Roll-to-roll paper tension synergic intelligent control method for multi-color label printer

Abstract

The application provides a roll-to-roll paper tension cooperative intelligent control method for a multi-color label printer, a multi-point synchronized tension data acquisition method based on a high sampling rate fiber tension sensor array, a variational time-frequency feature extraction method and a tension propagation time sequence fingerprint construction method, and in combination with a historical interference topological knowledge base, a light-weight graph neural network model is used to realize high-reliability dynamic identification of a spatial interference path; according to the identification result, a time sequence compensation channel containing amplitude scaling, phase correction and delay compensation parameters is automatically generated, and compensation is actively injected into a servo driver through a feedforward modulation signal, so that fine dynamic adjustment of tension control is realized, the application can effectively improve the spatial tracing and interference identification capability of tension fluctuation, enhance the active inhibition and real-time adaptive compensation capability of the system to complex disturbance, and realize a substantial improvement in printing tension stability.

Description

Technical Field

[0001] This invention relates to the field of automated control of printing equipment and integration of industrial cyber-physical systems, and particularly to a roll-to-roll paper force collaborative intelligent control method for multi-color label printing presses. Background Technology

[0002] Currently, in the automation control technology of multi-color label printing equipment, tension-coordinated control, as a core means to ensure high-speed printing quality and production capacity, has widely adopted mainstream technical routes such as multi-sensor data acquisition and fusion, control strategy optimization based on physical models, and multi-level compensation algorithms. In key sections of the printing press, such as unwinding, traction, and rewinding, traditional tension control systems typically use spatially distributed tension sensors to provide real-time feedback on the paper's stress state, combined with dynamic modeling of pulleys / guide rollers, segmented PID, or active disturbance rejection control methods, to achieve basic adaptation to the equipment's operating conditions. However, current technologies still have significant limitations: First, they rely heavily on physical modeling and prior parameters, lacking the ability to adapt to dynamic uncertainties such as changes in paper properties and mechanical wear, making them prone to model mismatch and parameter drift; second, the control strategies are mostly "fixed structure" designs, only able to switch weights online rather than dynamically adjust compensation channels, making them prone to coupling misjudgments when dealing with complex multi-source and multi-cascade interference; third, they ignore the "spatiotemporal propagation" law of tension fluctuations in the mechanical transmission chain, lack dynamic tracking of interference diffusion paths, and have difficulty distinguishing between dominant and secondary interferences, resulting in ambiguous compensation thresholds and delayed response. Furthermore, existing technologies generally rely on offline training or manual parameter calibration, making it difficult to achieve rapid end-to-end interference detection without a model. After long-term operation, these systems require frequent maintenance, hindering the long-term self-driving operation of automated printing production lines and failing to meet the high-precision requirements for tension consistency in high-speed label printing. These problems directly result in insufficient tension control stability and limited interference suppression effects under complex operating conditions, necessitating a breakthrough in the traditional parameter modeling and weight fusion technical paradigm. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a roll-to-roll paper force coordination intelligent control method for multi-color label printing presses.

[0004] The technical solution of this invention is implemented as follows: a roll-to-roll paper force coordination intelligent control method for multi-color label printing presses, comprising: S1: Multiple high sampling rate tension sensors are deployed in the unwinding section, the traction roller inlet of each color group, and the rewinding section of the multi-color label printing machine. The hardware time synchronization module provides unified timing for each sensor, and collects and generates synchronized multi-point tension timing signals. S2: Perform joint time-frequency analysis on the synchronized multi-point tension timing signal, extract the instantaneous phase and amplitude of the main frequency component of each channel, and calculate the relative phase difference and amplitude attenuation ratio between channels to generate instantaneous phase component and amplitude attenuation ratio. S3: Based on the instantaneous phase component and amplitude attenuation ratio, construct the tension propagation time-series fingerprint, wherein the tension propagation time-series fingerprint is defined as a time-series feature vector containing microsecond-level phase shift, attenuation gradient and path dependence features, used to uniquely identify the current dominant interference type and its spatial action path; S4: Obtain the interference propagation topology graph database, wherein the interference propagation topology graph database is constructed based on historical interference labels and past tension propagation time sequence fingerprints, and stores the mapping relationship between interference types and propagation paths and the joint distribution entropy of edge weights; S5: Input the tension propagation time-series fingerprint into the lightweight graph neural network model, and use the interference propagation topology graph database for online matching. Dynamically weight the contribution of each edge through the graph attention mechanism to generate interference path matching results. S6: Determine whether the confidence level in the interference path matching result meets the preset threshold condition. If the highest confidence level is greater than the first threshold and the second highest confidence level is less than the second threshold, then generate a compensation trigger command. S7: Based on the compensation trigger command and the interference path matching result, a timing compensation channel is defined, wherein the timing compensation channel is configured with precise delay, amplitude scaling and phase correction parameters, and a feedforward compensation signal is injected into the servo driver through the timing compensation channel to generate a compensated tension control signal. S8: Monitor the stability of the compensated tension control signal. If the residual error exceeds the allowable range, generate model optimization parameters based on the current tension propagation time sequence fingerprint and interference path matching results to update the interference propagation topology database.

[0005] The roll-to-roll paper force coordination intelligent control method for multi-color label printing presses provided by this invention has the following beneficial effects: (1) This invention constructs a tension signal full-path sensing network based on a high sampling rate fiber optic grating sensor and nanosecond-level time synchronization, and extracts the instantaneous phase and amplitude dynamic features of multi-channel tension time-series signals by combining improved synchronous compressed wavelet transform, forming a 36-dimensional “tension propagation time fingerprint” (TPTF) containing relative phase difference, amplitude attenuation ratio and its higher-order differential information, which significantly improves the identification accuracy of the starting position, propagation path and time-varying characteristics of interference sources under complex working conditions; Compared with the traditional method that relies on physical modeling or spectrum peak analysis, this invention does not require the prior establishment of equipment dynamic equations or vibration mode identification, effectively overcoming the misjudgment problem caused by model mismatch and parameter drift, and can stably extract distinguishable spatiotemporal features in 21 typical interference scenarios such as sudden temperature and humidity changes, guide roller eccentricity, and paper moisture absorption and expansion, so that the interference identification accuracy and response sensitivity are significantly improved; (2) This invention introduces a lightweight graph convolutional neural network (GCN) combined with a graph attention mechanism. Based on the offline TPTF-interference type-propagation topology mapping database, it realizes the embedding similarity learning and matching path retrieval of real-time TPTF on multi-interference dedicated topology graphs. It captures the differences in the propagation mode of tension disturbance in different structural paths only through data-driven methods, avoiding explicit modeling and parameter estimation of the physical mechanism of the interference. This method has strong robustness and generalization ability. It can accurately output the top-3 most likely interference propagation paths and confidence levels without modifying the original PID control architecture. By setting a dual threshold decision mechanism (the highest confidence level > 0.85 and the second highest < 0.6), the reliability of compensation triggering is ensured, which greatly reduces the risk of miscompensation. Compared with the traditional technical route that relies on interference observer reconstruction, registration error mapping or multi-objective weight allocation, this invention gets rid of the dependence on auxiliary observation variables and manual parameter tuning, and realizes model-free, adaptive interference path identification. It is especially suitable for high-speed printing scenarios with multi-variable coupling and frequent switching of working conditions. (3) In this invention, the compensation execution layer adopts a “time-sequence compensation channel” mechanism for specific propagation paths. Based on the identification results, a feedforward signal with precise delay, amplitude scaling and phase correction is injected into the target color group servo system. Through zero-phase Butterworth filter shaping and independent FPGA hard real-time scheduling, the compensation signal is ensured to complete the response within ≤5 μs without introducing additional phase distortion, thus realizing the time consistency and dynamic fidelity intervention of the disturbance propagation process. This design completely avoids the calculation delay and control lag problems caused by traditional methods such as horizontal curve statistical averaging, dual-mode prediction coupling and vibration spectrum identification. It not only retains the stability of the original control system, but also enhances the active suppression capability against external time-varying disturbances. Attached Figure Description

[0006] Figure 1This is a flowchart of the intelligent control method for roll-to-roll paper force coordination for a multi-color label printing machine according to the present invention. Figure 2 This is a sub-flowchart of the intelligent control method for roll-to-roll paper force coordination for a multi-color label printing machine according to the present invention. Figure 3 This is another sub-flowchart of the roll-to-roll paper force coordination intelligent control method for multi-color label printing presses according to 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. like Figure 1 As shown, this invention provides a roll-to-roll paper force coordination intelligent control method for multi-color label printing presses, specifically including: S1: Multiple high sampling rate tension sensors are deployed in the unwinding section, the traction roller inlet of each color group, and the rewinding section of the multi-color label printing machine. The hardware time synchronization module provides unified timing for each sensor, and collects and generates synchronized multi-point tension timing signals. S2: Perform joint time-frequency analysis on the synchronized multi-point tension timing signal, extract the instantaneous phase and amplitude of the main frequency component of each channel, and calculate the relative phase difference and amplitude attenuation ratio between channels to generate instantaneous phase component and amplitude attenuation ratio. S3: Based on the instantaneous phase component and amplitude attenuation ratio, construct the tension propagation time-series fingerprint, wherein the tension propagation time-series fingerprint is defined as a time-series feature vector containing microsecond-level phase shift, attenuation gradient and path dependence features, used to uniquely identify the current dominant interference type and its spatial action path; S4: Obtain the interference propagation topology graph database, wherein the interference propagation topology graph database is constructed based on historical interference labels and past tension propagation time sequence fingerprints, and stores the mapping relationship between interference types and propagation paths and the joint distribution entropy of edge weights; S5: Input the tension propagation time-series fingerprint into the lightweight graph neural network model, and use the interference propagation topology graph database for online matching. Dynamically weight the contribution of each edge through the graph attention mechanism to generate interference path matching results. S6: Determine whether the confidence level in the interference path matching result meets the preset threshold condition. If the highest confidence level is greater than the first threshold and the second highest confidence level is less than the second threshold, then generate a compensation trigger command. S7: Based on the compensation trigger command and the interference path matching result, a timing compensation channel is defined, wherein the timing compensation channel is configured with precise delay, amplitude scaling and phase correction parameters, and a feedforward compensation signal is injected into the servo driver through the timing compensation channel to generate a compensated tension control signal. S8: Monitor the stability of the compensated tension control signal. If the residual error exceeds the allowable range, generate model optimization parameters based on the current tension propagation time sequence fingerprint and interference path matching results to update the interference propagation topology database.

[0009] Step S1: Multiple high-sampling-rate tension sensors are deployed in the unwinding section, the traction roller inlet of each color group, and the rewinding section of the multi-color label printing machine. A hardware time synchronization module is used to synchronize the timing of each sensor, collecting and generating synchronized multi-point tension timing signals. Specifically, this includes: S1.1: Obtain the spatial topology of the mechanical transmission chain of the unwinding section, the inlet and outlet of the traction rollers of each color group and the winding section of the multi-color label printing press. Based on the attenuation characteristics and propagation delay requirements of tension fluctuations on the paper transmission path, determine the physical installation coordinates of no less than seven high-precision fiber Bragg grating tension sensors and generate a sensor spatial layout scheme. The spatial topology data of the mechanical transmission chain of the unwinding section, the inlet of each color group traction roller, the outlet of each color group traction roller, and the winding section of the multi-color label printing press are obtained. The spatial coordinate mapping of the roller group position, paper transport path, and tension propagation channel is realized by using a three-dimensional parametric modeling method (parameters: transmission shaft length, guide roller diameter, roller spacing, paper width). Furthermore, the attenuation characteristics of tension fluctuations on the paper transmission path are simulated using a finite element mechanical transmission model (parameters: Young's modulus of paper, coefficient of friction, ambient temperature and humidity coefficient), and the tension fluctuation attenuation coefficient matrix of each node is output. Furthermore, the propagation delay of the tension wave at each candidate sensor location is calculated using a time-series propagation delay analysis method (parameters: paper speed, roll diameter, roll gap). ,in For transmission distance, The paper running speed is determined, and a delay distribution table is obtained for precise positioning of the sensor mounting point; Furthermore, by using a sensor layout optimization method (parameters: attenuation coefficient matrix, delay distribution table, sensor quantity limit ≥ 7), the candidate installation locations are screened and their positions are optimized to ensure that the tension fluctuation characteristics of the entire path are captured and key attenuation nodes are covered. Furthermore, by installing coordinate mapping processing, the optimized candidate sensor locations are transformed into a physical installation coordinate dataset, forming a sensor spatial layout scheme; By combining three-dimensional parametric modeling and attenuation delay analysis, the structural features of the mechanical transmission chain nodes are transformed into physical installation coordinate data, forming an optimized sensor layout scheme, which maximizes the capture of tension propagation features while meeting the requirement of installing no less than seven high-precision fiber optic tension sensors. For example, on a multi-color label printing press with a width of 350 mm, a total paper travel path length of 18.6 m from the unwinding section to the rewinding section, and a travel speed of 1.25 m / s, the topology data of the mechanical transmission chain with a guide roller diameter of 0.1 m and an average roller spacing of 1.5 m was obtained, and 3D modeling was used to realize the spatial coordinate mapping of the guide rollers. With the Young's modulus of the paper set to 3.2 GPa and the coefficient of friction set to 0.27, and under the condition of an ambient temperature and humidity coefficient of 0.018, the attenuation coefficient matrix of each node was calculated using a finite element model, yielding an attenuation coefficient of 0.94 for the unwinding section, 0.82 for the second color group inlet, and 0.78 for the third color group outlet. The propagation delay formula was then used... Calculate the delay time from the unwinding section to the entrance of the third color group. The delay was calculated to be 5.04 seconds. Using the attenuation coefficient matrix and delay distribution table, a layout optimization algorithm was employed to select seven installation locations: unwinding section, color group 1 inlet, color group 2 outlet, color group 3 inlet, color group 4 outlet, winding guide roller, and winding section. This layout scheme was mapped to physical installation coordinates, verifying 100% coverage across the entire path. Actual sensor deployment testing showed that the captured tension fluctuation characteristics were highly consistent with the simulation results in both attenuation and delay distribution comparison analysis, ensuring the accuracy of subsequent timing signal synchronization and feature extraction. S1.2: Based on the sensor spatial layout scheme, deploy high sampling rate fiber optic tension sensors at the determined physical installation coordinates, configure the sampling frequency of each sensor to above 20 kHz, and perform high-frequency discretization acquisition of dynamic tension changes during paper movement to generate original asynchronous multi-point tension time-series data stream. S1.3: Obtain the hardware module of the precision clock protocol supported by the industrial Ethernet switch, establish a master-slave clock synchronization architecture based on the IEEE 1588v2 standard, distribute the master clock timing signal to the data acquisition terminal connected to each fiber optic tension sensor, perform nanosecond-level phase calibration on the internal clock source of each data acquisition terminal, and generate a unified global time reference. An industrial Ethernet switch hardware module with precision clock protocol support is used as the core device for time synchronization. The input objects are the deployed fiber optic tension sensor data acquisition terminals and their internal clock sources. The IEEE 1588v2 master-slave clock synchronization protocol (parameter: nanosecond-level time accuracy) is adopted to realize the real-time physical clock synchronization function across terminals and obtain the master clock signal. Furthermore, by using a hardware-level delay compensation method (parameters: link delay measurement value, round-trip delay compensation coefficient), the transmission delay of the master clock signal during its distribution to each acquisition terminal is accurately predicted and corrected, and the time reference signal after delay compensation is obtained. Furthermore, a phase deviation detection and correction method (parameters: sampling period, phase measurement window) is adopted to realize real-time estimation of the phase error of the internal clock source of each acquisition terminal and generate phase correction values; Furthermore, through nanosecond-level phase calibration processing, the time reference signal after delay compensation and the phase correction value are applied to the internal clock source of each acquisition terminal to output a unified global time reference signal. By combining IEEE 1588v2 protocol time synchronization with nanosecond-level phase calibration, the original time synchronization result from the previous step is transformed into a precise and consistent global time reference, achieving the expected technical effect of strict time alignment for multi-point tension signal acquisition. For example, in a multi-color label printing machine operating at a speed of 250 m / min, an industrial Ethernet switch model A-1588X supporting the IEEE 1588v2 protocol is selected. The time accuracy of the master clock module is configured to 1 nanosecond, the link delay measurement is 32 nanoseconds, and the round-trip delay compensation coefficient is set to [value missing]. The sampling period for each data acquisition terminal is fixed at [time]. The phase measurement window length is After performing delay compensation, the transmission delay model error of the timing signal was reduced to [value missing]. Within. After applying the phase deviation detection and correction algorithm, the maximum phase difference between the clock sources of each acquisition terminal is reduced from the original value. Significantly reduced to After finally outputting a unified global time reference signal, in the subsequent sliding window data alignment process, the data frames acquired by all tension sensors can be synchronized across multiple channels at a microsecond-level time resolution, which significantly improves the accuracy and stability of tension fluctuation time-series feature extraction. S1.4: Using the unified global time base, each data frame in the original asynchronous multi-point tension timing data stream is stamped with an absolute timestamp, and time axis alignment and reordering processing is performed to eliminate timing misalignment caused by transmission delay and clock drift, and generate a synchronized multi-point tension timing signal with strict time synchronization characteristics. S1.5: Based on the synchronized multi-point tension time-series signal, perform sliding window data buffer management, encapsulate the multi-channel tension data in the continuous time domain into a fixed-length time-series data block, and output a standardized synchronized multi-point tension time-series signal dataset as the direct input object for subsequent joint time-frequency analysis processing.

[0010] Step S2: Perform joint time-frequency analysis on the synchronized multi-point tension timing signal, extract the instantaneous phase and amplitude of the main frequency component of each channel, and calculate the relative phase difference and amplitude attenuation ratio between channels to generate instantaneous phase components and amplitude attenuation ratios. Specifically, this includes: S2.1: Acquire synchronized multi-point tension time-series signals, use a sliding window truncation algorithm to segment the continuous time domain signals, and generate synchronized tension data frames containing the latest 100 milliseconds of data to establish the real-time calculation object for joint time-frequency analysis; S2.2: Based on the synchronized tension data frame, an improved synchronous compressed wavelet transform algorithm is used to perform time-frequency energy aggregation processing on the signals of each channel, sharpening the blurred time-frequency map ridges into high-precision time-frequency ridges, so as to separate the main frequency components and eliminate the spectral dispersion effect caused by noise interference. S2.3: Based on the high-precision time-frequency ridge, the instantaneous phase sequence and instantaneous amplitude sequence of the main frequency components of each channel are extracted by the Hilbert-Huang transform analytical algorithm to generate a set of instantaneous phase components and a set of instantaneous amplitude components that characterize the dynamic characteristics of single-point tension. S2.4: For the instantaneous phase component set and the instantaneous amplitude component set, perform multi-channel differential operation and ratio operation to calculate the relative phase difference value and amplitude attenuation ratio value between any two sensor nodes, and generate a relative phase difference matrix and amplitude attenuation ratio matrix characterizing the propagation delay and energy loss of tension wave in the mechanical transmission chain. For the instantaneous phase component set and the instantaneous amplitude component set, a multi-channel differential operation method (parameters: number of sensor nodes ≥ 7, time resolution = 1 μs) is adopted to realize the point-by-point difference calculation of the corresponding phase components of any two channels and form an initial relative phase difference data sequence. Furthermore, by using the amplitude ratio calculation method (parameters: amplitude component set, amplitude sampling window length = 100ms), the corresponding point division operation of any two channel amplitude components is realized, and the amplitude attenuation ratio data sequence is obtained, reflecting the decreasing ratio of tension fluctuation energy in the spatial path of the mechanical transmission chain; Furthermore, using a matrix-based recombination process (parameters: node number index, phase difference data sequence, attenuation ratio data sequence), the relative phase difference values ​​of all pairs of nodes are filled into the corresponding positions of the phase difference matrix, and the amplitude attenuation ratio values ​​are filled into the corresponding positions of the attenuation ratio matrix, ensuring that the matrix row and column indices correspond one-to-one with the spatial coordinates of the sensor nodes. Furthermore, a delay quantization calculation method is employed (parameters: phase difference matrix, main frequency). The propagation delay index is calculated using the following formula: in, As a propagation delay indicator, This represents the phase difference between the two channels (in degrees). The main frequency value (unit: Hz) is used to quantize the time delay corresponding to the phase difference; Furthermore, using the loss coefficient calculation method (parameter: amplitude attenuation ratio matrix), the energy loss index is calculated using the following formula: in, As an indicator of energy loss, This represents the amplitude attenuation ratio, used to characterize the uneven propagation of tension fluctuation energy along the path. By using the phase difference matrix and amplitude attenuation ratio matrix generated by matrixing, the instantaneous phase component set and instantaneous amplitude component set of the previous step are transformed into spatial topological quantitative data that can directly characterize propagation delay and energy loss, thereby achieving a global quantitative description of the temporal characteristics of tension fluctuations. For example, in a multi-color label printing machine production line, seven fiber Bragg grating tension sensors are deployed with a sampling frequency of 20 kHz to obtain the instantaneous phase component set and instantaneous amplitude component set within a 100 ms time window. For node 1 and node 3, the phase difference is calculated as follows: The main frequency value is Hz, the delay time is calculated using the propagation delay formula. The result is The amplitude attenuation ratio was calculated. The energy loss value is calculated using the loss factor formula. The result is The delay and loss values ​​obtained by combining all nodes in pairs are filled into the phase difference matrix and the amplitude attenuation ratio matrix. It has been verified that this matrix can accurately capture the tension propagation delay and energy loss characteristics from the unwinding section to the third color group caused by temperature and humidity changes, and significantly improves the system's identification and response speed to interference in online control compensation. S2.5: Integrate the relative phase difference matrix and amplitude attenuation ratio matrix, and perform structured recombination according to the preset time-series feature vector encapsulation protocol. Output the final result of the instantaneous phase component and the final result of the amplitude attenuation ratio containing microsecond-level phase offset information and path-dependent attenuation information, as direct input data for constructing the tension propagation time-series fingerprint.

[0011] like Figure 2 As shown, step S3: Based on the instantaneous phase component and amplitude attenuation ratio, a tension propagation time-series fingerprint is constructed. This fingerprint is defined as a time-series feature vector containing microsecond-level phase shift, attenuation gradient, and path dependence features, used to uniquely identify the current dominant interference type and its spatial action path. Specifically, it includes: S3.1: Perform multi-channel pairwise combination processing on the instantaneous phase component and amplitude attenuation ratio to calculate the relative phase difference sequence and amplitude attenuation ratio sequence between any two sensor channels, generating a channel relative phase difference sequence and channel amplitude attenuation ratio sequence with microsecond-level time resolution; The set of numbers for all sensor channels is obtained from the final results of the instantaneous phase component and the final results of the amplitude attenuation ratio, and used as the index input object for pairwise combination processing of multiple channels; A combined enumeration method (parameter: number of sensor channels ≥ 7) is adopted to generate all possible channel pair combination indices, and the instantaneous phase component data and amplitude attenuation ratio data of the corresponding channels under the same time base are called in sequence to realize the data pairing function between channels; Furthermore, by using a phase difference calculation method (parameters: microsecond-level sampling time, phase value unit: degrees), the instantaneous phase components of each pair of channels are interpolated to obtain a data sequence of relative phase differences between channels. The phase difference calculation formula is as follows: in, Let be the instantaneous phase difference between channel i and channel j. Let i be the instantaneous phase of channel i. Let be the instantaneous phase of channel j; Furthermore, using the amplitude attenuation ratio calculation method (parameters: amplitude unit in Newtons, consistent sampling time base), a ratio operation is performed on the instantaneous amplitude components of each pair of channels to obtain the inter-channel amplitude attenuation ratio data sequence. The formula for calculating the amplitude attenuation ratio is as follows: in, Let i be the amplitude attenuation ratio between channel i and channel j. Let i be the instantaneous amplitude of channel i. Let be the instantaneous amplitude of channel j; Furthermore, a time stamp binding method (parameter: nanosecond-level unified global time reference) is adopted to attach an absolute timestamp to each phase difference data point and amplitude attenuation ratio data point, thereby achieving precise recording with microsecond-level time resolution; By using a serialization process, the relative phase difference sequence and amplitude attenuation ratio sequence of each pair of channels are merged into a structured inter-channel propagation feature pair and stored in a feature buffer as a direct input for constructing the tension propagation time-series fingerprint. By using the above multi-channel pairwise combination processing method, the instantaneous phase component and amplitude attenuation ratio results of the previous step are transformed into a channel relative phase difference sequence and channel amplitude attenuation ratio sequence containing microsecond-level time resolution, thereby realizing a fine characterization of the propagation characteristics of tension waves in the spatial topological path. For example, in a multi-color label printing machine, between the unwinding section and the entrance section of the third color group, sensors numbered S1 and S4 have instantaneous phase components of 32.8° and 15.6°, and instantaneous amplitude components of 78.4 N and 55.7 N, respectively. The relative phase difference is obtained using the phase difference formula. =17.2°. According to the amplitude attenuation ratio formula, we get... =0.71. By binding with a nanosecond-level global time base, both values ​​are marked at the same absolute timestamp of 42.300 ms, forming a feature point with microsecond-level resolution, which is stored in the propagation feature buffer of the channel combination (S1, S4) for subsequent time series fingerprint construction and interference path matching; S3.2: Based on the inter-channel relative phase difference sequence and the inter-channel amplitude attenuation ratio sequence, perform a sliding window feature extraction operation, and use the difference algorithm to calculate its first-order difference feature vector and second-order difference feature vector to generate a first-order difference feature vector and second-order difference feature vector characterizing the trend of tension fluctuation. For the inter-channel relative phase difference sequence and the inter-channel amplitude attenuation ratio sequence, a sliding window feature extraction method (window length: 10 ms, step size: 5 ms) is used to sample the dynamic changes of the time series sequence in the local time domain. Furthermore, using a difference algorithm (difference order: first order), the rate of change between adjacent sampling points is calculated, and a first-order difference feature vector is obtained. The first-order difference operation formula is as follows: in, These are first-order difference eigenvalues. The value of the sampling point in the current window. This is the value from the previous sampling point; Furthermore, the acceleration of the tension fluctuation rate of change is calculated using a difference algorithm (difference order: second order), and a second-order difference feature vector is generated. The second-order difference operation formula is as follows: in, These are second-order difference eigenvalues. , , These represent the values ​​of the current and the two previous sampling points, respectively. Furthermore, by using the characteristic statistical methods of the difference sequence (statistical indicators: mean, standard deviation, kurtosis, skewness), the difference characteristics are quantitatively described in the local time domain, and a difference characteristic statistical matrix is ​​generated as the basis for trend analysis. Furthermore, by using an abnormal fluctuation detection method based on differential features (detection threshold: mean ± 3 times standard deviation), the abnormal interference bands that may exist in the tension fluctuation trend can be automatically identified, and abnormal location index data can be obtained. By using the above differential algorithm and sliding window feature extraction method, the inter-channel relative phase difference sequence and amplitude attenuation ratio sequence from the previous step are transformed into first-order differential feature vectors and second-order differential feature vectors that characterize the trend of tension fluctuation, thereby achieving accurate quantification of the dynamic change rate and acceleration of tension propagation. For example, in the relative phase difference sequence between the unwinding section and the exit of the second color group of a multi-color label printing machine, the sampling frequency is set to 20 kHz, the sliding window length is set to 200 sampling points (10 ms), and the step size is 100 sampling points (5 ms). Applying the first-order difference formula to the phase difference data within the window, the calculated rate of change values ​​reaches an average of 0.85° / ms and a standard deviation of 0.12° / ms across multiple windows. Then, applying the second-order difference formula, the acceleration values ​​are calculated, with an average of 0.03° / (ms²) and a standard deviation of 0.005° / (ms²). Based on the anomaly detection threshold (mean ± 3 times standard deviation), peak phase difference rate of change values ​​of 1.21° / ms and 1.35° / ms are detected in windows 17 and 23, respectively, with corresponding interference band position index outputs as [17, 23]. Verification showed that the interference band corresponds to the paper's moisture absorption and expansion caused by sudden changes in temperature and humidity on the production line. The differential feature analysis results accurately located the interference interval, providing stable trend feature data for the subsequent construction of tension propagation time-series fingerprints. S3.3: Perform multi-dimensional data splicing and fusion processing on the inter-channel relative phase difference sequence, inter-channel amplitude attenuation ratio sequence, first-order differential feature vector and second-order differential feature vector, and generate an initial high-dimensional tension propagation feature vector according to the preset architecture rule of superimposing 12-dimensional basic features and 24-dimensional differential features. S3.4: Based on the initial high-dimensional tension propagation feature vector, perform normalization mapping and path-dependent encoding processing, introduce sensor topological location index as spatial constraint condition, map the feature data with unified physical dimensions to a standardized feature space, and generate a standardized tension propagation feature vector with spatial path-dependent attributes. Based on the initial high-dimensional tension propagation feature vector, normalization mapping and path-dependent encoding are performed. The sensor topological location index is introduced as a spatial constraint condition. The feature data with unified physical dimensions are mapped to a standardized feature space to generate a standardized tension propagation feature vector with spatial path-dependent properties. A Z-score-based normalization algorithm (parameters: mean μ and standard deviation σ are derived from global feature statistics) is used to achieve scale unification and centering of data in each dimension of the initial high-dimensional tension propagation feature vector, and to generate a normalized feature matrix. Furthermore, by using the minimum-maximum mapping transformation method (parameters: the global minimum value min and the maximum value max are taken from the corresponding dimensions of the normalized feature matrix), the numerical domain of the feature matrix is ​​compressed to [0,1], eliminating the difference in numerical amplitude caused by different physical dimensions, and obtaining the domain-standardized feature matrix; Furthermore, by using a path-dependent encoding method (parameter: sensor topology location index table), each feature quantity is spatially associated with its corresponding sensor node location to generate a feature node matrix containing node spatial index information. Furthermore, a spatial constraint embedding vector construction method (parameters: node position coordinates X, Y, Z and mechanical connection weights between sensors) is adopted to embed the node spatial index information into the standardized feature matrix, thereby forming an embedding feature matrix with spatial path dependency properties. Through the above normalization and path-dependent encoding processes, the results of the previous step are transformed into a standardized tension propagation feature vector containing unified physical dimensions and spatial topological constraints, achieving the expected technical effect of combining spatial path-dependent features with numerical consistency. For example, in the initial high-dimensional tension propagation feature vector of the seven fiber Bragg grating sensors in a multi-color label printing machine, the first dimension is the relative phase difference, with a mean of 0.12 radians and a standard deviation of 0.03 radians; the second dimension is the amplitude attenuation ratio, with a mean of 0.68 and a standard deviation of 0.05. The corresponding normalization formula is as follows: in These are the normalized eigenvalues. These are the original eigenvalues. Let be the mean of this feature dimension. The standard deviation of this feature dimension is given. After Z-score normalization, the first-dimensional feature value is mapped to the range of [-2,2], and then the minimum-maximum mapping method is applied for mapping. After mapping, the numerical domains of relative phase difference and amplitude attenuation ratio are compressed to [0,1]. In the path dependency encoding stage, the sensor topology location index table is read. Assuming that sensor 1 is located at the entrance of the unwinding section (coordinates X=0, Y=0, Z=0) and sensor 4 is located at the exit of the third color group (coordinates X=150, Y=0, Z=0), the node position coordinates and mechanical connection weights are encoded into the feature matrix to achieve spatial constraint embedding. After the embedding vector is constructed, the feature values ​​of this matrix between nodes 1 and 4 reflect the phase difference and attenuation gradient distribution along this path and have a clear spatial sequence identifier. In actual operation, this embedded feature matrix is ​​used for subsequent interference path matching, which can significantly improve the robustness and positioning accuracy of multi-source interference identification in complex topologies. S3.5: The standardized tension propagation feature vector with spatial path dependency is structurally encapsulated, and timestamp tags and interference scene context metadata are added to generate a tension propagation temporal fingerprint that uniquely identifies the current dominant interference type and its spatial action path.

[0012] like Figure 3 As shown, step S4 involves obtaining an interference propagation topology graph database. This database is constructed based on historical interference labels and past tension propagation time-series fingerprints, storing the mapping relationship between interference types and propagation paths, as well as the joint distribution entropy of edge weights. Specifically, this includes: S4.1: Perform association retrieval processing on the historical interference label dataset and the past tension propagation time series fingerprint dataset in the storage medium to extract sample pairs with clear causal labels and generate a set of original feature sequences with interference source attribute labels; The historical interference tag dataset and the past tension propagation time series fingerprint dataset in the storage medium are obtained as input objects. Both contain structured information such as timestamps, sensor location indexes and signal feature values, and meet the requirements of uniform formatting. An association retrieval method (parameters: timestamp matching tolerance ≤ 1μs, sensor index consistency comparison) is used to achieve accurate pairing of interference tag records with the corresponding tension propagation time sequence fingerprints; Furthermore, by using a causal labeling screening method (parameters: interference label category encoding, spatial action path node sequence consistency verification), noise samples caused by temporal misalignment or path inconsistency are eliminated, resulting in a sample pairing set with clear causal relationships. Furthermore, a feature sequence concatenation processing method (parameters: basic feature dimension 12, differential feature dimension 24, path dependency encoding index) is adopted to bind the temporal feature vectors in the pairing set with the interference label categories one by one, and generate a complete original feature sequence structure. Furthermore, by using a data integrity verification method (parameters: missing value ratio threshold 0.01, outlier judgment standard is three times the absolute deviation of the median), sample pairs that do not meet the quality requirements are removed, and the final set of original feature sequences with interference source attribute labels is obtained; Through the above-mentioned association retrieval and annotation processing methods, the tension propagation time sequence fingerprint and historical interference labels from the previous step are transformed into raw feature data with clear interference categories and action path attributes, thus providing basic data support for interference source identification training and topology map construction. For example, in a multi-color label printing machine production line, the storage medium contains 12,000 historical interference label records. Each record has an interference type code (0-20), a dominant position node sequence, and a timestamp accurate to nanoseconds; it also contains a tension propagation time-series fingerprint dataset for the corresponding time, with a feature vector length of 36 dimensions, normalized to [-1,1]. The timestamp matching tolerance is set to... The sensor index consistency comparison algorithm ensures that the consistency of path nodes in paired samples is ≥98%. A causal labeling filtering method was used to remove 840 noisy samples caused by path deviations, retaining 11,160 causally clear sample pairs. A feature sequence concatenation method was employed to combine the 12-dimensional basic features and 24-dimensional differential features of each sample into a 36-dimensional feature vector, with added interference type encoding and path node index. During data integrity verification, the missing value ratio threshold was set to... 172 samples were detected and removed because they exceeded the threshold. The final output was a set of 10,988 original feature sequences that had passed quality control. This set was used to significantly improve the stability and accuracy of interference path identification when applied to subsequent topology graph construction and graph convolutional neural network training. S4.2: Based on the original feature sequence set with interference source attribute labels, execute the graph structure reconstruction algorithm to map the sensor position as nodes and the statistical distribution of the relative phase difference and amplitude attenuation ratio as edges to generate an initial interference propagation topology graph instance; S4.3: Calculate the joint distribution entropy value for each connecting edge in the initial interference propagation topology instance to quantify the uncertainty of tension fluctuation propagation on a specific path and generate the edge weight joint distribution entropy parameter; For each connecting edge in the initial interference propagation topology instance, a joint entropy calculation method based on statistical probability distribution (parameters: sampling window length of 120 milliseconds, statistical resolution of 0.05) is used to estimate the joint probability distribution of the relative phase difference data sequence and the amplitude attenuation ratio data sequence. Furthermore, by combining two-dimensional histogram statistics with kernel density estimation (parameters: the number of two-dimensional histogram grid points is 64×64, and the kernel density bandwidth is adaptively calculated based on the Silverman algorithm), the joint probability density function of relative phase difference-amplitude attenuation ratio is constructed, and the probability distribution matrix corresponding to each connecting edge is obtained; Furthermore, the edge weight entropy value of each connected edge is calculated using the joint entropy formula: in, The edge weight entropy value. The relative phase difference This is the amplitude attenuation ratio. and These are the corresponding discrete bin center values. For joint probability mass function; Furthermore, by normalizing the joint entropy values ​​(parameters: minimum value corresponds to 0, maximum value corresponds to 1), the uncertainty between different connecting edges is compared, and a set of joint distribution entropy parameters for edge weights is generated. By using joint entropy quantization, the initial topological edge data from the previous step is transformed into an accurate entropy index that can be used for subsequent weighted updates, thereby achieving standardized quantification of the uncertainty of interference path propagation. For example, in a test scenario of a multi-color label printing machine, the connection edge between the inlet sensor of the first color group and the outlet sensor of the third color group is selected. A relative phase difference sequence of 2400 points and an amplitude attenuation ratio sequence of 2400 points are collected. The sliding window length is set to 120 milliseconds, and the step size is 60 milliseconds, resulting in 40 feature pairs. A joint probability density matrix is ​​constructed using a hybrid approach of 64×64 grid-based two-dimensional histogram statistics and Gaussian kernel density estimation, with 4096 matrix elements. The matrix is ​​then input into the joint entropy calculation formula, and the accumulated entropy value is... The bit, after normalization, corresponds to the joint distribution entropy parameter of the edge weights. This value is used in the subsequent S4.4 step to weight the connection edge, thereby enhancing the significance of the path in the interference propagation features and ultimately significantly improving the probability of the interference path being correctly identified during online matching. S4.4: The initial interference propagation topology instance is weighted and updated using the edge weights and distribution entropy parameters to enhance the saliency of typical interference paths and suppress noise path interference, thereby generating a standardized interference propagation topology. S4.5: Based on the standardized interference propagation topology map, establish an index mapping mechanism to bidirectionally bind and store the interference type identifier and the propagation path topology structure, and generate an interference propagation topology map database that can be queried online; Obtain the standardized interference propagation topology graph data object after weighted update processing by edge weight joint distribution entropy, establish an index mapping mechanism, and store the interference type identifier and the propagation path topology in a two-way binding manner; A bidirectional hash mapping method is used (parameters: interference type identifier set, topology node sequence set, and edge weight joint distribution entropy matrix) to generate the mapping relationship between interference type identifiers and topology node sequences, and to obtain an index table of interference type → node sequence. Furthermore, by using the topology sequence hash reverse matching method (parameters: node sequence, interference type hash key), the reverse mapping from the topology node sequence to the interference type identifier is realized, thereby generating a reverse index table of node sequence → interference type; Furthermore, by using a mapping relationship verification method (parameters: bidirectional index table, edge weight joint distribution entropy matrix entropy threshold), the consistency verification of the bidirectional binding mapping relationship is realized, and a consistency verification dataset is generated. Furthermore, by establishing an efficient retrieval index structure (parameters: bidirectional index table, consistency check dataset, interference propagation topology graph data ID), online query responses to interference types or propagation paths are achieved, and an interference propagation topology graph database index instance that can be queried online is generated; By using a two-way binding storage mechanism, the standardized interference propagation topology map from the previous step is transformed into a database index with two-way query capabilities, enabling rapid matching and retrieval of interference types and propagation path topology to achieve the expected technical effect. For example, in a multi-color label printing production line, the interference type identifier set contains 21 types of interference sources, the node sequence set corresponds to the path sequence combination composed of 7 sensor positions, and the entropy value of the joint distribution entropy matrix of edge weights ranges from 0.12 to 0.87. A bidirectional hash mapping algorithm is used to map interference type identifiers such as "temperature and humidity change" to node sequences [unwinding section, second color group entrance, second color group exit], generating an interference type → node sequence index table; simultaneously, a topological sequence hash reverse matching algorithm is used to map node sequences [rewinding section, third color group exit, third color group entrance] back to the interference type identifier "equipment resonance", generating a reverse index table. A consistency check algorithm is used to check when the entropy value is lower than... Paths identified as noise are identified and removed to ensure the database contains only salient paths. During the search structure setup, for queries targeting the interference type "guide roller eccentricity," the database can... Within seconds, it returns the corresponding propagation path node sequence and related edge weight entropy values, enabling rapid online matching and interference control strategy invocation.

[0013] Step S5: Input the tension propagation temporal fingerprint into a lightweight graph neural network model, and perform online matching using the interference propagation topology graph database. Dynamically weight the contributions of each edge using a graph attention mechanism to generate interference path matching results. Specifically, this includes: S5.1: Obtain the standardized tension propagation feature vector as the input object, convert the spatial path dependency attribute in the feature vector into the initial feature matrix of the graph node based on the graph structure mapping rule, and map the joint distribution entropy of edge weights in the interference propagation topology graph database into the graph adjacency matrix to generate a graph data instance with spatiotemporal topological attributes to be matched. A standardized tension propagation feature vector is obtained, and a spatial path-dependent attribute mapping method (parameters: sensor topology location index, path weight encoding rule) is used to establish a one-to-one association between each feature vector element and the corresponding sensor node. Furthermore, by using a node feature matrix construction method (parameters: number of nodes = number of sensors, feature dimension = length of the vector after path dependency encoding), the standardized feature vector is converted into an initial feature matrix of graph nodes, and a two-dimensional matrix data result with node spatial coordinates and path dependency encoding is obtained. Furthermore, the physical propagation paths between nodes are mapped to graph adjacency matrix cell values ​​through the edge weight joint distribution entropy mapping processing method (parameters: joint distribution entropy values ​​come from the interference propagation topology graph database, and edge connection rules come from the sensor topology structure), and an adjacency matrix containing topological connection relationships and propagation uncertainty weights is generated. Furthermore, by using a spatiotemporal topology fusion method (parameters: node initial feature matrix, edge weight adjacency matrix), node features and adjacency relationships are combined into graph data instances with spatiotemporal topology attributes, and graph structure input objects that can be processed by graph convolutional neural networks are generated. Through the above processing method, the standardized tension propagation feature vector of the previous step is transformed into a matching graph data instance with spatiotemporal topological attributes, realizing the structured correspondence between interference propagation features and sensor spatial layout, and providing a standardized and information-complete input object for the subsequent topological embedding calculation of graph convolutional neural network. For example, under the layout of seven high-precision fiber Bragg grating tension sensors, the standardized tension propagation feature vector has a length of 36 dimensions, including 12 basic features and 24 differential features. Using sensor topology location indices 1-7, the feature vector elements are mapped to the nodes corresponding to the rows of the initial node feature matrix in index order, forming a 7×36 node feature matrix with 36 columns. The joint distribution entropy of edge weights for each pair of sensor nodes in the interference propagation topology graph database ranges from 0.15 to 0.87. These entropy values ​​are combined according to node indices using a mapping algorithm and filled into a 7×7 graph adjacency matrix, with 0s on the main diagonal to indicate no self-loop connections. Using a spatiotemporal topology fusion method, the node feature matrix and adjacency matrix are simultaneously input into the graph data instance generation module to generate a structured graph object G, which contains 7 nodes, 36 feature dimensions, and adjacency matrix weights ranging from 0.15 to 0.87. This example can be used directly in the input stage of a lightweight graph convolutional neural network. It has been verified that the G structure can significantly improve the feature fidelity of interference path matching and reduce the graph structure matching error rate during the embedding similarity calculation of samples with different types of interference. S5.2: Perform lightweight graph convolutional neural network forward propagation processing on the graph data instance to be matched, use multi-layer graph convolution operators to aggregate the feature information of adjacent sensor nodes, extract high-dimensional node embedding vectors that characterize the propagation law of tension fluctuations in the mechanical transmission chain, and generate a set of node hidden state containing local topological context information. For the graph data instance to be matched, a lightweight graph convolutional neural network forward propagation algorithm (with parameters set to kernel size 3, number of layers 2, and node embedding dimension 64) is used to achieve feature aggregation based on the local neighborhood of nodes. Furthermore, with the support of the adjacency matrix, the initial feature matrix of each sensor node is linearly combined with the features of its directly adjacent nodes by the first-layer graph convolution operator to generate a node feature update matrix containing first-order neighborhood propagation information. Furthermore, the ReLU activation function is used to perform a nonlinear mapping on the node feature update matrix to enhance the expressive power of node features in the nonlinear space and obtain the set of hidden state of the first layer node. Furthermore, the hidden state set of the first layer node and the hidden state of its neighboring nodes are aggregated by the second layer graph convolution operator, and a layer normalization parameter is introduced during the aggregation process to reduce the risk of gradient vanishing and overfitting, resulting in a high-dimensional node embedding vector set containing second-order neighborhood propagation information. By employing a global feature preservation strategy, the second-order neighborhood propagation information and the original node input features are processed through residual connection, so that the high-dimensional node embedding vector retains both the local topological context and the original feature benchmark, thereby achieving multi-scale information fusion of nodes. For example, in a matching graph data instance of a multi-color label printing machine, the number of nodes is set to 7, corresponding to the key sensor positions from the unwinding to the rewinding section. The initial feature matrix has a dimension of 12×7, consisting of instantaneous phase components and amplitude attenuation ratios. The adjacency matrix is ​​derived from the joint distribution entropy of edge weights in the interference propagation topology graph database, with edge weights ranging from 0.12 to 0.87. During the calculation of the first-layer graph convolution operator, the node feature update formula is: in, This is the first-order neighborhood aggregation feature matrix output by the first layer graph convolution. The initial characteristic matrix, It is an adjacency matrix. This is the first layer weight matrix. The activation function is ReLU. The formula for calculating the second-layer graph convolution operator is: in, This is the second-order neighborhood aggregation feature matrix output by the second-layer graph convolution. This is the weight matrix for the second layer. After two layers of convolution and residual connection processing, the output 7×64-dimensional high-dimensional node embedding vector shows a significant improvement in the accuracy test of interference recognition. It can stably characterize the propagation law of paper rolling force in the mechanical transmission chain under multi-source interference, and provide high-quality local topological context information for the dynamic weighting of the subsequent graph attention mechanism. S5.3: Based on the set of hidden state of the nodes, a graph attention mechanism is introduced to calculate the attention coefficient score between any two sensor nodes, so as to dynamically quantify the contribution of different propagation paths to the current tension fluctuation characteristics and generate a dynamic edge weight allocation matrix that reflects the significance of interference propagation paths. The set of hidden state of the nodes is obtained as the input data object. The graph attention mechanism is used to initialize the algorithm (parameters: node feature dimension, initial weight of adjacency matrix, activation function type). Each sensor node is assigned a query vector, key vector and numerical vector to be trained, so as to realize the learnable representation of the attention coefficient between nodes. Furthermore, by using a multi-head attention calculation method based on the adjacency matrix (parameters: number of attention heads ≥ 4, normalization method is Softmax), the inner product operation is performed on the query vector of all nodes and the key vector of the adjacent nodes to obtain the unnormalized association strength score matrix, and normalization processing is performed in each attention head to capture the multi-scale feature contribution of tension waves under different propagation paths. Furthermore, the additive attention coefficient calculation formula is used to perform a nonlinear mapping on the above association strength scores, as follows: in, For sensor nodes To the node Attention coefficient It is a linear transformation matrix. For node feature vectors, Let be the high-dimensional embedding feature vector of node j. Let k be the high-dimensional embedding feature vector of the neighboring node k. For trainable weight vectors, This represents vector concatenation. For nodes The neighborhood group, This is the matrix transpose. Furthermore, by averaging or weighting the coefficient matrices output by each attention head, the fused attention coefficients can fully reflect the saliency characteristics of the tension propagation path between different sensor nodes, and generate a dynamic edge weight allocation matrix, where each element of the matrix corresponds to the real-time importance weight of a sensor node connection edge. Through the graph attention mechanism described above, the set of hidden state of nodes in the previous step is transformed into a dynamic edge weight allocation matrix containing spatiotemporal saliency weights, which realizes the accurate quantification of the contribution of different propagation paths and provides a weighted topological input for subsequent interference path matching. For example, a multi-color label printing machine deploys nine fiber Bragg grating tension sensors in the unwinding section, the four color group inlet and outlet sections, and the rewinding section. The node feature dimension is set to 64-dimensional. The initial weights of the adjacency matrix are configured as random values ​​in the range of 0.2-0.8 based on the joint distribution entropy result of step S4. The number of attention heads is set to 6, and each head uses the LeakyReLU activation function with a negative slope of 0.2. When executing the attention coefficient calculation formula, the linear transformation matrix... It is a 64×64 dimensional trainable matrix, with weight vectors. The length is 128. Assume the feature concatenation vector between node 3 and node 5 is... After transformation and The inner product score is 1.426. After LeakyReLU mapping, the exponent value is 4.165. The sum of the normalized denominators of the adjacent nodes is 14.732. =0.2828. After fusing the attention coefficients of each head, the dynamic edge weight of node 3→5 is 0.2941. This weight is used for the weighted node embedding calculation in step S5.4, which ultimately makes the confidence of the path "unrolled segment → third color group" in the interference matching result significantly improved, ensuring that the path saliency and interference recognition accuracy are improved simultaneously. S5.4: The dynamic edge weight allocation matrix is ​​used to perform weighted fusion processing on the hidden state set of nodes, and the graph-level features are compressed into fixed-dimensional graph embedding vectors through global pooling operation. The graph embedding vectors are then compared with various interference-specific propagation topology prototypes pre-stored in the interference propagation topology graph database to generate a candidate interference path similarity score sequence. S5.5: Perform normalization mapping and sorting processing on the candidate interference path similarity score sequence, convert the similarity score into a probability form of confidence score, and associate it with the corresponding interference type identifier and spatial action path node sequence, outputting an interference path matching result vector containing the interference type with the highest matching degree and its propagation path details.

[0014] Step S6: Determine whether the confidence level in the interference path matching result meets the preset threshold condition. If the highest confidence level is greater than the first threshold and the second highest confidence level is less than the second threshold, then generate a compensation trigger command. Specifically, this includes: S6.1: Obtain the interference path matching result vector output by the lightweight graph convolutional neural network model, extract the confidence score data corresponding to each candidate interference path in the vector, and sort the confidence score data in descending order to generate an ordered confidence sequence containing the highest confidence score and the second highest confidence score. The interference path matching result vector of the lightweight graph convolutional neural network model is obtained as the input object. The vector parsing method (parameters: matching result vector structure definition, field positioning rules) is used to identify and parse the fields of the matching result data structure one by one, including the candidate interference path identifier field and the corresponding confidence score field. Furthermore, by using a data filtering method (parameters: full set of confidence score fields, limit on the number of candidate paths), only the confidence score data corresponding to the candidate interference paths is retained, and a list of confidence scores is obtained as the original data object for subsequent sorting processing. Furthermore, by using a numerical sorting method (parameters: descending sorting mode, stability sorting rule), the confidence score list is arranged from high to low, and the confidence sequence data results are generated in descending order. Furthermore, by using the sequence index extraction method (parameters: descending sequence data structure, index position setting), the first and second values ​​in the descending confidence sequence are obtained and recorded as the highest confidence score and the second highest confidence score, respectively. By using a sequence construction process, the highest confidence score and the second highest confidence score from the previous step are combined into an ordered confidence sequence containing two elements, thus providing direct input data for the threshold comparison operations required in subsequent S6.2 and S6.3. For example, for a specific production batch, after the tension propagation time-series fingerprint of a multi-color label printing machine is processed in step S5, the lightweight graph convolutional neural network model outputs a data vector containing five candidate interference paths and their confidence scores. When parsing this data vector, the candidate path identifiers and corresponding scores are precisely extracted according to field location rules; for example, the score list is [0.902, 0.576, 0.432, 0.385, 0.241]. This score list is sorted in descending order to obtain a descending confidence sequence [0.902, 0.576, 0.432, 0.385, 0.241] (in this example, the original list is already arranged in descending order). Using an index extraction method, the highest confidence score of 0.902 and the second highest confidence score of 0.576 are extracted from this sequence and combined to form an ordered confidence sequence (0.902, 0.576). In subsequent threshold comparisons, this ordered confidence sequence is directly used to verify the reliability of the current dominant interference type identification and the uniqueness of the path, effectively improving the stability and accuracy of the compensation triggering logic. S6.2: Based on the highest confidence score in the ordered confidence sequence, compare it with a preset first confidence threshold. If the highest confidence score is greater than the first confidence threshold, it is determined that the current dominant interference type identification has high reliability, and a preliminary interference confirmation flag is generated. S6.3: Based on the second highest confidence score in the ordered confidence sequence, compare it with the preset second confidence threshold. If the second highest confidence score is less than the second confidence threshold, it is determined that there is no multi-solution fuzzy conflict in the current interference propagation path, and a path uniqueness verification flag is generated. S6.4: Perform a logical AND operation on the preliminary interference confirmation flag and the path uniqueness verification flag. If both flags are true at the same time, it is determined that the preset threshold condition is met, the compensation trigger logic is activated, and a compensation trigger command signal with a timestamp is generated. Perform a Boolean AND operation on the preliminary interference confirmation flag and the path uniqueness verification flag (parameter: all flag values ​​are Boolean type) to simultaneously verify the reliability of the interference type and the satisfaction of the path uniqueness condition; Furthermore, through a logic operation unit (parameter: the operation instruction is AND, and the input terminals are respectively connected to the preliminary interference confirmation flag and the path uniqueness verification flag), the condition combination judgment is realized, and the combination judgment result value is obtained; Furthermore, by utilizing the condition determination module (parameter: the valid state of the combined determination result value is true), the trigger determination when the preset threshold condition is met is realized, and trigger flag data is generated; Furthermore, by using the compensation trigger logic activation module (parameter: trigger flag data is true), the compensation command generation function is realized, and a draft compensation command containing the logic trigger state is obtained; Furthermore, by utilizing a high-precision timestamp generation unit (parameters: time synchronization accuracy at the nanosecond level, global clock signal sourced from the IEEE 1588v2 master clock), an absolute timestamp is added to the initial draft of the compensation command, and a final compensation trigger command signal with a timestamp is generated. By combining logical AND operation with timestamp generation processing, the flag determination result of the previous step is transformed into a compensation trigger instruction bound to time precision, thereby achieving the expected technical effect of instantly and accurately starting the compensation execution channel under the condition that the recognition result is reliable and the path is unique. For example, during the operation of a multi-color label printing machine, the highest confidence score of the interference path matching result output by the lightweight graph convolutional neural network model is 0.91, and the second highest confidence score is 0.55. The system sets the first threshold to 0.85 and the second threshold to 0.6. When performing numerical comparison operations, the preliminary interference confirmation flag and the path uniqueness verification flag are both determined to be true. The logic AND operation unit performs an AND operation on the two flag inputs, obtaining a combined judgment result of true. The condition judgment module detects that the condition is met and triggers the flag data to be true. The compensation trigger logic activation module generates a draft compensation command and adds an absolute timestamp to the command through the nanosecond-level timestamp generation unit, for example... Finally, a compensation trigger command signal with the timestamp is generated. This signal is immediately transmitted to the timing compensation channel definition unit in step S7, driving the corresponding color group servo driver to perform feedforward compensation actions for precise delay, amplitude scaling, and phase calibration, ensuring the immediacy and accuracy of compensation execution when the interference path is clearly determined and with high confidence.

[0015] Step S7: Based on the compensation trigger command and the interference path matching result, a timing compensation channel is defined, wherein the timing compensation channel is configured with precise delay, amplitude scaling, and phase correction parameters, and a feedforward compensation signal is injected into the servo driver through the timing compensation channel to generate a compensated tension control signal. Specifically, this includes: S7.1: Obtain the dominant interference type identifier and spatial action path node sequence from the interference path matching result, retrieve the pre-stored interference-specific propagation topology based on the spatial action path node sequence, extract the mean of the measured amplitude attenuation ratio distribution and the mean of the relative phase difference distribution of the corresponding path edge, and define the mean of the measured amplitude attenuation ratio distribution as the reference amplitude scaling factor and the mean of the relative phase difference distribution as the reference phase correction angle. S7.2: Read the propagation delay timestamp marked in the interference-specific propagation topology diagram, discretize and round the propagation delay timestamp in combination with the current system sampling clock cycle to generate an integer number of delay steps, and package the integer number of delay steps with the reference amplitude scaling factor and the reference phase correction angle to construct a timing compensation channel configuration vector containing precise delay parameters, amplitude scaling parameters and phase correction parameters; Obtain the propagation delay timestamp marked on the interference-specific propagation topology map, and use a clock cycle conversion algorithm (parameters: sampling frequency, sampling period) to map the delay time from the physical time domain to the sampling point number domain; Furthermore, by using a discretization rounding method (parameters: rounding mode, maximum tolerance ±1 sampling point), the number of sampling points is converted into an integer number of delay steps, and the discretized delay step size for hard real-time delay control is obtained. Furthermore, by using a parameter mapping method (parameters: mean distribution of measured amplitude attenuation ratio at the path edge, mean distribution of relative phase difference at the path edge), the reference amplitude scaling factor and the reference phase correction angle are extracted, and the corresponding control coefficient dataset is generated. Furthermore, by encapsulating the processing method (parameters: delay steps, amplitude scaling factor, phase correction angle), a structured combination of the three types of control parameters is achieved, and a configuration vector conforming to the timing compensation channel calling specification is generated; A symbolic tagging method (parameters: timestamp label, interference type identifier) ​​is used to annotate the metadata of the configuration vector and obtain a timing compensation channel configuration object with unique identification capability; The delay timestamp of the previous step is converted using a clock cycle conversion algorithm. Converted to the number of sampling points, where Indicates the propagation delay time. Indicates the length of a single sampling period; By using a discretization and rounding algorithm, the result of the previous step is... Mapped to integer time delay steps and ensure It meets the requirements for hard real-time buffer synchronization; By encapsulating the process, the results of the previous step are transformed into a data vector containing precise delay parameters, amplitude scaling parameters, and phase correction parameters, thereby enabling the executable configuration of the timing compensation channel. For example, in the actual operation of a multi-color label printing machine, the propagation delay timestamp of the interference-specific propagation topology icon annotation is... milliseconds, sampling frequency set to Hertz, i.e., sampling period Seconds. The number of sampling points is calculated using a clock cycle conversion algorithm: ≈ Points. The discretization and rounding algorithm is used to... Point mapping to integer time delay steps The mean of the measured amplitude attenuation ratio distribution along the path edge is... The mean of the relative phase difference distribution is Degree. During the packaging process, the delay steps will be... Amplitude scaling factor Phase correction angle The configuration vector [846, 0.71, -17.2] is combined and a timestamp tag and interference type identifier are attached to form a structured timing compensation channel configuration object. This object directly drives the delay buffer, amplitude normalization and phase rotation correction in subsequent execution, ensuring that the delay, intensity and phase of the compensation signal are accurately out of phase with the interference fluctuations, thereby significantly improving the robustness and stability of tension control; S7.3: Extract the feedforward compensation original signal sequence at the current moment, perform a shift register buffer operation on the feedforward compensation original signal sequence using the integer delay steps in the timing compensation channel configuration vector to generate a delay-aligned signal sequence, and perform point-by-point multiplication operation on the delay-aligned signal sequence based on the reference amplitude scaling factor to generate an amplitude-normalized signal sequence. S7.4: Perform zero-phase Butterworth filtering on the amplitude normalized signal sequence to eliminate high-frequency noise distortion, generate a smooth transition signal sequence, and perform phase rotation correction operation on the smooth transition signal sequence in the frequency domain according to the reference phase correction angle to generate a phase-corrected feedforward compensation modulation signal, ensuring that the phase-corrected feedforward compensation modulation signal and the target interference fluctuation maintain anti-phase superposition characteristics in the time domain. S7.5: The phase-corrected feedforward compensation modulation signal is injected into the torque command input terminal of the target color group servo driver through the independent FPGA hard real-time scheduling interface. It is linearly superimposed with the original closed-loop feedback control signal of the servo driver to generate a compensated tension control signal containing dynamic interference cancellation components, and drives the printing press traction roller motor to perform tension fine adjustment.

[0016] Step S8: Monitor the stability of the compensated tension control signal. If the residual error exceeds the allowable range, generate model optimization parameters based on the current tension propagation time sequence fingerprint and interference path matching results to update the interference propagation topology database. Specifically, this includes: S8.1: Obtain the real-time time-series data stream of the compensated tension control signal, perform sliding window root mean square error calculation on the real-time time-series data stream to generate a residual error statistic that characterizes the current control accuracy, and compare the residual error statistic with a preset allowable range threshold. If it is determined that the residual error statistic exceeds the preset allowable range threshold, the online model update process is triggered. S8.2: Based on the triggered online model update process, extract the tension propagation time-series fingerprint feature vector and the corresponding interference path matching result label at the current moment, and use the bias inversion algorithm to jointly analyze and process the tension propagation time-series fingerprint feature vector and the interference path matching result label to generate a model optimization parameter set containing node weight correction coefficients and edge connection probability adjustment. S8.3: Obtain historical graph structure data from the interference propagation topology graph database. Based on the node weight correction coefficients and edge connection probability adjustment amounts in the model optimization parameter set, perform dynamic reconstruction and update processing on the joint distribution entropy of edge weights and node topology connection relationships in the historical graph structure data to generate an updated version of the interference propagation topology graph database with the ability to adapt to the latest operating conditions.

[0017] 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.

[0018] 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 roll-to-roll paper force coordination intelligent control method for multi-color label printing presses, characterized in that, Includes the following steps: S1: Deploy multiple tension sensors in key sections of the multi-color label printing machine, and synchronize the timing of each sensor to collect and generate synchronized multi-point tension timing signals; S2: Perform joint time-frequency analysis on the synchronized multi-point tension timing signal, extract the instantaneous phase and amplitude of the main frequency component of each channel, and calculate the relative phase difference and amplitude attenuation ratio between channels to generate instantaneous phase component and amplitude attenuation ratio. S3: Construct a tension propagation time sequence fingerprint based on the instantaneous phase component and amplitude attenuation ratio; S4: Construct an interference propagation topology database based on historical interference labels and past tension propagation time sequence fingerprints; S5: Input the tension propagation time sequence fingerprint into the graph neural network model, and use the interference propagation topology graph database for online matching to generate interference path matching results; S6: Determine whether the confidence level in the interference path matching result meets the preset threshold condition. If the highest confidence level is greater than the first threshold and the second highest confidence level is less than the second threshold, then generate a compensation trigger command. S7: Based on the compensation trigger command and the interference path matching result, define a timing compensation channel, and inject a feedforward compensation signal into the servo driver through the timing compensation channel to generate a post-compensation tension control signal.

2. The roll-to-roll paper force coordination intelligent control method for a multi-color label printing machine according to claim 1, characterized in that, Following step S7, the following is also included: S8: Monitor the stability of the compensated tension control signal. If the residual error exceeds the allowable range, generate model optimization parameters based on the current tension propagation time sequence fingerprint and interference path matching results to update the interference propagation topology database.

3. The roll-to-roll paper force coordination intelligent control method for a multi-color label printing machine according to claim 1, characterized in that, The key sections of the multicolor label printing press include the unwinding section, the traction roller inlets for each color group, and the rewinding section.

4. The roll-to-roll paper force coordination intelligent control method for a multi-color label printing machine according to claim 1, characterized in that, Step S3 specifically includes: The instantaneous phase component and amplitude attenuation ratio are processed by multi-channel pairwise combination to calculate the relative phase difference sequence and amplitude attenuation ratio sequence between any two sensor channels, thereby generating the inter-channel relative phase difference sequence and inter-channel amplitude attenuation ratio sequence. Based on the inter-channel relative phase difference sequence and the inter-channel amplitude attenuation ratio sequence, a sliding window feature extraction is performed to calculate its first-order difference feature vector and second-order difference feature vector; The inter-channel relative phase difference sequence, the inter-channel amplitude attenuation ratio sequence, the first-order difference feature vector, and the second-order difference feature vector are spliced ​​and fused in multiple dimensions to generate an initial high-dimensional tension propagation feature vector. Based on the initial high-dimensional tension propagation feature vector, normalization mapping and path-dependent encoding are performed to generate a standardized tension propagation feature vector; The standardized tension propagation feature vector is structured and encapsulated, and timestamp tags and interference scene context metadata are added to generate a tension propagation temporal fingerprint.

5. The roll-to-roll paper force coordination intelligent control method for a multi-color label printing machine according to claim 4, characterized in that, The tension propagation temporal fingerprint is defined as a temporal feature vector containing microsecond-level phase shift, decay gradient, and path dependence features.

6. The roll-to-roll paper force coordination intelligent control method for a multi-color label printing machine according to claim 1, characterized in that, Step S4 specifically includes: The historical interference label dataset and the past tension propagation time series fingerprint dataset in the storage medium are correlated and retrieved to extract sample pairs with clear causal labels and generate the original feature sequence set. Based on the original feature sequence set, a graph structure is reconstructed, mapping the sensor position as a node and the statistical distribution of the relative phase difference and amplitude attenuation ratio as an edge, generating an initial interference propagation topology graph instance. For each connecting edge in the initial interference propagation topology graph instance, calculate the joint distribution entropy value and generate the edge weight joint distribution entropy parameter; The initial interference propagation topology graph instance is weighted and updated using the edge weight joint distribution entropy parameter to generate a standardized interference propagation topology graph. An index mapping mechanism is established based on the standardized interference propagation topology map to bidirectionally bind and store interference type identifiers with propagation path topology structures, thereby generating an interference propagation topology map database.

7. The intelligent control method for roll-to-roll paper force coordination for a multi-color label printing machine according to claim 1, characterized in that, Step S5 specifically includes: Based on the graph structure mapping rules, the spatial path dependency attribute in the standardized tension propagation feature vector is converted into the initial feature matrix of the graph node, and the joint distribution entropy of the edge weights in the interference propagation topology graph database is mapped into the graph adjacency matrix to generate the graph data instance to be matched. The graph convolutional neural network forward propagation is performed on the graph data instance to be matched. The feature information of adjacent sensor nodes is aggregated by multi-layer graph convolution operators, high-dimensional node embedding vectors are extracted, and a set of node hidden state is generated. Based on the node hidden state set, a graph attention mechanism is introduced to calculate the attention coefficient score between any two sensor nodes, dynamically quantify the contribution of different propagation paths to the current tension fluctuation characteristics, and generate a dynamic edge weight allocation matrix. The dynamic edge weight allocation matrix is ​​used to perform weighted fusion processing on the set of hidden state of the node, and the graph-level features are compressed into graph embedding vectors of fixed dimension. The graph embedding vectors are then compared with various interference-specific propagation topology prototypes pre-stored in the interference propagation topology graph database to generate a candidate interference path similarity score sequence. Normalization mapping and sorting processing are performed on the candidate interference path similarity score sequence to convert the similarity score into a probability form of confidence score, and the corresponding interference type identifier and spatial action path node sequence are associated to output the interference path matching result vector.

8. The roll-to-roll paper force coordination intelligent control method for a multi-color label printing machine according to claim 7, characterized in that, The node hidden state set contains local topological context information.

9. The intelligent control method for roll-to-roll paper force coordination for a multi-color label printing machine according to claim 7, characterized in that, The interference path matching result vector contains the interference type with the highest matching degree and its propagation path details.

10. The roll-to-roll paper force coordination intelligent control method for a multi-color label printing machine according to claim 1, characterized in that, The timing compensation channel is configured with precise delay, amplitude scaling and phase correction parameters.

Images