Adaptive and Coordinated Control Method for Tension in Cable Stranding Process

Abstract

This invention provides a method for adaptive and coordinated tension control in cable stranding processes, comprising: deploying a high-sampling-rate distributed fiber optic strain sensor array at key equipment nodes to acquire multi-point micro-strain signals in real time; performing sliding window phase space reconstruction and persistent cohomology calculation on the strain difference sequence through topological data analysis to dynamically generate a topological perturbation index characterizing structural disturbances; employing an adaptive threshold determination method to promptly detect potential instability; and utilizing a feedforward compensation algorithm based on graph attention networks to generate and project compensation commands to the servo driver. The method also includes a topological consistency verification and online fine-tuning mechanism to improve response robustness and adaptability. This invention enables early warning, accurate identification, and effective intervention of tension field structural instability during cable stranding, improving product processing stability and equipment operational reliability.

Description

Technical Field

[0001] This invention relates to the field of cable stranding tension collaborative control and industrial intelligent manufacturing technology, and in particular to a method for adaptive collaborative control of cable stranding process tension. Background Technology

[0002] Tension control in cable stranding is a core component of high-precision manufacturing, directly impacting the structural uniformity and mechanical properties of the finished cable. Currently, mainstream technologies in cable stranding tension control include tension prediction and feedback control based on physical modeling, data-driven time-series deep learning-assisted feedforward regulation, and multi-physics numerical simulation optimization strategies. Some advanced systems utilize multi-point tension sensor data, employing PID or adaptive control algorithms to perform closed-loop adjustments on the equipment drive side, or leveraging LSTM / GRU structures to predict tension changes under disturbances, thereby achieving limited feedforward compensation. Other technologies focus on Koopman operator data fusion, PINN physical information neural networks, and multi-loop feedforward-feedback superposition systems, all aiming to improve feedforward response speed and reduce the impact of disturbances on tension stability through modeling. These technologies have wide applications in high-end continuous traction equipment such as cable manufacturing and fiber processing, driving the overall intelligent and precision progress of the industry. However, the fundamental limitation of existing technologies lies in the fact that their feedforward compensation mechanisms generally rely on the parameterization of physical equations, empirical formulas, or fitting of time-series data, resulting in limited ability to characterize the dynamic evolution of tension under multi-source disturbances (such as material elastic hysteresis, changes in environmental temperature and humidity, and equipment inertial coupling). On the one hand, physical modeling struggles to accurately encompass dynamic anomalies such as nonlinearity, bifurcation, and singularities in high-dimensional complex process chains, easily leading to model mismatch and control lag. On the other hand, deep learning and data-driven methods lack generalization beyond disturbance boundaries, and feedforward compensation commands are often superimposed on feedback loops, making it impossible to avoid the inherent phase delay of mechanical transmission chains and control systems, resulting in difficulty in overcoming bottlenecks in on-site response speed. Furthermore, under the influence of multi-source coupled disturbances, existing tension control systems lack the ability to proactively identify potential structural instability risks, and feedforward control is mostly passive error compensation, making it difficult to achieve active intervention in the complex dynamic processes of the tension field, which can easily induce high-frequency tension fluctuations or even process failure risks. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for adaptive and coordinated control of tension in cable stranding process.

[0004] The technical solution of this invention is implemented as follows: a cable stranding process tension adaptive and coordinated control method, comprising: S1: Deploy a high sampling rate distributed fiber optic strain sensor array at key nodes of the cable stranding equipment to synchronously collect the micro-strain time-series signals of the wire at the outlet of the wire feeding frame, each traction point of the winding cage, and the front end of the take-up tension sensor, and generate a multi-source strain differential sequence as the original input data. S2: Perform sliding window topological embedding processing on the strain difference sequence between each group of adjacent sensing points. The window length matches the system inertial time constant calibration value of 8-12 milliseconds. The embedding dimension is determined to be 4-6 dimensions according to the Cao method. Construct the local phase space trajectory of the tension state manifold. S3: Based on the time derivatives of the number of 0-dimensional connected branches and the number of 1-dimensional weighted ring structures calculated from the local phase space trajectory, a topological perturbation index is generated, which characterizes the dynamic evolution characteristics of the spatiotemporal topology of the tension field. S4: Compare the topological disturbance index with the dynamic baseline threshold, which is updated on a rolling basis by the 99.5th percentile of the topological disturbance index under historical undisturbed conditions. When the topological disturbance index exceeds the threshold for three consecutive sampling periods, it is determined that the tension field has a potential structural instability. S5: After determining structural instability, the topology guidance feedforward module is triggered, and the strain-position map structure of all sensing points in the current window is input into the pre-trained graph attention network model. The model takes minimizing the peak increase of the topology perturbation index in the following 20 milliseconds as the training objective, and outputs the torque correction direction and relative amplitude weight of each winch drive unit. S6: Geometrically project the weight vector output by the graph attention network onto the velocity-tension reference trajectory planned by the main control system, calculate the feedforward compensation command increment of each winch drive unit, and directly inject the increment into the position loop feedforward channel of the servo driver to achieve phase delay bypass. S7: After the compensation instruction is executed, perform counterfactual reconstruction on the Betti number evolution path of the next sliding window, and generate topology consistency verification results by comparing the difference between the topology perturbation index of the reconstructed path and the actual path. S8: If the topology consistency check result shows that the topology perturbation index has not fallen back to the baseline ±15% safety zone, then reduce the confidence of the graph attention network and start small-sample online fine-tuning, and retrain the last layer attention head of the graph attention network using only the abnormal topology features of the most recent 5 windows.

[0005] The cable stranding process tension adaptive and coordinated control method provided by this invention has the following beneficial effects: (1) This invention models the dynamic behavior of the wire tension field during stranding as the evolution of topological manifolds in a high-dimensional phase space, introduces an adaptive persistent homology method to extract the rate of change of the 0-dimensional and 1-dimensional Betti numbers in real time, and constructs a "topological perturbation index" as a sensitive criterion for the stability of the system structure, which significantly improves the sensitivity to early tension instability. This mechanism can capture the global connectivity and loop structure mutations hidden in multi-point strain signals, and identify the dynamic topological distortions caused by wire shaking, winch slippage or material inhomogeneity before obvious tension deviation occurs. It effectively overcomes the response lag problem caused by existing technologies relying on accurate physical models or historical fault samples, and shows stronger robustness and generalization ability, especially in the scenarios of working condition switching, start-stop transition or sudden interference. (2) This invention designs a topology-guided feedforward compensation architecture for detected topology disturbances. It uses a graph attention network to directly input a space-strain map composed of distributed sensing nodes, outputs the torque correction weights required by each drive unit, and injects them into the position loop feedforward channel of the servo system through geometric projection. This bypasses the phase delay bottleneck of the traditional PID feedback loop and achieves closed-loop feedforward intervention within hundreds of milliseconds. This strategy avoids the control quantity conflict and integral saturation risk caused by superimposing the feedforward signal on the feedback control output. At the same time, the training objective of the graph attention network focuses on suppressing the growth trend of the topology disturbance exponent in the short period of time, rather than fitting the specific tension value. This makes it more concerned with the essential stability of the system rather than local measurement noise, which greatly improves the physical consistency and anti-interference ability of the control command. (3) To further ensure the reliability of the system under long-term operation and evolving conditions, this invention introduces a topological consistency counterfactual verification and small-sample online fine-tuning mechanism: after each compensation, the intervention effect is evaluated by reconstructing the Betti number evolution path. If the disturbance does not converge, the confidence of the GAT model is dynamically adjusted and only the data from the five most recent outlier windows is used to perform a lightweight update of the last layer of attention head, so as to quickly adapt to non-steady-state conditions such as new line gauges, changes in environmental temperature and humidity, or equipment wear. This mechanism does not require retraining the parameters of the entire network or relying on a large-scale labeled dataset, which significantly reduces the operation and maintenance cost and ensures that the system recovers high-precision disturbance recognition performance within 72 hours. It forms a complete adaptive closed loop of "perception-decision-execution-verification-evolution" and has good interpretability and engineering deployability. Attached Figure Description

[0006] Figure 1 This is a flowchart of the cable stranding process tension adaptive and coordinated control method of the present invention; Figure 2 This is a sub-flowchart of the cable stranding process tension adaptive and coordinated control method of the present invention; Figure 3This is another sub-flowchart of the cable stranding process tension adaptive collaborative control method of the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, the present invention provides a method for adaptive and coordinated tension control in cable stranding processes, specifically including: S1: Deploy a high sampling rate distributed fiber optic strain sensor array at key nodes of the cable stranding equipment to synchronously collect the micro-strain time-series signals of the wire at the outlet of the wire feeding frame, each traction point of the winding cage, and the front end of the take-up tension sensor, and generate a multi-source strain differential sequence as the original input data. S2: Perform sliding window topological embedding processing on the strain difference sequence between each group of adjacent sensing points. The window length matches the system inertial time constant calibration value of 8-12 milliseconds. The embedding dimension is determined to be 4-6 dimensions according to the Cao method. Construct the local phase space trajectory of the tension state manifold. S3: Based on the time derivatives of the number of 0-dimensional connected branches and the number of 1-dimensional weighted ring structures calculated from the local phase space trajectory, a topological perturbation index is generated, which characterizes the dynamic evolution characteristics of the spatiotemporal topology of the tension field. S4: Compare the topological disturbance index with the dynamic baseline threshold, which is updated on a rolling basis by the 99.5th percentile of the topological disturbance index under historical undisturbed conditions. When the topological disturbance index exceeds the threshold for three consecutive sampling periods, it is determined that the tension field has a potential structural instability. S5: After determining structural instability, the topology guidance feedforward module is triggered, and the strain-position map structure of all sensing points in the current window is input into the pre-trained graph attention network model. The model takes minimizing the peak increase of the topology perturbation index in the following 20 milliseconds as the training objective, and outputs the torque correction direction and relative amplitude weight of each winch drive unit. S6: Geometrically project the weight vector output by the graph attention network onto the velocity-tension reference trajectory planned by the main control system, calculate the feedforward compensation command increment of each winch drive unit, and directly inject the increment into the position loop feedforward channel of the servo driver to achieve phase delay bypass. S7: After the compensation instruction is executed, perform counterfactual reconstruction on the Betti number evolution path of the next sliding window, and generate topology consistency verification results by comparing the difference between the topology perturbation index of the reconstructed path and the actual path. S8: If the topology consistency check result shows that the topology perturbation index has not fallen back to the baseline ±15% safety zone, then reduce the confidence of the graph attention network and start small-sample online fine-tuning, and retrain the last layer attention head of the graph attention network using only the abnormal topology features of the most recent 5 windows.

[0010] Step S1: Deploy a high sampling rate distributed fiber optic strain sensor array at key nodes of the cable stranding equipment to synchronously collect the micro-strain time-series signals of the wire at the cable release frame outlet, each strand traction point, and the section before the take-up tension sensor, generating a multi-source strain differential sequence as the raw input data. Specifically, this includes: S1.1: Based on the mechanical structure parameters and tension propagation path model of the cable stranding equipment, perform key node location determination processing to generate a list of deployment locations for the cable release frame outlet, each strand traction point, and the front end of the take-up tension sensor, ensuring coverage of the entire dynamic evolution of tension and meeting the sensing spatial resolution requirements; Based on the three-dimensional mechanical structure parameter set of the cable stranding equipment (including the spatial coordinates, installation angles, and transmission chain moments of inertia of the wire feeding frame, winding cage, traction mechanism, and take-up unit), the structural topology mapping method (parameters: node identifier set, component connection matrix) is used to realize the correlation modeling of the geometric coordinates of key nodes of the equipment with the tension transmission path. Furthermore, by using a tension propagation path modeling algorithm (parameters: wire elastic modulus E, cross-sectional diameter d, path length L, friction coefficient μ of each node), the time delay and attenuation characteristics of tension waves propagating between different nodes are calculated, and the tension propagation matrix between nodes is obtained. Furthermore, through the path coverage evaluation method (parameter: sensor point spacing) (Node index set, tension change spatial resolution threshold), to achieve comprehensive determination of the entire process of dynamic tension evolution in the sensing deployment scheme, and generate a set of candidate deployment locations that meet the spatial resolution requirements; Furthermore, by using a node priority sorting algorithm (parameters: the position of the maximum tension gradient in the propagation matrix, the position of the highest value of the structural vibration modal frequency, and the influence coefficient of temperature and humidity), the sorting of the deployment location set is optimized, and a sorted list of key node deployments is generated. By using the deployment list generation process, the results of the previous step are transformed into the final installation position coordinates of the wire feeding frame outlet, each winch traction point, and the front end of the take-up tension sensor, thereby determining and accurately calibrating the physical installation points of the sensors. For example, on a certain model of six-cage stranding machine, the mechanical structural parameters are input as follows: the radius of the pay-off frame is 0.8 m, the diameter of the stranding cage is 1.2 m, the rotational speed range is 0–800 rpm, and the distance from the traction unit to the front end of the take-up tension sensor is 22 m. A connection matrix is ​​constructed using a structural topology mapping method, with a total of 14 nodes and a component connection density of 0.65. Using a tension propagation path modeling algorithm, the elastic modulus of the wire is set... The cross-sectional diameter d = 4.5 mm, the path length L = 52 m, and the friction coefficient at each node are... = 0.05, the maximum tension propagation delay is calculated to be 9.5 ms. The path coverage evaluation method is used, with the sensor point spacing... With a resolution of 0.5 m and a spatial resolution threshold of 1.0 N, 8 candidate deployment locations were identified. After processing using a node priority ranking algorithm, the top 5 locations included the wire feeder exit (node ​​1), the first winch traction point (node ​​4), the third winch traction point (node ​​9), the fifth winch traction point (node ​​12), and the section before the take-up tension sensor (node ​​14). A final deployment list was generated, with physical coordinate errors controlled within ±2 mm. Coverage calculations showed that the system could completely capture the dynamic evolution of tension throughout the entire winding process. S1.2: Based on the list of key node deployment locations, perform the selection and configuration of distributed fiber optic strain sensing units, and set the sampling rate parameter to 10 kHz or higher to build a high sampling rate distributed fiber optic strain sensing array to ensure that the requirements for high-frequency tension fluctuation detection are met and that the equipment mechanical interface is compatible. Based on the list of key node deployment locations, a sensor selection and matching method (parameters: node mechanical installation space size, equipment interface type, sensing area length range) is used to achieve preliminary screening of distributed fiber optic strain sensing unit models. Furthermore, through a sensing performance parameter mapping method (parameter: strain resolution ≤ 1) Linearity ≥ 99.8%, temperature drift coefficient ≤ 0.5 This allows for the verification of the performance compliance of the selected models and the generation of a list of candidate high-precision strain sensing units. Furthermore, a sampling rate setting algorithm (parameters: target sampling rate ≥ 10 kHz, signal bandwidth ≥ 4 kHz, dynamic tension response frequency band coverage ≥ 95%) is adopted to achieve precise setting of sampling rate parameters and generate a parameter configuration table that meets high-frequency sampling requirements; Furthermore, through a mechanical-optical interface matching method (parameters: fiber type SMF-28, interface structure ST / FC / APC, sensor unit external dimensions and mounting bracket fit ≥98%), the compatibility verification of the sensor unit and the device's mechanical interface is achieved, and interface matching verification data is generated. By integrating the above screening results with performance parameters, sampling rate configuration and interface verification data through system integration configuration, a high sampling rate distributed fiber optic strain sensor array is constructed to achieve seamless compatibility between the high-frequency tension fluctuation detection function and the system mechanical interface. For example, in a 12-reel cable stranding machine, the list of key node deployment locations includes one pay-off frame exit node, eight reel traction nodes, and three take-up front nodes. The corresponding strain sensing unit selected is model FBG-S100, with a strain resolution of 0.5. The linearity is 99.85%, and the temperature drift coefficient is 0.3. The sampling rate was set to 12 kHz, and the signal bandwidth was configured to 5 kHz to cover the maximum tension fluctuation frequency of the equipment. The mechanical interface uses ST fiber optic connectors, and the measured bracket fit was 99.2%. During configuration, through... The sampling rate margin calculation ensures no aliasing effect in the signal frequency domain. In actual operation, this array can capture micro-strain changes in rapid tension fluctuation events, shortening the dynamic response time to the 1 / 12000th of a second level, realizing the high-frequency input conditions required by the tension feedforward compensation control module, and significantly improving the real-time control accuracy of the system. S1.3: Based on the configured high sampling rate distributed fiber optic strain sensor array, synchronous installation and signal acquisition and processing are performed to obtain the micro-strain time sequence signal of the wire at the wire feeder outlet, each winding point and the front end of the take-up tension sensor, to ensure the time alignment of multi-source signals and eliminate the influence of external electromagnetic interference. S1.4: Based on the acquired original wire micro-strain time-series signal sequence, perform first-order differential calculation processing to generate strain differential sequences between each group of adjacent sensing points, eliminate DC offset and highlight the dynamic change characteristics of tension, and form a multi-source strain differential sequence as the original input data. S1.5: Based on the generated multi-source strain differential sequence, perform synchronization verification and integrity verification processing, detect data packet loss rate and timestamp alignment error, and output the verified original input data to ensure that the reliability requirements of subsequent sliding window topology embedding processing are met.

[0011] Step S2: Perform sliding window topological embedding processing on the strain difference sequence between each group of adjacent sensing points. The window length matches the system inertial time constant calibration value of 8-12 milliseconds, and the embedding dimension is determined to be 4-6 dimensions according to the Cao method, constructing the local phase space trajectory of the tension state manifold. Specifically, this includes: S2.1: Perform a step response test on the cable stranding equipment, calibrate the maximum system inertial time constant based on the dynamic attenuation characteristics of the equipment, and obtain a calibration value of 8 to 12 milliseconds to determine the matching range of the sliding window length; To assess the overall dynamic response of the cable stranding equipment, a step response test method was used (parameters: the input step amplitude was selected as 10% of the rated torque, and the duration of the step signal was greater than five times the expected inertial constant) to collect the dynamic attenuation characteristics of the equipment. Furthermore, through the real-time acquisition link of a high sampling rate distributed fiber optic strain sensor array (sampling rate: ≥10 kHz), the micro-strain response sequences of each key node under the action of a step signal are synchronously acquired, and a node-level time series dataset is obtained. Furthermore, an exponential curve fitting method is employed (model form: This enables parameter identification of single-node micro-strain decay curves, where... This represents the micro-strain response value. Let A be the initial strain bias, A be the step response amplitude, and t be the time variable. It is a time constant; Furthermore, multi-node parameter aggregation processing is performed (method: taking the parameters fitted from each node). (Meaning and removing outliers exceeding twice the standard deviation) to achieve a robust estimate of the maximum system inertial time constant and generate a calibration range; Furthermore, a boundary detection algorithm (threshold: fitting residual less than 5% of the mean within the calibration range) is used to confirm the validity of the inertial time constant, and a calibration value of 8~12 milliseconds is obtained as the basis for subsequent sliding window length matching; By using step response testing and dynamic attenuation characteristic modeling, the time series micro-strain results from the previous step are transformed into calibration data for the maximum system inertial time constant, achieving the expected technical effect of matching the sliding window length setting with the physical dynamic characteristics of the equipment. For example, in a cable stranding production line with six parallel-operating twisted cages, the torque amplitude of the input step signal is set to 10% of the rated torque, the duration is set to 0.1 seconds, and the sensor array sampling rate is set to 10 kHz. The time constants obtained by fitting the micro-strain response curves collected at each node are 8.5 ms, 9.1 ms, 11.8 ms, 8.7 ms, 9.0 ms, and 12.0 ms, respectively. After removing outliers exceeding twice the standard deviation, the calculated mean is 9.85 ms, and the standard deviation is 1.34 ms. The final calibrated maximum system inertial time constant range is 8-12 ms. Under these parameters, the subsequent sliding window length setting is highly matched with the system inertial characteristics, enabling accurate capture of dynamic response in the topology embedding processing stage and significantly improving the early identification capability of tension field topology disturbance characteristics. S2.2: Obtain the strain differential sequence between each group of adjacent sensing points generated by S1 as the original input data, and extract the differential features of the wire micro-strain time series signal based on the synchronous acquisition results of the distributed fiber grating strain sensing array. S2.3: Based on the calibrated maximum system inertia time constant, set the sliding window length to 8 to 12 milliseconds to ensure that the window covers the system dynamic response cycle in order to match the inertial characteristics of the cable stranding process; S2.4: The minimum embedding dimension is calculated by applying the Cao method to the strain difference sequence. Based on the phase space reconstruction theory of time series, the embedding dimension is determined to be 4 to 6 dimensions to fully unfold the geometric structure of the tension state manifold. Based on the sliding window length setting result obtained in S2.3 and the strain difference sequence extracted in S2.2, the embedding dimension calculation requirements of the current time series are determined. The Cao method (parameters: strain difference sequence, sliding window length 8-12 milliseconds, correlation delay time τ is calculated by the average mutual information method) is used to realize the minimum embedding dimension estimation function of time series. Furthermore, using the first statistic in the Cao method With the second statistic The numerical trend of change is calculated to determine the degree of dynamic structural unfolding of the sequence as the embedding dimension m increases, and the results are obtained. The m containing the saturation value is used as a candidate for the minimum embedding dimension; Furthermore, through the Cao method Curve analysis identifies whether the time series contains significant periodic components, avoiding overestimation or underestimation of the embedding dimension due to periodic effects, and generates a corrected minimum embedding dimension estimate. The first statistic is calculated using the following formula. Perform the calculation:

[0012] in, The total number of samples, To delay time, It is a time series vector. Let be the nearest neighbor index of the i-th vector when the embedding dimension is m, and let ||·| represent the Euclidean distance between the vectors; The second statistic is calculated using the following formula. Perform the calculation:

[0013] The meanings of the symbols are the same as before, through... Value convergence analysis identifies potential cycles; By calculating the inertial time constant in S2.3 using the Cao method, the final embedding dimension is determined to be 4 to 6 dimensions. The embedding dimension configuration parameters are then output for the subsequent topological embedding construction process in S2.5, achieving a fully unfolded time series geometric structure and a high-fidelity representation of the tension state manifold. For example, in the strain differential sequence acquired by five fiber Bragg grating sensing points deployed between the terminal block and the take-up frame of a three-strand cable stranding machine, the sampling rate is 12 kHz, the sliding window length is set to 10 milliseconds, and the delay time τ is calculated to be 0.8 milliseconds using the average mutual information method. The Cao method is then executed to calculate... The statistic shows that when m increases from 1 to 4, E1(m) rapidly decreases and tends to stabilize; the saturation point is determined to be m=4. (Execution) Statistical calculations showed no significant non-monotonic jumps in the curves, ruling out strong periodicity effects, and the final embedding dimension was determined to be 4. Under different operating conditions (e.g., 20% and 60% humidity environments), this embedding dimension maintained the complete unfolding of the time series phase space structure. The noise suppression capability of the local phase space trajectory constructed in S2.5 was significantly improved in Betti number evolution analysis, and the stability of the topological perturbation index under high-frequency fluctuations was enhanced, resulting in a significant improvement in the performance of tension state trend prediction. S2.5: Perform sliding window topological embedding processing to convert the strain difference sequence into a delayed coordinate vector space. Based on the set window length and the determined embedding dimension, construct the local phase space trajectory of the tension state manifold to provide a dynamic evolution characterization for topological perturbation analysis.

[0014] like Figure 2 As shown, step S3 involves calculating the time derivatives of the number of 0-dimensional connected branches and the number of 1-dimensional weighted loop structures based on the local phase space trajectory, generating a topological perturbation index. This index characterizes the dynamic evolution of the spatiotemporal topological structure of the tension field. Specifically, this includes: S3.1: An adaptive persistent homology algorithm is constructed based on topological data analysis theory. Taking the local phase space trajectory as input, the persistent graph is generated by calculating the persistent homology group of point cloud data, and the real-time estimates of the number of 0-dimensional connected components and the number of 1-dimensional weighted ring structures are output, providing an algorithmic basis for the accurate calculation of Betti number. Based on local phase space trajectory point cloud data, an adaptive persistent cohomology algorithm (parameters: filter scale σ dynamically adjusted, scale search range 0.1 to 2.0, adjustment step size 0.05) is adopted to achieve automatic extraction of multi-scale topological features; Furthermore, by constructing Rips complexes (parameters: distance metric is Euclidean distance, critical radius sequence is set according to 50% of the maximum point spacing of the sampling window), simplex sets at different scales are generated, and multi-level nested relational data of the connection structure are obtained; Furthermore, a persistent computation method is employed (parameters: connectivity judgment based on disjoint-set data structure optimization and ring structure detection based on shortest cycle search) to achieve persistent homology groups during self-scale growth. and The time evolution sequence is generated and the structural lifecycle information is output in the form of a persistent graph; Furthermore, an adaptive threshold update mechanism (parameters: Betti number change rate momentum coefficient 0.85, update period once every 50 windows) is used to adjust the homology group truncation condition, thereby achieving automatic removal of spurious features caused by noise and generating real-time stable Betti number estimates. This algorithm transforms the local phase space trajectory from the previous step into a sequence of 0-dimensional connected branches and 1-dimensional weighted ring structures over continuous time, enabling a high-confidence measurement of the dynamic evolution characteristics of the spatiotemporal topology of the tension field. For example, in strain monitoring of a certain type of cable stranding equipment, a local phase space trajectory containing 100 sampling points in a single window is selected. The initial value of the filter scale σ is set to 0.2, and a dynamic adjustment mechanism is enabled. The Rips complex is calculated within the scale growth range of 0.1 to 2.0, and the critical radius sequence is set to three levels: 0.15, 0.30, and 0.45 mm. Fast connectivity detection using a disjoint-set data structure yields the results for the first radius level. The initial value is 15, which decreases to 3 under the third radius; the shortest cycle search is used to obtain the value under the second radius. The initial value is 5, and it stabilizes at 2 under the third radius. The above... and The change process is plotted as a persistence graph, and a truncation condition is applied to remove spurious features with a lifetime less than 0.05, ultimately yielding a real-time estimate. 3. The value is 2. This output is directly input into the subsequent time derivative calculation module, which significantly improves the noise robustness and structural instability prediction accuracy in the topological perturbation exponent generation process; S3.2: Apply the adaptive persistent cohomology algorithm to the local phase space trajectory, process the phase space point cloud data in each sliding window in real time, calculate the time series of the number of 0-dimensional connected components and the number of 1-dimensional weighted ring structures, and obtain the dynamic evolution data of the Betti number. S3.3: Perform central difference numerical differentiation on the time series of zero-dimensional connected components, calculate the absolute value of its time derivative to quantify the rate of change of connected components, and output the sequence of absolute values ​​of the time derivative of the zero-dimensional connected components. S3.4: Perform central difference numerical differentiation on the time series of the 1-dimensional weighted ring structure number, calculate the absolute value of its time derivative to quantify the rate of change of the ring structure, and output the sequence of absolute values ​​of the time derivative of the 1-dimensional weighted ring structure number. For the time series of 1-dimensional weighted ring structure numbers, the central difference numerical differentiation method (parameter: time step size is consistent with sliding window sampling period) is used to realize the quantitative calculation of the rate of change of the ring structure over time. Furthermore, by introducing a symmetrical sampling mechanism for forward and backward time points into the central difference operator (parameter: the index offset of the phase space trajectory point cloud sequence is ±1 sampling period), a smooth estimation of the instantaneous change trend is achieved, and a numerical expression for the time derivative is obtained. Furthermore, an absolute value transformation algorithm (parameter: the absolute value operator operates on each element of the derivative vector) is used to achieve unified amplitude quantization of the positive and negative changes of the time derivative and generate an amplitude sequence of the rate of change of the ring structure. Furthermore, by matching the timestamps of the change rate data of the zero-dimensional connected components generated in the previous sub-step S3.3, we ensure that the two change rate sequences are strictly aligned in the time domain, providing a basis for subsequent weighted summation calculations. By using the above-mentioned central difference numerical differentiation and absolute value transformation processing method, the changing trend of the ring structure in the local phase space trajectory is transformed into high time resolution amplitude quantization data, thereby realizing the accurate rate characterization of the dynamic evolution of the tension field topology. For example, under the condition that the inertial time constant of the winding equipment is calibrated to 10ms and the sliding window sampling period is configured to 1ms, the time series sequence of the input 1D weighted ring structure number is {2,3,5,4,3,5}. Using the central difference method, with the sampling period set to 1ms, the time derivative of the third sampling point is calculated: The absolute value is then taken to obtain the rate of change amplitude of 1 unit / ms. Performing this operation on the entire sequence generates a sequence of absolute rate of change of {1,1,0.5,...}. This sequence is then weighted and merged with the rate of change sequence of the zero-dimensional connected component in the subsequent calculation of the topological perturbation index, realizing a multi-dimensional rate representation of the dynamic evolution of the tension topology. S3.5: The absolute value sequence of the time derivative of the number of 0-dimensional connected branches and the absolute value sequence of the time derivative of the number of 1-dimensional weighted ring structures are weighted and summed with a weighting coefficient of 0.7 to generate a topological perturbation exponential time series signal, which characterizes the dynamic evolution characteristics of the spatiotemporal topology of the tension field. Based on the absolute value sequence of the time derivatives of the 0-dimensional connected component number output by S3.3 and the absolute value sequence of the time derivatives of the 1-dimensional weighted ring structure number output by S3.4, a weighted summation algorithm (with the weight coefficient fixed at 0.7) is used to achieve a fusion representation of the change rates of the two types of Betti numbers. By performing a sequence element-level pairing operation before numerical accumulation, the data at corresponding sampling times are strictly consistent to prevent topological perturbation exponential deviation caused by time mismatch. Furthermore, by using a point-by-point weighted summation method, the absolute value of the time derivative of the number of 1D weighted ring structures at each sampling time is multiplied by a coefficient of 0.7, and then arithmetically summed with the absolute value of the time derivative of the number of 0D connected components at the same sampling time to generate the instantaneous value of the topological perturbation exponent at that sampling time. ; Furthermore, the topological perturbation index sequence is calculated using the following formula:

[0015] in, The rate of change of the number of 0-dimensional connected components. This represents the rate of change over time of the number of 1-dimensional weighted ring structures. The time sampling interval; Furthermore, by serializing and storing the above instantaneous values ​​over the entire sliding window range, a complete time-series signal of the topological perturbation index is formed, ensuring that the signal can continuously reflect the dynamic evolution characteristics of the spatiotemporal topology of the tension field. By using a weighted summation algorithm, the change rates of multiple dimensions of S3.3 and S3.4 are fused into a single scalar index, thereby improving the sensitivity of interference identification and providing a quantifiable description of the topological state. For example, under a certain cable stranding condition, the absolute value sequence of the time derivative of the number of 0-dimensional connected branches output by S3.3 is [0.15, 0.22, 0.18, 0.27], and the absolute value sequence of the time derivative of the number of 1-dimensional weighted ring structures output by S3.4 is [0.09, 0.14, 0.11, 0.16]. Using a weighted summation algorithm, each element of the 1-dimensional sequence is first multiplied by 0.7 to obtain [0.063, 0.098, 0.077, 0.112], and then added to the corresponding 0-dimensional sequence elements to obtain the topological perturbation index sequence [0.213, 0.318, 0.257, 0.382]. The continuous change curve of this sequence in the time domain can intuitively characterize the dynamic evolution trend of the tension field from stability to instability. In this embodiment, the disturbance index of the fourth sampling point reaches 0.382, which is significantly higher than that of the first three sampling points, triggering the instability judgment logic processing in the subsequent S4 step, thus realizing a rapid early warning for tension feedforward compensation control.

[0016] like Figure 3 As shown, step S4 involves comparing the topological disturbance index with a dynamic baseline threshold. This dynamic baseline threshold is updated on a rolling basis using the 99.5th percentile of the topological disturbance index under historical undisturbed conditions. When the topological disturbance index exceeds this threshold for three consecutive sampling periods, it is determined that potential structural instability has occurred in the tension field. Specifically, this includes: S4.1: Perform quantile calculation on the topology disturbance index sequence under historical undisturbed operating conditions to obtain the 99.5th quantile as the benchmark value for the dynamic baseline threshold; implement a rolling window update mechanism based on this benchmark value to generate the real-time dynamic baseline threshold. S4.2: Based on the real-time dynamic baseline threshold generated in S4.1, the topology perturbation index of the current sampling period is numerically compared to generate a threshold comparison result signal; Based on the real-time dynamic baseline threshold generated by S4.1, a numerical comparison algorithm (parameters: input signal is the topological perturbation index of the current sampling period, comparison benchmark is the real-time dynamic baseline threshold) is used to realize the quantitative calculation of the difference between the single-period perturbation amplitude and the baseline level. Furthermore, by using the difference operation method (parameter: difference = current topological disturbance index − dynamic baseline threshold), the symbolic result of the disturbance amplitude is extracted, and comparative label data that periodically exceeds or falls below the baseline is obtained; Furthermore, a threshold sign determination algorithm is adopted (parameters: threshold deviation tolerance is set to zero, and the sign domain determination rule is to output an over-threshold flag when the value is greater than zero and an under-threshold flag when the value is less than or equal to zero) to realize the sign domain mapping of the difference result and generate a Boolean threshold comparison flag. Furthermore, through state encoding processing (parameters: input is a Boolean threshold comparison flag bit, encoding rule is that the threshold exceedance indicator corresponds to state 1, and the non-threshold exceedance indicator corresponds to state 0), the state encoding of the threshold exceedance behavior in the current sampling period is realized, and a single-period threshold comparison result signal is generated; By using numerical comparison and state coding, the dynamic baseline threshold generated in the previous step and the current topological disturbance index are transformed into periodic threshold comparison results, thereby achieving the technical effect of real-time identification of whether a single-period disturbance exceeds the dynamic baseline. For example, during the operation of a cable stranding device, the topology disturbance index of the distributed fiber optic grating sensor array output for the current sampling period is 0.0185, and the dynamically updated baseline threshold calculated by S4.1 is 0.0156. The period difference is calculated using the difference operation method as 0.0185 - 0.0156 = 0.0029. Applying the threshold sign determination algorithm, with the deviation tolerance set to 0, an over-threshold flag is output because the difference is greater than zero. After state coding processing, the over-threshold flag is mapped to state code 1, which is used as the current period output of the threshold comparison result signal. During continuous operation of the device, this signal remains at state code 1 for the next two consecutive sampling periods. Combined with the continuity detection in S4.3, continuous over-threshold behavior can be further determined, entering the structural instability determination link, and achieving rapid early warning of abnormal tension field conditions. S4.3: Perform continuity detection algorithm processing on the threshold comparison result signal generated in S4.2 to identify topological perturbation exponential overthreshold behavior for three consecutive sampling periods; S4.4: Based on the continuous over-threshold behavior identified in S4.3, perform instability judgment logic processing to output a potential structural instability judgment signal for the tension field; Based on the continuous over-threshold behavior detection results generated by S4.3, a multi-condition logic judgment method (parameters: number of continuous sampling periods = 3, over-threshold type = dynamic baseline threshold) is adopted to realize the function of identifying potential structural instability in the tension field; Furthermore, by using a state machine decision algorithm (state set: stable state, pre-instability state, unstable state; transition condition: number of times η(t) exceeds the threshold and time interval), the mapping from detection data to stable state is realized, and the unstable state trigger signal data is obtained; Furthermore, by constructing a decision function, the threshold comparison result signal and the continuity detection result are logically ANDed to generate an explicit instability confirmation Boolean variable and a corresponding trigger flag bit; Furthermore, an event-triggered output mechanism is adopted (trigger condition: instability confirmation Boolean variable = true) to realize the real-time output of potential structural instability judgment signal and generate input event packets for downstream feedforward compensation module to call; Through the above logical judgment and event triggering processing, the overthreshold behavior detection result of the previous step is transformed into a formatted tension field potential structural instability judgment signal, so as to achieve the expected technical effect of providing accurate and low-latency triggering conditions for the topology guidance feedforward module. For example, on a stranding production line equipped with 12 fiber optic strain sensing units, the continuous sampling frequency is 10 kHz, and the dynamic baseline threshold is determined to be 0.042 by S4.1 rolling calculation. In three consecutive sampling periods, the actual topology perturbation index η(t) sequences are 0.051, 0.054, and 0.053, respectively. A decision function is used...

[0017] in The dynamic baseline threshold is 0.042. This represents the number of consecutive threshold exceedances. The first sub-condition indicates the current... If the threshold is exceeded, the second sub-condition indicates that the number of consecutive threshold exceedances is equal to 3. When I is true, an instability judgment signal is output. Calculation shows that when I is true, the system state machine directly transitions from the stable state to the unstable state. The event trigger module outputs a structural instability judgment signal, which is then called by the topology-guided feedforward module of S5 within 1 ms. This enables early intervention against high-frequency mechanical disturbances, significantly improving the response speed and stability of the control system under complex operating conditions.

[0018] Step S5: After determining structural instability, the topology-guided feedforward module is triggered, inputting the strain-position map structure of all sensing points within the current window into a pre-trained graph attention network model. This model aims to minimize the peak increase of the topology perturbation exponent within the subsequent 20 milliseconds, and outputs the torque correction direction and relative amplitude weight of each winch drive unit. Specifically, this includes: S5.1: Based on the time series data of topology perturbation index under historical undisturbed working conditions, a training sample set for the graph attention network model is constructed. The optimization objective is to minimize the peak increase of the topology perturbation index within the next 20 milliseconds. The network parameters are trained through the backpropagation algorithm to generate a pre-trained graph attention network model. The input conditions include obtaining time series data of the topological disturbance index under historical undisturbed operating conditions. This data is acquired by a distributed fiber optic strain sensor array under stable tension and calculated in step S3. sequence; A time series segmentation method (parameters: window length 20 ms, sliding step size 5 ms) is used to divide the topological perturbation index time series into equal-length analysis segments to achieve a balanced temporal distribution of the samples; Furthermore, through a graph structure mapping algorithm (parameters: node represents the coordinates of the sensing point, edge weight is the strain difference value), the strain-location information corresponding to each time segment is used to generate an initial graph structure representation, and a topological perturbation index is added as a global graph feature to achieve joint encoding of the spatial relationship and temporal variation of the tension field; Furthermore, a feature normalization method (parameters: mean 0, variance 1) is used to normalize all node features and global graph features, reducing the impact of feature scale differences on training stability. Using a training sample annotation method, the target output is set as the relative increase of the peak value of the topological perturbation exponent within the next 20 milliseconds, which serves as the supervision signal, and a loss function is constructed accordingly:

[0019] in, The output value of the loss function. The total number of samples, It is the exponential time function of topological perturbation; Furthermore, a graph attention network training method (parameters: 8 attention heads, 64 hidden layer dimensions, and 0.2 dropout rate) is adopted. The training sample set is input into the network model, forward propagation is performed to calculate the attention weights between nodes, and the target amplification value is predicted in the output layer. Using the backpropagation algorithm (parameters: optimizer is Adam, learning rate...) To minimize the above loss function, iteratively optimize the network parameters during training until the validation set loss converges; The optimal parameter combination is selected through performance evaluation of the training set and validation set. The trained network model is saved as a pre-trained graph attention network model for subsequent calculation of feedforward compensation instructions after instability determination. For example, on a cable stranding device containing 12 distributed fiber Bragg grating sensing nodes, a 600-second time series of topological perturbation exponents under historical undisturbed operating conditions was acquired at a sampling rate of 10 kHz. The data was divided into sample segments with a window length of 20 ms and a sliding step size of 5 ms. Each segment corresponds to the strain difference matrix of 12 nodes and the corresponding... Values. Through feature normalization, node features are compressed to a standard normal distribution with a mean of 0 and a variance of 1. The target output is set to be within 20 ms after the segment ends. The difference between the peak value and the current value generates supervisory annotations. In model construction, a graph attention network structure with 8 attention heads and 64 hidden dimensions is used, with a Dropout rate of 0.2 to prevent overfitting, and the Adam optimizer learning rate is set to 0.0001. During training, the loss function is calculated iteratively batch by batch according to the above formula, until the loss on the validation set stabilizes at [value missing] after 250 rounds. The final pre-trained model showed a significant reduction in prediction error on the test set, and was able to accurately output predictions for the peak increase of the topological perturbation exponent within the next 20 ms, which could be used to guide the feedforward correction decision of the winch drive torque. S5.2: After receiving the instability judgment signal of the tension field structure, activate the execution process of the topology guidance feedforward module and switch the module state to the ready mode to prepare to receive input data. After receiving the instability determination signal of the tension field structure, an event-driven module state management algorithm (parameter: the instability determination signal source is the data trigger output by S4.4) is used to realize the state switching control of the topology guidance feedforward module; Furthermore, by using a finite state machine modeling method (parameters: state set = {idle, ready, running}, transition condition = instability judgment signal is valid and timestamp matches the current control cycle), the state transition from idle state to ready mode is realized, and the module is locked to prevent interference from concurrent tasks; Furthermore, through signal channel initialization processing (parameters: input buffer size = current sliding window sampling capacity, communication protocol = high-speed data bus, FIFO queue length = 128), the data path pre-configuration and synchronous access preparation of the sensor array are realized, and an input buffer structure with non-blocking read and write capabilities is obtained. Furthermore, the parameter file of the pre-trained graph attention network is loaded through the real-time task scheduler (parameters: storage path is output by S5.1, model version number is matched by hash check), realizing the preloading of the model in the execution environment and generating a running instance with acceptable strain-position graph structure input; By using a system resource self-check and load assessment algorithm (parameters: CPU usage threshold = 65%, memory usage threshold = 70%, I / O latency threshold = 2ms), the state switching results of the previous step are transformed into performance preparation indicators for module operation, thereby enabling rapid activation and stable execution of the feedforward compensation control process. For example, in a cable stranding equipment control server with a 12-core processing unit and 8 GB of memory, upon receiving an instability determination signal output by S4.4 (timestamped at the 5th millisecond of the current period, signal strength at high priority), the event-driven module state management algorithm immediately switches the topology-guided feedforward module from idle to ready mode. In the finite state machine's state transition diagram, the idle → ready transition condition is that the instability determination signal is valid and matches the period. After the system meets this condition, it performs signal channel initialization processing, configuring the input buffer size to a sliding window sampling value of 96, the communication protocol to a PCIe high-speed bus, and the FIFO queue length to 128, ensuring that the strain differential sequence enters the module cache without delay. The task scheduler loads the pre-trained graph attention network parameters stored in the path " / models / GAT_v3.chk", and verifies the version number using SHA-256 hash verification to ensure model consistency. During the resource self-check, the CPU utilization rate was 52%, the memory utilization rate was 64%, and the I / O latency was 1.2ms, all below the set thresholds. The performance readiness indicators were deemed satisfactory, and the module entered a stable, ready state capable of accepting input data. This implementation activated the topology guidance feedforward module in 3.4 milliseconds, significantly improving the startup response speed of the feedforward compensation control and effectively supporting the real-time performance and stability of the subsequent S5.3 data input and inference computation links. S5.3: Obtain the strain-position map structure information output by the distributed fiber optic grating sensing array within the current sliding window, and format it as the input feature tensor of the graph attention network model; S5.4: Input the formatted input feature tensor into the pre-trained graph attention network model, perform graph convolution and attention weight calculation operations to generate the torque correction direction vector and amplitude weight coefficient of each axle drive unit; Based on the formatted input feature tensor, a graph convolutional network algorithm (parameters: the node feature dimension is a combination of strain amplitude and displacement coordinates, and the adjacency matrix is ​​defined by the spatial topological relationship of the sensing points) is used to achieve weighted aggregation processing of local strain modes among the nodes of the sensing array. Furthermore, by using the attention weight calculation method (parameters: multi-head attention mechanism, 8 attention heads, and Softmax function for normalization coefficients), the relative importance of different sensing points and their neighborhoods in the strain feature propagation process is calculated, and the attention coefficient matrix between node pairs is obtained. Furthermore, a weighted message passing mechanism (parameter: adjacency matrix reconstruction based on attention coefficient matrix) is adopted to sum the neighborhood feature vectors of each node according to the attention weights and generate an updated node representation vector to integrate local geometric structure and strain dynamic change information. Furthermore, by using the output layer mapping function of the graph attention network (parameter: the dimension of the fully connected weight matrix matches the number of the winch drive units), the updated node representation is mapped to the torque correction direction vector and amplitude weight coefficient of the corresponding drive unit, while keeping the output dimension consistent with the control interface parameters. By using the above graph convolution and attention weight joint calculation algorithm, the input feature tensor of the previous step is transformed into torque correction direction and amplitude weight data of each winch drive unit, so as to achieve the expected technical effect of real-time feedforward compensation parameter generation based on the spatiotemporal topology of the tension field. For example, for a cable stranding device with a four-cage structure, the strain-position map structure output by 12 distributed fiber optic grating sensor points within the current window is obtained. The node features include two-dimensional features: standardized strain values ​​and distance coordinates along the stranding axis. The adjacency matrix is ​​constructed by the spatial arrangement of the sensor points. The input feature tensor has a dimension of 12×2, which is updated to a 12×16 feature matrix after processing by a graph convolutional layer. The attention weight calculation process sets the number of heads to 8, uses the Softmax function for normalization, and the weight matrix size of each attention is 16×16. After weighted message passing, a new node feature of 12×32 is generated. The output mapping layer weight matrix is ​​configured with a dimension of 32×8, corresponding to the positive and negative directions of the four cage drive units, finally obtaining the torque correction direction vector (0.12, -0.08, 005, -0.03) and amplitude weight coefficients (0.7, 0.6, 0.65, 0.62). The peak increase of the topology perturbation index was significantly reduced in the subsequent 20-millisecond period, proving that the feedforward compensation parameters generated in this step achieved the expected fast response and stability improvement effect. S5.5: The calculated torque correction direction vector and amplitude weighting coefficient are output to the cooperative control module as the decision basis for feedforward compensation commands to drive the actuator.

[0020] Step S6: Geometrically project the weight vector output by the graph attention network onto the velocity-tension reference trajectory planned by the main control system, calculate the feedforward compensation command increment for each winch drive unit, and directly inject this increment into the position loop feedforward channel of the servo driver to achieve phase delay bypass. Specifically, this includes: S6.1: Perform data synchronization processing on the torque correction direction and relative amplitude weight vector output by the graph attention network, as well as the velocity-tension reference trajectory planned by the main control system, to generate the input dataset for geometric projection; The torque correction direction vector and relative amplitude weight vector output by the graph attention network, as well as the velocity-tension reference trajectory planned by the main control system, are aligned using a time synchronization alignment method (parameters: sampling frequency 10 kHz, timestamp accuracy ±0.1 ms) to achieve alignment of each data source under a unified time base. Furthermore, by using a signal resampling algorithm (parameters: cubic spline interpolation, interpolation order 3), unified processing of data with different sampling frequencies is achieved, and a precisely matched time series matrix is ​​obtained. Furthermore, a normalization method (parameters: zero mean, unit variance standardization) is adopted to achieve scale unification between data of different dimensions, and to generate a standardized torque correction vector and a velocity-tension reference vector; Furthermore, based on the data structure mapping algorithm (parameters: two-dimensional matrix dimension matching, column alignment rules), the structural coupling between the torque correction direction vector and the velocity-tension reference trajectory is realized, forming an input dataset that meets the requirements for geometric projection calculation; Through the above data synchronization and structural coupling processing method, the network output results of the previous step and the reference trajectory of the main control system are transformed into a standardized input matrix required for geometric projection calculation, thereby achieving the accuracy and stability of subsequent orthogonal projection. For example, in a certain type of cable stranding production line, the torque correction direction vector output by the graph attention network has a dimension of 6, corresponding to 6 winch drive units. The relative amplitude weight vector is a 6-dimensional floating-point array with a sampling frequency of 10240 Hz. The reference trajectory of the main control system is a 12-dimensional vector group composed of velocity and tension components with a sampling frequency of 9800 Hz. Timing synchronization alignment is performed, resampling both types of data at a frequency of 10 kHz and using cubic spline interpolation. After interpolation, zero-mean unit variance normalization is performed. Finally, the two types of vectors are merged into an input matrix of size (6×2) according to column alignment rules. Each row of the matrix corresponds to a winch drive unit, containing the correction direction and target velocity-tension pair of that unit. Verification shows that this processing method maintains significantly improved matrix matching accuracy even with a high rate of change in relative amplitude weights, ensuring the stability of subsequent geometric projection and compensation command calculations. S6.2: Based on the vector space projection algorithm, perform orthogonal projection calculation on the input dataset to generate projection coefficient vector; Based on the vector space projection algorithm (parameter: the input dataset contains the torque correction direction vector output by the graph attention network). With relative magnitude weight vector The speed-tension reference trajectory vector planned by the main control system This enables orthogonal decomposition calculation between torque correction information and the reference trajectory; Furthermore, the Gram-Schmidt orthogonalization method (parameter: vector set) is used. This allows us to construct orthogonal bases for each input vector within a common subspace, resulting in independent sets of basis vectors. ; Furthermore, by solving the projection coefficient formula This achieves a linear projection of the torque correction direction vector onto the reference trajectory basis vector, and obtains the first projection coefficient. Similarly, the linear projections of the amplitude weight vector onto the second basis vector and the linear projections of the master control system reference trajectory vector onto the third basis vector are calculated to obtain the second projection coefficients. With the third projection coefficient ; Furthermore, through combination Together with the basis vector set, they form a complete projection coefficient vector. This vector is used for incremental calculation of subsequent feedforward compensation instructions; By using the vector space projection algorithm, the input dataset processed in the previous step is transformed into a projection coefficient vector, realizing the geometric mapping relationship between torque correction information and reference trajectory, and providing an accurate parameter basis for subsequent torque-velocity mapping and incremental solution. For example, in a cable stranding device embodiment, the torque correction direction vector output by the graph attention network is (0.8, 0.3, 0.5), the relative amplitude weight vector is (1.2, 0.9, 0.4), and the main control system reference trajectory vector is (0.7, 0.5, 0.6). After Gram-Schmidt orthogonalization, the basis vector set is obtained as (1,0,0), (0,1,0), and (0,0,1). Using the projection formula... achievable Similarly, calculate , The final generated projection coefficient vector (0.8, 0.9, 0.4) significantly improves the calculation accuracy of the feedforward compensation command in the subsequent torque-speed mapping, ensuring that the winch drive unit maintains stable tension control performance under complex disturbance conditions; S6.3: Based on the projection coefficient vector and the preset torque-speed mapping relationship, calculate the feedforward compensation command increment of each winch drive unit; S6.4: The calculated feedforward compensation command increment is directly injected into the position loop feedforward channel of the servo driver to bypass the PID control loop; S6.5: Through the real-time injection operation of the position loop feedforward channel, the inherent phase delay of the mechanical transmission chain is compensated, ensuring the immediate effect of the tension control command; The incremental signals of the feedforward compensation command of each winch drive unit obtained by S6.4 are input into the position loop feedforward channel control interface of the servo drive. The high-speed digital bus direct writing method (data bandwidth ≥ 5 Mbps, update cycle ≤ 0.5 ms) is used to realize the real-time coverage of the control variables inside the position loop of the servo drive and ensure that the signal is not buffered and accumulated in the transmission link. Furthermore, through the hardware-level priority scheduling mechanism of the position ring (the priority setting value is higher than the request level of the speed ring and the current ring), the compensation command is preemptively inserted into the loop execution queue, and the loop execution delay is guaranteed to be less than the matching value of the system's maximum inertial time constant, so as to meet the requirements of high dynamic response; Furthermore, a phase response correction algorithm (with parameters calibrated based on the inherent delay characteristics of the transmission chain) is adopted to calculate the phase advance amount that the compensation signal needs to apply when entering the position loop, and generate a phase correction coefficient. This coefficient is then directly multiplied into the time axis mapping function of the compensation command to realize the advance effect of the signal in the mechanical drive link. Furthermore, by utilizing a dual-loop consistent sampling mechanism, the instantaneous phase difference between the position loop feedforward input and the spindle encoder feedback is synchronously acquired. An adaptive filtering algorithm (with a bandwidth of 500 Hz and a damping factor of 0.9) is used to correct the residual delay within the loop in real time, and the closed-loop execution status index after delay compensation is output. Through the real-time injection, priority preemption, phase advance and delay filtering of the above position loop feedforward channel, the compensation command calculated in the previous step is transformed into a drive signal that is executed by the hardware in real time, so as to realize the effective compensation of the phase delay of the mechanical transmission chain, and ensure that the tension control command takes effect within the target time window and is not affected by the phase lag of the traditional PID loop. For example, on a cable stranding production line with a rated speed of 3 m / s and an inertial time constant of 10 ms, the calculated compensation command increment of a certain winch drive unit is... rad / s. This signal is input to the position loop feedforward channel via the CANopen bus, with a data update period of 0.5 ms and an interface priority parameter of 9 (out of 10) to ensure that commands instantly overwrite the position loop's internal buffer. A phase response correction algorithm is used, and based on the measured inherent delay of 3ms in the drive train, the phase advance coefficient is calculated as follows: The compensation signal time axis should be adjusted in advance accordingly. The residual phase difference between the position loop feedforward input and the encoder feedback was measured to be -0.3 ms during execution. The residual phase difference was 0.02 rad after filtering, reducing it to 0.005 rad, thus verifying the accuracy of the compensation execution. In the above implementation, the tension response time was reduced from the original 13 ms to 9.5 ms, significantly improving the system's stability and real-time performance under high-speed winding conditions.

[0021] Step S7: After the compensation instruction is executed, a counterfactual reconstruction is performed on the Betti number evolution path of the next sliding window. By comparing the difference in topological perturbation exponent between the reconstructed path and the actual path, a topological consistency verification result is generated. Specifically, this includes: S7.1: Perform feature engineering on the time series data of Betti number evolution path stored under historical undisturbed conditions, extract the state features of the number of zero-dimensional connected branches and the number of one-dimensional weighted ring structures in the current sliding window, and perform lightweight linear regression algorithm training based on the extracted state features to build a counterfactual reconstruction model that predicts the Betti number evolution path of the next window under the condition of no compensation instruction, and output a counterfactual reconstruction model with real-time prediction capability. S7.2: Using a counterfactual reconstruction model, the evolution path of the Betti number in the next sliding window after the compensation instruction is executed is reconstructed using counterfactual methods. The number of zero-dimensional connected branches and the number of one-dimensional weighted loop structures in the current window are input into the counterfactual reconstruction model to generate a predicted topological perturbation index sequence as the reconstruction path. The counterfactual reconstruction model based on S7.1 with real-time prediction capability receives the number of zero-dimensional connected branches and the number of one-dimensional weighted ring structures of the current sliding window as input feature vectors, performs normalization processing (the parameter range is determined by the statistical range of historical undisturbed working conditions), and realizes feature dimension consistency to eliminate the impact of scale difference on the model calculation accuracy. Furthermore, by using the feature vector mapping algorithm (parameter: mapping weight coefficients are derived from the training results of the counterfactual reconstruction model), the normalized number of zero-dimensional connected branches and the number of one-dimensional weighted ring structures are mapped to the internal state space of the model, thereby realizing the construction of the basic state for predicting the evolution trend of the Betti number in the next sliding window; Furthermore, through the regression calculation module (algorithm type: lightweight linear regression, parameters: coefficient matrix solved by least squares), the predicted values ​​for the number of zero-dimensional connected components and the number of one-dimensional weighted loop structures in the next sliding window are generated. The calculation process follows the following formula:

[0022] in, To predict the output vector, The coefficient matrix, The current sliding window feature input vector; Furthermore, using a topological perturbation index generation algorithm (with a fixed weight coefficient of 0.7), the predicted Betti number evolution values ​​are converted into a topological perturbation index prediction sequence, achieving a seamless mapping from the Betti number domain to the topological perturbation index domain. The calculation formula is as follows:

[0023] in, This is the predicted topological perturbation index sequence. and These are the time derivatives for predicting the number of zero-dimensional connected branches and the number of one-dimensional weighted loop structures, respectively. By using sequence splicing, the continuously predicted topological disturbance indices are arranged in chronological order to form a complete reconstructed path dataset, thereby enabling a counterfactual description of the system's dynamic response after the execution of compensation instructions. Through this counterfactual reconstruction process, the feature input of the previous step is transformed into the topological perturbation index prediction sequence of the next sliding window, thereby generating basic data for quantitative evaluation of the effect of structural instability intervention. For example, in a cable stranding production line, the current sliding window has 12 zero-dimensional connected branches and 3 one-dimensional weighted ring structures, which, after normalization, are 0.48 and 0.25 respectively. The weight matrix input to the counterfactual reconstruction model is... The mapping calculation yielded a predicted zero-dimensional connected component number of 11.5 and a predicted one-dimensional weighted ring structure number of 3.4. Applying the central difference method to the derivative of the predicted values, we obtain... and The predicted value of the topology disturbance index, calculated using the formula, is 0.5 + 0.28 = 0.78. Within the sampling period, a length of 5 is formed. The reconstructed sequence was used for difference comparison, and the results showed that the deviation from the actual path was significantly reduced within the dynamic safety threshold range, proving that the compensation command has a significant effect on improving control accuracy. S7.3: Perform point-by-point difference calculation between the topology disturbance index sequence of the reconstructed path and the actual measured topology disturbance index sequence, and perform absolute value accumulation processing based on the difference results to generate the topology disturbance index difference value; S7.4: Compare the topology disturbance index difference value with the dynamic safety threshold, which is determined by the range of ±15% of the baseline value of the topology disturbance index under historical undisturbed operating conditions. When the difference value exceeds the safety threshold, an inconsistency verification result is generated; otherwise, a consistency verification result is generated. The topology consistency verification result is output for system stability assessment.

[0024] Step S8: If the topology consistency check result shows that the topology perturbation index has not fallen back to the baseline ±15% safety zone, then reduce the confidence of the graph attention network and start small-sample online fine-tuning, retraining the last layer attention head of the graph attention network using only the abnormal topology features of the most recent 5 windows. Specifically, this includes: S8.1: Perform a percentage deviation calculation on the actual value of the topology disturbance index and the dynamic baseline value in the topology consistency verification result to determine whether the topology disturbance index exceeds the safety zone range of ±15% of the baseline, and output a condition trigger signal as the basis for subsequent operations. S8.2: Based on the conditional trigger signal, perform exponential decay processing on the current confidence value of the graph attention network to generate a reduced confidence value, which is used for weight decay control in the subsequent small sample fine-tuning process; S8.3: Based on the conditional trigger signal, the topological perturbation index sequence and Betti number evolution characteristics of the most recent 5 sliding windows are obtained from the historical data cache to generate an abnormal topological feature dataset, which serves as the training input source for small-sample online fine-tuning; S8.4: Using the abnormal topology feature dataset and the reduced confidence value, perform few-sample gradient descent optimization on the last attention head of the graph attention network to minimize the prediction error and output the updated attention head parameters; S8.5: Replace the corresponding parameters in the original graph attention network model with the updated attention head parameters to generate an updated graph attention network model for real-time decision support in subsequent tension feedforward compensation control.

[0025] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0026] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for adaptive and coordinated tension control in cable stranding process, characterized in that, Includes the following steps: S1: Deploy fiber optic strain sensor arrays at key nodes of cable stranding equipment to synchronously collect the micro-strain time-series signals of the wire at the outlet of the pay-off frame, each strand traction point, and the front end of the take-up tension sensor, and generate multi-source strain differential sequences. S2: Perform sliding window topological embedding processing on the strain difference sequence between each group of adjacent sensing points to construct the local phase space trajectory of the tension state manifold; S3: Calculate the time derivatives of the number of connected branches and the number of weighted loop structures based on the local phase space trajectory, and generate the topological perturbation index; S4: Compare the topological perturbation index with the dynamic baseline threshold. When the topological perturbation index exceeds the dynamic baseline threshold for multiple consecutive sampling periods, it is determined that the tension field has experienced potential structural instability. S5: After determining structural instability, the topology-guided feedforward module is triggered, which inputs the strain-position map structure of all sensing points in the current window into the pre-trained graph attention network model and outputs the torque correction direction and relative amplitude weight of each winch drive unit. S6: Geometrically project the torque correction direction, the relative amplitude weight, and the speed-tension reference trajectory planned by the main control system to calculate the feedforward compensation command increment of each winch drive unit, and inject the feedforward compensation command increment into the position loop feedforward channel of the servo driver to achieve phase delay bypass.

2. The cable stranding process tension adaptive and coordinated control method according to claim 1, characterized in that, Step 6 is followed by: S7: After the compensation instruction is executed, perform counterfactual reconstruction on the Betti number evolution path of the next sliding window, and generate topology consistency verification results by comparing the difference between the topology perturbation index of the reconstructed path and the actual path. S8: If the topology consistency check result shows that the topology perturbation index has not fallen back to the baseline safety zone, then reduce the confidence of the graph attention network and start small-sample online fine-tuning.

3. The cable stranding process tension adaptive and coordinated control method according to claim 1, characterized in that, The distributed fiber optic strain sensing array has a strain resolution of ≤1 microstrain, linearity of ≥99.8%, temperature drift coefficient of ≤0.5 microstrain / ℃, and a sampling rate of 10kHz to 20kHz.

4. The cable stranding process tension adaptive and coordinated control method according to claim 1, characterized in that, In the sliding window topology embedding process, the window length matches the system inertial time constant calibration value of 8 to 12 milliseconds, and the embedding dimension is determined to be 4 to 6 dimensions according to the Cao method.

5. The cable stranding process tension adaptive and coordinated control method according to claim 1, characterized in that, Step S3 specifically includes: An adaptive persistent homology algorithm is constructed based on topological data analysis theory. It takes the local phase space trajectory as input, generates a persistent graph by calculating the persistent homology group of point cloud data, and outputs real-time estimates of the number of 0-dimensional connected components and the number of 1-dimensional weighted loop structures. An adaptive persistent cohomology algorithm is applied to the local phase space trajectory to process the phase space point cloud data in each sliding window in real time, calculate the time series of the number of 0-dimensional connected components and the number of 1-dimensional weighted ring structures, and obtain the dynamic evolution data of the Betti number. Perform central difference numerical differentiation processing on the time series of the 0-dimensional connected component number, calculate the absolute value of its time derivative, and output the sequence of absolute values ​​of the time derivative of the 0-dimensional connected component number; The central difference numerical differentiation process is performed on the time series of the 1D weighted ring structure number to calculate the absolute value of its time derivative and output the sequence of the absolute value of the time derivative of the 1D weighted ring structure number. The absolute value sequence of the time derivative of the number of 0-dimensional connected branches and the absolute value sequence of the time derivative of the number of 1-dimensional weighted ring structures are weighted and summed to generate the topological perturbation exponential time series signal.

6. The cable stranding process tension adaptive and coordinated control method according to claim 5, characterized in that, In the adaptive persistent coherence algorithm, the filtering scale is set to be dynamically adjusted, the scale search range is set to 0.1 to 2.0, and the change step size is set to 0.

05.

7. The cable stranding process tension adaptive and coordinated control method according to claim 1, characterized in that, Step S4 specifically includes: The topology disturbance index sequence under historical undisturbed operating conditions is processed by quantile calculation to obtain the 99.5th quantile as the benchmark value of the dynamic baseline threshold. Based on the benchmark value, a rolling window update mechanism is implemented to generate the real-time dynamic baseline threshold. The topology perturbation index of the current sampling period is numerically compared based on the real-time dynamic baseline threshold to generate a threshold comparison result signal. The threshold comparison result signal is processed by a continuity detection algorithm to identify over-threshold behavior of the topological perturbation index; Based on the above-mentioned topological perturbation index exceeding the threshold behavior, the instability judgment logic is executed, and the potential structural instability judgment signal of the tension field is output.

8. The cable stranding process tension adaptive and coordinated control method according to claim 7, characterized in that, Step 4 further includes entering the instability determination process when the topology disturbance index exceeds the dynamic baseline threshold for three consecutive sampling periods. The instability confirmation signal is triggered by an event to wake up the feedforward compensation module.

9. The cable stranding process tension adaptive and coordinated control method according to claim 1, characterized in that, Step S5 specifically includes: Based on the time series data of topology perturbation index under historical undisturbed operating conditions, a training sample set for the graph attention network model is constructed. The optimization objective is to minimize the peak increase of the topology perturbation index within the next 20 milliseconds. The network parameters are trained through the backpropagation algorithm to generate a pre-trained graph attention network model. After receiving the instability determination signal of the tension field structure, the execution process of the topology guidance feedforward module is activated, and the module state is switched to ready mode to prepare to receive input data. Obtain the strain-position map structure information output by the distributed fiber optic grating sensing array within the current sliding window, and format it as the input feature tensor of the graph attention network model; The input feature tensor is input into the pre-trained graph attention network model, and graph convolution and attention weight calculation operations are performed to generate the torque correction direction vector and amplitude weight coefficient of each winch drive unit. The torque correction direction vector and amplitude weighting coefficient are output to the collaborative control module as the decision basis for feedforward compensation commands to drive the actuator.

10. The cable stranding process tension adaptive and coordinated control method according to claim 9, characterized in that, The graph attention network model employs a multi-head attention mechanism with 8 attention heads, 64 hidden dimensions, and a Dropout rate of 0.

2. It is pre-trained using a loss function that minimizes the peak increase in topological perturbation exponent within 20 milliseconds. Training utilizes the Adam optimizer with a learning rate of [missing information]. .

Images