then the number of particles mutated by each operator is calculated. Then, the selection rate of each mutation operator is applied to the population of particles, and finally, the fitness of the resulting offspring is evaluated.

[0104] The mutation operator with a higher offspring fitness value will be selected compared to another mutation operator with a lower offspring fitness value. The steps for selecting the optimal mutation operator are as follows:

[0105] Evaluation of progress values ​​for each generation of particle swarm operations:

[0106]

[0107] in, and M represents the fitness of the parents and offspring produced by the mutation operator i in generation t. i Let i be the number of particles mutated by the mutation operator i.

[0108] Evaluation of particle swarm feedback values ​​for each generation:

[0109]

[0110] Among them, Z i K is the number of particles whose offspring have better fitness than themselves after operator i mutates. i Let η be the selectivity of the mutation operator in generation t, η be the random weight coefficient between (0,1), N be the number of mutation operators, and c be the selectivity of the mutation operator in generation t. i The penalty factor for mutation operator i is specifically defined as:

[0111]

[0112] The selection update rate of the mutation operator for the next generation:

[0113]

[0114] Where λ is the minimum selection ratio for each mutation operator i.

[0115] The innovation of traditional PSO (Particle Swarm Optimization) intelligent optimization algorithms stems from the social and cognitive behavior of the swarm. Tamako's theory suggests that particles are largely influenced by their previous best particles and local particles. Once the local optimum of the best example remains unchanged, all other particles will rapidly converge to the position of the best particle. The adaptive hybrid swarm optimization algorithm proposed in this paper searches for the neighbors of the globally best particle to mutate in each generation of the search, rather than selecting the globally best particle for mutation. Therefore, it is very helpful for particles to escape local minima, and the entire particle swarm will move to a better position. The framework based on the adaptive hybrid swarm optimization algorithm is given below:

[0116] 1. Randomly generate the initial position and velocity of each particle in the particle swarm.

[0117] 2. Evaluate the fitness of each particle and determine the local and global optimal fitness of each particle in the population.

[0118] 3. Set the initial selection ratio to 1 / 3.

[0119] 4. Update each particle according to the formula.

[0120] 5. For each particle i, if its fitness is less than its previous best position P i If the fitness is such that P is updated, then P is updated. i .

[0121] 6. If there exists a fitness value less than the current best fitness value P g Then update the fitness of the best position for all particles.

[0122] 7. Apply each mutation operator to the number of particles according to the particle selection ratio.

[0123] 8. Evaluate the progress value and feedback value, select the best one from the three mutation operators, and then update the selection ratio of each operator for the next generation.

[0124] 9. Use the optimal mutation operator to mutate the best neighboring particles of the globally optimal particle, and select the best neighborhood from the mutated best neighborhood to generate

[0125] 10. Comparison With P g Choose the better ones to reproduce in the next generation.

[0126] 11. If the stopping criteria are met, stop; otherwise, proceed to step 4.

[0127] Beneficial effects:

[0128] 1. To address the instability of the tension control system in a winding machine caused by data loss at both the output and input ends of the networked control system during tension control, this paper analyzes that the output error of the networked control system depends on the data loss rate and the accuracy of fitting the lost data. Machine learning can improve the generalization ability of the model, correctly identifying even unobserved data not included in the training data. Therefore, this invention proposes using a machine learning algorithm to fit the data loss of the networked winding machine system. Simulation results demonstrate the effectiveness of the machine learning algorithm in fitting the data loss system of the winding machine.

[0129] 2. Building upon the proposed machine learning-fitted tension system with data loss, this paper further proposes a fuzzy iterative controller to address the problem of low tension control accuracy caused by the time-varying radius of the coil, which leads to time-varying radii of the 5 take-up shafts and 1 unwind shaft. Since the coil radius changes relatively little within the same winding cycle, it can be treated as an initial value change in the iterative system. The system output error and error change rate are used as inputs, and an adjustment factor is output through fuzzy inference. In each iteration, the controller extracts the adjustment factor output by the fuzzy controller and multiplies it by a proportional factor to construct an adjustment coefficient matrix with the same number of bits. This allows the time-varying performance of the system to be ignored in a single winding, thereby improving the tension control accuracy. Attached Figure Description

[0130] Figure 1 This is a flowchart of a machine learning-based optimized tension network iterative learning control for a winding machine.

[0131] Figure 2 This is a structural diagram of the winding system proposed in this invention;

[0132] Figure 3 This invention presents a networked system control structure diagram;

[0133] Figure 4 This is an output error curve diagram under the condition that the data packet loss rate is 25% and the data loss is not processed;

[0134] Figure 5 This is an output error curve diagram under the condition that the data packet loss rate is 50% and the data loss is not processed;

[0135] Figure 6 This is a diagram showing the system output tension tracking effect under a data loss rate of 25%;

[0136] Figure 7 This is the system output error curve when the data packet loss rate is 25% and data loss occurs.

[0137] Figure 8 This is a diagram showing the system output tension tracking effect under a data loss rate of 50%;

[0138] Figure 9 This is the system output error curve when the data packet loss rate is 50% and data is lost. Detailed Implementation

[0139] As described above, to address the problems of loose and uneven enameled wire winding and easy breakage during winding caused by data loss in the network control system and time-varying radius of the winding material in the tension control system of a winding machine, this invention designs a networked iterative learning control method for winding machine tension based on machine learning optimization, in order to improve the dynamic and steady-state performance of the tension control system and reduce the impact of tension disturbances caused by data loss and changes in the radius of the winding material on the system. The invention will be described in more detail below with reference to the accompanying drawings.

[0140] See Figure 1 The flowchart shown illustrates the iterative learning control of winding machine tension network based on machine learning optimization, which specifically includes the following steps:

[0141] Step S1: Modeling the tension control system:

[0142] See Figure 2 As shown, the winding system structure of the present invention is as follows:

[0143] Wherein, L - length of the wound material after deformation (m), V11 unwinding shaft speed (m / s), V25 take-up shaft speed (m / s), D11 unwinding shaft diameter (m / s), D25 take-up shaft diameter (m / s)

[0144] According to Hooke's Law, the tension of the winding system can be expressed as:

[0145]

[0146]

[0147]

[0148] Where σ is the surface stress of the winding material, ε is the elastic deformation of the winding material, E is the elastic modulus of the winding material, T is the tension, A is the cross-sectional area of ​​the winding material, L0 is the absolute deformation of the winding material, and L is the length of the winding material deformed by tension in the current state.

[0149] When V1≠V2, the tension change over time t+dt is as follows:

[0150]

[0151]

[0152]

[0153] V=πDn (7)

[0154]

[0155] Where n is the winding speed of the drive motor; D is the diameter of the winding shaft.

[0156] Traditional winding systems primarily achieve winding by controlling a motor to drive a rotating shaft. The following discussion assumes that the speed of the unwinding motor in this system is effectively controlled and its rotational speed can be measured using a speed sensor. Therefore, a mathematical model for a five-rewinding motor is established below, but this model can also be applied to the unwinding motor.

[0157] A dynamic model of the speed loop of the tension control system of the winding machine is established, and its equations are as follows:

[0158]

[0159] Where J is the moment of inertia of the 5th winding shaft; K i is the torque constant under unit input voltage; u is the control input; B is the equivalent coefficient combining friction and model; F f (ω1) is the static friction coefficient; △2(t,T,ω1) is an uncertain nonlinearity; ω is the angular velocity of the reel; r1 is the radius of the reel.

[0160] F f (ω1)=a1 tanh(c1ω1)+a2[tanh(c2ω1)-tanh(c3ω1)] (10)

[0161] Where a1 and a2 are different friction coefficient levels of the system, c i i = 1, 2, 3 are shape parameters that approximate various friction effects.

[0162]

[0163]

[0164] Step S2: Design an iterative controller for winding tension with initial value variation:

[0165] As shown in formula (8), in the system mathematical model, the change in the radius of the winding material will cause a change in the diameter D of the spool. Therefore, the winding radius is a time-varying parameter, and it will cause a change in the initial value of the system control parameters in each winding cycle. Currently, there are two methods to measure the change in the spool diameter. One is to configure a dedicated ultrasonic sensor, image detector or other spool diameter measuring device in the winding and unwinding system to measure the spool diameter in real time. This method is mostly used in situations where high measurement accuracy is required and the spool diameter changes significantly over time. The other method is to obtain the actual motor rotation angle or COUNT command from the data fed back by the motor encoder and calculate the change in the spool diameter. This method is mostly used in situations where the spool diameter changes little and the speed is fast. In industrial composite material winding systems, due to the small radius of the winding material, most literature ignores the measurement of intermediate change values. Therefore, this paper will use the second method to obtain the initial value of the spool diameter. The specific calculation formula is as follows:

[0166] D un =D u(n-1) -θ u(n-1) A / 2π (13)

[0167] D wn =D w(n-1) -θ w(n-1) A / 2π (14)

[0168] Among them, D un and D wn Let D be the initial diameter of the unwinding spool 1 and the take-up spool 5 before the nth winding. u(n-1) and D w(n-1) Let θ be the initial value of the diameter of the spool before the (n-1)th winding. u(n-1) and θ w(n-1) This is the angular displacement of the two shafts after the (n-1)th winding, calculated using the encoder feedback value.

[0169] Fuzzy tension iterative controller with initial value variation:

[0170]

[0171] Γ k (t)=Γ k-1 (t)+△ k (16)

[0172] △ k =ξ k ×k p (17)

[0173]

[0174] η k (t)=η k-1(t)+Θ k (19)

[0175] Θ k =γ k ×k i (20)

[0176] e k (t)=y d (t)-y k (t) (21)

[0177] Among them, u (k) (t) represents the control variable of the system in the k-th iteration, x k (0) represents the initial state of the system during the k-th iteration, y d (t) represents the expected output value of the system at the k-th iteration, y k (t) represents the actual output value of the system at the k-th iteration. The fuzzy controller uses the system output error e as the reference value. k (t) and rate of change of error As input, the adjustment factor ζ is output through fuzzy inference. k In each iteration, the controller extracts the adjustment factor ζ from the output of the fuzzy controller. k Multiply by the scaling factor k p , construct a with Γ k Adjustment matrix with the same number of digits △ k Similarly, the fuzzy controller uses the system input error e k (0) and the rate of change of input error As input, through fuzzy inference factor γ k In each iteration, the controller extracts the adjustment factor γ from the output of the fuzzy controller. k Multiply by the scaling factor k i , construct a with η k Same adjustment matrix Θ k .

[0178] Step S3: Design of a tension network iterative control system for the winding machine:

[0179] The winding system can be simplified into a repetitive SISO nonlinear discrete system as follows:

[0180]

[0181] Where x(t)∈R n ,y(t)∈R m ,u(t)∈R r Let these represent the system's state variables, output variables, and control variables, respectively. For the aforementioned repetitive nonlinear system, this section will provide the following assumptions.

[0182] Assumption 1: The nonlinear functions f(·), b(·), g(·), d(·) satisfy the global Lipschitz condition for the system's state variable x(t), that is, for all t∈[0,N], there exists a finite constant k. f k b k g k d satisfy:

[0183] ||f(x1(t))-f(x2(t))||≤k f ||x1(t)-x2(t)|| (23)

[0184] ||b(x1(t))-b(x2(t))||≤k b ||x1(t)-x2(t)|| (24)

[0185] ||g(x1(t))-g(x2(t))||≤k g ||x1(t)-x2(t)|| (25)

[0186] ||d(x1(t))-d(x2(t))||≤k d ||x1(t)-x2(t)|| (26)

[0187] Here, x1(t) and x2(t) are two arbitrary state variables.

[0188] Assumption 2: For a given desired trajectory curve y d (t), where there is a control input u d (t) and state variable x d (t) satisfies the following conditions:

[0189]

[0190] Among them, u d (t) and x d (t) is called the expected input and expected state.

[0191] Assumption 3: The initial conditions of the above nonlinear discrete system satisfy the following equation:

[0192] x (k+1) (0)=x k (0)+ψe k (0) (28)

[0193] See Figure 3 The diagram shown is a networked system architecture diagram of the present invention:

[0194] Depend on Figure 3The networked iterative learning control graph shows that the system output y k (t) and the control output of the iterative control unit in iterative control. Data loss can occur in network control systems due to network congestion, signal loss, and other reasons. To analyze this data loss, a mathematical description of data packet loss in networked iterative learning control of a winding system is presented here.

[0195] Assuming a known variable ω(t) is used to represent the data transmission situation at time t in iterative control, this paper assumes that the variable ω(t) follows the following 0-1 distribution:

[0196]

[0197]

[0198]

[0199] Where X[ω(t)=1] represents the probability that the variable ω(t)=1; Let be a known constant representing the probability of normal data transmission. Therefore, for any k-th iteration, the packet loss matrix can be defined as follows:

[0200]

[0201]

[0202] The same principle can be used to define the control output of the iterative control unit. Transmission status in the network:

[0203]

[0204] Networked iterative control of winding machine tension:

[0205] For nonlinear winding systems that satisfy assumptions 1-3, the aforementioned ILC algorithm is used. When data loss occurs, the control system must satisfy the following equation:

[0206] ||IL(dI+ΦJ)||1<1 (35)

[0207] Then we can obtain:

[0208]

[0209] in,

[0210]

[0211]

[0212]

[0213] in

[0214] (1)y k (t) represents the output data of the tension control system;

[0215] (2) The k-th iteration output is the fitted value obtained by machine learning optimization fitting of the output data of the tension control system during the transmission process.

[0216] (3) Indicates by y k (t) and The combined output vector, with the data remaining intact, is represented by y. k (t), the data loss part is taken

[0217]

[0218] (4) This represents the difference between the fitted data and the actual data for the k-th iteration;

[0219] (5) This represents the actual input data for the iterative learning unit;

[0220] (6) This represents the fitted value of the k-th iteration obtained by machine learning optimization fitting of the part lost during the transfer of control variables;

[0221] (7)u k (t) represents the expression by and The output vector is composed of the parts of the vector that have not lost data. Data loss part

[0222]

[0223] (8) This represents the difference between the fitted data and the actual data for the k-th iteration;

[0224] Step S4: Design of an improved tension network-based iterative control system optimized by particle swarm optimization:

[0225]

[0226] The formula shows that the error norm 1 of the networked iterative control system depends on the data loss rate and the data fitting approximation error. On the other hand, it can be shown that, under the same data loss rate, finding a reasonable error fitting approximation strategy can make the system convergence performance better.

[0227] Improved Particle Swarm Optimization Algorithm:

[0228]

[0229] Among them, P i V represents the current position of particle i; i Let P be the current velocity of particle i. a Let P be the optimal position of particle i at the current moment. b The optimal position is defined within the entire particle swarm. c1 and c2 are cognitive coefficients, r1 and r2 are random real numbers belonging to (0,1), and β... i The inertia weight is used to maintain balance in the exploration and development of the search space by the particle swarm. The inertia weight dynamically decreases from 1.0 to near 0 during each generation, and its specific expression is as follows:

[0230]

[0231] Among them, iter max The maximum number of iterations is given by `iter`, which is currently the number of iterations.

[0232] Different types of mutation operators can be used to increase population diversity and help particle swarm optimization (PSO) algorithms escape local minima. The type of mutation operator can be more or less effective, depending on the stage of the optimization process. In this paper, three different types of mutation operators are applied to different stages of the problem to further explore the search space. An adaptive method is proposed to select the appropriate mutation operator for each stage of the problem.

[0233] Cauchy mutation operator:

[0234] V g =V g exp(δ) (43)

[0235] P g =P g +V g δ g (44)

[0236] Among them, P g and V g Position and velocity for the globally optimal example. δ and δ g It is a Cauchy random number with a scaling parameter of 1.

[0237] Gaussian Mutation Operator:

[0238] V g =V g exp(θ) (45)

[0239] P g =P g +V g θ g (46)

[0240] Among them, P g and V g Position and velocity for the globally optimal example. δ and δ g is the Gaussian distribution number with a mean of 0 and a variance of 1.

[0241] Levy mutation operator:

[0242] V g =V g exp(L(α)) (47)

[0243] P g =P g +V g L g (α) (48)

[0244] Where, L(α) and L g (α) is a random real number described by the Levy distribution, with the parameter α set to 1.3.

[0245] The proposed method for adaptive mutation utilizes the three mutation operators mentioned above. Initially, the selection ratio is set to 1 / 3 using this ratio, and then the number of particles mutated by each operator is calculated. Then, the selection rate of each mutation operator is applied to the population of particles, and finally, the fitness of the resulting offspring is evaluated.

[0246] The mutation operator with a higher offspring fitness value will be selected compared to another mutation operator with a lower offspring fitness value. The steps for selecting the optimal mutation operator are as follows:

[0247] Evaluation of progress values ​​for each generation of particle swarm operations:

[0248]

[0249] in, and M represents the fitness of the parents and offspring produced by the mutation operator i in generation t. i Let i be the number of particles mutated by the mutation operator i.

[0250] Evaluation of particle swarm feedback values ​​for each generation:

[0251]

[0252] Among them, Z i K is the number of particles whose offspring have better fitness than themselves after operator i mutates. i Let η be the selectivity of the mutation operator in generation t, η be the random weight coefficient between (0,1), N be the number of mutation operators, and c be the selectivity of the mutation operator in generation t. i The penalty factor for mutation operator i is specifically defined as:

[0253]

[0254] The selection update rate of the mutation operator for the next generation:

[0255]

[0256] Where λ is the minimum selection ratio for each mutation operator i.

[0257] See Figure 4 As shown, the tension network control method for the winding machine of this invention was established using MATLAB and simulated. The tension control system of the winding machine mainly consists of 1. unwinding shaft, 2. unwinding wheel, 3. tension detector, 4. winding wheel, and 5. winding shaft. The specific parameters of the tension control system of the winding machine are as follows: Young's modulus E / (N / m) 2 1.6×10 -9 Material cross-sectional area A / (A / m) 2 1.2×10 -6 Unwinding speed v2 / (m / s) 0.5; Take-up shaft diameter D1 / (m) 0.2; Take-up shaft moment of inertia ε(N / m) 0.0146; Voltage torque constant K i 1; Friction damping coefficient B 0.25; Friction grade coefficients a1, a2 0.02 0.01; Friction shape coefficients c1, c2, c3 700 15 15; Learning gain L = 0.3.

[0258] See Figure 4 , 5 As shown, the tension output error (1-norm) curves of the unprocessed tension control system of the winding machine with packet loss rates of 25% and 50% using MATLAB simulation data are based on... Figure 4 , 5It is known that in a networked iterative control tension control system for winding, if the lost data of the networked iterative learning system is not processed, the system will become extremely unstable and unable to achieve normal tracking performance, even if the packet loss rate is small. Therefore, an optimized iterative control algorithm is designed to achieve tension control in the winding machine system when data loss occurs.

[0259] See Figure 6 , 7 As shown in Figures 8 and 9, by appropriately selecting the learning gain factor, when the network packet loss rate is low, the machine learning optimization algorithm can accurately approximate the output and input ends of the application device, resulting in good system tracking. However, when the network packet loss rate is high, the system error will fluctuate within a small range, leading to system instability. Therefore, in engineering practice, it is essential to use channels with high transmission accuracy to reduce the probability of data packet loss and avoid system instability.

[0260] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the embodiments are only for explaining the invention and are not intended to limit the scope of protection of the invention. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the invention; therefore, the scope of protection of the invention should be determined by the scope defined in the claims.

Claims

1. A networked iterative learning control method for tension in a winding machine based on machine learning optimization, characterized in that: Using the tension difference of the winding machine as input, this paper designs an iterative tension controller with initial value changes and a BP neural network tension control system optimized by a genetic algorithm. The slow change in the winding radius is set as the initial state change problem in iterative learning, and an iterative learning control algorithm with an initial state learning law is proposed. Considering the data packet loss and time delay phenomena in the networked control system of the winding machine in actual industrial production, the data packet loss problem in industrial production is described as a stochastic Bernoulli process with known probability. A BP neural network optimized by a genetic algorithm in machine learning is used to predict the system output under data loss conditions. The method includes the following steps: S1: Modeling of the tension control system; S2: Design an iterative controller for winding tension with varying initial values; S3: Design of a networked iterative control system for tension in a winding machine; S4: Design of a tension network-based iterative control system optimized by an improved particle swarm optimization algorithm; A tension network-based iterative learning control method for winding machines based on machine learning optimization, characterized in that, in step S4, the tension network-based iterative control system optimized by the improved particle swarm optimization algorithm is designed as follows: The formula shows that the error norm 1 of the networked iterative control system depends on the data loss rate and the data fitting approximation error. On the other hand, it can be shown that, under the same data loss rate, finding a reasonable error fitting approximation strategy can make the system convergence performance better. Improved Particle Swarm Optimization Algorithm: in, For particles Current location; For particles Current speed, For particles At the optimal position at the current moment To be the optimal position in the entire particle swarm, and For cognitive coefficient, and Belonging to random real numbers, Inertial weights are used to maintain balance in the exploration and development of the search space by the particle swarm; the inertial weights are adjusted in each generation from... to approach It decreases dynamically during runtime, and its specific expression is as follows: in, The maximum number of iterations, Current iteration number; Evaluation of progress values ​​for each generation of particle swarm operations: in, and Mutation operator In the The fitness of the parents and offspring produced in each generation. Mutation operator The number of mutated particles; Evaluation of particle swarm feedback values ​​for each generation: in, For operators The number of particles whose offspring have better fitness than their own after a mutation. For the mutation operator in the th The selection rate of the generation For the middle Random weighting coefficients between them The number of mutation operators. Mutation operator The penalty factor is specifically defined as: The selection update rate of the mutation operator for the next generation: in, For each mutation operator The minimum selection ratio.

2. The method for networked iterative learning control of winding machine tension based on machine learning optimization according to claim 1, characterized in that, In step S1, the tension control system is modeled: A dynamic model of the tension loop in the tension control system of the winding machine is established, and its equations are as follows: According to Hooke's Law, the tension of the winding system can be expressed as: in, For the surface stress of the wound material, For the elastic deformation of the wound material, The elastic modulus of the wound material. For tension, The cross-sectional area of ​​the wound material. The absolute deformation of the wound material. This is the length of the winding material deformed due to tension in the current state. when At that time, The tension changes over time are as follows: in, The winding speed of the drive motor's reel; The diameter of the reel; A dynamic model of the speed loop of the tension control system of the winding machine is established, and its equations are as follows: in, The moment of inertia of the take-up shaft is 5. The torque constant under unit input voltage; For control input; The equivalent coefficient for friction combined with the model; The coefficient of static friction; It is uncertain and nonlinear; The angular velocity of the reel; The radius of the take-up reel is 5. in, For different friction coefficient levels of the system, The shape parameters are approximations of various friction effects; 。 3. The method for networked iterative learning control of winding machine tension based on machine learning optimization according to claim 1, characterized in that, In step S2, the design includes an iterative winding tension controller with initial value variations: The actual motor rotation angle or COUNT command is obtained from the data fed back by the motor encoder. The change in roll diameter is then calculated using the following equation: in, and The first The initial diameter values ​​of the first unwinding shaft and the fifth winding shaft before the next winding. and For the first The initial value of the diameter of the spool before the next winding. and For the first The angular displacement of the two shafts after the second winding is calculated using the encoder feedback value; Fuzzy tension iterative controller with initial value variation: in, For the first The system's control variables during the next iteration For the first The initial state of the system during the next iteration. For the first The expected output value of the system in the next iteration For the first The actual output value of the system during the next iteration; the fuzzy controller uses the system output error. and error change rate As input, output adjustment factors through fuzzy inference. During each iteration, the controller extracts the adjustment factor from the output of the fuzzy controller. Multiply by the scaling factor , build a with Adjustment matrix with the same number of bits Similarly, the fuzzy controller uses the system input error and the rate of change of input error As input, through fuzzy inference factors During each iteration, the controller extracts the adjustment factor from the output of the fuzzy controller. Multiply by the scaling factor , build a with Same adjustment matrix .

4. The networked iterative learning control method for winding machine tension based on machine learning optimization according to claim 1, characterized in that, In step S3, the design of the winding machine tension network iterative control system is as follows: The winding system can be simplified into a repetitive SISO nonlinear discrete system as follows: in, These represent the system's state variables, output variables, and control variables, respectively. Using the above ILC algorithm, when data packet loss occurs in the system, as long as the control system satisfies the following formula: Then we can obtain: in, 。

Images