Synchronous control system and method for vibration suppression and quality monitoring of twisted wire

Abstract

The invention belongs to the technical field of synchronous control, and particularly relates to a synchronous control system and method for vibration suppression and quality monitoring of twisted pairs, and the method comprises the following steps: S1, obtaining the measurement data of an acceleration sensor and a laser displacement sensor array on a twisted pair production line, a real-time vibration vector field of the twisted pair in the monitoring area is constructed through data fusion; based on the real-time vibration vector field, utilizing a Hamiltonian agent model to predict the vibration energy state of the twisted pair in a future prediction time window; and S2, adjusting a weight coefficient for minimizing the weighted sum of the predicted vibration energy and the control increment norm in real time according to the twisting quality characteristic index calculated in the previous control period. According to the method, conversion from passive response to active predictive suppression is realized, optimal control measures can be taken before vibration deterioration, and the timeliness of vibration suppression is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of synchronous control, and particularly relates to a synchronous control system and method for suppressing and monitoring the vibration of a pair of twisted wires. BACKGROUND

[0002] Due to the influence of multiple complex factors such as core wire pay-off tension fluctuation, guide wheel slight defects, equipment mechanical vibration and air flow disturbance, the pair of twisted wires will inevitably vibrate in the high-speed movement path from the twisting point to the take-up reel. The vibration will not only seriously affect the uniformity and stability of the twisting pitch, causing fluctuations in key electrical performance indicators such as characteristic impedance, but also may cause micro-damage to the core wire insulation layer under deformation, burying long-term reliability risks.

[0003] Suppressing the cable vibration in the production process is a core technical problem for improving the product quality and consistency of the pair of twisted wires. Moreover, vibration suppression and quality monitoring are usually two independent links, and quality evaluation mostly relies on offline sampling after production, which cannot provide real-time guidance for the vibration control strategy in the production process, thus missing the best opportunity for online optimization and adjustment. The prior art lacks a synchronous control method that integrates the prediction of distributed vibration of the pair of twisted wires, the stability of the twisting process and the optimization of the control strategy, and it is difficult to realize real-time monitoring and active protection of the quality of the pair of twisted wires while ensuring production efficiency. SUMMARY

[0004] The application provides a synchronous control system and method for suppressing and monitoring the vibration of a pair of twisted wires, to solve the technical problem of lacking real-time control of the integration of prediction-stability-optimization of the vibration of the pair of twisted wires in the prior art.

[0005] In a first aspect, the application provides a synchronous control method for suppressing and monitoring the vibration of a pair of twisted wires, comprising the following steps: S1, obtaining the measurement data of an acceleration sensor and a laser displacement sensor array arranged on a pair of twisted wire production line, constructing a real-time vibration vector field of the pair of twisted wires in a monitoring area through data fusion; based on the real-time vibration vector field, predicting the vibration energy state of the pair of twisted wires in a future prediction time window using a Hamiltonian proxy model; S2, adjusting a weight coefficient aiming to minimize the weighted sum of the predicted vibration energy and the control increment norm in real time according to the twisting quality characteristic index calculated in the last control period, and calculating the optimal suppression control sequence of the current control period based on the weight coefficient; driving the vibration suppression device according to the optimal suppression control sequence; S3, collecting residual vibration data of the pair of twisted wires after the vibration suppression effect, calculating the bispectrum entropy of the residual vibration data of the pair of twisted wires, and taking the bispectrum entropy as a twisting quality characteristic index representing the stability of the current twisting process, which is used for adjusting the weight coefficient in the next control period; S4, using the deviation between the stranded wire residual vibration data and the Hamiltonian proxy model prediction value, updating the internal state parameters of the Hamiltonian proxy model online through extended Kalman filtering; when the stranded mass characteristic index exceeds the preset threshold, the prediction time window of the Hamiltonian proxy model is shortened.

[0006] Further, the measurement data of the acceleration sensor and the laser displacement sensor array arranged on the stranded wire production line are acquired, and a real-time vibration vector field of the stranded wire in the monitoring area is constructed through data fusion, including: Five acceleration sensors are equidistantly arranged along the stranded wire axial direction at an interval of 50 cm, and a set of mutually orthogonal laser displacement sensors are arranged at the center position of the monitoring area; the sensor data is synchronously collected at a sampling frequency of 1000 Hz, the displacement after integration of the acceleration data is fused with the displacement measured by the laser displacement sensor through a multi-rate Kalman filtering algorithm; the discrete measurement point data after fusion is constructed into a real-time vibration vector field with a spatial resolution of 5 mm in the monitoring area through cubic spline interpolation.

[0007] Further, based on the real-time vibration vector field, the vibration energy state of the stranded wire in the future prediction time window is predicted using the Hamiltonian proxy model, including: The real-time vibration vector field is input into a pre-trained deep neural network proxy model for fitting the Hamiltonian energy function of the stranded wire as the initial generalized coordinates and generalized momenta; the deep neural network proxy model is time-integrated using symplectic geometry algorithm, the integration step is set to 1 ms, and the vibration state in the future 100 ms is predicted forward; the kinetic energy and potential energy of all spatial nodes in the prediction time window are summed to calculate the total vibration energy state.

[0008] Further, according to the stranded mass characteristic index calculated in the last control period, a weight coefficient aiming to minimize the weighted sum of the predicted vibration energy and the control increment norm is adjusted in real time, and the optimal suppression control sequence of the current control period is calculated based on the weight coefficient, including: According to the formula Adjust the weight coefficient, where W(t) is the weight coefficient of the current control period, is the basic weight, Q(t-1) is the stranded mass characteristic index of the last control period, is the mass index threshold, and S is the sensitivity coefficient; the objective function is constructed and solved, where is the predicted vibration energy, is the control increment sequence of the current control period; the interior point method is used to complete the solution within 10 ms to obtain the optimal suppression control sequence.

[0009] Further, the residual vibration data of the pair of twisted wires after the vibration suppression is collected, and a bispectrum entropy of the residual vibration data of the pair of twisted wires is calculated, and the bispectrum entropy is taken as a twisted quality characteristic index representing the stability of the current twisting process, including: At the end of each control cycle, a residual vibration time series with a length of 1024 data points is collected from the laser displacement sensor at the center position; the residual vibration time series is divided into 4 overlapping 50% sub-sections and a Hamming window is added, and the third-order cumulant is calculated by using an indirect method; the bispectrum estimation is obtained by two-dimensional fast Fourier transform of the third-order cumulant ; after the bispectrum amplitude is normalized as a probability density function, the bispectrum entropy is calculated according to the Shannon entropy formula.

[0010] Further, the deviation between the residual vibration data of the pair of twisted wires and the predicted value of the Hamiltonian proxy model is used to update the internal state parameters of the Hamiltonian proxy model online by using an extended Kalman filter, including: The neural network weights and biases of the Hamiltonian proxy model are taken as the state vector of the extended Kalman filter; the collected residual vibration data of the pair of twisted wires are taken as the measurement; the difference between the predicted value of the Hamiltonian proxy model based on the state at the last moment and the measurement of the residual vibration of the pair of twisted wires at the current moment is taken as the innovation; the measurement equation is linearized by calculating the Jacobian matrix of the Hamiltonian proxy model output to the state vector; and the Kalman gain is calculated and the state vector is updated according to the standard update steps of the extended Kalman filter, so as to complete the online adjustment of the neural network weights and biases.

[0011] Further, when the twisted quality characteristic index exceeds a preset threshold, the prediction time window of the Hamiltonian proxy model is shortened, including: When the calculated bispectrum entropy value is greater than the preset threshold of the twisted quality characteristic index for 3 consecutive control cycles, the length of the prediction time window of the Hamiltonian proxy model is shortened from 100 ms to 50 ms; and when the bispectrum entropy value returns to below the preset threshold of the twisted quality characteristic index for 10 consecutive control cycles, the length of the prediction time window is restored to 100 ms.

[0012] Further, in S2, the vibration suppression device is an electromagnetic actuator or a piezoelectric ceramic sheet installed on the path of the pair of twisted wires.

[0013] Further, in S4, the internal state parameters of the Hamiltonian proxy model include an equivalent tension and a damping coefficient.

[0014] In a second aspect, the present application provides a synchronous control system for the vibration suppression and quality monitoring of a pair of twisted wires, including a memory and a processor, and the memory stores computer program instructions, which, when executed by the processor, realize the above-mentioned synchronous control method for the vibration suppression and quality monitoring of a pair of twisted wires.

[0015] The beneficial effects are: the application can obtain the vibration form of the cable in the whole monitoring area by constructing the real-time vibration vector field of the stranded wire, predict the future vibration energy by using the Hamiltonian proxy model, realize the transformation from passive response to active predictive suppression, take the optimal control measures before the vibration deteriorates, and improve the timeliness of vibration suppression. The bi-spectrum entropy is used as the quality characteristic index of stranding, the instability precursor in the stranding process can be obtained, and the quality index is directly fed back to optimize the weight of the control strategy, a direct closed-loop correlation between vibration suppression and stranding quality is established, and the control target directly serves the improvement of product quality. In addition, by updating the model parameters online and adjusting the prediction range according to the quality state, the stranding wire vibration suppression and product quality are simultaneously monitored and ensured. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The flowchart of the synchronous control method for stranded wire vibration suppression and quality monitoring. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0018] The embodiments of the synchronous control method for stranded wire vibration suppression and quality monitoring provided by the application include the following steps: As shown in the accompanying drawings, Figure 1 The synchronous control method for stranded wire vibration suppression and quality monitoring includes the following steps: S1, the measurement data of the acceleration sensor and the laser displacement sensor array arranged on the stranded wire production line are obtained, and the real-time vibration vector field of the stranded wire in the monitoring area is constructed through data fusion; based on the real-time vibration vector field, the vibration energy state of the stranded wire in the future prediction time window is predicted by using the Hamiltonian proxy model.

[0019] Specifically, a plurality of laser displacement sensors and acceleration sensors are arranged along the length direction of the paired wire on the path from the stranding point to the take-up reel, the laser displacement sensors are used to measure the transverse displacement of a plurality of discrete points on the paired wire, and the acceleration sensors are used to measure the transverse acceleration of the discrete points; a controller collects real-time data of all sensors at a fixed sampling frequency; a mathematical function representing the continuous vibration pattern of the paired wire in the entire monitoring area is fitted by using modal analysis or a spatial interpolation algorithm such as cubic spline interpolation, the independent variables of the mathematical function are time and spatial position, and the dependent variables are the displacement and velocity vectors of the point, thereby forming a continuously changing real-time vibration vector field; the real-time vibration vector field constructed above is input as an initial condition into a pre-trained Hamiltonian agent model, the Hamiltonian agent model is preferably a deep neural network capable of learning and complying with the law of conservation of energy, the Hamiltonian agent model includes two parts, one is a deep neural network used as an energy function approximator, the input is the current state, and the output is a scalar value, i.e., the Hamiltonian value of the system, and the other is a numerical integrator satisfying the Hamiltonian canonical equation, which calculates and predicts the state sequence and energy change in the future, for example, in a time window, by using the energy gradient. More specifically, the Hamiltonian agent model is a Hamiltonian neural network, which iteratively calculates forward in time by learning the dynamics of the paired wire, outputs the spatial vibration pattern of the paired wire at each time in the future, for example, in the next 100 ms, and then calculates the total kinetic energy and total potential energy at each time according to the predicted displacement and velocity, and the sum of the two is the predicted vibration energy state.

[0020] In an alternative embodiment, the measurement data of the acceleration sensor and the laser displacement sensor array arranged on the paired wire production line are acquired, and a real-time vibration vector field of the paired wire in the monitoring area is constructed through data fusion, including: Five acceleration sensors are equidistantly arranged along the axial direction of the paired wire at an interval of 50 cm, and a set of mutually orthogonal laser displacement sensors are arranged at the center position of the monitoring area; the data of each sensor is synchronously collected at a sampling frequency of 1000 Hz, the displacement after integration of the acceleration data is fused with the displacement measured by the laser displacement sensor by using a multi-rate Kalman filtering algorithm; and the fused discrete measurement point data is constructed into a real-time vibration vector field with a spatial resolution of 5 mm in the monitoring area by cubic spline interpolation.

[0021] In a monitoring area with a length of 2 m, an acceleration sensor is installed every 50 cm from the starting point, a total of 5, for measuring the normal acceleration of the paired wire. At the center of the area, i.e., 1 m, two mutually perpendicular laser displacement sensors are installed for measuring the displacement of the point. All sensors collect data at a frequency of 1000 times per second through a synchronous clock, ensuring that the data are aligned in time.

[0022] Next, multi-source data fusion is performed. The acceleration signals collected by the five acceleration sensors are integrated twice to obtain displacement signals, but the signals have low-frequency drift errors. The displacement signals measured by the laser displacement sensor are very accurate in the low-frequency band, but the response is insufficient in the high-frequency band. Using a multi-rate Kalman filtering algorithm, the acceleration integrated displacement is taken as the high-frequency component of the system state, and the laser displacement measurement is taken as the low-frequency calibration reference. The advantages of the two are combined to output displacement time series on five discrete measuring points with high and low frequency accuracy.

[0023] In order to obtain continuous vibration morphology, a cubic spline interpolation method is used to process the fused displacement data of the five discrete measuring points. This method can generate a smooth curve between measuring points, thereby calculating the displacement and velocity of any position in the monitoring area. By performing interpolation calculation with a step size of 5 mm, a high-resolution real-time vibration vector field is constructed, which can finely depict the complete vibration morphology of the paired wire at any time in the entire monitoring area.

[0024] In an optional embodiment, based on the real-time vibration vector field, the Hamiltonian proxy model is used to predict the vibration energy state of the paired wire in the future prediction time window, including: The real-time vibration vector field is input into a pre-trained deep neural network proxy model for fitting the Hamiltonian energy function of the paired wire as the initial generalized coordinates and generalized momentum. The deep neural network proxy model is time-integrated using symplectic geometry algorithm, with an integration step size of 1 ms, to predict the vibration state in the future 100 ms. The kinetic energy and potential energy of all spatial nodes in the prediction time window are summed to calculate the total vibration energy state.

[0025] The real-time vibration vector field data at the current time is processed, and the position information of each spatial node in the monitoring area is taken as the generalized coordinates of the system, and the velocity information is taken as the generalized momentum. The data together constitute the complete phase space state at the current time. The complete phase space state is input into a deep neural network that has been trained with a large amount of simulation or experimental data as the initial condition. The deep neural network can quickly calculate the Hamiltonian, i.e. total energy, under any given state.

[0026] In order to derive future changes, a symplectic geometry integration algorithm with energy conservation characteristics is used. Starting from the initial state at the current time, the algorithm uses the energy gradient calculated by the deep neural network proxy model to iteratively solve the Hamiltonian equation with a fixed step size of 1 ms. Each iteration step obtains the generalized coordinates and generalized momentum of all spatial nodes in the next millisecond. This process is repeated 100 times to obtain the vibration state sequence of the paired wire every millisecond in the future 100 ms from the current time.

[0027] Based on the prediction sequence containing 100 future time state, the total predicted vibration energy is calculated, for each time in the prediction sequence, the kinetic energy and potential energy of all spatial nodes are calculated and summed up to get the total energy of this time. Then the total energy of 100 times is accumulated or integrated to get a scalar value, which is the total vibration energy state in the prediction time window, as the target of subsequent optimization control.

[0028] S2, according to the twisting mass characteristic index calculated in the last control period, real-time adjust a weight coefficient aiming to minimize the weighted sum of predicted vibration energy and control increment norm, and calculate the optimal suppression control sequence of the current control period based on this; drive the vibration suppression device according to the optimal suppression control sequence.

[0029] A cost function is established, the integral of vibration energy in the future prediction time window, that is, the sum of the square of the applied control force, represents the control energy consumption, wherein the weight coefficient of the vibration energy term is not fixed, but is adjusted according to the twisting mass characteristic index obtained in the last period, if the mass index is poor, indicating that the vibration is severe, then increase the weight of the vibration energy term, so that the control is more focused on suppressing vibration; otherwise, reduce it to save energy, then use the model predictive control algorithm to solve a set of control force sequence that minimizes the cost function under this weight through numerical optimization, as the optimal suppression control sequence of the current period.

[0030] The first or first few control instruction values in the calculated optimal suppression control sequence are converted into analog voltage or current signals through a digital-to-analog converter, the signal is sent to a power amplifier for amplification, and then drives the electromagnetic actuator or piezoelectric ceramic sheet and other vibration suppression devices installed on the pair-twisted wire path, the vibration suppression device will generate a physical force or displacement proportional to the instruction value, acting on the pair-twisted wire to offset the impending vibration.

[0031] In an alternative embodiment, according to the twisting mass characteristic index calculated in the last control period, a weight coefficient aiming to minimize the weighted sum of predicted vibration energy and control increment norm is real-time adjusted, and the optimal suppression control sequence of the current control period is calculated based on this, including: According to the formula Adjust the weight coefficient, where W(t) is the weight coefficient of the current control period, is the base weight, Q(t-1) is the twisting mass characteristic index of the last control period, is the mass index threshold, S is the sensitivity coefficient; construct and solve the minimum value of the objective function , where is the predicted vibration energy, The control increment sequence for the current control period; the interior point method is used to solve within 10 ms to obtain the optimal suppression control sequence.

[0032] At the beginning of each control period, read the twisted mass characteristic index Q(t-1) calculated at the end of the last period, for example, a value of 0.8 bispectrum entropy. The preset basic weight is 0.5, the quality index threshold is 0.7, and the sensitivity coefficient S is 0.1. The weight of the current control period is calculated according to the formula , since 0.8>0.7, the exponential term is positive, making greater than the basic weight 0.5. It means that when the twisted mass of the last control period decreases, the punishment for the change of control action is automatically increased, and the output of more gentle control instructions is inclined to ensure stability.

[0033] A target function to be optimized is constructed, which is the weighted sum of two parts, the first part is the future vibration energy predicted by the Hamiltonian proxy model, and the second part is the two-norm square of the control increment sequence to be solved in the current control period, which represents the degree of control action. The calculated weight coefficient is used to balance the two goals of minimizing future vibration and keeping control smooth.

[0034] The convex optimization problem with constraints needs to be solved in a very short time, and the interior point algorithm is used, which can find the optimal control increment sequence that makes the target function J reach the minimum value within the hard real-time constraint of 10 ms.

[0035] S3, collect the residual vibration data of the twisted wire after the vibration suppression action, calculate the bispectrum entropy of the residual vibration data of the twisted wire, and take the bispectrum entropy as the twisted mass characteristic index representing the stability of the current twisting process, which is used for weight coefficient adjustment in the next control period.

[0036] While the vibration suppression device is in action, the sensor array continuously collects vibration data of the twisted wire, which is residual vibration. High-order spectrum analysis is performed on the collected residual vibration time series signal, bispectrum is calculated, which can obtain the nonlinear coupling characteristics existing in the signal, then information entropy is calculated based on the two-dimensional bispectrum distribution, and a single numerical value, bispectrum entropy, is obtained. The higher the value, the stronger the nonlinear and random components in the residual vibration, the more unstable the twisting process, and the worse the quality.

[0037] In an alternative embodiment, the residual vibration data of the pair of twisted wires after the vibration suppression is collected, the bispectrum entropy of the residual vibration data of the pair of twisted wires is calculated, and the bispectrum entropy is taken as a twisting quality characteristic index representing the stability of the current twisting process, comprising: At the end of each control cycle, a residual vibration time series with a length of 1024 data points is collected from the laser displacement sensor at the center position; the residual vibration time series is divided into 4 overlapping 50% sub-sections and a Hamming window is added, and the third-order cumulant is calculated by an indirect method; the third-order cumulant is subjected to two-dimensional fast Fourier transform to obtain a bispectrum estimate ; after the bispectrum amplitude is normalized into a probability density function, the bispectrum entropy is calculated according to the Shannon entropy formula.

[0038] Specifically, immediately after the control instruction is executed, data collection is started. A laser displacement sensor located at the center of the monitoring area continuously records 1024 sampling points of the residual vibration displacement data of the pair of twisted wires, forming an original time series. In order to improve the reliability of subsequent spectrum estimation and reduce the variance, the Welch method is used for data preprocessing, and the 1024 data points are divided into 4 sub-sections with a length of 512 points, and there are 256 overlapping points between adjacent sub-sections. At the same time, a Hamming window is added to each sub-section to suppress spectral leakage.

[0039] For each windowed sub-section data, the third-order cumulant is calculated by an indirect method. The third-order cumulant can reveal the nonlinearity and non-Gaussian characteristics of the signal, and is crucial for judging the stability of the twisting process. The calculated third-order cumulant is subjected to two-dimensional fast Fourier transform to obtain the bispectrum estimate of the sub-section. The bispectrum estimate is a two-dimensional frequency domain function, which can represent the phase coupling strength between different frequency components and . The bispectrum estimates of the 4 sub-sections are averaged to obtain a smoother and more stable bispectrum.

[0040] All elements of the obtained bispectrum amplitude matrix are summed up, and each element value is divided by the sum, so as to normalize the bispectrum amplitude into a two-dimensional probability density function. A flat and divergent bispectrum morphology represents a complex and disordered system, while a sharp and concentrated morphology represents a simple and ordered system. According to the definition of Shannon entropy, the information entropy of the probability density function is calculated to obtain a single value, i.e. the bispectrum entropy. The higher the value, the more unstable the system, which is taken as the twisting quality characteristic index of the current cycle.

[0041] S4, the deviation between the residual vibration data of the pair of twisted wires and the predicted value of the Hamiltonian proxy model is used to update the internal state parameters of the Hamiltonian proxy model online by extended Kalman filtering; when the twisting quality characteristic index exceeds a preset threshold, the prediction time window of the Hamiltonian proxy model is shortened.

[0042] Specifically, the actual measurement of the residual vibration data of the pair of stranded wires is compared with the vibration prediction value of the Hamiltonian surrogate model for the same time period without applying the control action, and the difference between the two is the model prediction error. The error is input into an extended Kalman filter, which takes the parameters representing the physical characteristics of the pair of stranded wires, such as the equivalent tension, the damping coefficient, etc., in the Hamiltonian surrogate model as state variables to be estimated, and automatically calculates and corrects the internal parameters according to the size of the prediction error, so that the prediction result of the model is closer to the actual physical process, thereby realizing online self-correction of the model.

[0043] An alarm threshold of bispectrum entropy is set in advance, for example, 0.8. At the end of each control cycle, the calculated stranded wire quality characteristic index, i.e., bispectrum entropy, is compared with the threshold. If the current bispectrum entropy value is greater than 0.8, it indicates that the stranded process has a serious instability, and the long-term vibration prediction is unreliable at this time. The control system will automatically shorten the prediction time window of the surrogate model from the regular 100 ms to, for example, 30 ms, forcing the model predictive control algorithm to only focus on the short-term suppression effect and adopt a faster and more conservative control strategy. When the bispectrum entropy falls below the threshold, the regular prediction time window is restored.

[0044] In an optional embodiment, the deviation between the residual vibration data of the pair of stranded wires and the prediction value of the Hamiltonian surrogate model is used to update the internal state parameters of the Hamiltonian surrogate model online through the extended Kalman filter, including: The neural network weights and biases of the Hamiltonian surrogate model are taken as the state vector of the extended Kalman filter. The collected residual vibration data of the pair of stranded wires are taken as the measurement. The difference between the prediction value of the Hamiltonian surrogate model based on the state at the last time and the measurement of the residual vibration of the pair of stranded wires at the current time is taken as the innovation. The measurement equation is linearized by calculating the Jacobian matrix of the Hamiltonian surrogate model output to the state vector. The Kalman gain is calculated and the state vector is updated according to the standard update steps of the extended Kalman filter, thereby completing the online adjustment of the neural network weights and biases.

[0045] All the weight and bias parameters of the deep neural network constituting the Hamiltonian surrogate model are extracted and connected into a one-dimensional vector, which is the state vector of the Kalman filter (EKF) and is the object to be estimated and corrected in real time. The actual measurement of the residual vibration data of the pair of stranded wires measured by the sensor is defined as the measurement of the EKF, which represents the true output.

[0046] At each update time, EKF performs a prediction step, i.e. the Hamiltonian proxy model uses the current weight and bias parameters to predict what the residual vibration should be at the current time, given the state at the previous time. The difference between the Hamiltonian proxy model prediction and the actual measurement collected by the sensor is called innovation or residual. The size of the innovation directly reflects the reliability of the current model parameters, if the difference is large, it means that the model needs to be corrected.

[0047] Since the neural network is a nonlinear model, the standard Kalman filter formula cannot be directly applied. The partial derivative of the model output with respect to all weight and bias parameters needs to be calculated to form the Jacobian matrix, which is used to approximate the nonlinear relationship. According to the standard update equation of EKF, the Kalman gain is calculated by combining the prediction error covariance, measurement noise and Jacobian matrix. The Kalman gain is multiplied by the innovation to correct the state vector, i.e. the weight and bias of the neural network, so that it adjusts in the direction of reducing the prediction error, thereby realizing online update.

[0048] In an optional embodiment, when the twisting quality characteristic index exceeds the preset threshold, the prediction time window of the Hamiltonian proxy model is shortened, including: When the calculated bispectrum entropy value is greater than the preset threshold of the twisting quality characteristic index for 3 consecutive control periods, the length of the prediction time window of the Hamiltonian proxy model is shortened from 100ms to 50ms; when the bispectrum entropy value returns to below the preset threshold of the twisting quality characteristic index for 10 consecutive control periods, the length of the prediction time window is restored to 100ms.

[0049] Preferably, the preset threshold of the twisting quality characteristic index, i.e. the bispectrum entropy, is 0.75, and higher than this value indicates that the stability of the twisting process is deteriorating. In order to avoid overreaction to fluctuations in a single period, a continuous counter is used to continuously monitor the bispectrum entropy value calculated in each control period. Only when the value remains above 0.75 for three consecutive periods, for example, 0.77, 0.80, 0.76, will the adjustment mechanism be triggered.

[0050] When the trigger condition is met, an adaptive adjustment action is immediately performed, which shortens the time span of the Hamiltonian proxy model prediction, i.e. the prediction time window, from the regular 100ms to 50ms. Long-term prediction becomes unreliable and risky, and shortening the prediction time window allows the model prediction control MPC controller to focus more on the short-term trend in the near future, which is more certain, and adopt a more conservative and fast response control strategy, with the priority task being to restore stability as soon as possible, rather than pursuing long-term optimal performance.

[0051] After switching to the short prediction window mode, the twisting quality is continuously monitored. To recover from the conservative mode to the regular mode, a recovery condition is also set. It requires that the bispectrum entropy value must be steadily below the 0.75 threshold for 10 consecutive control cycles. Once this condition is met, for example, the measured value is between 0.6 and 0.7 for 10 consecutive times, it is judged that the twisting process has returned to a stable state, and then the length of the prediction window is recovered from 50 ms to the standard 100 ms, so as to return to the normal control mode which pursues the long-term performance optimization.

[0052] Embodiments of the synchronization control system for suppressing wire vibration and monitoring quality provided by the present application: The synchronization control system for suppressing wire vibration and monitoring quality comprises a processor and a memory, and the memory stores computer program instructions which, when executed by the processor, implement the synchronization control method for suppressing wire vibration and monitoring quality described above.

[0053] The synchronization control system for suppressing wire vibration and monitoring quality further comprises a communication interface and other components known to those skilled in the art, and the settings and functions thereof are known in the art, so they will not be described here.

[0054] In the present application, the aforementioned memory can be any tangible medium that contains or stores a program, which can be used by or in conjunction with an instruction execution system, device or apparatus. For example, the computer readable storage medium can be any suitable magnetic storage medium or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random-Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, a module or both. Any such computer storage media can be part of a device or accessible or connectable to a device. Any application or module described in the present application can be implemented using computer readable / executable instructions stored or otherwise held by such computer readable media.

[0055] The above are preferred embodiments of the present application, and do not limit the protection scope of the present application, so: any equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A method for simultaneous control of twist wire vibration suppression and quality monitoring, characterized by, The method comprises the following steps: S1, obtaining measurement data of an acceleration sensor and a laser displacement sensor array arranged on a pair-twisted wire production line, constructing a real-time vibration vector field of the pair-twisted wire in a monitoring area through data fusion; based on the real-time vibration vector field, predicting a vibration energy state of the pair-twisted wire in a future prediction time window by using a Hamiltonian agent model; S2, adjusting a weight coefficient aiming at minimizing a weighted sum of the predicted vibration energy and a control increment norm in real time according to a twisting quality characteristic index calculated in the last control cycle, and calculating an optimal suppression control sequence of the current control cycle based on the weight coefficient; driving a vibration suppression device according to the optimal suppression control sequence; S3, collecting residual vibration data of the pair-twisted wire after the vibration suppression, calculating a bispectrum entropy of the residual vibration data of the pair-twisted wire, and taking the bispectrum entropy as a twisting quality characteristic index representing stability of the current twisting process, which is used for adjusting the weight coefficient in the next control cycle; S4, using a deviation between the residual vibration data of the pair-twisted wire and a predicted value of the Hamiltonian agent model to update internal state parameters of the Hamiltonian agent model online through an extended Kalman filter; when the twisting quality characteristic index exceeds a preset threshold, shortening the prediction time window of the Hamiltonian agent model.

2. The method of claim 1, wherein the method further comprises: Obtaining measurement data of an acceleration sensor and a laser displacement sensor array arranged on a pair-twisted wire production line, constructing a real-time vibration vector field of the pair-twisted wire in a monitoring area through data fusion, comprising: Five acceleration sensors are equidistantly arranged along the axial direction of the pair-twisted wire at an interval of 50 cm, and a set of mutually orthogonal laser displacement sensors are arranged at the center position of the monitoring area; sensor data is synchronously collected at a sampling frequency of 1000 Hz, and displacement after integration of acceleration data and displacement measured by the laser displacement sensor are fused by using a multi-rate Kalman filtering algorithm; the fused discrete measurement point data is constructed into a real-time vibration vector field with a spatial resolution of 5 mm in the monitoring area through cubic spline interpolation.

3. The method of claim 1, wherein the method further comprises: Based on the real-time vibration vector field, the vibration energy state of the pair-twisted wire in the future prediction time window is predicted by using the Hamiltonian agent model, comprising: The real-time vibration vector field is taken as initial generalized coordinates and generalized momenta, and is input into a pre-trained deep neural network agent model used for fitting a Hamiltonian energy function of the pair-twisted wire; the deep neural network agent model is time-integrated by using a symplectic geometry algorithm, the integration step is set to 1 ms, and the vibration state in the future 100 ms is predicted forward; the kinetic energy and potential energy of all spatial nodes in the prediction time window are summed to calculate the total vibration energy state.

4. The method of claim 1, wherein the method further comprises: According to the twisting quality characteristic index calculated in the last control cycle, a weight coefficient aiming at minimizing a weighted sum of the predicted vibration energy and a control increment norm is adjusted in real time, and an optimal suppression control sequence of the current control cycle is calculated based on the weight coefficient, comprising: According to the formula Adjust the weight coefficient, wherein W(t) is the weight coefficient of the current control period, is the basic weight, Q(t-1) is the stranding quality characteristic index of the last control period, is the quality index threshold, and S is the sensitivity coefficient; construct and solve the objective function of the minimum value, wherein is the predicted vibration energy, is the control increment sequence of the current control period; the solution is completed within 10 ms by using the interior point method to obtain the optimal suppression control sequence.

5. The method of claim 1, wherein the method further comprises: Residual vibration data of the pair-twisted wire after the vibration suppression is collected, a bispectrum entropy of the residual vibration data of the pair-twisted wire is calculated, and the bispectrum entropy is taken as a twisting quality characteristic index representing stability of the current twisting process, comprising: At the end of each control cycle, a residual vibration time series with 1024 data points is collected from the laser displacement sensor at the center position; the residual vibration time series is divided into 4 overlapping 50% segments and added with a Hamming window, and the third-order cumulant is calculated by using an indirect method; the bispectrum estimation is obtained by two-dimensional fast Fourier transform of the third-order cumulant ; after the bispectrum amplitude is normalized into a probability density function, the bispectrum entropy is calculated according to the Shannon entropy formula.

6. The method of claim 1, wherein the method further comprises: The deviation between the residual vibration data of the paired strand and the predicted value of the Hamiltonian proxy model is used to update the internal state parameters of the Hamiltonian proxy model online by extended Kalman filtering, including: The neural network weights and biases of the Hamiltonian proxy model are taken as the state vector of the extended Kalman filtering; the collected residual vibration data of the paired strand are taken as the measurement; the difference between the predicted value of the Hamiltonian proxy model based on the state at the last time and the measurement of the residual vibration of the paired strand at the current time is taken as the innovation; the Jacobian matrix of the state vector output by the Hamiltonian proxy model is calculated to linearize the measurement equation; the Kalman gain is calculated and the state vector is updated according to the standard update steps of the extended Kalman filtering, so as to complete the online adjustment of the neural network weights and biases.

7. The method of claim 1, wherein the method further comprises: When the twisting quality characteristic index exceeds the preset threshold, the prediction time window of the Hamiltonian proxy model is shortened, including: When the calculated bispectrum entropy value is greater than the preset threshold of the twisting quality characteristic index for 3 consecutive control periods, the prediction time window length of the Hamiltonian proxy model is shortened from 100ms to 50ms; when the bispectrum entropy value returns to below the preset threshold of the twisting quality characteristic index for 10 consecutive control periods, the prediction time window length is restored to 100ms.

8. The method of claim 1, wherein the method further comprises: In S2, the vibration suppression device is an electromagnetic actuator or a piezoelectric ceramic sheet installed on the paired strand path.

9. The method of claim 1, wherein the method further comprises: In S4, the internal state parameters of the Hamiltonian proxy model include the equivalent tension and the damping coefficient.

10. A system for simultaneous control of twist wire vibration suppression and quality monitoring, characterized by The device comprises a memory and a processor, and the memory stores computer program instructions, which, when executed by the processor, realize the synchronous control method for paired strand vibration suppression and quality monitoring according to any one of claims 1-9.

Images