An adaptive control system for the operating state of a twisting apparatus

Abstract

This application belongs to the field of twisting equipment control technology, and particularly relates to an adaptive control system for the operating state of twisting equipment. The system includes: acquiring operating signals from the twisting equipment; constructing a state observation sequence using a sliding window and extracting phase lag, vibration envelope skewness, tension gradient, spectral energy difference, sign change density, and multi-channel correlation disorder; dividing the operating state to establish a discrete state set; statistically analyzing the dwell length of each discrete state to construct a semi-Markov transfer model; combining the winding diameter range and tension gradient with piecewise correction to obtain a time-varying state transfer model; adjusting the prediction time domain length based on the expected remaining dwell time of each discrete state; predicting future state paths and solving the objective function to obtain the control increment; and outputting the control increment to the actuator to complete the operating state control. This application can improve the adjustment accuracy and reliability of the operating state control of twisting equipment.

Description

Technical Field

[0001] This application relates to the field of twisting equipment control technology, and in particular to an adaptive control system for the operating status of twisting equipment. Background Technology

[0002] Twisting equipment plays a core role in yarn twisting and forming during spinning production. The stability of process parameters such as spindle speed, tension, and twist directly determines the yarn quality. In actual production, as the yarn bobbin winding diameter gradually increases from an empty bobbin to a full bobbin, the yarn air pocket shape continuously changes, resulting in tension fluctuations and mechanical vibrations that constitute typical multi-source time-varying disturbances. The phase lag relationship between spindle speed and tension signals differs at different winding stages, and the envelope characteristics of vibration amplitude exhibit different distribution patterns under normal operation and abnormal impact conditions. These characteristics make the operating state of twisting equipment non-stationary, placing high demands on the real-time adaptability of the control system.

[0003] Combining Markov transition system theory with model predictive control (MMC) techniques for the operational state control of mechanical equipment is one of the main technological approaches in this field. This type of technology utilizes Markov chains to represent the stochastic switching process of a system between multiple discrete operating states and performs pre-emptive compensation control by predicting future state paths.

[0004] In the prior art, Chinese patent document CN107829181B discloses an electric tensioner, a twisting machine, and a tension control method. This method achieves yarn tension adjustment through the damped operation and reverse torque control of the electric tensioner, gradually increasing the reverse torque to reduce the diameter of the air ring after the air ring stabilizes. This solution focuses on open-loop and semi-closed-loop switching control at the tension actuator level, without addressing the modeling and prediction of equipment operating status driven by multi-dimensional signal feature fusion.

[0005] However, conventional Markov models typically assume that state transition probabilities remain constant throughout operation, failing to reflect the differences in processing stages across different winding diameter ranges and failing to track short-term disturbances during sudden changes in tension gradients. This leads to deviations between state transition predictions and actual operating conditions under long-term operating conditions. In predictive control, existing schemes use a fixed-length prediction time domain, which cannot be dynamically adjusted according to the actual dwell expectations of different discrete states. When the equipment is about to undergo a state transition, an excessively long prediction time domain introduces distortion information; when the equipment is running smoothly, an excessively short prediction time domain limits the global optimization accuracy, resulting in lagging or conservative control actions. Under extreme high-vibration conditions, the equipment vibration envelope exhibits an impact-driven asymmetric distribution, and the cooperative constraint capability between multi-channel signals decreases. Existing control schemes, when solving objective functions that include multiple constraints such as spindle speed, tension, and twist deviation, use a fixed control increment rate constraint strength. This cannot dynamically tighten the control action amplitude according to the severity of vibration. The rigid constraint mechanism may produce abrupt control outputs during equipment state switching, which not only fails to suppress vibration but also exacerbates tension instability and system oscillations, thus restricting the control accuracy and operational reliability of the twisting equipment. Summary of the Invention

[0006] To address the technical problems of insufficient adaptability of existing state transition models for twisting equipment, control lag due to fixed prediction time domain, and constraint rigidity under high vibration conditions, this application provides an adaptive control system for the operating state of twisting equipment, including: A module is established to collect the operating signals of the twisting equipment. A state observation sequence is constructed according to a sliding window. Based on the time-domain statistical characteristics and frequency-domain energy distribution of the operating signals, phase lag, vibration envelope skewness, tension gradient, spectral energy difference, sign change density, and multi-channel correlation disorder are extracted. The operating states of the twisting equipment are divided by combining the spectral energy difference, sign change density, multi-channel correlation disorder, phase lag, and vibration envelope skewness to establish a discrete state set. The solution module is used to construct a semi-Markov transition model by statistically analyzing the dwell length of each discrete state in the discrete state set. It combines the winding diameter range and tension gradient to perform piecewise correction on the state transition probability to obtain a time-varying state transition model. It adjusts the prediction time domain length according to the expected remaining dwell time of each discrete state, uses the time-varying state transition model to predict the future state path, and solves the objective function that satisfies the spindle speed constraint, tension constraint, twist deviation constraint and control increment change rate constraint to obtain the control increment. The control module is used to adjust the constraint strength of the control increment change rate based on the vibration envelope skewness and multi-channel correlation disorder when the future state path contains a high vibration state, and output the control increment to the actuator to complete the control of the twisting equipment's operating state.

[0007] This application acquires multidimensional operating signals from a twisting device and extracts phase hysteresis, vibration envelope skewness, tension gradient, spectral energy difference, sign change density, and multi-channel correlation disorder. It then uses these multidimensional features to classify operating states and establish a discrete state set. The state identification process simultaneously incorporates vibration envelope asymmetry and multi-channel degradation features, expanding the identification coverage for non-stationary conditions. A semi-Markov transition model is constructed by statistically analyzing the discrete state dwell time. This model combines the winding diameter range and tension gradient to piecewise correct the state transition probability, resulting in a time-varying state transition model. This ensures that the transition probability is updated synchronously with the gradual change in processing stage and the sudden change in tension. Based on the remaining dwell time... The prediction time domain length is adjusted according to the expected condition. When the expected remaining dwell time is lower than the preset threshold, the prediction time domain is shortened to improve the control response speed. The prediction time domain is kept at a longer level when the expected remaining dwell time is greater than or equal to the preset threshold. When the future state path contains a high vibration state, the control increment change rate constraint strength is tightened according to the vibration envelope skewness and the multi-channel correlation disorder, so that the change rate of the control increment is limited to the upper bound after tightening, and the output control increment is smooth, reducing the risk of secondary impact on the transmission system caused by high frequency adjustment. The control increment after the above processing is output to the actuator to complete the feedback adjustment of spindle speed, tension and twist deviation.

[0008] Preferably, the operating signals include spindle speed signal, tension signal, twist deviation signal, current signal, vibration amplitude signal, and winding diameter signal.

[0009] Preferably, when the establishment module calculates the phase hysteresis, vibration envelope skewness, and tension gradient: Cross-correlation analysis is performed on the tension signal and spindle speed signal within the sliding window. The time delay corresponding to the peak value of the cross-correlation function is extracted and converted into the phase lag between the tension signal and the spindle speed signal by combining the main frequency period of the spindle speed signal. The envelope of the vibration amplitude signal within the sliding window is extracted using Hilbert transform. The third central moment of the envelope amplitude distribution is calculated, and the vibration envelope skewness is obtained by dimensionless processing of the cube of the standard deviation. Perform a first-order forward difference operation on the tension signal sequence within the sliding window, and divide the difference result by the sampling time interval to obtain the tension gradient.

[0010] The phase lag was calculated by cross-correlation analysis, the vibration envelope skewness was extracted by Hilbert transform, and the tension gradient was obtained by first-order forward difference operation. These three indicators respectively characterize the temporal coupling delay between signals, the impact asymmetry of the vibration envelope, and the degree of instantaneous change in tension, providing differentiated numerical ranges for different fault modes in subsequent state classification.

[0011] Preferably, when the module calculates the spectral energy difference, symbol change density, and multi-channel correlation disorder and establishes the discrete state set: Perform a Fast Fourier Transform on the signals of adjacent sliding windows, and use the difference in the integral area of ​​the power spectral density of the corresponding frequency band as the spectral energy difference; The symbol change density is defined as the ratio of the number of symbol flips in the differential sequence of each dimension of the signal to the sequence length. A multidimensional state matrix is ​​constructed from the normalized ingot speed signal, tension signal and vibration amplitude signal, and the Shannon entropy of the covariance matrix eigenvalues ​​is used to represent the multi-channel correlation disorder. The above features, along with phase hysteresis and vibration envelope skewness, are used to construct a feature vector and cluster them. Discrete states where the vibration envelope skewness exceeds a set threshold are marked as high vibration states.

[0012] The frequency domain energy transfer between adjacent time periods is characterized by calculating the spectral energy difference using Fast Fourier Transform. The symbol change density, which reflects the signal oscillation frequency, is obtained through differential sequence symbol flipping statistics. The multi-channel correlation disorder is calculated using the Shannon entropy of the covariance matrix eigenvalues, which numerically represents the degree of cooperative degradation between multiple physical channels. Furthermore, a discrete state set is established by clustering the phase lag and vibration envelope skew using a Gaussian mixture model. This allows the state division to simultaneously perceive frequency domain energy changes and multi-channel coupling degradation characteristics, thereby improving the sensitivity of abnormal state identification.

[0013] Preferably, when the solver module constructs a semi-Markov transition model and obtains a time-varying state transition model: The continuous dwell time of each discrete state in the discrete state set in the statistical observation history is used to fit the probability density function of the dwell time of each discrete state using the Weibull distribution, and a basic semi-Markov transfer model is constructed. The winding diameter signal is divided into multiple preset winding diameter intervals, and a slow-varying transfer probability benchmark matrix is ​​set for different winding diameter intervals. Calculate the ratio of the absolute value of the tension gradient at the current moment to a preset threshold to generate a rapid change correction factor; By multiplying the off-diagonal elements of the slow-varying transition probability baseline matrix with the fast-varying correction factor and normalizing the result, the state transition probabilities in the basic semi-Markov transition model are updated with weights to obtain the time-varying state transition model.

[0014] A semi-Markov transition model is constructed by fitting the dwell length using a Weibull distribution. A slow-varying transition probability baseline matrix is ​​set based on the winding diameter range, and a fast-varying correction factor is generated using the ratio of the tension gradient to a preset threshold to weight and update the off-diagonal elements. The state transition probability follows the switching of the winding diameter range in the long-period dimension and responds to the sudden change of the tension gradient in the short-time dimension. The two types of corrections are applied to the slow-varying baseline matrix and the fast-varying correction factor, respectively.

[0015] Preferably, when the solution module adjusts the prediction time domain length and solves the objective function to obtain the control increment: The conditional expectation of the remaining dwell time is calculated based on the current dwell time and Weibull distribution of the discrete state. If the conditional expectation is less than a preset threshold, the prediction time domain length is reduced, and vice versa. The future state path is generated by Monte Carlo uniform sampling through a time-varying state transition model, and the future state path is mapped to a continuous prediction sequence based on the mean rate of change of each discrete state. The objective function for minimizing the control increment is constructed by weighting the sum of squares of spindle speed deviation, sum of squares of tension deviation, sum of squares of twist deviation, and penalty term for the rate of change of control increment. The control increment is obtained by solving the problem within the constraints of physical limits and rate of change.

[0016] The prediction time domain length is adaptively adjusted by the remaining residence time condition expectation. The future state path is generated by Monte Carlo sampling and mapped to a continuous prediction sequence. The control increment is obtained by solving the minimization objective function within the physical limit and rate of change constraints. This allows the predictive controller to shorten the prediction time domain to improve the response speed when the remaining residence time condition expectation is low, and to use a longer prediction time domain when the remaining residence time is sufficient to obtain better multi-objective solution results.

[0017] Preferably, when the control module adjusts the constraint strength of the control increment change rate and completes the control of the twisting equipment's operating status: when the probability of a high vibration state in the future state path exceeds the safety threshold, the vibration envelope skewness and multi-channel correlation disorder of the current sliding window are extracted. Based on vibration envelope skewness and multi-channel correlation disorder, a mapping function in the form of negative exponential decay is constructed, and constraint coefficients are calculated. The constraint coefficients are used to reduce the allowable upper bound of the rate of change of control increment in the minimization objective function, thereby increasing the constraint strength and resolving to obtain the smoothed control increment. The control increment is converted into a voltage or current drive signal for the actuator, and output to the motor and tension adjustment device of the twisting equipment to complete the control of the twisting equipment's operating status.

[0018] When the future state path contains a high vibration state, the constraint coefficient is calculated by a mapping function in the form of negative exponential decay and the allowable upper bound of the control increment change rate is reduced. This allows the controller to tighten the control action amplitude according to the constraint coefficient under high vibration conditions and output a smoothed control increment, thereby reducing the risk of mechanical resonance caused by violent acceleration and deceleration.

[0019] Preferably, in the minimization objective function, the sum of squares of spindle speed deviation, the sum of squares of tension deviation, the sum of squares of twist deviation, and the penalty term for the rate of change of control increment correspond to differentiated penalty weights. The penalty weight of the sum of squares of twist deviation is greater than the penalty weight of the sum of squares of tension deviation, and the penalty weight of the sum of squares of tension deviation is greater than the penalty weight of the sum of squares of spindle speed deviation. This ensures that the process indicators of twist and tension are kept stable when solving the objective function.

[0020] Preferably, in the calculation of the constraint coefficient, the difference between the vibration envelope skewness and the multi-channel correlation disorder degree exceeding their respective basic thresholds is calculated by the maximum value function. The difference exceeding their respective basic thresholds is multiplied by the corresponding penalty adjustment coefficient and then fused by negative exponential. The value of the constraint coefficient is limited to between a preset protection lower bound and 1.

[0021] The constraint coefficient is calculated by stripping the basic threshold through the maximum value function and then weighting it with a negative exponent after penalty adjustment coefficient. The constraint coefficient is limited to between the preset protection lower bound and 1, which ensures that the constraint adjustment only intervenes when the vibration envelope skewness and multi-channel correlation disorder exceed the safe range. At the same time, the setting of the protection lower bound ensures that the controller retains a minimum adjustment capability under extreme and harsh conditions.

[0022] Preferably, the probability of a high-vibration state occurring in the future state path is the ratio of the number of paths containing high-vibration states to the total number of paths generated by Monte Carlo sampling.

[0023] The technical solution of this application has the following beneficial technical effects: This application establishes a discrete state set by collecting multi-dimensional operating signals of the twisting equipment and extracting phase lag, vibration envelope skewness, tension gradient, spectral energy difference, sign change density, and multi-channel correlation disorder. A semi-Markov transition model based on Weibull distribution is constructed and a time-varying state transition model is obtained by combining the winding diameter range and tension gradient for piecewise correction. This allows the state transition probability to follow the gradual trend of the processing stage and the short-term tension disturbance update, improving the prediction accuracy of the future state path. The prediction time domain length is expected to be adjusted according to the remaining residence time condition, and the objective function is solved under multiple physical constraints to obtain the control increment, thereby improving the adjustment accuracy and stability of the twisting equipment's operating state control.

[0024] Furthermore, when the future state path includes a high vibration state, the constraint coefficient is calculated by negative exponential decay mapping based on the vibration envelope skewness and multi-channel correlation disorder. The allowable upper bound of the control increment change rate is tightened and the smoothed control command is obtained by resolving the problem. This reduces the cumulative duration of the high vibration state and the yarn breakage rate, and improves the reliability of the twisting equipment's operating status control and the yarn quality. Attached Figure Description

[0025] Figure 1 The waveform diagram of the vibration signal and its envelope in the time domain is shown. Figure 2 This is a graph showing the mapping between the tension gradient and the rapid change correction factor. Figure 3 This is a comparison chart of performance indicators for multiple schemes. Detailed Implementation

[0026] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0027] An adaptive control system for the operating status of a twisting device includes a setup module, a solution module, and a control module.

[0028] A module is established to collect spindle speed, tension, twist deviation, current, vibration amplitude, and winding diameter signals from the twisting equipment. An incremental photoelectric encoder acquires the spindle speed signal, a triaxial piezoelectric accelerometer acquires the vibration amplitude signal, a non-contact laser displacement sensor acquires the winding diameter signal, a Hall effect closed-loop current sensor acquires the current signal, and a high-frequency tension sensor acquires the tension signal. A machine vision camera captures images of the yarn surface at a frame rate of 500 frames per second. Based on the Canny algorithm, the pixel coordinates of the yarn stem edge between adjacent image frames are extracted. The number of angular changes of edge pixels per unit length is counted, and the number of periodic changes in the edge direction is converted into the actual twist count. The difference between this and the process-set twist count yields the twist deviation signal. All continuous physical quantity signals are converted from analog to digital by a 24-bit data acquisition card and input to the host industrial control computer at a sampling rate of 1000Hz. The control update cycle is 500ms, meaning that feature extraction, state prediction, and incremental control solution are performed every 500 sampling points.

[0029] The continuous time series is sliced ​​and truncated according to the rule of an outer main sliding window length of 2000 sampling points and a step size of 500 sampling points to generate a main sliding window state observation sequence. The main sliding window is used to cover the statistical analysis of overall non-stationary features. Inside the main sliding window, inner sub-sliding windows are set according to different needs of subsequent feature extraction: the phase lag calculation uses an inner cross-correlation analysis window with 1024 sampling points; the spectral energy difference and symbol change density calculation uses an adjacent inner spectral analysis window with 512 sampling points and a 50% overlap rate. In this embodiment, the main sliding window length is 2000 sampling points, covering a 2-second time span. In other embodiments, it can be adjusted between 1000 and 4000 sampling points depending on the device rotation speed range. A step size of 500 sampling points corresponds to a 75% window overlap rate, which is sufficient to capture the short-term non-stationary features of the signal while avoiding excessive computational load.

[0030] Cross-correlation analysis was performed on the tension and spindle speed signals within the sliding window. The time delay corresponding to the peak value of the cross-correlation function was extracted and converted into the phase lag of the tension and spindle speed signals by combining it with the signal's dominant frequency period. When calculating the phase lag of the tension and spindle speed signals, the length of the inner cross-correlation analysis window was set to 1024 sampling points, and the sampling frequency was set to 1000Hz. Discrete cross-correlation calculations were performed on the mean-reduced tension and spindle speed signals respectively. The sampling offset at which the cross-correlation function reaches its maximum value was determined through global optimization.

[0031] The main frequency of the spindle speed signal is extracted using Fast Fourier Transform (FFT). In this embodiment, the main frequency is 50Hz, resulting in a corresponding signal main frequency period of 0.02s. The time delay is obtained by calculating the ratio of the sampling offset corresponding to the cross-correlation peak value to the sampling frequency. The time delay is then modulo the main frequency period and converted into a phase lag within a 2π period. This phase lag is used to characterize the hysteresis effect of tension following spindle speed changes. An increase in the phase lag value indicates that the delay in tension response relative to spindle speed changes is aggravated. When the phase lag exceeds π / 2, the tension adjustment's ability to follow spindle speed fluctuations is insufficient, and the equipment enters a weakly coupled operating range. In other embodiments, the main frequency can be selected within the range of 30Hz to 80Hz according to the actual working conditions of the spindle speed.

[0032] The envelope of the vibration amplitude signal within a sliding window is extracted using the Hilbert transform. The third central moment of the envelope amplitude distribution is calculated, and the vibration envelope skewness is obtained by dimensionless processing using the cube of the standard deviation. A complex analytic signal is constructed from the vibration amplitude signal sequence, where the imaginary part is generated by the Hilbert transform of the original signal. The magnitude of the analytic signal is taken to obtain the envelope. The mean and standard deviation of this envelope sequence are calculated. The third central moment, i.e., the asymmetry index of the envelope amplitude distribution, is obtained by summing the cubes of the deviations from the mean at each sampling point and calculating the mean. This is then divided by the cube of the standard deviation to achieve dimensionless processing, yielding the vibration envelope skewness. The time-domain waveform of the vibration signal under typical working conditions and the envelope extracted by the Hilbert transform are shown in the figure. Figure 1 A vibration envelope skewness greater than 0 indicates the presence of an impulse component in the vibration envelope. Under normal operation, the vibration envelope skewness is between 0.1 and 0.5, while it exceeds 1.5 under abnormal impact.

[0033] A first-order forward differencing operation is performed on the tension signal sequence within the sliding window. The tension gradient is obtained by dividing the difference result by the sampling time interval. The difference between the tension value of the current sampling point and the next sampling point is taken as the first-order forward difference component, which is divided by the sampling interval of 0.001s to obtain the tension gradient value, in cN / s, used to characterize the degree of abrupt change in yarn tension during twisting. When a sampling point is lost or a duplicate timestamp occurs, the most recent valid sampling interval is used instead.

[0034] The power spectral density is obtained by performing Fast Fourier Transform (FFT) on the signals within two adjacent sliding windows. The difference in the integral area of ​​the power spectral density within the corresponding frequency bands of the two sliding windows is calculated as the spectral energy difference. In calculating the spectral energy difference, adjacent inner-layer spectral analysis windows with 512 sampling points and a 50% overlap rate are used. The power spectral density function is calculated using FFT, with a frequency resolution of 1.95 Hz. A core frequency band sensitive to the mechanical state of the equipment is selected; in this embodiment, the core frequency band is from 10 Hz to 150 Hz. The power spectral density of the previous window and the current window within this core frequency band is numerically integrated, and the difference in the integral area is the spectral energy difference, characterizing the time-varying fluctuation amplitude of the system's kinetic energy. As an optional approach, the core frequency band can be adjusted within the range of 5 Hz to 200 Hz according to the equipment's inherent frequency.

[0035] The symbol flip count is calculated for each dimension of the first-order difference sequence within adjacent sliding windows. The ratio of the flip count to the sequence length is used as the symbol change density. For each point-by-point difference sequence of the observed signal within two adjacent windows, the number of times the product of adjacent difference values ​​is less than 0 is counted as the symbol flip count, which is then divided by the total sequence length. Under normal conditions, the symbol change density typically remains between 0.15 and 0.25, but may increase to above 0.4 during abnormal oscillations.

[0036] A multidimensional state matrix containing normalized ingot speed, tension, and vibration amplitude signals is constructed. The eigenvalues ​​of the covariance matrix of the multidimensional state matrix are calculated, and the Shannon entropy of the eigenvalues ​​is used to represent the multi-channel correlation disorder. A higher multi-channel correlation disorder value indicates a weaker cooperative constraint ability among the physical quantities. The ingot speed, tension, and vibration amplitude signals normalized to the [-1, 1] interval are constructed into a 3×2000 multidimensional state matrix. The covariance matrix is ​​solved by matrix multiplication and eigenvalue decomposition is performed to obtain three non-negative eigenvalues. The normalized proportion of each eigenvalue in the sum of eigenvalues ​​is calculated, and the multi-channel correlation disorder is obtained using the Shannon entropy formula.

[0037] The spectral energy difference, sign change density, multi-channel correlation disorder, phase lag, and vibration envelope skewness are used to construct feature vectors. A Gaussian mixture model (GMM) clustering algorithm is employed for unsupervised classification of these feature vectors, mapping the centers of the clusters to different discrete states. Discrete states with vibration envelope skewness exceeding a set threshold at the cluster centers are marked as high-vibration states, thus establishing a discrete state set. The extracted five indicators are then fused to construct a 5-dimensional feature vector. At least 5000 sets of offline observation samples are collected and input into the GMM. The K-means pre-clustering results are used as the initial mean for the GMM expectation-maximization algorithm. The initial covariance of each Gaussian component is taken from the offline sample covariance matrix, and the initial mixing weights are evenly distributed at 1 / 6. The expectation-maximization algorithm is used for 50 to 100 iterations to converge, yielding the parameter distributions of the six clusters. Six cluster centers are used to form a discrete state set. The vibration envelope skewness dimension of each cluster center in the offline observation samples is statistically analyzed. Under normal operating conditions, the vibration envelope skewness is concentrated between 0.1 and 0.5, while under abnormal impact conditions it exceeds 1.5. A value of 1.2, near the midpoint between the upper bound of the normal interval and the lower bound of the abnormal interval, is taken as a set threshold. If the vibration envelope skewness dimension value corresponding to a certain cluster center is greater than the set threshold of 1.2, the discrete state is marked as a high vibration state. During online operation, the 5-dimensional feature vector of the current sliding window is calculated in real time. Based on the six trained Gaussian component parameters, the posterior probability of the current feature vector belonging to each Gaussian component is calculated using the Bayesian posterior probability formula. The index of the Gaussian component corresponding to the maximum posterior probability is used as the current real-time discrete state identifier. In other embodiments, the high vibration state threshold can be adjusted within the range of 0.8 to 1.5 according to the equipment vibration intensity level. A lower threshold makes high vibration state identification more sensitive but may introduce false alarms, while a higher threshold increases the risk of missed detections.

[0038] The historical state sequence is traversed, and the number of time steps in which the device is continuously in the same discrete state is recorded as the dwell time. The continuous dwell time of each discrete state within the historical observation set is statistically analyzed. The probability density function of the dwell time of each discrete state is fitted using a Weibull distribution to construct a basic semi-Markov transition model. The Weibull distribution is a continuous probability distribution that can flexibly model increasing, constant, or decreasing failure rates. The dwell time sequences of six discrete states are extracted from the historical observation dataset. The probability density function of each state is fitted using the maximum likelihood estimation method to obtain the shape parameter and scale parameter. Based on these, the cumulative distribution function is calculated, thus constructing the basic state transition probability matrix of the semi-Markov transition model. Considering the differences in dwell time of the twisting equipment at different operating stages, the shape parameter of the Weibull distribution ranges from [0.8, 2.5], and the scale parameter ranges from [10, 50] time steps. A shape parameter less than 1 indicates that the dwell time has a decreasing failure rate distribution, meaning that the longer the state dwells, the less likely it is to jump; a shape parameter greater than 1 indicates that the dwell time has an increasing failure rate distribution, meaning that the longer the state dwells, the more likely it is to jump to a new state.

[0039] The winding diameter signal is divided into multiple preset winding diameter intervals, and a slow-varying transition probability benchmark matrix is ​​set for each winding diameter interval. The measured winding diameter is divided into three preset intervals: empty tube period [0, 60) mm, forming period [60, 150) mm, and full tube period [150, 250] mm. The long-term transition frequency between states within the three physical intervals is statistically analyzed, and corresponding three-dimensional 6×6 slow-varying transition probability benchmark matrices are established. Since the yarn air ring shape gradually changes during the winding process from empty tube to full tube, the transition law of the equipment operating state in different winding diameter intervals is significantly different. Therefore, a segmented modeling method is used to characterize the state transition benchmark characteristics of each stage.

[0040] The ratio of the tension gradient to a preset threshold is transformed into a rapid change correction factor using a mapping function. Considering that the rapid change correction factor needs to be sensitive to tension abrupt changes and has a saturation upper limit, 1+tanh is used. The mapping method makes the range of the fast-changing correction factor smoothly limited to the interval [1, 2).

[0041] Reference Figure 2 The mapping curve between the tension gradient and the rapid change correction factor exhibits an S-shaped monotonically increasing trend. The tension gradient tolerance threshold is 5 cN / s in this embodiment, but in other embodiments, it can be selected within the range of 3 cN / s to 8 cN / s depending on the yarn material and process specifications. When the tension gradient is much smaller than the tolerance threshold, the rapid change correction factor approaches 1, meaning no additional correction is made to the basic transition probability; when the tension gradient is much larger than the tolerance threshold, the rapid change correction factor approaches 2, indicating a tendency for amplified state transitions.

[0042] In each control cycle update of the model, the slow-varying transition probability baseline matrix corresponding to the current winding diameter is matched, and all off-diagonal transition probability elements are multiplied by a fast-varying correction factor. The diagonal self-transition probability elements are not modified at this stage. The corrected transition matrix is ​​summed row-wise and normalized to ensure that the sum of state transition probabilities in each row equals 1. During row normalization, diagonal elements are scaled proportionally to the off-diagonal elements. This mechanism amplifies the state transition tendency under high-tension abrupt changes, fusing the slow-varying changes of the equipment over a long timescale with the fast-varying disturbances over a short timescale, outputting a time-varying state transition model characterizing the current non-stationary operating state of the twisting equipment.

[0043] During the operation of the predictive controller, the dwell timestamps of the current discrete states of the device are recorded. A conditional probability integral is performed on the future time domain using a preset Weibull distribution model to calculate the expected remaining dwell time of the current state. A warning transition threshold of 5 sampling steps is set. This threshold is derived from the lower quartiles of the dwell length distribution of each discrete state in offline historical data. The lower quartiles of the dwell lengths of the six discrete states range from 3 to 8 steps. The median of 5 steps is taken as the warning transition threshold. When the expected remaining dwell time is less than 5 steps, it indicates that the device is likely to undergo a state transition. In this case, the prediction time domain length is reduced to 3 steps to improve the computation frame rate and system response speed. When the expected remaining dwell time is greater than or equal to 5 steps, the prediction time domain length is increased to 10 or 15 steps to achieve a balance between global optimization accuracy and computational load. If the warning transition threshold is too small, the system may shorten the prediction time domain before the state transition occurs, and there will not be enough time to switch the control strategy; if the warning transition threshold is too large, the frequent shortening of the prediction time domain will cause the controller to be in short-sighted mode for a long time, and the overall control effect will decrease.

[0044] Within the locked prediction time-domain steps, taking the current operating state as the initial node, and based on the row-normalized time-varying state transition matrix, the Monte Carlo uniform random sampling method is invoked to perform 50 predictions in parallel, generating multiple possible future state transition paths. Based on the mean change rate parameters of the spindle speed signal, tension signal, and twist deviation signal under various discrete states statistically obtained during the offline training phase, and using the measured spindle speed signal value, tension signal value, and twist deviation signal value of the current control cycle as the initial iteration values, the first-order Euler forward iteration method is used to map the pure logic state path derivation into a continuous prediction sequence of the spindle speed signal, tension signal, and twist deviation signal. Taking the twisting equipment used in this embodiment as an example, the operating parameters under a certain high vibration state are verified: the tension decrease rate is 0.5 cN / s, the spindle speed fluctuation rate is 10 rpm / s, the current tension is 25 cN, and the spindle speed is 17000 rpm. Therefore, after one forward prediction step, the predicted tension value is 24.75 cN, and the predicted spindle speed value is 17005 rpm or 16995 rpm. In this embodiment, the number of Monte Carlo samplings is 50. In actual engineering deployments, the number of samplings can also be adjusted from 30 to 100 depending on the control cycle duration and the performance of the computing platform.

[0045] Each deviation term undergoes normalization of its respective physical range before being substituted into the objective function, ensuring that deviations of different dimensions are uniformly mapped to a dimensionless space. This minimizes the objective function, constructing a multi-objective quadratic cost function. From the perspective of twisting process priority, twist deviation directly determines yarn twist strength and appearance uniformity, thus receiving the highest penalty weight of 1.5. Since tension deviation affects yarn breakage rate and package quality, its penalty weight is set to 1.2. The influence of spindle speed deviation on yarn quality is indirectly transmitted through tension and twist, and its lowest value of 0.3 is chosen among the four penalty weights. The control increment change rate penalty term is used to suppress drastic fluctuations in control output, typically set between 0.3 and 0.8 in engineering practice; in this embodiment, it is set to 0.5. The penalty weight ranges from [0.1, 2]. If the weight is too small, the core process indicator deviation is ignored, leading to an imbalance in the control objective. If the weight is too large, the controller becomes overly sensitive to this deviation, sacrificing the adjustment capabilities of other channels. In other implementations, the weight can be adjusted within this range according to yarn type and process priority.

[0046] The feasible speed range of the twisting main motor [0, 20000 rpm] is used as the upper and lower limits of the absolute output constraint. The upper limit of motor acceleration (100 rpm / s) and the limit of the current rise rate of the tension adjustment mechanism are set as the rate of change constraint band. A sequential quadratic programming algorithm is used to iteratively descent the minimization objective function within this constraint space to obtain the optimal first term value of the control increment sequence that satisfies the multi-physical quantity decoupling requirement, i.e., the control increment of the current control cycle. The sequential quadratic programming algorithm is a well-known iterative algorithm for handling nonlinear constraint optimization problems, and will not be elaborated upon here.

[0047] The controller evaluates all future state paths generated by Monte Carlo sampling in real time. If the ratio of paths containing high-vibration states to the total number of paths exceeds a set safety threshold of 20%, it indicates a risk of future mechanical instability. The controller then retrieves the vibration envelope skewness and multi-channel correlation disorder from the current data processing process. In this embodiment, the safety threshold is 20%. When the safety threshold is too low, constraint adjustments are frequently triggered, leading to a conservative system response and reduced production efficiency. When the safety threshold is too high, constraint adjustments are not easily activated, resulting in a lack of system protection under high-vibration conditions. In other embodiments, the threshold can be set within the range of 10% to 30% based on the yarn breakage tolerance.

[0048] Considering that the constraint coefficient should monotonically tighten as the equipment degradation index increases and possess a [0, 1] limiting characteristic, the basic threshold of vibration envelope skewness is determined based on offline statistical data under normal operating conditions. During normal operation, the vibration envelope skewness is distributed between 0.1 and 0.5, and the upper bound of the distribution, 0.5, is taken as the basic threshold. The basic threshold of multi-channel correlation disorder is determined based on offline statistical data when the signals of each channel are highly coupled, and 0.3 is taken as the basic threshold. The difference between the vibration envelope skewness and the multi-channel correlation disorder exceeding their respective basic thresholds is calculated using a maximum value function. The difference exceeding the basic threshold is multiplied by penalty adjustment coefficients of 0.5 and 0.3 respectively for negative exponential fusion. This exponential mapping mechanism ensures that when the equipment degradation trend intensifies, the calculated constraint coefficient will rapidly decay and be limited to the protection range of [0.3, 1]. The penalty adjustment coefficients of 0.5 and 0.3 correspond to the sensitivity weighting of the vibration envelope skewness channel and the multi-channel correlation disorder channel, respectively. Since vibration envelope skewness directly reflects the intensity of mechanical impact and has a more direct impact on the constraint of control increment rate of change, a higher penalty adjustment coefficient is assigned. Multi-channel correlation disorder reflects the degree of coordinated degradation between signal channels, which is an indirect degradation indicator, and a lower penalty adjustment coefficient is assigned. The protection lower bound is 0.3, meaning the constraint coefficient will not be lower than 0.3, ensuring that the controller retains a minimum adjustment capability under extreme conditions without completely locking up.

[0049] The results were verified using actual operational data on vibration envelope skewness and multi-channel correlation disorder: the current vibration envelope skewness is 1.5, and the base threshold is 0.5; the multi-channel correlation disorder is 0.9, and the base threshold is 0.3. The difference between the vibration envelope skewness and the base threshold is 1, which is multiplied by a penalty adjustment coefficient of 0.5 to obtain 0.5; the difference between the multi-channel correlation disorder and the base threshold is 0.6, which is multiplied by a penalty adjustment coefficient of 0.3 to obtain 0.18. The sum of the two is 0.68, and the negative exponent is taken to obtain the constraint coefficient. It falls within the protection range of [0.3, 1] and does not require truncation.

[0050] The physical upper bound of the original control increment rate of change in the sequential quadratic programming solver is multiplied by the constraint coefficient. Following the previous verification results, when the constraint coefficient is 0.507, the original upper bound of 100 rpm / s shrinks to 50.7 rpm / s, and the constraint strength of the control system's rate of change increases accordingly. The control model initiates a quadratic optimization iteration within this tightened feasible region, causing the system to abandon the high-gain response and output a smoothed control increment.

[0051] The spindle speed component in the control increment sequence is added to the current actual spindle speed command to generate the target absolute speed command; the tension component in the control increment sequence is added to the current actual tension command to generate the target tension command. These commands are processed by a digital-to-analog converter module. The target absolute speed command is mapped to a 0V to 10V servo voltage drive signal and output to the AC frequency converter motor of the twisting spindle. The target tension command is mapped to a 4mA to 20mA industrial standard current drive signal and output to the piezoelectric electronic tension regulating device.

[0052] The experiment used a standard twisting equipment to build the test environment, with the spindle rated speed set at 18,000 rpm and the sensor sampling frequency set at 1,000 Hz. The continuous test cycle was 720 hours. During the test, the measured winding diameter gradually increased from 0 mm to 250 mm at full capacity. Every 2 hours, random airflow disturbances and yarn defects were artificially applied to the twisting equipment to create unstable and harsh working conditions. Three comparative test groups were established. The first group was the baseline group, using a conventional model predictive control scheme without dynamic state transition and rate of change constraint adjustment capabilities. The second group was the partial scheme group, using the scheme of this application with the constraint adjustment module removed, using only a time-varying semi-Markov transition model for prediction and basic control. The third group was the complete scheme group, using the complete scheme of this application, including vibration envelope skewness extraction and control increment constraint adjustment mechanisms.

[0053] After the complete testing cycle, core operational data for the three sets of equipment were extracted and statistically analyzed. The first set had a yarn tension mean square error of 3.42 cN, a twist deviation rate of 4.5%, a cumulative duration of high vibration (86 hours), and a yarn breakage rate of 1.2% per 10,000 meters. The second set had a yarn tension mean square error reduced to 1.85 cN, a twist deviation rate decreased to 2.1%, a cumulative duration of high vibration (42 hours), and a yarn breakage rate of 0.6% per 10,000 meters. The third set had a yarn tension mean square error of 0.76 cN, a twist deviation rate controlled at 0.8%, a cumulative duration of high vibration (only 11 hours), and a yarn breakage rate of 0.15% per 10,000 meters. Figure 3The performance comparison of the control group, ablation group, and complete scheme shows that the prediction mechanism based on the conditional expectation of the remaining residence time of discrete states and the time-varying semi-Markov transfer model can detect the slow-changing gradual characteristics and transient tension rapid-changing disturbances caused by the increase in twisted package weight, thus improving the basic adjustment accuracy of tension and twist. The comparison data between the second and third groups show that the constraint adjustment mechanism constructed by vibration envelope skewness and multi-channel correlation disorder plays a core protective role. When there is a high risk of instability in the future state path, the exponential mapping mechanism shrinks the upper bound of the incremental change rate of control, reducing the duration of high vibration state and the abrupt amplitude of control output, compressing the duration of high vibration state to about 1 / 8 of the control group, and improving the reliability of the twisting equipment's operating state control and yarn quality.

[0054] It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this patent application shall be determined by the appended claims.

Claims

1. An adaptive control system for the operating status of a twisting device, characterized in that, include: A module is established to collect the operating signals of the twisting equipment. A state observation sequence is constructed according to a sliding window. Based on the time-domain statistical characteristics and frequency-domain energy distribution of the operating signals, phase lag, vibration envelope skewness, tension gradient, spectral energy difference, sign change density, and multi-channel correlation disorder are extracted. The operating states of the twisting equipment are divided by combining the spectral energy difference, sign change density, multi-channel correlation disorder, phase lag, and vibration envelope skewness to establish a discrete state set. The solution module is used to construct a semi-Markov transition model by statistically analyzing the dwell length of each discrete state in the discrete state set. It combines the winding diameter range and tension gradient to perform piecewise correction on the state transition probability to obtain a time-varying state transition model. It adjusts the prediction time domain length according to the expected remaining dwell time of each discrete state, uses the time-varying state transition model to predict the future state path, and solves the objective function that satisfies the spindle speed constraint, tension constraint, twist deviation constraint and control increment change rate constraint to obtain the control increment. The control module is used to adjust the constraint strength of the control increment change rate based on the vibration envelope skewness and multi-channel correlation disorder when the future state path contains a high vibration state, and output the control increment to the actuator to complete the control of the twisting equipment's operating state.

2. The adaptive control system for the operating status of a twisting device according to claim 1, characterized in that, The operating signals include spindle speed signal, tension signal, twist deviation signal, current signal, vibration amplitude signal, and winding diameter signal.

3. The adaptive control system for the operating status of a twisting device according to claim 2, characterized in that, When the module calculates the phase hysteresis, vibration envelope skewness, and tension gradient: Cross-correlation analysis is performed on the tension signal and spindle speed signal within the sliding window. The time delay corresponding to the peak value of the cross-correlation function is extracted and converted into the phase lag between the tension signal and the spindle speed signal by combining the main frequency period of the spindle speed signal. The envelope of the vibration amplitude signal within the sliding window is extracted using Hilbert transform. The third central moment of the envelope amplitude distribution is calculated, and the vibration envelope skewness is obtained by dimensionless processing of the cube of the standard deviation. Perform a first-order forward difference operation on the tension signal sequence within the sliding window, and divide the difference result by the sampling time interval to obtain the tension gradient.

4. The adaptive control system for the operating status of a twisting device according to claim 3, characterized in that, When the module calculates the spectral energy difference, symbol change density, and multi-channel correlation disorder and establishes a discrete state set: Perform a Fast Fourier Transform on the signals of adjacent sliding windows, and use the difference in the integral area of ​​the power spectral density of the corresponding frequency band as the spectral energy difference; The symbol change density is defined as the ratio of the number of symbol flips in the differential sequence of each dimension of the signal to the sequence length. A multidimensional state matrix is ​​constructed from the normalized ingot speed signal, tension signal and vibration amplitude signal, and the Shannon entropy of the covariance matrix eigenvalues ​​is used to represent the multi-channel correlation disorder. The above features, along with phase hysteresis and vibration envelope skewness, are used to construct a feature vector and cluster them. Discrete states where the vibration envelope skewness exceeds a set threshold are marked as high vibration states.

5. The adaptive control system for the operating status of a twisting device according to claim 1, characterized in that, When the solver module constructs a semi-Markov transition model and obtains a time-varying state transition model: The continuous dwell time of each discrete state in the discrete state set in the statistical observation history is used to fit the probability density function of the dwell time of each discrete state using the Weibull distribution, and a basic semi-Markov transfer model is constructed. The winding diameter signal is divided into multiple preset winding diameter intervals, and a slow-varying transfer probability benchmark matrix is ​​set for different winding diameter intervals. Calculate the ratio of the absolute value of the tension gradient at the current moment to a preset threshold to generate a rapid change correction factor; By multiplying the off-diagonal elements of the slow-varying transition probability baseline matrix with the fast-varying correction factor and normalizing the result, the state transition probabilities in the basic semi-Markov transition model are updated with weights to obtain the time-varying state transition model.

6. The adaptive control system for the operating status of a twisting device according to claim 5, characterized in that, When the solution module adjusts the prediction time domain length and solves the objective function to obtain the control increment: The conditional expectation of the remaining dwell time is calculated based on the current dwell time and Weibull distribution of the discrete state. If the conditional expectation is less than a preset threshold, the prediction time domain length is reduced, and vice versa. The future state path is generated by Monte Carlo uniform sampling through a time-varying state transition model, and the future state path is mapped to a continuous prediction sequence based on the mean rate of change of each discrete state. The objective function for minimizing the control increment is constructed by weighting the sum of squares of spindle speed deviation, sum of squares of tension deviation, sum of squares of twist deviation, and penalty term for the rate of change of control increment. The control increment is obtained by solving the problem within the constraints of physical limits and rate of change.

7. The adaptive control system for the operating status of a twisting device according to claim 1, characterized in that, When the control module adjusts the constraint strength of the control increment change rate and completes the control of the twisting equipment's operating status: when the probability of high vibration state in the future state path exceeds the safety threshold, the vibration envelope skewness and multi-channel correlation disorder of the current sliding window are extracted. Based on vibration envelope skewness and multi-channel correlation disorder, a mapping function in the form of negative exponential decay is constructed, and constraint coefficients are calculated. The constraint coefficients are used to reduce the allowable upper bound of the rate of change of control increment in the minimization objective function, thereby increasing the constraint strength and resolving to obtain the smoothed control increment. The control increment is converted into a voltage or current drive signal for the actuator, and output to the motor and tension adjustment device of the twisting equipment to complete the control of the twisting equipment's operating status.

8. The adaptive control system for the operating status of a twisting device according to claim 6, characterized in that, In the minimization objective function, the sum of squares of spindle speed deviation, the sum of squares of tension deviation, the sum of squares of twist deviation, and the penalty term for the rate of change of control increment correspond to differentiated penalty weights. The penalty weight of the sum of squares of twist deviation is greater than the penalty weight of the sum of squares of tension deviation, and the penalty weight of the sum of squares of tension deviation is greater than the penalty weight of the sum of squares of spindle speed deviation.

9. The adaptive control system for the operating status of a twisting device according to claim 7, characterized in that, In the calculation of the constraint coefficient, the difference between the vibration envelope skewness and the multi-channel correlation disorder degree exceeding their respective basic thresholds is calculated by the maximum value function. The difference exceeding their respective basic thresholds is multiplied by the corresponding penalty adjustment coefficient and then fused by negative exponential. The value of the constraint coefficient is limited to between the preset protection lower bound and 1.

10. An adaptive control system for the operating state of a twisting device according to claim 7, characterized in that, The probability of a high-vibration state occurring in the future state path is the ratio of the number of paths containing high-vibration states to the total number of paths generated by Monte Carlo sampling.

Images