Monofilament production line tension cooperative control system and method based on neural network fuzzy adaptive PID

Abstract

The invention discloses a monofilament production line tension cooperative control system and method based on neural network fuzzy adaptive PID. The system is composed of a distributed sensing layer, an execution driving layer and an intelligent control layer. The intelligent control layer adopts a neural network fuzzy adaptive PID control algorithm, optimizes a membership function and a rule weight of a fuzzy inference device on line through a neural network, dynamically sets parameters of a corresponding PID controller according to tension real-time deviation and a deviation change rate of each key process point, and executes an embedded multi-transmission-point coordination rule, so as to realize the optimal control of the tension of each key process point. And intelligent cooperative control of each transmission point is realized. According to the method, the self-learning ability of the neural network and the interpretability of fuzzy logic are fused, the problems that a traditional PID is poor in adaptability, a fuzzy controller depends on fixed expert experience, and multi-transmission-point collaboration is insufficient are solved, and the tension stability and uniformity and product consistency in the monofilament production process are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of polymer fiber material processing equipment and intelligent control technology, specifically to a system and method for intelligent and stable tension control in monofilament production, which integrates neural networks and fuzzy adaptive PID control. Background Technology

[0002] In the monofilament production process, precise and stable tension control is a core technological element that determines the physical properties of the final product (such as linear density uniformity, breaking strength, and elongation consistency) and production economy (such as filament breakage rate and raw material utilization rate). A typical production line consists of multiple independently driven transmission points connected in series. The tension at each point is transmitted and dynamically coupled through the moving monofilaments, forming a complex controlled object with multiple variables, strong coupling, nonlinearity, and time delay.

[0003] Currently widely used traditional PID control strategies have fixed parameters, making them difficult to adapt to dynamic operating conditions. While simple fuzzy control has some adaptability, its rules and membership functions rely on expert experience for design, making it difficult to guarantee optimality and limiting steady-state accuracy. Existing methods combining fuzzy control and PID (fuzzy adaptive PID) have improved adaptability to some extent, but the performance of its fuzzy component is still limited by the initial manual design and cannot be self-improved through process data. Control quality may degrade when facing complex, variable, or unverified production conditions. Furthermore, existing technologies for collaborative control between multiple drive points often employ master-slave control or speed synchronization strategies, lacking intelligent collaborative mechanisms based on tension coupling relationships, making it difficult to achieve global dynamic equilibrium.

[0004] Neural networks possess powerful nonlinear mapping capabilities and the ability to learn from data, but their decision-making processes have poor interpretability, making them unsuitable for embedding explicit process knowledge and control rules. Fuzzy logic, on the other hand, excels at handling uncertainty problems. Its rule-based reasoning is intuitive and highly interpretable, facilitating the integration of domain expert experience, but it lacks the ability to learn and optimize from data itself.

[0005] This invention deeply integrates neural networks and fuzzy logic to construct a neural network fuzzy system. This integrated system fully leverages the self-learning and adaptive capabilities of neural networks, optimizing core parameters (such as membership functions and rule weights) of the fuzzy system online or offline through a data-driven approach. Simultaneously, the system retains the rule-based reasoning framework of fuzzy logic, ensuring transparency and interpretability in the control decision-making process and facilitating the embedding of prior technological knowledge, such as multi-drive point coordination. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing tension control technologies for monofilament production lines, such as limited adaptive capabilities, reliance on expert experience in fuzzy controllers, and insufficient coordination among multiple transmission points. This invention provides a tension collaborative control system and method for monofilament production lines based on a neural network-based fuzzy adaptive PID controller. This system and method optimize the core parameters of the fuzzy inference engine through online learning via a neural network, enabling the fuzzy adaptive PID controller to possess stronger self-learning and adaptive capabilities. Combined with embedded collaborative rules, it achieves high-precision, high-stability, intelligent collaborative control of the tension at each transmission point.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A tension collaborative control system for a monofilament production line based on neural network fuzzy adaptive PID is characterized by: tension detection units set at at least four key process points along the monofilament production line process flow. The first detection unit is located between the cooling water tank and the first traction roller, used to measure the initial introduction tension N1 of the filament bundle; the second unit is located between the key drafting rollers, used to measure the main drafting tension N2; the third unit is located between the final hot roller and the winding machine, used to measure the tension N3 of the filament bundle after shaping, and a high-precision laser scanning diameter gauge is installed near this unit to continuously and in real time measure the monofilament diameter Φ at a frequency of not less than 100Hz; the fourth unit is located at the guide of the winding device, measuring the final winding tension N4. Each detection unit preferably uses a frictionless, high dynamic response contact or non-contact tension sensor. Simultaneously, a high-resolution absolute encoder is installed at the rear end of the servo motor of each drive roller to accurately feedback the surface linear velocity V_k of each roller. An encoder is installed on the winding shaft or the winding diameter D is calculated in real time using a torque-current model.

[0008] The system's actuators consist of high-performance AC servo motors and matching servo drivers, used to independently drive each process roller; the winding device employs a vector frequency converter motor with torque control mode. All drivers support high-speed fieldbus communication protocols to receive speed or torque commands from the controller.

[0009] The system's control core is an industrial-grade programmable logic controller (PLC), which internally runs a multi-loop neural network fuzzy adaptive PID collaborative control algorithm. The controller's central processing module is responsible for the synchronous acquisition of global data, process logic sequence control, and system safety interlocking. Each core control loop is configured with an independent neural network fuzzy inference engine. Each neural network fuzzy inference engine takes the real-time deviation e_k and the rate of change of deviation ec_k of the corresponding tension as input. The neural network part first processes the input and outputs the adjustment amount used to adjust the membership function parameters in the associated fuzzy inference engine. δcij , ssijSubsequently, the fuzzy inference engine, based on the adjusted membership function, fuzzifies the precise quantity and performs inference according to the preset fuzzy rule base (whose rule weights can also be adjusted with the assistance of neural networks), ultimately outputting the online dynamic correction amount for the PID controller parameters of this loop. Kp_k , Ki_k , Kd_k The collaborative decision-making logic of the system also exists in the form of embedded rules in the fuzzy rule base.

[0010] A tension collaborative control method for a monofilament production line based on neural network fuzzy adaptive PID is characterized by comprising the following steps: Step 1: Synchronous Data Acquisition and Preprocessing The system synchronously samples all sensor signals at a fixed control period Ts (typically 1-10ms) and performs preprocessing on the raw signals, including digital filtering, unit conversion, and validity verification.

[0011] Step 2: Control State Calculation For the k-th tension control loop (k=1, 2, 3, 4), calculate the tension deviation e_k(i) and the rate of change of deviation ec_k(i):

[0012] Step 3: Neural Network Fuzzy Inference and Parameter Tuning The precise values ​​e_k and ec_k are input into the neural network module. This neural network can be a simple three-layer feedforward structure (input layer, hidden layer, output layer). The number of neurons in the output layer corresponds to the number of fuzzy parameters that need to be adjusted. For example, if the number of fuzzy subsets for each input variable e and ec is 7, and a Gaussian membership function is used, then the center of each subset... c Both the width σ and the parameter width σ are adjustable, and the neural network can output the adjustment amount δ of these parameters. c , δσ. The hidden layer of the neural network uses the Tanh function as the activation function, and its expression is:

[0013] This function maps the input to a continuous, smooth output space of (-1, 1). The Tanh function was chosen primarily based on the following considerations: The Tanh function has a zero-mean output, a characteristic that benefits the learning of subsequent neurons, accelerates the convergence speed of gradient descent, and makes weight updates smoother and more directional during training, avoiding the accumulation of biases towards positive or negative directions. Compared to the Sigmoid function, the Tanh function has a steeper gradient range, with a larger gradient when the input is close to zero, providing a stronger learning signal; while when the absolute value of the input is large, its output smoothly saturates within ±1, with the gradient approaching zero, preventing the parameters from oscillating violently due to excessively large single adjustments, thus enhancing the stability of the online learning process of the control system. The tension deviation e and its rate of change ec are typical continuous signals with both positive and negative values. The Tanh function's symmetry about the origin allows it to naturally and unbiasedly handle positive and negative inputs, which aligns with the principle of symmetrical distribution of positive and negative linguistic variables in fuzzy logic, facilitating the neural network's learning and output of symmetrical and balanced adjustments to the membership function's center and width parameters. The output layer of the neural network uses a linear activation function to directly output the continuous adjustment δ of the fuzzy parameter. c δσ has an unrestricted range and can be adapted to the needs of actual parameter adjustment.

[0014]

[0015]

[0016] in, For the above Tanh activation function, It is a linear activation function. , , , For network weights and biases.

[0017] Update the current membership function parameters using the neural network output:

[0018] in, , This is the learning rate.

[0019] Using the updated membership function, the precise values ​​e_k and ec_k are fuzzified, and the activation degree of each fuzzy rule is calculated. The Mamdani inference method and the centroid method are used to defuzzify the data, yielding the precise PID parameter corrections. Kp_2 , Ki_2 , Kd_2 .

[0020] Step 4: Parameter Update, Collaborative Decision Making, and Closed-Loop Optimization Update PID controller parameters online:

[0021] The fuzzy rule base contains cross-loop cooperative rules. For example, when N1 is too low and N2 is slightly high, the cooperative decision should significantly enhance the proportional effect of the first traction roller to rapidly increase N1, while slightly weakening the integral effect of the main drafting roller to allow N2 to decrease gradually, thereby stabilizing the system. This is expressed as "IF e_1 is NB AND e_2 is PS THEN". Kp_1 is PB AND Ki_2is NS". Each adaptive PID controller calculates and outputs control quantities to the driver. Based on the measured value of diameter Φ, the main tensile tension setpoint N2_set is slowly and with small amplitude closed-loop feedback correction to stabilize the diameter at the target value.

[0022] Step 5: Online Neural Network Learning The neural network weights are updated every L control cycles (e.g., L=100). A loss function J is constructed using recent control performance metrics, such as the sum of squared tension deviations.

[0023] The backpropagation algorithm is used to calculate the gradient of the loss function with respect to the neural network weights. And update the weights with a smaller learning rate:

[0024] in, The learning rate for online learning is typically set to be much smaller than the parameter update learning rate. , The value is used to ensure the smoothness of the control process.

[0025] This invention deeply integrates neural networks and fuzzy logic to construct a neural network fuzzy system. This integrated system fully leverages the self-learning and adaptive capabilities of neural networks, optimizing core parameters (such as membership functions and rule weights) of the fuzzy system online or offline through a data-driven approach. Simultaneously, the system retains the rule-based reasoning framework of fuzzy logic, ensuring transparency and interpretability in the control decision-making process and facilitating the embedding of prior technological knowledge, such as multi-drive point coordination. Compared to existing technologies, the advantages of this invention are: 1. By optimizing the membership function and rule weights of the fuzzy system online through neural networks, the controller can learn from actual operating data, continuously self-optimize, and better adapt to unforeseen changes in operating conditions, overcoming the limitation of traditional fuzzy controllers that rely on fixed expert experience.

[0026] 2. Fuzzy parameter optimization with neural network assistance makes the tuning of PID parameters more precise and closer to the actual process, further reducing system overshoot, oscillation and steady-state error, and improving the tension control accuracy and stability.

[0027] 3. The core control logic is still based on fuzzy rules, which facilitates the embedding of process knowledge and collaborative rules (such as multi-drive point coupling compensation rules), thus maintaining the interpretability of the system and its global collaborative decision-making capability.

[0028] 4. By combining diameter closed-loop optimization, the system has further improved compared with traditional fuzzy adaptive PID control in key indicators such as tension stability, wire breakage rate, and uniformity of single wire diameter. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall hardware architecture and signal connections of the system of the present invention.

[0030] Figure 2 This is a block diagram illustrating the principle structure of the neural network fuzzy adaptive PID controller in this invention.

[0031] Figure 3 This is an overall flowchart of the control method of the present invention.

[0032] Figure 4 The graph shows a comparison of the response curves of tension N2 during the acceleration process of the production line with that of traditional PID control after implementing the present invention. Detailed Implementation

[0033] To further understand the present invention, the following embodiments use a production line for producing PA6 (nylon 6) monofilament with a diameter of 0.20 mm as an example, and the present invention will be described in a non-limiting manner with reference to the accompanying drawings.

[0034] According to the system design requirements of this invention, four tension detection units are installed on the PA6 production line, located respectively at: between the cooling water tank and the first traction roller, between the first and second drafting rollers, after the final heat setting roller (near which a laser diameter gauge is installed), and at the yarn guide of the winding device. A high-resolution absolute encoder is installed at the rear end of the servo motor of each drive roller for feedback of linear speed. An encoder is installed on the winding shaft to calculate the roll diameter. The actuators use high-performance AC servo motors to drive each process roller, and the winding uses a vector frequency converter motor. All equipment is connected to a programmable automation controller via an EtherCAT bus.

[0035] Taking the main tensile tension N2 control loop as an example, the initial fuzzy system settings are as follows: e_2 basic universe of discourse: [-4,+4]N, quantization factor Ke = 1.5. Fuzzy universe of discourse: [-6,+6]; ec_2 basic universe of discourse: [-1.5,+1.5]N / s, quantization factor Kec = 4. Fuzzy universe of discourse: [-6,+6]; Kp_2 Output universe of discourse: [-0.5, +0.5]; Ki_ 2 Output universe of discourse: [-0.02, +0.02]s⁻¹; Kd_2 Output universe of discourse: [-0.005, +0.005]s.

[0036] For the input and output variables, seven Gaussian membership functions (NB, NM, NS, ZE, PS, PM, PB) are uniformly distributed across their respective fuzzy universes. The typical initial center value set is {-6, -4, -2, 0, +2, +4, +6}, and the initial width is set to 2. The table below shows... Kp_2 Fuzzy Rule Table Design:

[0037] The input layer has 2 neurons (e, ec), the hidden layer has 5 neurons (with the activation function tanh), and the output layer has 28 neurons (adjusting the center and width of the membership functions for 7 each of e and ec, with linear activation functions). Initialization is performed using small random numbers. In this embodiment, the hidden layers of the neural network explicitly use the Tanh function. Compared to the Sigmoid function, Tanh's output range is (-1, 1) with a mean of 0. This makes the mean of the inputs to each layer closer to 0 in the early stages of training, helping to alleviate the vanishing gradient problem and accelerate network convergence. Simultaneously, its symmetry allows it to handle positive and negative signals of tension bias more evenly. Compared to the ReLU function, Tanh is differentiable and smooth across its entire domain. Although computationally more complex, in this system's application of online, small-step, continuous fine-tuning scenarios, it provides a more stable and smoother parameter adjustment output, avoiding the "neuron death" problem that ReLU may cause and the resulting model update stagnation risk, thus ensuring the robustness of the learning process.

[0038] Parameter adjustment learning rate = =0.01, online weight learning rate =0.001. Basic parameter setting value: =2.5, =0.05s⁻¹, =0.01s. The system sampling period Ts is set to 5ms.

[0039] Taking sampling time T1 as an example, the measured N2_pv = 27.2N, the setpoint N2_sv = 30.0N, and the previous cycle deviation e_2(i-1) = 2.7985N. Using a quantization factor for scaling, the universe of discourse values ​​used in fuzzy inference are obtained: E = 4.2, EC = 1.2. Where E, EC ∈ [-6, 6]. Inputting (E, EC) = (4.2, 1.2) into the neural network module, assuming that the network calculates and outputs a set of adjustments to the current membership function parameters. For example, for the PM fuzzy subset of the input variable E, the network suggests the center value C of its Gaussian membership function. E PM increased by 0.15, width σ E PM decreases by 0.08; for the PS fuzzy subset of the input variable EC, a center value C is recommended. EC PS increases by 0.1, width σ EC PM decreases by 0.05. Parameters are updated based on the network output and the current learning rate. Assuming before the update... =4.0, =2.0; =2.0, =2.0. After the update: ≈4.0015, =1.9992; =2.001, =1.9995. The membership degree of the current input (E, EC) = (4.2, 1.2) to each relevant fuzzy subset is calculated using the updated membership function. Gaussian membership function is used. μ(x)=exp( ） For example: For E=4.2, calculate its subset of PM (c≈4.0015, ≈1.9992) and the PB subset (assuming c=6.0, Membership degree of (=2.0): m PM (E) ≈0.995, m PB (E) ≈0.641. For EC=1.2, calculate its approximation on the subset of ZE (c=0, =2.0) and PS subset (c≈2.001, Membership degree ≈1.9995: m ZE (EC) ≈0.835, m PS (EC) ≈0.98.

[0040] Query the fuzzy rule base to output Kp Here are a few example rules:

[0041]

[0042]

[0043] in, This indicates the output fuzzy value. The activation degree of each rule is calculated using the minimum operator:

[0044]

[0045]

[0046] The centroid method is used to calculate the accurate output values. Assume that the single-point values ​​of the output fuzzy subsets “PS”, “ZE”, and “NS” on the output universe of discourse [-6, 6] are respectively O ps =+2, O ZE =0, O NS =-2. Then Kp The sharpness value in the fuzzy domain is:

[0047] Results can be obtained ≈0.158. Map the output value of the fuzzy universe of discourse to the actual PID parameter correction universe of discourse. For Kp2 Its actual universe of discourse is [-0.5, 0.5], and the corresponding output scaling factor is approximately 0.0833. Therefore Kp2 =0.0132. The PID controller parameters are updated using the obtained correction.

[0048] Based on the updated PID parameters, the digital PID controller of this loop calculates and outputs control commands to the main drafting roll driver according to the current deviation.

[0049] During operation, the controller uses the sum of squared tension deviations from the most recent 100 cycles as the loss function every L=100 control cycles. By fine-tuning the weights of the neural network modules through the backpropagation algorithm, the controller can continuously optimize its parameter adjustment strategy to adapt to changes in the production process.

[0050] On the PA6 production line with a diameter of 0.2mm, the control effects of traditional fixed-parameter PID control, fuzzy adaptive PID control, and the neural network fuzzy adaptive PID control of this invention were compared and tested. The test conditions included the start-up process from a standstill to 500m / min and subsequent steady-state operation. The comparison results are as follows:

[0051] Where OS is the maximum overshoot. The peak value of tension N2, =30N. Traditional PID: OS≈25%; Fuzzy adaptive PID: OS≈8%; This invention: OS≈4.5%.

[0052] Ts enters and remains in response. The time within ±2% of the band is as follows: Traditional PID: Ts≈12s; Fuzzy adaptive PID: Ts≈3s; This invention: Ts≈2s; 10 wire breakage tests: Traditional PID: 4 times; Fuzzy adaptive PID: 0 times; This invention: 0 times.

[0053]

[0054] To measure the standard deviation of tension fluctuation, one minute of high-frequency data (Z=60000) was collected. Traditional PID: ≈0.45N; Fuzzy adaptive PID: ≈0.15N; This invention: ≈0.09N.

[0055]

[0056] CV is the coefficient of variation of the diameter of a single filament. Traditional PID: CV≈1.5%; Fuzzy adaptive PID: CV≈0.8%; This invention: CV≈0.5%.

Claims

1. A tension collaborative control system for a monofilament production line based on neural network fuzzy adaptive PID, characterized by: A distributed detection device acquires in real-time tension measurements, linear speeds of each drive roller, monofilament diameters, and winding diameters at at least four key process points on the production line. An actuator independently drives the traction roller, drafting roller, heat-setting roller, and winding device. An intelligent controller communicates with the detection device and the actuator. The intelligent controller is configured to: calculate the real-time tension deviation *e* and the rate of change of deviation *ec* at each key process point based on the acquired tension measurements and their setpoints, and input *e* and *ec* into a neural network fuzzy inference engine. The neural network fuzzy inference engine performs inference based on its internally trained and optimized fuzzy rule base and membership functions, outputting corrections to the PID controller parameters, including proportional gain corrections. Kp, integral coefficient correction amount Ki and differential coefficient correction amount Kd; the parameters of the corresponding PID controller are dynamically updated online using the correction amount, and control commands for the actuator are generated based on the updated parameters to adjust the torque or speed of the corresponding transmission unit.

2. The tension collaborative control system for a monofilament production line based on neural network fuzzy adaptive PID according to claim 1, characterized in that, The neural network fuzzy inference engine consists of a feedforward neural network coupled with a fuzzy inference engine. The inputs of the feedforward neural network are the tension deviation e and the deviation change rate ec, and its output is used to dynamically adjust the parameters of the membership function or the rule weights in the fuzzy inference engine. The fuzzy inference engine performs inference based on the adjusted membership function and the rule base, and outputs the PID parameter correction amount.

3. The tension collaborative control system for a monofilament production line based on neural network fuzzy adaptive PID according to claim 2, characterized in that, The feedforward neural network module has 1 to 3 hidden layers. The activation function of the hidden layers is the Tanh function, and the activation function of the output layer is a linear function.

4. The tension collaborative control system for a monofilament production line based on neural network fuzzy adaptive PID according to claim 2, characterized in that, The feedforward neural network module adopts a training method that combines offline and online training: in the offline training stage, historical production process data is used to pre-train the network and initially establish the mapping relationship between the input (e, ec) and the fuzzy parameter adjustment amount; During the online operation phase, the feedforward neural network module continuously fine-tunes its network weights at a preset learning rate using the error backpropagation algorithm based on real-time control performance indicators.

5. The tension collaborative control system for a monofilament production line based on neural network fuzzy adaptive PID according to claim 1, characterized in that, The distributed detection device includes at least four tension sensors, which are respectively set at the following process locations: between the cooling device and the first traction roller, between the key drafting rollers, after the final heat setting roller, and at the entrance of the winding device.

6. The tension collaborative control system for a monofilament production line based on neural network fuzzy adaptive PID according to claim 1, characterized in that, The distributed detection device also includes a laser diameter gauge for online continuous measurement of the diameter of a single filament. The intelligent controller is further configured to adjust the tension setting value of the drawing zone by feedforward or feedback compensation based on the deviation between the real-time diameter value measured by the laser diameter gauge and the target diameter value.

7. The tension collaborative control system for a monofilament production line based on neural network fuzzy adaptive PID according to claim 1, characterized in that, The rule base of the neural network fuzzy inference engine also includes cooperative control rules. The fuzzy antecedent of the cooperative control rules includes the tension deviation e and the deviation change rate ec of two different tension control loops. The fuzzy consequent of the cooperative control rules outputs the correction amount of the PID controller parameters corresponding to the two loops respectively.

8. A tension collaborative control method for a monofilament production line based on neural network fuzzy adaptive PID, characterized in that, Includes the following steps: Step 1: The distributed detection device is used to collect tension signals at each key process point in real time, and calculates the real-time deviation e and the real-time deviation change rate ec between the actual tension value and the set value at each point. Step 2: Input the e and ec corresponding to each control loop into the neural network fuzzy inference engine; the feedforward neural network module in the neural network fuzzy inference engine processes the input and outputs the adjustment amount used to adjust the internal parameters of the associated fuzzy inference engine, including the center value and width of the membership function and the weight of the fuzzy rule; Step 3: Based on the adjusted parameters, the fuzzy inference engine fuzzifies the precise quantities e and ec, and performs inference and defuzzification operations according to the optimized fuzzy rule base to obtain the correction amount for the PID controller parameters of each loop. Kp , Ki , Kd ; Step 4: Update PID controller parameters online: ； in, , , These are the initial parameters of the corresponding PID controller or the parameters of the previous control cycle; Step 5: Using the updated parameters, each PID controller calculates and outputs control commands to the corresponding actuators; Step 6: Based on the deviation between the real-time diameter of the monofilament measured online and the target diameter, dynamically adjust the tension setting value of the drawing zone to achieve closed-loop control of the monofilament diameter; Step 7: Periodically construct a loss function based on recent control performance indicators, and fine-tune the weights of the feedforward neural network module online using the error backpropagation algorithm to achieve continuous self-learning and self-adaptation of the controller.

9. The tension collaborative control method for a monofilament production line based on neural network fuzzy adaptive PID according to claim 8, characterized in that, The online learning described in step seven employs an incremental learning strategy, where the learning rate η adaptively decreases as the control system stabilizes. Its update formula is: ； in, The learning rate for the current control cycle. The attenuation coefficient is... It is the moving average of the absolute values ​​of tension deviation over several recent control cycles.

Images