Paper machine roll set dynamic balancing analysis method based on edge computing

Abstract

The application belongs to the technical field of edge computing, and particularly relates to a paper machine roll set dynamic balance analysis method based on edge computing. The method receives roll body rotation speed sequences, vibration response sequences and running state identifiers at roll set side edge computing nodes, completes the same analysis window alignment according to roll body numbers, collection positions and time stamps, converts the vibration response into an angular domain sequence with roll body rotation angle as a reference, and establishes a roll set coupling graph according to a transmission relationship, a paper web contact sequence and a rack transmission relationship. The method extracts a first-order amplitude, phase stability and cross-roll coherence characteristics, combines graph edge propagation delay, attenuation coefficients and phase offset ranges, and decomposes the same frequency response into a roll self-excitation component, an adjacent roll transmission component and a common disturbance component. The method can retain transient phase and cross-roll transmission relationship on the edge side, reduce the situation that a passive response roll is misjudged as a source roll, and improve the traceability and reviewability of roll set dynamic balance analysis results.

Description

Technical Field

[0001] This invention belongs to the field of edge computing technology, specifically relating to a dynamic balance analysis method for paper machine rolls based on edge computing. Background Technology

[0002] In continuous forming, pressing, drying, sizing, and rewinding processes, paper machine rolls bear the functions of supporting, traction, and transmission of the paper web. Uneven mass distribution of the rolls, uneven wear of the covering layer, or changes in installation status can all generate first-order vibration responses within the rotation cycle. Existing dynamic balancing analysis typically relies on existing speed and vibration response data from the production line, sending sampling sequences from bearing housings, frames, or roll ends to on-site monitoring terminals or host computers for time and frequency domain processing. The most readily available conventional approach involves capturing a fixed-length time window when the rolls are at a relatively stable speed, performing detrending, filtering, envelope, or spectral transformation on the vibration response sequence, extracting the amplitude, root mean square value, peak-to-peak value, first-order frequency peak value, and phase near the rotational frequency, and comparing the extracted results with historical thresholds, empirical thresholds, or benchmark values ​​for similar rolls. If the energy at a certain measuring point increases near the rotational frequency, the corresponding roll is usually listed as a suspected dynamic imbalance object. Further analysis, combined with trend curves, maintenance records, and manual inspection results, determines whether speed reduction observation, machine shutdown verification, or counterweight correction is necessary. The above processing flow uses single-point anomaly as the main basis for judgment. The calculation chain is relatively short and easy to deploy on site. However, it lacks a quantitative description of the source of the same frequency response in the continuous operation of multiple rolls. The judgment result is easily affected by the response transmitted by adjacent rolls and frames. Under the conditions of dense roll arrangement and continuous paper web traction, the peak value of the rotational frequency of a single measuring point cannot directly correspond to the mass eccentricity position of a single roll.

[0003] To reduce the burden of transmitting raw vibration data and shorten the on-site analysis link, some online monitoring solutions for paper machines have moved data preprocessing to the edge. Edge computing nodes typically receive the rotational speed sequence, vibration response sequence, and operating status identifiers of each roll. They perform timestamp correction, anomaly sampling removal, buffer queue management, data compression, and feature extraction locally, before uploading statistical features, alarm status, or trend segments to the upper-level platform. Due to fluctuations in speed increase / decrease, gradual load changes, and paper web traction during paper machine operation, simply performing spectral analysis at fixed time intervals can lead to frequency diffusion, resulting in unstable first-order peak amplitude and phase. To address this issue, conventional solutions construct corner trajectories based on the rotational speed sequence, perform order tracking or corner domain resampling on the vibration response sequence, convert the fixed time interval sequence into a fixed corner interval sequence, and then extract the first-order amplitude and phase. Corner domain processing allows the vibration response of a single roll to correspond to its own rotational period, and the edge side can also generate alarm results with shorter cycles locally. However, most solutions still focus on calculations for single rolls, single measuring points, or a small number of measuring points, and the results mainly indicate whether the measuring point itself is abnormal. Even if multiple rollers complete angular domain resampling, without a common phase reference and cross-roller propagation constraints, the first-order components of different rollers are still difficult to use for source determination.

[0004] In a multi-roll continuously arranged paper machine, when one roll becomes dynamically unbalanced, the vibration response does not remain only at the source roll's measuring point. The paper web continuously wraps around or contacts multiple rolls, creating synchronous or near-synchronous motion relationships on the drive side. The frame and foundation supports also transmit periodic forces, causing the source roll's first-order response to exhibit the same or nearly the same frequency on adjacent rolls. Different propagation paths have different delays, attenuations, and phase shifts; the source roll's amplitude is not necessarily higher than the passive response roll's amplitude. Existing multi-measuring-point diagnostic methods typically use synchronous increases in amplitude, similar spectral peaks, small phase differences, or high coherence as correlation criteria for judgment, followed by manual inference of potential sources based on roll arrangement and historical fault experience. Other solutions upload multi-measuring-point data to the cloud or a centralized server for model analysis; however, the upload process often involves sampling compression, fragment truncation, or feature processing. The data received by the remote model often leans towards long-term trends and statistical indicators, making it difficult to fully preserve transient phase, propagation delay, and cross-roll correspondence within the same time window. When the source roll response propagates simultaneously through the paper web contact path and the frame support path, the passive response measurement point will be superimposed with multiple co-frequency components from the same source. Conventional coherence judgment cannot distinguish the direction of the superimposed components.

[0005] When multiple paths of coupling and slow speed fluctuations coexist in paper machine roll sets, existing dynamic balancing analyses struggle to distinguish the first-order vibration responses of the same frequency appearing on multiple rolls from self-excited responses caused by the dynamic imbalance of the source roll itself, transmitted responses propagating along the paper web or frame from adjacent rolls, and common responses formed by joint disturbances in the operating conditions. Spectral peak determination only indicates an energy increase near the rotational frequency at a certain measuring point, but cannot determine whether the energy originates from the corresponding roll; angular domain resampling only stabilizes the order characteristics of a single roll, but cannot explain the propagation path of that order component in the roll set; coherence analysis only indicates a correlation between multiple measuring points, but cannot determine the direction and component attribution of the correlation. Due to the lack of a cross-roll comparison mechanism based on the same angular domain phase, and the lack of a propagation model incorporating paper web contact, drive synchronization, and frame transmission into a unified constraint, conventional solutions may treat the passive response roll as the source roll of dynamic imbalance when the passive response roll experiences frame transmission or paper web coupling, causing the source roll positioning and phase interval judgment to deviate from the true imbalance position. This deviation will cause the roller number, phase direction, and abnormal segment used for subsequent verification to be inconsistent with the true source, thus making the dynamic balance analysis results lack verifiable propagation path evidence. Summary of the Invention

[0006] The purpose of this invention is to provide a dynamic balance analysis method for paper machine rolls based on edge computing, which can effectively solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The edge computing-based dynamic balancing analysis method for paper machine rolls includes: receiving the speed sequence, vibration response sequence, and operating status identifier associated with each roll at edge computing nodes arranged on the side of the paper machine rolls; The rotational speed sequence, the vibration response sequence, and the operating status identifier are aligned to the same analysis window based on the roller number, acquisition location, and timestamp. The vibration response sequence is converted into an angular domain sequence based on the corresponding roller rotation angle; Establish a roller group coupling diagram based on the roller body transmission relationship, paper web contact sequence, and frame transmission relationship; The first-order amplitude, phase stability, and cross-roll coherence features are extracted from the angular domain sequence, and the same-frequency response is decomposed into the self-excited component of this roll, the component transmitted from the neighboring roll, and the common disturbance component in combination with the roll group coupling diagram. Based on the decomposition results, the dynamic imbalance source roller and its phase range are generated.

[0008] Preferably, the alignment of the same analysis window includes: determining candidate segments in stable operation based on the running status identifier, and generating continuous turning trajectories from the rotation speed sequence within the candidate segments; The vibration response sequence is resampled based on the continuous angular trajectory to obtain the angular domain sequence with equal angular intervals; In the presence of multiple roller speed ratios, the angular trajectory of a preset reference roller is used as a common phase reference. The angular domain sampling points of non-reference rollers are mapped to the phase coordinates under the common phase reference, so that the first-order components of different rollers have corresponding phase indices within the same analysis window.

[0009] Preferably, the establishment of the roller coupling diagram includes: setting the rollers involved in the dynamic balance analysis and their vibration acquisition positions as graph nodes, and setting the vibration propagation association formed by the paper web contact sequence, transmission synchronization relationship, frame transmission path and adjacent arrangement relationship as graph edges; Configure a propagation delay, attenuation coefficient, and phase offset range for each of the graph edges; The propagation delay, the attenuation coefficient, and the phase offset range are calculated from the cross-roll coherence characteristics, order phase difference, and amplitude attenuation sequence in historical stable operation segments, and stored according to the propagation association category corresponding to the graph edge.

[0010] Preferably, the decomposition of the same frequency response includes: taking the propagation delay matching degree and phase offset matching degree between adjacent graph nodes in the roller coupling diagram as constraints, first extracting the common disturbance component from the first-order components of multiple graph nodes; Then, estimate the adjacent roller transfer component based on the adjacent graph edge transfer item received by each graph node; The remaining vector after subtracting the common disturbance component and the adjacent roll transmission component from the first-order component is determined as the self-excited component of this roll, and the phase direction and amplitude continuity of the self-excited component of this roll within the continuous analysis window are recorded.

[0011] Preferably, the generation of the continuous turning trajectory includes: performing clock drift correction on the rotation speed sequence, and writing the rotation speed sequence and the vibration response sequence into the same edge buffer queue according to the acquisition timestamp; When there are missing or jittering points between adjacent sampling points, a segmented rotation angle interpolation curve is constructed based on the local variation trend of the rotation speed sequence; The intersection of the segmented corner interpolation curve and the stable operating interval corresponding to the operating status identifier is clipped to obtain the continuous corner trajectory covered by the same time period as the vibration response sequence, and the original time index corresponding to each corner sampling point is retained.

[0012] Preferably, the phase coordinate mapping includes: establishing a phase conversion table based on the speed ratio and transmission phase relationship between the preset reference roller and the non-reference roller; The angular domain sampling points of the non-reference roller are first mapped to their own first-order periodic coordinates, and then converted to the common phase reference according to the phase conversion table. When there is a gradual change in speed within the same analysis window, the phase conversion table is updated in segments according to the continuous turning trajectory, and the order vectors in the overlapping area of ​​the phase coordinates before and after the update are checked for consistency, and phase jump sampling points are removed.

[0013] Preferably, the calculation of the propagation delay, the attenuation coefficient, and the phase offset range includes: extracting first-order vector pairs of adjacent graph nodes from historical stable operation segments; The first-order vector pairs are subjected to coherence screening, and vector pairs that meet the coherence requirements and have consistent operating status identifiers are retained. The phase difference distribution, magnitude ratio distribution, and time index difference distribution are calculated based on the retained vector pairs; Parameter statistical intervals are generated according to the propagation association categories corresponding to the graph edges, and the common disturbance components shared by different propagation association categories are separated from the parameter statistical intervals.

[0014] Preferably, the extraction of the common disturbance component includes: selecting a set of graph nodes in the roller coupling diagram that have a common operating state identifier and meet the requirements of cross-roller coherence characteristics; Phase normalization is performed on the first-order vector of the graph node set to generate candidate common vectors; The candidate common vector is substituted back to each graph node along different propagation paths of the roller coupling graph, and the propagation components that only exist between local adjacent graph nodes are excluded based on the substitution residual, so as to obtain the common disturbance component that is not bound to the phase direction of a single roller.

[0015] Preferably, the generation of the parameter statistical interval further includes: establishing sample clusters for the phase difference distribution, the amplitude ratio distribution, and the time index difference distribution according to paper web contact propagation, transmission synchronous propagation, and frame transfer propagation, respectively; When the same graph edge falls into multiple sample clusters in different stable running segments, a conditional index is established for the multiple sample clusters based on the running state identifier and order coherence feature. During online analysis, the propagation delay, attenuation coefficient, and phase offset range under the corresponding condition index are invoked based on the current analysis window's running status indicator to participate in the same-frequency response decomposition.

[0016] Preferably, the determination of the self-excited component of the local roller further includes: inputting the remaining vector after deducting the common disturbance component and the neighboring roller transmitted component into the source roller determination sequence; The source roller determination sequence records the phase direction, amplitude continuity, and residual relationship between the remaining vector and the edge propagation constraint of the adjacent graph according to the continuous analysis window; When the residual vector of the same graph node maintains the same phase direction in multiple consecutive analysis windows, and the residual after reverse interpretation along the adjacent graph edge is lower than the residual after forward interpretation, the roller body corresponding to the same graph node is written into the dynamic unbalanced source roller set and a phase interval is generated.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention receives the rotational speed sequence, vibration response sequence, and operating status identifier of each roller within the edge computing node. It aligns these sequences within the same analysis window according to roller number, acquisition location, and timestamp. The vibration response sequence is then converted into an angular domain sequence based on the corresponding roller rotation angle, allowing the first-order components of different rollers to participate in comparison under a common phase reference. The roller coupling diagram transforms the paper web contact sequence, transmission synchronization relationship, frame transmission path, and adjacent arrangement relationship into propagation constraints between graph nodes and graph edges, and configures propagation delay, attenuation coefficient, and phase offset range for the graph edges. The angular domain sequence provides a phase-consistent order vector, and the roller coupling diagram provides the constraint boundary of the propagation path. When combined, the same-frequency response is no longer directly equated to the fault source response of the corresponding roller, but is decomposed according to the common disturbance component, the neighboring roller transmission component, and the self-excited component of the current roller. For peak values ​​propagating along adjacent graph edges, edge computing nodes can classify them into neighboring roll propagation components based on propagation delay matching and phase offset matching. For responses occurring together by multiple graph nodes and not bound to the phase direction of a single roll, they can be classified into common disturbance components through candidate common vector back substitution and residual exclusion. The remaining vectors after deducting the above components are then used to generate the dynamic imbalance source roll and phase interval. The continuous analysis window further records the phase direction, amplitude continuity, and residual relationship between the remaining vectors and the propagation constraints of adjacent graph edges, ensuring that a single amplitude mutation does not directly lead to source roll judgment. Since source roll judgment is based on angular domain phase consistency, propagation path constraints, and residual component attribution, even if a passive response roll exhibits a higher-order amplitude, it must undergo verification through propagation delay, phase offset, and reverse interpretation residuals before entering the source roll set. The deviation between the source roll location and the actual imbalance position is compressed to the range that the propagation model can explain. The phase interval is generated by the remaining vector sequence, and the source relationship corresponds to the propagation path.

[0018] 2. This invention also retains the original time index corresponding to the corner sampling points on the edge side, and forms a continuous corner trajectory through clock drift correction, segmented corner interpolation, and stable operating interval trimming. This limits the impact of sampling loss, time jitter, and gradual speed changes on the order vector to within a traceable analysis window. The corner sampling points of non-reference rolls are mapped from their own first-order periodic coordinates to a common phase reference. During gradual speed changes, the phase conversion table is updated segmentally according to the continuous corner trajectory. Phase jump sampling points are eliminated through consistency verification of overlapping areas, preventing phase mapping errors from entering subsequent frequency decomposition. Historical stable operating segments are used to calculate the phase difference distribution, amplitude ratio distribution, and time index difference distribution of the graph edges. Paper web contact propagation, drive synchronous propagation, and frame transmission propagation each form parameter statistical intervals with conditional indices. During online analysis, matching propagation constraints can be called according to the current operating status. Conditional indices allow the propagation delay, attenuation coefficient, and phase offset range of the same graph edge in different stable operating segments to switch according to the operating status, avoiding the direct use of historical parameters under mismatched conditions in the current window. The continuous analysis window records the phase direction, amplitude continuity, and residual relationship between the remaining vectors and the propagation constraints of adjacent graph edges, ensuring that the generation of the source roll set does not depend on abrupt amplitude changes in a single window. Edge computing nodes perform data alignment, phase mapping, propagation decomposition, and residual determination locally. The original transient phase and cross-roll correspondence are preserved before centralized processing at the remote end. The analysis results can be traced back to specific time indices, graph edge constraints, and remaining vector sequences. Subsequent verification can directly correspond to stable operating segments and source roll phase intervals, and the data received by the remote platform also has a clear component source. Simultaneously, the preserved original time index allows for segment-by-segment verification of abnormal segments, phase mapping results, and graph edge parameter calls. This forms a continuous data chain from the sampling window to source roll determination. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the paper machine roll dynamic balance analysis method based on edge computing according to the present invention; Figure 2 This is a flowchart of the edge-side data alignment and corner domain synchronization processing of the present invention; Figure 3 This is a flowchart of the roller coupling diagram establishment and propagation parameter generation process of the present invention; Figure 4 This is a flowchart of the same-frequency response decomposition and dynamic imbalance source roller determination of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please refer to Figure 1 This embodiment provides a dynamic balancing analysis method for paper machine rolls based on edge computing. The method is executed by edge computing nodes deployed on the side of the paper machine roll assembly. Each edge computing node receives the rotational speed sequence, vibration response sequence, and operating status identifier associated with each roll. Data from different sources is written to the same analysis cache according to roll number, acquisition location, and timestamp. The operating status identifier distinguishes between stable operation, speed increase, speed decrease, paper type change transition, gradual load change, and abnormal sampling to be rejected. The vibration response sequence is time-domain response data acquired along existing measurement points on the roll assembly. The rotational speed sequence is used to establish the rotational trajectory of the corresponding roll. Within an analysis window, the edge computing node first performs time alignment and segment clipping, and then converts the vibration response sequence into a time-domain response. An angular domain sequence is established based on the corresponding roll rotation angle, and a roll group coupling diagram is established according to the roll transmission relationship, paper web contact sequence, and frame transmission relationship. Then, the first-order amplitude, phase stability, and cross-roll coherence characteristics are extracted from the angular domain sequence. Under the propagation constraints of the roll group coupling diagram, the same-frequency response is decomposed into the self-excited component of this roll, the transmission component of the adjacent roll, and the common disturbance component. The working process is manifested by using angular domain synchronization to eliminate the order diffusion caused by the gradual change of speed, using graph edge constraints to describe the propagation source of the same-frequency vibration in the roll group, and then using the component residuals to generate the dynamic unbalanced source roll and its phase interval. The advantage of this embodiment is that the judgment of the same-frequency peak value is transformed into a source separation process with propagation path constraints, so that there is a traceable correspondence between the source roll number, abnormal phase, and analysis window.

[0022] In this embodiment, reference Figure 2 The data received by the edge computing node is not directly mixed in the form of raw files. Instead, a unified data record is established according to the roller number, collection position and timestamp. Each record contains necessary fields related to subsequent corner domain synchronization and coupling decomposition. The specific fields and processing methods can be organized according to Table 1. Table 1 is used to explain the meaning of the fields, processing methods and output destination of the edge side analysis data before entering the dynamic balance analysis process.

[0023] Table 1. Edge-side analysis data fields and processing methods

[0024] In one embodiment, when writing to the cache queue, the edge computing node establishes a time-incrementing speed cache and vibration cache for each roller. When the window starts, it retrieves data segments with overlapping time ranges from the cache. If there is a sampling start point offset between the speed sequence and vibration response sequence of a certain roller, the overlapping time stamp interval is used as a candidate interval, and data that does not enter the candidate interval is retained as the front cache for subsequent windows. If the running status indicator shows that there are sudden acceleration / deceleration changes or abnormal sampling in the candidate interval that need to be removed, the edge computing node trims the candidate interval into continuous stable sub-intervals and records the trimming boundary. If a sampling point has a timestamp repetition, short-term reverse order, or obvious missing time stamps, it is marked as a data point that cannot be used for phase mapping by checking the monotonicity of adjacent timestamps. The remaining data points maintain a corresponding relationship according to the roller number, sampling position, and original time index. After the above processing, the data of different rollers in the same analysis window have a common time coverage range and a traceable time source. The advantage of this embodiment is that data alignment, status filtering, and traceable index are fixed at the edge side, avoiding the misclassification of data from different running segments as co-frequency related data in the subsequent order decomposition process.

[0025] Continuous turning trajectories can be generated using the following formula: ; in, Indicates the roller body number. This represents the time variable within the analysis window. and These represent the start and end points of the analysis window, respectively. Indicates the roller body In time The corresponding continuous turning trajectory, This represents the angular velocity sequence after edge-side clock correction. Indicates the roller body Clock drift correction amount corresponding to the data channel, This represents the time variable during the integration process. Physically, it means accumulating the rotational speed sequence into an angular reference that can be used for angular domain resampling. For example, when... , , , hour, This result indicates that the normalized angle rotated by the rollers cumulatively within the corresponding time interval is 6.4.

[0026] In one embodiment, the edge computing node associates the continuous corner trajectory with the vibration response sequence. Specifically, it generates target sampling points with equal angular intervals on the continuous corner trajectory, looks up the original time index corresponding to each target sampling point through the corner trajectory, and then reads or interpolates the corner domain response value from the vibration response sequence. If the original time corresponding to the target sampling point falls within the missing interval, a missing marker is used to prevent the point from entering the first-order calculation. If the target sampling point is located between two valid original sampling points, linear interpolation is performed according to the original time distance. If the target sampling point crosses the boundary of the stable operating interval, the target sampling point is not written into the current corner domain sequence. The processing converts the vibration response from a time-arranged sequence to a corner-arranged sequence. The frequency drift caused by the gradual change in speed is converted into an angular interval change. The first-order component in the corner domain sequence can correspond to the complete rotation cycle of the roller. The advantage of this embodiment is that it binds the vibration response of a single roller to its own rotation phase, providing a stable input for cross-roller phase comparison of the same order.

[0027] Corner resampling can be performed using the following formula: ; in, Indicates the roller body In the Vibration response values ​​at each sampling point in the corner region Indicates the roller body In time The time-domain vibration response at that point, Represents a continuous turning trajectory The inverse mapping function, Indicates the starting angle of the current angle domain sequence. This represents the angular interval between adjacent angular domain sampling points. The index is a non-negative integer, and its physical meaning is to obtain the corresponding response from the time-domain vibration sequence at a specified angular position. For example, when , , When the target angle is 2.0, if ,and ,but This result indicates that the response value of the fourth corner sampling point is 0.32.

[0028] In one embodiment, when there are multiple speed ratios in the paper machine roll group, the edge computing node selects the rotational trajectory of a preset reference roll as a common phase reference, and establishes a phase conversion table based on the speed ratio and transmission phase relationship between the reference roll and the non-reference roll. The angular domain sampling points of the non-reference roll are first mapped to their own first-order periodic coordinates, and then converted to the phase coordinates under the common phase reference through the phase conversion table. If a gradual change in speed occurs within the same analysis window, the edge computing node divides the window into several phase continuous sub-segments according to the continuous rotational trajectory, and updates the phase conversion table in segments. The overlapping area before and after the update is used for order vector consistency verification. If the phase difference of the same order vector in the overlapping area exceeds the phase continuity range, the corresponding sampling point is marked as a phase jump sampling point and removed from the first-order calculation. After the above processing, the first-order components of rolls with different speed ratios have the same phase index and comparable angular domain positions. The advantage of this embodiment is that it avoids false coherence relationships between rolls with different speeds due to inconsistent phase coordinates.

[0029] In a preferred embodiment, the generation of continuous corner trajectories further includes edge-side clock drift correction and segmented corner interpolation. Specifically, the edge computing nodes first compare the local interval changes of the rotation speed sequence timestamps with the vibration response sequence timestamps, and establish a drift correction amount according to the reference time in the same buffer queue. The corrected rotational speed sequence and vibration response sequence are then written into the same edge buffer queue according to timestamps. When there is a short-term missing or jitter between adjacent sampling points, the edge computing node does not directly discard the entire data segment. Instead, it constructs a segmented angle interpolation curve based on the local rotational speed change trend before and after the missing point. If the interpolation curve intersects with the stable operating interval corresponding to the operating state identifier, only the angle trajectory within the intersection is retained and the original time index corresponding to each angle domain sampling point is recorded. If the interpolation interval crosses the operating state switching boundary, the interpolation interval is split and enters the corresponding state buffer. After the above processing, the continuous angle trajectory can not only cover the analyzable data in the stable segment, but also retain the boundary information of missing and jitter. The advantage of this embodiment is that it reduces the phase error introduced by clock offset and sampling jitter, so that the subsequent propagation path judgment is based on the same time.

[0030] Further, refer to Figure 3The roller coupling diagram is stored in the edge computing node as graph nodes and graph edges. Graph nodes correspond to the rollers participating in the dynamic balance analysis and their vibration acquisition positions. Graph edges correspond to the vibration propagation associations formed by the paper web contact sequence, transmission synchronization relationship, frame transmission path, and adjacent arrangement relationship. Graph edges do not represent new mechanical structures, but rather the propagation constraints of existing roller group operation relationships in data analysis. Each graph edge is configured with propagation delay, attenuation coefficient, and phase offset range. The propagation delay is used to represent the time index difference corresponding to the first-order response of one graph node reaching another graph node through the propagation path. The attenuation coefficient is used to represent the amplitude ratio before and after propagation. The phase offset range is used to represent the first-order phase difference boundary caused by the propagation path. If the same pair of graph nodes has both paper web contact propagation and frame transmission propagation, the edge computing node establishes different types of graph edge records for the same node pair. During online analysis, matching graph edges are selected to participate in decomposition based on the operating status identifier and order coherence characteristics. The advantage of this embodiment is that it unifies the roller group geometric adjacency, transmission association, and frame transmission into computable propagation constraints, avoiding the inference of the response source based solely on the proximity of the measurement point position.

[0031] In this embodiment, the edge parameters in the roller coupling graph are calculated from historical stable operation segments. The edge computing node reads segments with consistent operating status identifiers and no known anomaly markers from the historical cache, extracts first-order vector pairs of adjacent graph nodes, and calculates cross-roll coherence features, phase difference distribution, amplitude ratio distribution, and time index difference distribution. Only vector pairs with coherence features that meet the requirements and consistent operating status are used for graph edge parameter updates. If the vector pairs exhibit different propagation characteristics in multiple operating states, the edge computing node stores parameter statistical intervals according to the operating status identifier. If some graph nodes show in-phase changes in multiple propagation categories, the edge computing node marks the change as a candidate common disturbance component and separates it from the single graph edge parameter statistics. The graph edge parameters formed by the above processing can be called according to changes in operating status without mixing all historical segments into a single propagation parameter. The advantage of this embodiment is that the propagation delay, attenuation, and phase offset ranges have a working condition correspondence, reducing the decomposition deviation caused by the mismatch between historical parameters and the current window.

[0032] The cross-roll coherence characteristics can be calculated using the following formula: ; in, Represents graph nodes With graph nodes First-order cross-roll coherence characteristics between them This represents the set of stability analysis windows used for statistical analysis. Represents a set One of the analysis windows, Represents graph nodes In the analysis window The first-order complex vector extracted from the interior. express The conjugate vector, Represents the vector magnitude, which physically measures the degree of consistency in the magnitude and phase of the first-order responses of two graph nodes across multiple stability windows. For example, when two windows... They are respectively and , They are respectively and When the numerator is |1×0.5+2×1|=2.5, the denominator is... Calculations yielded .

[0033] The graph edge propagation parameters can be generated using the following formula:

[0034]

[0035] ; in, Represents graph nodes To the graph node The delay in transmission, Represents graph nodes To the graph node The attenuation coefficient, Represents graph nodes To the graph node The phase offset range, Analysis window The time index difference between the responses of the two graph nodes. Analysis window The ratio of the first-order magnitudes of the two inner graph nodes. Analysis window First-order phase difference between two nodes in the inner graph. This indicates taking the median. and These represent taking the least significant quantile and the most significant quantile, respectively. Physically, this means using the statistical distribution within the stable window to describe the propagation path's delay, amplitude variation, and phase variation. For example, when... When the value is 1, 2, 2, 3, or 4, ,when When the value is 0.5, 0.6, 0.6, 0.7, or 0.8, ,when When arranged in ascending order as 0.10, 0.15, 0.20, 0.25, and 0.30, An approximate interval [0.10, 0.30] can be taken.

[0036] In a preferred embodiment, the graph edge propagation parameters can also establish a condition index according to the propagation association category. The edge computing node establishes sample clusters for paper web contact propagation, transmission synchronous propagation, and frame transmission propagation respectively. Each sample cluster stores the corresponding phase difference distribution, amplitude ratio distribution, and time index difference distribution. If the same graph edge falls into multiple sample clusters in different stable operation segments, a condition index is established according to the operation status identifier and order coherence characteristics. The condition index is used to call the propagation delay, attenuation coefficient, and phase offset range that match the current window during online analysis. The specific processing method can be organized according to Table 2. Table 2 is used to illustrate the correspondence between propagation association category, formation basis, statistical content, and online call basis.

[0037] Table 2. Graph Edge Propagation Association Categories and Conditional Index Content

[0038] In this embodiment, reference Figure 4 In the online analysis phase, the same-frequency response decomposition uses the first-order vector as the basic object. The edge computing node forms a first-order complex vector for each graph node within the current analysis window, and decomposes the first-order complex vector into the self-excited component of the local roll, the neighboring roll transmitted component, and the common disturbance component. The common disturbance component is first generated by a set of graph nodes with common operating state identifiers and cross-roller coherence characteristics that meet the requirements. The edge computing node performs phase normalization processing on the first-order vectors in the graph node set to form candidate common vectors. Then, the candidate common vectors are substituted back to each graph node along different propagation paths of the roller coupling graph. If a candidate common vector can only explain local adjacent graph nodes but cannot explain other graph nodes under the same operating state, then the candidate common vector is excluded from the common disturbance candidate and transferred to the neighboring roll transmitted component candidate. The remaining vectors that are not bound to the phase direction of a single roll are determined as common disturbance components. The advantage of this embodiment is that it separates the common disturbance of the operating condition from the source roll phase judgment, so that the source roll judgment is not directly affected by the global same-frequency change.

[0039] The decomposition of the same frequency response can be solved by the following constraint: ; in, This represents the decomposition objective function. Represents graph nodes In the current analysis window, a first-order complex vector Represents graph nodes The corresponding self-excited component of this roller, Represents graph nodes Received adjacent roller transfer component, Represents graph nodes The corresponding common disturbance component, The set of edges in the roller coupling diagram. Indicates the graph nodes Propagate to graph nodes The edges of the graph, Represents graph edges Propagation matching weights, Represents graph edges The attenuation coefficient, Indicates according to phase offset The physical meaning of performing vector rotation is to ensure that the transmission components of adjacent rollers are consistent with the propagation parameters of the image edge, given that the first-order vector is equal to the sum of its three components. For example, when... , , hour, This can make the residual of the first term zero, if , , If so, the propagation residual in the second term is also 0.

[0040] In one embodiment, the estimation of the adjacent roll transfer component is driven by the adjacent graph edges of the roll group coupling graph. The edge computing node calls the propagation delay, attenuation coefficient, and phase offset range under the corresponding condition index according to the current operating state identifier, and performs delay compensation, amplitude scaling, and phase rotation on the first-order vector of the adjacent graph node along the propagation direction of the graph edge to obtain one or more candidate transfer vectors. If multiple candidate transfer vectors point to the same graph node at the same time, the edge computing node assigns weights to the candidate transfer vectors according to the cross-roll coherence characteristics and back-substitution residuals. The weights are not used as fixed empirical parameters, but are calculated by the propagation delay matching degree, phase offset matching degree, and condition index matching of the current window. If a candidate transfer vector can still explain the first-order vector of the target graph node after deducting common disturbances, the candidate vector enters the adjacent roll transfer component. If the explanation residual is continuously large, the candidate vector does not participate in the decomposition of this window. The advantage of this embodiment is that the same frequency peak of the passive response roll can be found to its source along the graph edge, instead of being directly included in the source roll set.

[0041] Furthermore, the self-excited component of this roll is determined by the remaining vector after deducting the common disturbance component and the neighboring roll propagation component. The edge computing node writes the remaining vector of each graph node into the source roll judgment sequence and records the phase direction, amplitude continuity, and residual relationship between the remaining vector and the propagation constraint of the adjacent graph edge according to the continuous analysis window. If the remaining vector of the same graph node has the same phase direction in multiple continuous windows, and the residual after reverse interpretation along the adjacent graph edge is lower than the residual after forward interpretation, then the roll body corresponding to the graph node is written into the dynamic unbalanced source roll set. Forward interpretation means interpreting the target graph node as propagated from the adjacent graph node, and reverse interpretation means taking the target graph node as the propagation starting point and searching for an interpretable path along the reverse number of the adjacent graph edge. When writing into the source roll set, the edge computing node synchronously generates a phase interval. The phase interval is formed by the phase boundary of the remaining vector in the continuous window and is saved at the same time as the original time index and the angular domain sampling index. The advantage of this embodiment is that the source roll judgment simultaneously satisfies the three conditions of phase continuity, propagation interpretability, and residual traceability.

[0042] The source roller judgment quantity can be calculated using the following formula: ; in, Represents graph nodes The corresponding source roller judgment quantity, This indicates the proportion of the self-excited component of this roller in the first-order vector. This indicates the consistency of the phase direction of the remaining vectors within the continuous analysis window. This indicates that the graph nodes This is interpreted as the positive interpretation residual obtained when adjacent graph nodes transmit responses. This indicates that the graph nodes The reverse interpretation residual is obtained when interpreting other graph nodes by reversing the numbering of adjacent graph edges from the propagation starting point. , , These represent the weights of the three types of decision quantities. This indicates preventing small positive quantities with a denominator of 0. Physically, it describes the probability of a source roll using the proportion of self-excitation, phase consistency, and residual direction difference. For example, when... , , , , , , , hour, .

[0043] In a preferred embodiment, after the edge computing node generates the dynamic unbalanced source roll set, it not only saves the roll number, but also the analysis window identifier, original time index range, angular domain sampling index range, phase interval, graph edge involved in the interpretation, called condition index, and remaining vector sequence corresponding to each source roll candidate. If the same roll continuously enters the source roll set in adjacent windows, the edge computing node merges the phase intervals. The merging is constrained by the consistency of the phase direction of the remaining vector. If there is a jump in the phase interval of adjacent windows and the jump position corresponds to the update boundary of the phase conversion table, the edge computing node first performs a consistency check of the overlapping area, and then decides whether to merge. If multiple rolls enter the source roll set at the same time, a candidate ranking is formed according to the proportion of the self-excited component of this roll, the reverse interpretation residual, and the remaining amplitude after deducting the common disturbance. The ranking result does not replace the component record and still retains the complete propagation interpretation chain of each candidate roll. The advantage of this embodiment is that the review process can return to the specific window, graph edge, and phase interval, rather than relying on only a single alarm value.

[0044] In one embodiment, the extraction of common disturbance components by edge computing nodes adopts a combination of graph node set filtering and path back substitution. The node set must simultaneously satisfy the requirements of consistent operating status identification and matching cross-roll coherence characteristics. The edge computing node performs phase normalization processing on the first-order vectors within the set, so that the phase representations of different nodes are transformed to a common phase reference. Then, candidate common vectors are generated using the normalized vectors. The candidate common vectors are substituted back to each graph node along the paper web contact propagation, transmission synchronization propagation, and frame transmission propagation in the roll coupling graph. If the back substitution residual remains interpretable in multiple propagation categories, the corresponding candidate vector is retained as a common disturbance component. If the back substitution result is concentrated only on a single adjacent node pair, the corresponding candidate vector is marked as a local propagation component and enters the estimation of adjacent roll transmission components. If the back substitution result is bound to the phase direction of a certain roll body residual vector for a long time, the corresponding candidate vector does not enter the common disturbance component. The advantage of this embodiment is that it avoids incorrectly classifying the first-order response of the source roll itself into the common disturbance, and also avoids incorrectly classifying global operating condition disturbances into the source roll residual vector.

[0045] In one embodiment, the updates of propagation delay, attenuation coefficient, and phase offset range are performed incrementally. After completing online analysis, the edge computing node writes windows that are determined to be stable and have clear residual relationships into the candidate sample library. The candidate sample library does not immediately cover the graph edge parameters, but participates in the statistics together with existing historical stable segments after the sample continuity requirement is met. If the phase difference distribution, amplitude ratio distribution, and time index difference distribution of the new sample fall within the existing condition index range, only the distribution record of the corresponding sample cluster is updated. If the new sample continues to form new distribution concentration areas, a new condition index is established and bound to the running status identifier. If the new sample simultaneously affects multiple propagation categories and presents a common phase change, it is first processed as a common disturbance candidate and is not directly written into a single graph edge. After the above processing, the graph edge parameters can be updated with the accumulation of the paper machine's running status, and the propagation constraints can be avoided by a single abnormal window. The advantage of this embodiment is that it keeps the statistical boundary of the propagation model consistent with the field data, providing a sustainable source of parameters for subsequent same-frequency decomposition.

[0046] In one embodiment, when the edge computing node performs component decomposition, it retains residual fragments that fail to decompose or are not fully interpreted. If a graph node's first-order response still has a residual vector after deducting the common disturbance component and the neighboring roll propagation component, but the phase direction of the residual vector is unstable within a continuous window, then the fragment is marked as a residual fragment to be reviewed. If multiple graph nodes in a fragment have high residual vectors and their propagation paths conflict with each other, the edge computing node saves the original time index, corner domain sequence, phase conversion table version, condition index version, and graph edge residual of the fragment. Subsequently, when a similar residual relationship appears again in a stable running fragment, it is included in the graph edge parameter update or source roll determination sequence. This processing ensures that the edge computing node will not directly determine a one-time uninterpretable same-frequency response as a source roll, nor will it discard data containing propagation model update value. The advantage of this embodiment is that it maintains the interpretability of source roll determination, while retaining data basis for subsequent correction of propagation parameters and abnormal components.

[0047] In a preferred embodiment, the edge computing node can perform conflict resolution on multi-source propagation paths based on the roll coupling graph. Specifically, when the target graph node simultaneously receives candidate transmission vectors from multiple adjacent graph nodes, the edge computing node performs delay compensation, amplitude scaling, and phase rotation along each graph edge, and then calculates the interpretation residual of each path for the target first-order vector. If the interpretation residuals of multiple candidate paths are close, the edge computing node checks whether the corresponding candidate paths share an upstream graph node or a common disturbance candidate vector. If they share an upstream graph node, the upstream path is retained and the allocation weight of repeated paths is reduced. If the corresponding candidate paths belong to different propagation categories, the corresponding condition index is called for re-matching according to the current running status identifier. If they still cannot be distinguished, the remaining vector of the target graph node does not enter the source roll set but enters the residual waiting-to-review segment. The advantage of this embodiment is that it handles the path overlap problem when paper web contact propagation and frame transmission propagation coexist, and reduces the situation of repeatedly interpreting the same frequency response.

[0048] In one embodiment, the edge computing node generates the phase interval using a continuous window merging method. Each window in the source roll determination sequence stores the phase direction of the remaining vector and the angular domain sampling index. The edge computing node maps the phase direction of the same source roll candidate in adjacent windows to the same common phase reference, and uses phase continuity as the merging constraint. If the phase intervals of two windows overlap, their intersection or weighted center interval is taken as the merging result. If the phase intervals of two windows do not overlap but the boundary misalignment is caused by the segmented update of the phase conversion table, the overlap area consistency verification result is used to correct the misalignment before merging. If the phase direction difference between two windows comes from the change in the common disturbance deduction result, the window is temporarily suspended from merging and enters the residual waiting-to-review segment. The phase interval is output together with the source roll number, graph edge interpretation relationship and original time index. The advantage of this embodiment is that the phase interval comes from the continuous remaining vector rather than a single point peak, which facilitates subsequent back-checking of the anomaly source by time and angle.

[0049] In one embodiment, the analysis results output by the edge computing node include a set of dynamic imbalance source rolls, the phase interval corresponding to each source roll, the self-excited component sequence of each source roll, the subtracted neighboring roll transmission component sequence, the subtracted common disturbance component sequence, the graph edge category involved in the interpretation, the condition index version, and the analysis window index. Each item in the output results can be traced back to the original time index and corner domain sampling index in the edge cache. If the remote platform needs to receive the analysis results, the edge computing node can upload only the fractionation results, source roll candidates, and necessary residual fragments, without uploading the complete original vibration sequence. If subsequent review finds that a certain source roll candidate is inconsistent with the on-site maintenance results, the edge computing node can check whether the error is caused by common disturbance separation deviation, neighboring roll transmission path conflict, or phase mapping jump based on the saved condition index version and residual fragments. The advantage of this embodiment is that it transforms the dynamic balance analysis from a single outlier output to a traceable data chain output, maintaining the data continuity between real-time processing on the edge side and subsequent review.

[0050] In one embodiment, the method can also be applied to segmented analysis between different roll segments of the same paper machine. The edge computing node establishes a local roll group coupling graph for each roll segment, and associates multiple local roll group coupling graphs with the same analysis period through a shared common phase reference and operating status identifier. Corner domain synchronization, graph edge propagation decomposition, and source roll determination are completed within the local graph. Only common disturbance candidate vectors, condition index identifiers, and first-order vectors of boundary graph nodes are exchanged between local graphs. The entire original vibration sequence is not exchanged. If the same-frequency response of the boundary graph node can be interpreted by the source roll residual vector in the adjacent local graph, the response is treated as a cross-segment adjacent roll transmission component. If it cannot be interpreted by the adjacent local graph and appears in phase in multiple local graphs, it is treated as a cross-segment common disturbance component. The advantage of this embodiment is that it maintains the consistency of decomposition logic when there are many roll groups, and the influence of cross-segment propagation on source roll determination is incorporated into a unified constraint.

[0051] In one embodiment, when the edge computing node executes all analysis processes, it establishes a processing record for each data pruning, corner domain resampling, phase coordinate mapping, graph edge condition index call, common disturbance back substitution, neighbor roll transfer estimation, local roll self-excited component generation, and source roll set writing. The processing record includes input data identifier, output data identifier, formula or statistical interval identifier used, index of the removed sampling points, and residual judgment result. If it is necessary to reproduce a certain analysis window later, the edge computing node reads the corresponding speed sequence, vibration response sequence, operating status identifier, phase conversion table, and roll group coupling diagram version in sequence according to the processing record, and can obtain the same corner domain sequence, component decomposition result, and source roll judgment sequence. The advantage of this embodiment is that each calculation link of the method has a definite input-output relationship, preventing the analysis results from being unreproducible due to missing data fragments, graph edge versions, or phase mapping records.

[0052] In one embodiment, the edge computing node performs a consistency check before outputting the source roller candidate. The check includes whether the phase direction of the remaining vector is stable within the continuous window, whether the source roller candidate can interpret the same frequency response of the adjacent graph nodes along the roller group coupling graph, whether the target graph node after common disturbance deduction still retains the remaining vector bound to its own phase, whether the condition index call is consistent with the current running status identifier, and whether the phase conversion table update boundary coincides with the phase jump sampling point. If any of these items cannot form a traceable record, the corresponding source roller candidate is marked as pending check rather than determined source roller. If all records exist and the residual relationship meets the source roller determination sequence requirements, the source roller number and phase interval are output. The advantage of this embodiment is that it limits the source roller determination to the range with a complete data chain and propagation interpretation chain, reducing misjudgments caused by isolated peaks, phase jumps, or mismatched historical graph edge parameters.

[0053] In one embodiment, the overall engineering implementation of the method can be completed sequentially and collaboratively by an edge buffer layer, a corner domain synchronization processing layer, a roll group coupling graph processing layer, a same-frequency component decomposition layer, and a source roll output layer. The edge buffer layer stores the rotation speed sequence, vibration response sequence, operating status identifier, and original time index. The corner domain synchronization processing layer generates a corner domain sequence based on the continuous rotation trajectory and common phase reference. The roll group coupling graph processing layer maintains graph nodes, graph edges, propagation parameters, sample clusters, and condition indexes. The same-frequency component decomposition layer solves the remaining vector according to the relationship between the common disturbance component, the neighboring roll transmitted component, and the self-excited component of the current roll. The source roll output layer generates a source roll set and phase interval based on the source roll judgment sequence. The data objects with indexes are transmitted between the layers rather than isolated judgment values. The output of the current layer is used as the reference basis for the next layer. The advantage of this embodiment is that it enables the edge-side dynamic balance analysis to form a closed process from original sampling, phase synchronization, propagation interpretation to source roll output. The closed process can address the main technical problems when multi-path coupling and gradual speed changes coexist in the paper machine roll group.

Claims

1. A dynamic balancing analysis method for paper machine rolls based on edge computing, characterized in that, include: The edge computing nodes located on the side of the paper machine roll assembly receive the speed sequence, vibration response sequence and operating status identifier associated with each roll. The rotational speed sequence, the vibration response sequence, and the operating status identifier are aligned to the same analysis window based on the roller number, acquisition location, and timestamp. The vibration response sequence is converted into an angular domain sequence based on the corresponding roller rotation angle; Establish a roller group coupling diagram based on the roller body transmission relationship, paper web contact sequence, and frame transmission relationship; The first-order amplitude, phase stability, and cross-roll coherence features are extracted from the angular domain sequence, and the same-frequency response is decomposed into the self-excited component of this roll, the component transmitted from the neighboring roll, and the common disturbance component in combination with the roll group coupling diagram. Based on the decomposition results, the dynamic imbalance source roller and its phase range are generated.

2. The dynamic balancing analysis method for paper machine rolls based on edge computing according to claim 1, characterized in that, The alignment of the same analysis window includes: determining candidate segments in stable operation based on the running status identifier, and generating continuous rotational trajectory from the rotational speed sequence within the candidate segments; The vibration response sequence is resampled based on the continuous angular trajectory to obtain the angular domain sequence with equal angular intervals; In the presence of multiple roller speed ratios, the angular trajectory of a preset reference roller is used as a common phase reference. The angular domain sampling points of non-reference rollers are mapped to the phase coordinates under the common phase reference, so that the first-order components of different rollers have corresponding phase indices within the same analysis window.

3. The dynamic balancing analysis method for paper machine rolls based on edge computing according to claim 1, characterized in that, The establishment of the roller coupling diagram includes: setting the rollers involved in the dynamic balance analysis and their vibration acquisition positions as graph nodes, and setting the vibration propagation association formed by the paper web contact sequence, transmission synchronization relationship, frame transmission path and adjacent arrangement relationship as graph edges; Configure a propagation delay, attenuation coefficient, and phase offset range for each of the graph edges; The propagation delay, the attenuation coefficient, and the phase offset range are calculated from the cross-roll coherence characteristics, order phase difference, and amplitude attenuation sequence in historical stable operation segments, and stored according to the propagation association category corresponding to the graph edge.

4. The dynamic balancing analysis method for paper machine rolls based on edge computing according to claim 1, characterized in that, The decomposition of the same frequency response includes: taking the propagation delay matching degree and phase offset matching degree between adjacent graph nodes in the roller coupling diagram as constraints, first extracting the common disturbance component from the first-order components of multiple graph nodes; Then, estimate the adjacent roller transfer component based on the adjacent graph edge transfer item received by each graph node; The remaining vector after subtracting the common disturbance component and the adjacent roll transmission component from the first-order component is determined as the self-excited component of this roll, and the phase direction and amplitude continuity of the self-excited component of this roll within the continuous analysis window are recorded.

5. The dynamic balancing analysis method for paper machine rolls based on edge computing according to claim 2, characterized in that, The generation of the continuous turning trajectory includes: performing clock drift correction on the rotation speed sequence, and writing the rotation speed sequence and the vibration response sequence into the same edge buffer queue according to the acquisition timestamp; When there are missing or jittering points between adjacent sampling points, a segmented rotation angle interpolation curve is constructed based on the local variation trend of the rotation speed sequence; The intersection of the segmented corner interpolation curve and the stable operating interval corresponding to the operating status identifier is clipped to obtain the continuous corner trajectory covered by the same time period as the vibration response sequence, and the original time index corresponding to each corner sampling point is retained.

6. The method for dynamic balancing analysis of paper machine rolls based on edge computing according to claim 2, characterized in that, The phase coordinate mapping includes: establishing a phase conversion table based on the speed ratio and transmission phase relationship between the preset reference roller and the non-reference roller; The angular domain sampling points of the non-reference roller are first mapped to their own first-order periodic coordinates, and then converted to the common phase reference according to the phase conversion table. When there is a gradual change in speed within the same analysis window, the phase conversion table is updated in segments according to the continuous turning trajectory, and the order vectors in the overlapping area of ​​the phase coordinates before and after the update are checked for consistency, and phase jump sampling points are removed.

7. The dynamic balancing analysis method for paper machine rolls based on edge computing according to claim 3, characterized in that, The calculation of the propagation delay, the attenuation coefficient, and the phase offset range includes: extracting first-order vector pairs of adjacent graph nodes from historical stable operation segments; The first-order vector pairs are subjected to coherence screening, and vector pairs that meet the coherence requirements and have consistent operating status identifiers are retained. The phase difference distribution, magnitude ratio distribution, and time index difference distribution are calculated based on the retained vector pairs; Parameter statistical intervals are generated according to the propagation association categories corresponding to the graph edges, and the common disturbance components shared by different propagation association categories are separated from the parameter statistical intervals.

8. The dynamic balancing analysis method for paper machine rolls based on edge computing according to claim 4, characterized in that, The extraction of the common disturbance component includes: selecting a set of graph nodes in the roller coupling diagram that have a common operating state identifier and meet the requirements of cross-roller coherence characteristics; Phase normalization is performed on the first-order vector of the graph node set to generate candidate common vectors; The candidate common vector is substituted back to each graph node along different propagation paths of the roller coupling graph, and the propagation components that only exist between local adjacent graph nodes are excluded based on the substitution residual, so as to obtain the common disturbance component that is not bound to the phase direction of a single roller.

9. The dynamic balancing analysis method for paper machine rolls based on edge computing according to claim 7, characterized in that, The generation of the parameter statistical interval also includes: establishing sample clusters for the phase difference distribution, the amplitude ratio distribution, and the time index difference distribution according to paper web contact propagation, transmission synchronous propagation, and frame transmission propagation respectively; When the same graph edge falls into multiple sample clusters in different stable running segments, a conditional index is established for the multiple sample clusters based on the running state identifier and order coherence feature. During online analysis, the propagation delay, attenuation coefficient, and phase offset range under the corresponding condition index are invoked based on the current analysis window's running status indicator to participate in the same-frequency response decomposition.

10. The dynamic balancing analysis method for paper machine rolls based on edge computing according to claim 8, characterized in that, The determination of the self-excited component of this roller also includes: inputting the remaining vector after deducting the common disturbance component and the neighboring roller transmitted component into the source roller determination sequence; The source roller determination sequence records the phase direction, amplitude continuity, and residual relationship between the remaining vector and the edge propagation constraint of the adjacent graph according to the continuous analysis window; When the residual vector of the same graph node maintains the same phase direction in multiple consecutive analysis windows, and the residual after reverse interpretation along the adjacent graph edge is lower than the residual after forward interpretation, the roller body corresponding to the same graph node is written into the dynamic unbalanced source roller set and a phase interval is generated.

Images