A method for automatic tracking of a rolling mill drive center line

Abstract

The present application relates to the technical field of rolling mill transmission fault diagnosis, and discloses a kind of automatic tracking determination method of rolling mill transmission center line.The method includes the original information of equipment parameters, motion data and roll data, respectively obtains structured working condition profile and component motion trajectory spectrum by state profile analysis and dynamic trajectory reconstruction.Periodic disturbance feature stripping is carried out on motion trajectory spectrum, disturbance feature profile is generated, and it is registered and superimposed with working condition profile in feature space, to form composite working condition feature map.Extract multi-dimensional feature cluster related to roll from the map, and input prediction model to calculate causal strength list.According to this list, reverse source positioning is carried out in the time-space coordinate system defined by composite feature map, to accurately lock the source working condition stage and specific transmission component causing center line deviation, and generate tracking report containing deviation source identification and mode.The method realizes the automatic, accurate root diagnosis of transmission center line deviation.

Description

Technical Field

[0001] This invention relates to the field of rolling mill transmission fault diagnosis technology, specifically to an automatic tracking and determination method for the center line of a rolling mill transmission. Background Technology

[0002] Currently, the monitoring and diagnosis of deviations in the centerline of rolling mill transmission generally rely on the separate acquisition and monitoring of equipment operating parameters, vibration or displacement signals of transmission components, and the geometric dimensions of the rolls. Conventional technical solutions typically involve parallel display of these data streams after time synchronization and threshold comparison, or the use of signal processing methods to analyze energy changes in specific frequency bands, attempting to establish a correlation between the time of occurrence of abnormal signals and changes in operating parameters.

[0003] The main drawback of these methods is their failure to effectively isolate the inherent periodic disturbances in the transmission chain that are related to the rolling rhythm. This results in characteristic information being drowned out by background noise, making it difficult to distinguish motion changes caused by normal load fluctuations from actual component deterioration or installation deviations. Furthermore, existing analyses often remain at the correlation level. Even if abnormal roll geometry data and synchronous changes in certain operating parameters are detected, it is impossible to quantitatively distinguish the intensity of each influencing factor, let alone reverse-engineer the initial source of the deviation and the specific responsible component within the complex operating sequence and component correlation network. The existing technical system lacks an analytical framework that can integrate macroscopic operating conditions and microscopic periodic disturbances, and can accurately trace the source based on quantitative causal inference. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic tracking and determination method for the center line of a rolling mill drive, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides an automatic tracking and determination method for the center line of a rolling mill drive, the method comprising:

[0006] The system receives raw monitoring information from the online monitoring device and historical database of the rolling mill. The raw monitoring information includes equipment operating parameters, motion data of transmission components, and geometric data of the rolls.

[0007] The working parameters of the equipment are analyzed by state profile to obtain a structured rolling mill working condition profile. The motion data of the transmission components are dynamically reconstructed to obtain the component motion trajectory spectrum. The rolling mill working condition profile and the component motion trajectory spectrum correspond to each other in the time dimension.

[0008] Periodic disturbance feature stripping is performed on the motion trajectory spectrum of the component to generate a disturbance feature profile. The disturbance feature profile is then registered and superimposed with the rolling mill working condition profile in feature space to form a composite working condition feature map.

[0009] Extract multidimensional feature clusters related to the roll geometry data from the composite working condition feature map, input the multidimensional feature clusters into the centerline offset trend prediction model, and calculate the causal strength list between roll geometry anomalies and each working condition dimension.

[0010] Based on the causal intensity list, in the spatiotemporal coordinate system defined by the composite working condition feature map, the roll geometry data is subjected to reverse tracing and positioning processing to lock the source working condition stage and transmission component that caused the centerline deviation, and a transmission centerline tracking report containing deviation source identification and deviation mode is generated.

[0011] Preferably, the step of performing state profile analysis on the equipment operating parameters to obtain a structured rolling mill operating condition profile includes:

[0012] The operating parameters of the equipment are categorized by parameter type, separating load parameters, speed parameters, temperature parameters, and lubrication parameters;

[0013] For each type of parameter's time series, trend inflection point detection is performed to identify the stable state phase, state transition phase, and state abnormal phase in the time series.

[0014] The mean feature vector of the stable state phase, the trend slope feature vector of the transition state phase, and the fluctuation amplitude feature vector of the abnormal state phase are extracted respectively.

[0015] The mean feature vector, trend slope feature vector, and fluctuation amplitude feature vector are assembled into a three-dimensional feature matrix according to the time sequence and stage category to form the rolling mill working condition profile.

[0016] Preferably, the step of dynamically reconstructing the motion data of the transmission component to obtain the component motion trajectory spectrum includes:

[0017] An integral transformation is performed on the vibration acceleration signal in the motion data of the transmission component to obtain a velocity signal sequence, and an integral transformation is performed again on the velocity signal sequence to obtain a displacement signal sequence;

[0018] In the displacement-time coordinate system, the displacement signal sequence, velocity signal sequence, and original vibration acceleration signal are synchronized and aligned.

[0019] Singularity detection is performed on the aligned multidimensional motion signal to identify discontinuous jump points in the motion trajectory;

[0020] The signal segments corresponding to the discontinuous jump points are removed, and spline interpolation based on adjacent signals is used to repair the removed positions to generate a smoothed motion signal sequence.

[0021] The smoothed motion signal sequence is divided into multiple motion cycle units according to a preset trajectory segmentation rule, and each motion cycle unit is mapped to a trajectory line in a high-dimensional phase space. The set of trajectory lines of all motion cycle units constitutes the motion trajectory spectrum of the component.

[0022] Preferably, the step of periodically stripping the motion trajectory spectrum of the component to generate a perturbation feature profile includes:

[0023] Spectral analysis is performed on the trajectory line of each motion cycle unit in the motion trajectory spectrum of the component, decomposing it into a fundamental frequency trajectory component and multiple harmonic trajectory components;

[0024] The fundamental frequency trajectory component is identified and filtered out, and the remaining harmonic trajectory components are synthesized to obtain the periodic disturbance trajectory.

[0025] Calculate the instantaneous energy density of the periodic perturbation trajectory at each sampling time to form a perturbation energy sequence;

[0026] Extract the geometric morphological features of the periodic perturbation trajectory in phase space to form a perturbation morphology vector;

[0027] The disturbance energy sequence and the disturbance morphology vector are merged according to the timestamp to form the disturbance feature profile.

[0028] Preferably, the step of performing feature space registration and superposition of the disturbance feature profile and the rolling mill operating condition profile to form a composite operating condition feature map includes:

[0029] A unified time grid based on the time axis is established, and the data points of the disturbance feature profile and the rolling mill condition profile are resampled to the nodes of the unified time grid.

[0030] Calculate the coupling coefficient matrix between each working condition feature vector in the mill working condition profile and the disturbance feature vector at the corresponding time in the disturbance feature profile;

[0031] Based on the coupling coefficient matrix, tensor product operation is performed on the working condition feature vector and the disturbance feature vector to generate a high-order coupling feature tensor;

[0032] The higher-order coupling feature tensor is expanded along the time dimension to form a two-dimensional feature image with time as the horizontal axis and coupling features as the vertical axis. The two-dimensional feature image is the composite working condition feature map.

[0033] Preferably, the step of extracting multidimensional feature clusters related to the roll geometry data from the composite working condition feature map includes:

[0034] Geometric error pattern recognition is performed on the geometric data of the rolls, and the diameter deviation, roundness error, and cylindricity error of the rolls are quantified into geometric error vectors;

[0035] On the composite working condition feature map, slide a time window that matches the length of the geometric error vector;

[0036] At each time window location, calculate the local feature map block of the composite working condition feature map within the time window;

[0037] Calculate the projection correlation score between the geometric error vector and each local feature patch, and filter out local feature patches whose projection correlation scores exceed a threshold;

[0038] The features corresponding to all the selected local feature patches are aggregated and redundancy removed to form the multidimensional feature cluster.

[0039] Preferably, the step of inputting the multidimensional feature cluster into the centerline offset trend prediction model to calculate the causal strength list between roll geometric anomalies and each working condition dimension includes:

[0040] The centerline offset trend prediction model includes a feature importance evaluation network and a causal graph inference network;

[0041] The multidimensional feature clusters are input into the feature importance evaluation network, which outputs the initial importance weights for each feature dimension.

[0042] The multidimensional feature clusters with the initial importance weights are input into the causal graph inference network. The causal graph inference network is based on a preset rolling mill transmission causal knowledge graph and activates and calculates the weights of the causal edges between feature nodes.

[0043] The causal graph inference network outputs a set of causal path strengths with the geometric error vector as the result node and each feature in the multidimensional feature cluster as the cause node. The causal path strength set is normalized and sorted to form the causal strength list.

[0044] Preferably, the step of performing reverse tracing and localization processing on the roll geometric data in the spatiotemporal coordinate system defined by the composite working condition feature map based on the causal intensity list includes:

[0045] Select the top N causal association pairs with the highest strength from the causal strength list. Each causal association pair contains a roll geometry anomaly term and a composite working condition feature.

[0046] On the composite operating condition feature map, locate all active feature regions that are completely time-corresponding to the first N composite operating condition features;

[0047] For each active feature region, trace back its corresponding original equipment operating parameters and transmission component motion data to identify specific abnormal physical parameter values ​​and abnormal motion state patterns;

[0048] The abnormal values ​​of the physical parameters are compared with the preset normal working range, and the abnormal motion state patterns are matched with the preset standard motion pattern library, thereby associating the active feature area with specific transmission components and working conditions.

[0049] Preferably, generating the transmission centerline tracking report, which includes deviation source identification and deviation pattern, includes:

[0050] Create a standardized report template, which includes fields for deviation source component, deviation condition stage, geometric anomaly type, correlation feature strength, and time series evolution.

[0051] Fill the identified specific transmission component identifier into the deviation source component field;

[0052] Fill the identified specific operating condition stage identifier into the deviation operating condition stage field;

[0053] Fill the geometric error pattern code identified in the roll geometry data into the geometric anomaly type field;

[0054] Extract the corresponding causal strength value from the causal strength list and fill it into the association feature strength field;

[0055] Extract the intensity change curve of the active feature region during the entire monitoring period from the composite working condition feature map and fill it into the time series evolution field;

[0056] Integrate all field information to generate a structured drive centerline tracking report.

[0057] Preferably, the construction steps of the centerline offset trend prediction model include:

[0058] Collect historical monitoring data, which includes historical equipment operating parameters, historical transmission component motion data, historical roll geometry data, and historical centerline offset records;

[0059] The historical equipment operating parameters are analyzed to obtain a historical rolling mill operating condition profile. The motion data of the historical transmission components are dynamically reconstructed to obtain the motion trajectory spectrum of the historical components.

[0060] Periodic disturbance features are stripped from the historical component motion trajectory spectrum to generate a historical disturbance feature profile. The historical disturbance feature profile is then registered and superimposed with the historical rolling mill operating condition profile to form a historical composite operating condition feature map.

[0061] Extract historical multidimensional feature clusters related to the historical roll geometry data from the historical composite working condition feature map;

[0062] Using the historical multidimensional feature cluster as input features and the historical centerline offset record as label, a feature importance evaluation network is trained. The feature importance evaluation network outputs initial importance weights by calculating the weights of each feature dimension.

[0063] A causal graph structure is constructed based on prior knowledge in the field of rolling mill transmission. The causal graph structure includes feature nodes, causal edges, and weight parameters. Historical multidimensional feature clusters with the initial importance weights are input into the causal graph inference network for training. The causal graph inference network learns the causal relationships between features by activating causal edges and updating weight parameters.

[0064] By iteratively optimizing the model, the error between the predicted output of the model and the historical centerline offset records is minimized, thus completing the construction of the centerline offset trend prediction model.

[0065] Compared with the prior art, the beneficial effects of the present invention are:

[0066] By performing periodic disturbance feature stripping on the trajectory spectrum reconstructed from the motion data of transmission components, independent disturbance feature profiles were generated. This separated the repetitive disturbances inherent in the transmission system and synchronized with the process rhythm from the overall motion signal. Subsequently, the stripped disturbance feature profiles were registered and deeply superimposed with the structured chemical condition profiles obtained from equipment operating parameters in a unified high-dimensional feature space. This resulted in a novel composite operating condition feature map, integrating macroscopic operating states and microscopic periodic dynamic disturbances within the same analytical framework. This allowed regular anomaly patterns closely related to the phase of specific operating conditions, previously masked by background noise, to be clearly revealed, providing a fundamental improvement in addressing the problems of incomplete feature extraction and low signal-to-noise ratio under complex interference environments. It also established a reliable data foundation carrying multi-dimensional, structured information for subsequent analysis.

[0067] Multidimensional feature clusters extracted from composite feature maps are input into the prediction model. The core output of the model is a list of quantified causal strengths between roll geometric anomalies and various operating conditions. This clarifies the ranking of the contributions of different influencing factors to the deviation results. Based on this quantified causal guidance, reverse tracing is performed on the observed geometric deviations within the spatiotemporal coordinate system defined by the composite feature maps. Following the inferred strong causal path, reverse reasoning and precise tracking are performed in a correlation network integrating time series, operating conditions, and component behavior. The final output clearly identifies the initial source operating condition stage and specific responsible component that caused the centerline deviation, and elucidates its operational mode. This represents a leap from identifying phenomenon correlations to determining data-driven causal roots, transforming the handling of transmission system deviations from experience-based fuzzy diagnosis to precise intervention based on quantitative causal inference, thus improving the accuracy and efficiency of fault tracing in complex systems. Attached Figure Description

[0068] Figure 1 This is a schematic diagram illustrating the working principle of the automatic tracking and determination method for the center line of the rolling mill transmission described in this invention.

[0069] Figure 2 A flowchart for analyzing the status profile of equipment operating parameters;

[0070] Figure 3 A flowchart for stripping periodic perturbation features;

[0071] Figure 4 Train the loss function variation curve for the centerline offset prediction model;

[0072] Figure 5 Identify active regions and time-series fluctuation feature maps for feature F7. Detailed Implementation

[0073] 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 embodiments of the present invention, and not all embodiments. 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.

[0074] Please see Figure 1This invention provides an automatic tracking and determination method for the centerline of a rolling mill drive. The method includes: receiving raw monitoring information from the online monitoring device and historical database of the rolling mill, the raw monitoring information including equipment operating parameters, motion data of transmission components, and roll geometric data; performing state profile analysis on the equipment operating parameters to obtain a structured rolling mill operating condition profile, and simultaneously performing dynamic trajectory reconstruction on the motion data of the transmission components to obtain a component motion trajectory spectrum, wherein the rolling mill operating condition profile and the component motion trajectory spectrum correspond to each other in the time dimension; periodically removing perturbation features from the component motion trajectory spectrum to generate a perturbation feature profile, and performing feature space registration and superposition of this perturbation feature profile and the rolling mill operating condition profile to form a composite operating condition feature map; extracting multidimensional feature clusters related to the roll geometric data from the composite operating condition feature map, and inputting these multidimensional feature clusters into a centerline offset trend prediction model to calculate a causal strength list between roll geometric anomalies and each operating condition dimension. Based on this causal strength list, in the spatiotemporal coordinate system defined by the composite working condition feature map, the roll geometry data is subjected to reverse tracing and localization processing to lock the source working condition stage and transmission component that caused the centerline deviation, and finally generate a transmission centerline tracking report containing deviation source identification and deviation mode.

[0075] Example 1: See Figure 2 In one implementation, a state profile analysis is performed on the equipment operating parameters to obtain a structured mill operating condition profile. This process includes classifying the equipment operating parameters by type, separating load parameters, speed parameters, temperature parameters, and lubrication parameters. For the time series of each type of parameter, trend inflection point detection is performed to identify the stable state phase, the transitional state phase, and the abnormal state phase in the time series. The mean feature vector of the stable state phase, the trend slope feature vector of the transitional state phase, and the fluctuation amplitude feature vector of the abnormal state phase are extracted respectively. The mean feature vector, trend slope feature vector, and fluctuation amplitude feature vector are assembled into a three-dimensional feature matrix according to the chronological order and phase category to form the mill operating condition profile.

[0076] In one implementation, dynamic trajectory reconstruction is performed on the motion data of the transmission component to obtain the component's motion trajectory spectrum. This process includes performing an integral transform on the vibration acceleration signal in the motion data of the transmission component to obtain a velocity signal sequence, and then performing an integral transform on the velocity signal sequence again to obtain a displacement signal sequence. In the displacement-time coordinate system, the displacement signal sequence, the velocity signal sequence, and the original vibration acceleration signal are synchronously aligned. Singularity detection is performed on the aligned multi-dimensional motion signal to identify discontinuous jump points in the motion trajectory. Signal segments corresponding to discontinuous jump points are removed, and spline interpolation based on adjacent signals is used to repair the removed positions to generate a smoothed motion signal sequence. The smoothed motion signal sequence is divided into multiple motion cycle units according to a preset trajectory segmentation rule, and each motion cycle unit is mapped to a trajectory line in a high-dimensional phase space. The set of trajectory lines of all motion cycle units constitutes the component's motion trajectory spectrum.

[0077] In practical implementation, consider an example scenario where, during the strip rolling process, the online monitoring device continuously collects equipment operating parameters including load current, spindle speed, bearing temperature, and lubricating oil pressure, as well as motion data of transmission components including vibration acceleration signals. A historical database stores similar data sets from the same rolling mill within historical rolling cycles. By comparing the data from the current monitoring period with historical normal periods, deviation patterns between equipment operating parameters and transmission component motion data can be identified. During condition profile analysis, equipment operating parameters are categorized into load parameters, speed parameters, temperature parameters, and lubrication parameters. For the time series of each parameter category, trend inflection point detection identifies stable, transitional, and abnormal states. For example, the load parameter time series exhibits a stable state during steady rolling, a transitional state during acceleration or deceleration, and an abnormal state during overload fluctuations. The mean feature vector for the stable state stage, the trend slope feature vector for the transitional state stage, and the fluctuation amplitude feature vector for the abnormal state stage are extracted respectively. These feature vectors are then assembled into a three-dimensional feature matrix according to chronological order and stage category to form the rolling mill condition profile. During dynamic trajectory reconstruction, the vibration acceleration signal in the motion data of the transmission component undergoes integral transformation to obtain a velocity signal sequence, and the velocity signal sequence undergoes another integral transformation to obtain a displacement signal sequence. In the displacement-time coordinate system, the displacement signal sequence, velocity signal sequence, and the original vibration acceleration signal are synchronously aligned. Singularity detection is performed on the aligned multi-dimensional motion signal to identify discontinuous jump points in the motion trajectory, such as signal spikes caused by mechanical impact. Signal segments corresponding to discontinuous jump points are removed, and spline interpolation based on adjacent signals is used to repair the removed positions, generating a smoothed motion signal sequence. The smoothed motion signal sequence is divided into multiple motion cycle units according to a preset trajectory segmentation rule, and each motion cycle unit is mapped to a trajectory line in a high-dimensional phase space. The set of trajectory lines of all motion cycle units constitutes the component motion trajectory spectrum. Data comparison shows that the component motion trajectory spectrum under normal rolling conditions is compared with that under abnormal vibration conditions. The normal trajectory spectrum exhibits regular periodicity, while the abnormal trajectory spectrum shows enhanced harmonic components or distortion of the phase space trajectory.

[0078] In some embodiments, the specific operations of state profile analysis include classifying the equipment operating parameters by parameter type, performing trend inflection point detection on the load parameter time series, identifying the mean value of 700 amps corresponding to the stable load state phase, the alternating positive and negative slope corresponding to the transitional load state phase, and the fluctuation amplitude exceeding 100 amps corresponding to the abnormal load state phase; in the speed parameter time series, the stable state phase corresponds to a constant speed of 300 rpm, the transitional state phase corresponds to linear acceleration, and the abnormal state phase corresponds to a sudden drop in speed. The extracted mean feature vector includes the load mean and speed mean, the trend slope feature vector includes the load change rate and acceleration, and the fluctuation amplitude feature vector includes the load fluctuation range and speed fluctuation range; when assembling the three-dimensional feature matrix, the time dimension corresponds to the sampling point sequence, the feature dimension corresponds to the parameter category, and the phase category dimension identifies stable, transitional, or abnormal. Data comparison shows that under normal operating conditions, the proportion of abnormal phases in the three-dimensional feature matrix is ​​less than 5%, while the proportion of abnormal phases in the early stage of a fault rises to 20%.

[0079] Optionally, the integral transform in dynamic trajectory reconstruction employs a numerical integration method. The vibration acceleration signal is converted into a velocity signal sequence using the trapezoidal integration method, as expressed by the formula:

[0080]

[0081] in: Indicates the first The velocity value at each sampling point Indicates the first Acceleration values ​​at each sampling point This indicates the sampling time interval. The velocity signal sequence is then used again to obtain the displacement signal sequence using the same integration method. Synchronization alignment ensures that the displacement, velocity, and acceleration signals are aligned at the same timestamp, and the displacement-time coordinate system is established with millisecond-level precision. Singularity detection is based on the local standard deviation threshold method to identify positions where the standard deviation of discontinuous jump points exceeds three times the baseline value; spline interpolation repair uses a cubic spline function to fit adjacent signal points to generate a smoothed motion signal sequence. The trajectory segmentation rule divides the motion signal sequence into motion cycle units based on the principal axis rotation period, with 1024 sampling points per cycle; high-dimensional phase space mapping converts each motion cycle unit into a trajectory line through a time delay embedding method, and the set of trajectory lines constitutes the component motion trajectory spectrum. In the data comparison, the trajectory lines of normal motion cycle units are closed and regular in phase space, while the trajectory lines of abnormal motion cycle units show bifurcation or divergence.

[0082] In some embodiments, the parallel execution of state profile analysis and dynamic trajectory reconstruction ensures that the mill operating condition profile and the component motion trajectory spectrum correspond to each other in the time dimension, with timestamp alignment accuracy reaching the millisecond level. In the example scenario, when the mill is rolling strip steel of different thicknesses, the state transition phase of the equipment operating parameters and the acceleration change phase of the transmission component motion data are perfectly matched in time, and data comparison verifies the matching consistency; during the abnormal state phase, when the load parameters fluctuate, discontinuous jump points appear synchronously in the component motion trajectory spectrum, indicating the correlation between the operating condition and the motion.

[0083] It is understandable that the 3D feature matrix assembly of the state profile analysis is stored in tensor format for easy access to subsequent features; the high-dimensional phase space mapping dimension of dynamic trajectory reconstruction can be adjusted according to signal complexity, for example, the embedding dimension can be set to 6 to capture motion dynamics. Data comparison uses visualization tools to show the time alignment effect between the mill working condition profile and the component motion trajectory spectrum. Normal data shows coordinated changes, while abnormal data shows mismatched areas. Optionally, trend inflection point detection uses a sliding window algorithm with a window size of 100 sampling points to balance sensitivity and stability; the mean feature vector is calculated as the arithmetic mean of the data within the window, the trend slope feature vector is obtained through linear regression fitting, and the fluctuation amplitude feature vector is calculated as the difference between the maximum and minimum values ​​within the window. In singularity detection, after removing discontinuous jump points, spline interpolation repair uses 10 adjacent sampling points for fitting to ensure smoothness; the motion cycle unit division is based on the rotational speed pulse signal to ensure that each unit corresponds to a complete mechanical cycle. In data comparison, historical normal data is used to calibrate the trend inflection point detection threshold and the singularity detection baseline, and the current monitoring data uses the same threshold and baseline to achieve consistency processing.

[0084] Example 2: See Figure 3 In one implementation, periodic disturbance features are stripped from the component's motion trajectory spectrum to generate a disturbance feature profile. This process includes spectral analysis of the trajectory line of each motion cycle unit in the component's motion trajectory spectrum, decomposing it into a fundamental frequency trajectory component and multiple harmonic trajectory components. The fundamental frequency trajectory component is identified and filtered out, and the remaining harmonic trajectory components are synthesized to obtain a periodic disturbance trajectory. The instantaneous energy density of the periodic disturbance trajectory at each sampling time is calculated to form a disturbance energy sequence. The geometric morphological features of the periodic disturbance trajectory in phase space are extracted to form a disturbance morphology vector. The disturbance energy sequence and the disturbance morphology vector are merged according to timestamps to form a disturbance feature profile.

[0085] In practical implementation, consider an example scenario where the component's motion trajectory spectrum consists of a set of trajectory lines mapped to a six-dimensional phase space. Each trajectory line represents the motion state of the main shaft within one complete rotation cycle. Data comparison shows that under normal operating conditions, the component's motion trajectory spectrum has regular trajectory lines with good repeatability. However, under conditions where the transmission gears experience localized wear, the trajectory lines of the component's motion trajectory spectrum exhibit periodic distortion. Spectral analysis is performed on the trajectory lines of each motion cycle unit in the component's motion trajectory spectrum. The coordinate sequence of the trajectory lines in each dimension of the phase space is then subjected to a Fast Fourier Transform, decomposing it into a fundamental frequency trajectory component and multiple harmonic trajectory components. The fundamental frequency trajectory component corresponds to the theoretical rotation frequency of the main shaft, for example, 50 Hz; the harmonic trajectory components have frequencies that are integer multiples of the fundamental frequency, such as 100 Hz and 150 Hz components. The fundamental frequency trajectory component is identified and filtered out, and the remaining harmonic trajectory components are synthesized in each dimension to obtain the periodic disturbance trajectory. The instantaneous energy density of the periodic perturbation trajectory at each sampling moment is calculated. This instantaneous energy density is determined by the sum of squares of the velocity components of the perturbation trajectory in each phase space dimension, forming a perturbation energy sequence. The geometric morphological features of the periodic perturbation trajectory in phase space are extracted. These features include the mean trajectory curvature, the trajectory bounding box volume, and the dispersion after principal component analysis, forming a perturbation morphology vector. The perturbation energy sequence and the perturbation morphology vector are merged according to timestamps to construct a perturbation feature profile.

[0086] In some embodiments, the spectrum analysis employs a Fast Fourier Transform (FFT) algorithm to process the trajectory line of each motion cycle unit. The fundamental frequency trajectory component is identified based on the principle of maximum energy, meaning that the frequency component with the largest amplitude in the spectrum is identified as the fundamental frequency component. Filtering out the fundamental frequency trajectory component involves setting the spectrum line corresponding to the fundamental frequency to zero in the frequency domain, followed by inverse Fourier transform to reconstruct the signal. Synthesizing the remaining harmonic trajectory components involves vector superposition of the time-domain signal components corresponding to the second-order and higher harmonic frequencies at each sampling point. The instantaneous energy density of the periodic disturbance trajectory at each sampling moment is calculated using the following formula:

[0087]

[0088] in: Represents the instantaneous energy density at sampling time t. This represents the total number of dimensions in the phase space. This represents the perturbation velocity component in the d-th dimension of the phase space at sampling time t. The perturbation energy sequence is composed of the values ​​from all sampling times. A one-dimensional array arranged in chronological order. In the data comparison scenario, under normal operating conditions, the amplitude of the disturbance energy sequence fluctuates within the range of 5 to 10 units, while when the bearing inner ring fails, the disturbance energy sequence exhibits impactful peaks with intervals equal to integers of the rotation period, and the peak amplitude can exceed 30 units.

[0089] Optionally, extracting the geometric morphological features of the periodic disturbance trajectory in phase space involves calculating the mean trajectory curvature, which is obtained by averaging the curvature formed by three consecutive points on the trajectory line. The trajectory bounding box volume is obtained by calculating the product of the differences between the maximum and minimum values ​​of the trajectory points in each phase space dimension. The dispersion after principal component analysis of the trajectory is obtained by calculating the sum of the variances of the trajectory points projected onto the plane spanned by the first and second principal components. These calculated values ​​together constitute a multidimensional disturbance morphology vector. In the example scenario, during the sudden load change phase of the rolling mill, the periodic disturbance trajectory becomes more complex, and the trajectory bounding box volume in the disturbance morphology vector increases by 15% to 20% compared to the stable phase. The data comparison clearly records this change.

[0090] In some embodiments, when merging the disturbance energy sequence and the disturbance morphology vector according to timestamps, it is necessary to ensure that both have the same time resolution and length. The merging operation generates a two-dimensional array, where the rows of the array correspond to the sampling time, and the columns of the array are, in order, the instantaneous energy density value and the values ​​of each component of the disturbance morphology vector. This two-dimensional array is the disturbance feature profile. Data comparison can be performed horizontally, aligning and comparing disturbance feature profiles of the same rolling mill on different days. For example, comparing profiles from Monday and Wednesday, the persistent high energy density region appearing in the afternoon of Wednesday indicates potential fatigue accumulation in the components. It is understood that the periodic disturbance feature stripping process relies entirely on the component motion trajectory spectrum as input, and the quality of the component motion trajectory spectrum directly affects the effectiveness of the disturbance feature profile. In the example scenario, if the component motion trajectory spectrum is not sufficiently smoothed in dynamic trajectory reconstruction due to excessive signal acquisition noise, the generated disturbance feature profile will contain a large number of high-frequency noise features. These noise features will appear as irregular spikes in the data comparison, which are significantly different from the actual periodic disturbance pattern. It is understandable that the disturbance feature profile, as a structured intermediate data carrier, has each row of data strictly corresponding to a sampling time point, and simultaneously carries information on the disturbance energy intensity and the disturbance geometry. In data comparison and analysis, operators can simultaneously observe the disturbance energy sequence curve and a certain component curve in the disturbance morphology vector, such as whether the trajectory dispersion and instantaneous energy density show coordinated changes at the same time point. This coordinated change pattern is often associated with specific types of mechanical faults.

[0091] Example 3: In one embodiment, the disturbance feature profile and the rolling mill operating condition profile are registered and superimposed in feature space to form a composite operating condition feature map. This process includes establishing a unified time grid based on the time axis, and resampling the data points of the disturbance feature profile and the rolling mill operating condition profile to the nodes of the unified time grid. The coupling coefficient matrix between each operating condition feature vector in the rolling mill operating condition profile and the disturbance feature vector at the corresponding time in the disturbance feature profile is calculated. Based on the coupling coefficient matrix, a tensor product operation is performed on the operating condition feature vector and the disturbance feature vector to generate a higher-order coupled feature tensor. The higher-order coupled feature tensor is expanded along the time dimension to form a two-dimensional feature image with time as the horizontal axis and coupled features as the vertical axis; this two-dimensional feature image is the composite operating condition feature map.

[0092] In one implementation, a multidimensional feature cluster related to the roll geometry data is extracted from the composite condition feature map. This process includes performing geometric error pattern recognition on the roll geometry data, quantizing the roll diameter deviation, roundness error, and cylindricity error into a geometric error vector. A time window matching the length of the geometric error vector is slid across the composite condition feature map. At each time window location, local feature patches of the composite condition feature map within the time window are calculated. The projection correlation score between the geometric error vector and each local feature patch is calculated, and local feature patches with projection correlation scores exceeding a threshold are selected. The features corresponding to all selected local feature patches are aggregated and deredundant to form a multidimensional feature cluster.

[0093] In practical implementation, consider an example scenario where the mill operating condition profile is a three-dimensional feature matrix, with dimensions representing time points, parameter categories, and state stage characteristics, respectively. The disturbance feature profile is a two-dimensional array, with rows representing time points and columns representing instantaneous energy density and various components of the disturbance morphology vector. Data comparison shows that during periods of healthy transmission system operation, the load parameters in the mill operating condition profile are in a stable state, and the instantaneous energy density value in the disturbance feature profile is at a low level. However, during periods of slight wear on the transmission gears, while the mill operating condition profile still shows a stable state, the specific harmonic energy density in the disturbance feature profile exhibits a periodic increase. This inconsistency requires alignment and fusion analysis through feature space registration. A unified time grid based on the time axis is established, with the node interval set to 10 milliseconds to meet the accuracy requirements of mill dynamic monitoring. Data points from both the disturbance feature profile and the mill operating condition profile are resampled to nodes of the unified time grid using linear interpolation, ensuring that the two data sources have feature values ​​at strictly identical timestamps. The coupling coefficient matrix between each working condition feature vector in the rolling mill working condition profile and the corresponding time-time disturbance feature vector in the disturbance feature profile is calculated. This coupling coefficient matrix expresses the degree of linear correlation between the working condition features and the disturbance features. Based on the coupling coefficient matrix, a tensor product operation is performed on the working condition feature vector and the disturbance feature vector to generate a higher-order coupling feature tensor. This higher-order coupling feature tensor is expanded along the time dimension to form a two-dimensional feature image with time as the horizontal axis and coupling features as the vertical axis; this two-dimensional feature image is the composite working condition feature map. In the data comparison, the composite working condition feature map during the healthy period has a uniform texture, while the composite working condition feature map corresponding to the wear period shows bright stripes in specific time regions. These bright stripes reflect an abnormal enhancement of the coupling features.

[0094] In some embodiments, multidimensional feature clusters related to roll geometry data are extracted from the composite condition feature map. This process includes geometric error pattern recognition of the roll geometry data, obtaining the roll diameter sequence, roundness error value, and cylindricity error value through a profile measuring instrument, and quantizing these errors into a geometric error vector, for example, a vector in the form of [maximum diameter deviation, roundness error, cylindricity error]. A time window matching the length of the geometric error vector is slid across the composite condition feature map. The length of the time window is set according to the evolution rate of the geometric error, for example, the window covers monitoring data from the past 5 minutes. At each time window location, local feature patches of the composite condition feature map within the time window are calculated. A local feature patch is a two-dimensional array containing the values ​​of all coupled features within that time period. The projection correlation score between the geometric error vector and each local feature patch is calculated. The projection correlation score measures the degree of correlation between the overall pattern of the local feature patch and the current roll geometry error pattern. Local feature patches with projection correlation scores exceeding a preset threshold are selected. The threshold is obtained through statistical analysis of historical normal data. The features corresponding to all selected local feature patches are aggregated and redundancy is removed using principal component analysis to form the final multidimensional feature cluster. Data comparison shows that when the roll experiences uniform wear, the local feature patches with high projection correlation scores are evenly distributed over time; when the roll experiences local spalling, the local feature patches with high projection correlation scores are concentrated in a few specific time windows corresponding to the spalling event.

[0095] Optionally, calculate the coupling coefficient matrix between the characteristic vector of each operating condition in the rolling mill operating condition profile and the disturbance characteristic vector at the corresponding time in the disturbance characteristic profile, defined by the formula:

[0096]

[0097] in: This represents the element in the i-th row and j-th column of the coupling coefficient matrix. This represents the value of the i-th operating condition feature in the rolling mill operating condition profile at time t. This represents the mean of the i-th working condition feature over the time series. Let represent the value of the j-th perturbation feature in the perturbation feature profile at time t. Let represent the mean of the j-th disturbance feature over the time series. A tensor product operation is performed on the operating condition feature vector and the disturbance feature vector. Each element in the higher-order coupled feature tensor is the product of the operating condition feature component and the disturbance feature component, multiplied by the corresponding coupling coefficient. In the example scenario, if the operating condition feature vector contains the load mean and the disturbance feature vector contains the second harmonic energy, then their tensor product represents the coupling strength between the load and the second harmonic vibration. This strength value occupies an independent vertical axis position in the composite operating condition feature map.

[0098] In some embodiments, the projection correlation score between the geometric error vector and each local feature patch is calculated using a combination of vector projection and matrix Frobenius inner product. The geometric error vector is treated as a query vector, and the local feature patches, averaged over the time dimension, are considered as a target feature vector. Projection correlation score The cosine of the angle between the query vector and the target feature vector is calculated, and then multiplied by the normalized result of the product of the magnitudes of the two vectors. A dynamic threshold is set for the filtering operation; this threshold is the median of the projection relevance scores across all time windows plus twice the standard deviation. Features corresponding to all local feature patches exceeding the threshold are aggregated to form a temporary feature set. Principal component analysis is performed on the temporary feature set, and principal components with a cumulative contribution rate exceeding 95% are retained as the final multidimensional feature clusters.

[0099] The multidimensional feature cluster is a subset of features extracted from this spatiotemporal coordinate system that is most relevant to the current roll geometry. In the data comparison of the example scenario, the curves of each feature in the multidimensional feature cluster changing over time on the composite condition feature map can be plotted and superimposed on the evolution curve of the roll geometry error on the same time axis, allowing for a visual observation of the lead-lag relationship between feature changes and error development over time. Optionally, the sliding step size of the time window can be set to half the window length to achieve dense scanning of the composite condition feature map, avoiding the omission of short-lived but important active feature regions. Geometric error pattern recognition includes not only diameter, roundness, and cylindricity, but also roll crown and roll profile deviation, making the geometric error vector a higher-dimensional vector. In the calculation of the projection correlation score, different weights can be assigned to different components of the geometric error vector to reflect the differences in importance of different geometric error types. Data comparison can be achieved by drawing a heatmap. The horizontal axis of the heatmap represents time, and the vertical axis represents the different feature indices extracted from the composite working condition feature map. The color intensity represents the magnitude of the feature value. On the heatmap, it is clear which time regions the features that are highly correlated with the geometric error vector have clustered.

[0100] Example 4: In one implementation, a multi-dimensional feature cluster is input into a centerline offset trend prediction model to calculate a causal strength list of roll geometric anomalies and each working condition dimension. This process involves a centerline offset trend prediction model comprising a feature importance evaluation network and a causal graph inference network. The multi-dimensional feature cluster is input into the feature importance evaluation network, which outputs the initial importance weight for each feature dimension. The multi-dimensional feature cluster with the initial importance weights is then input into the causal graph inference network. Based on a pre-defined causal knowledge graph of the mill drive, the causal graph inference network activates and weights the causal edges between feature nodes. The causal graph inference network outputs a set of causal path strengths with the geometric error vector as the result node and each feature in the multi-dimensional feature cluster as the cause node. This set of causal path strengths is normalized and sorted to form a causal strength list.

[0101] In one implementation, the construction steps of the centerline offset trend prediction model include collecting historical monitoring data, which includes historical equipment operating parameters, historical transmission component motion data, historical roll geometry data, and historical centerline offset records. The historical equipment operating parameters are analyzed to obtain a historical mill operating condition profile. The historical transmission component motion data are dynamically reconstructed to obtain a historical component motion trajectory spectrum. Periodic disturbance features are stripped from the historical component motion trajectory spectrum to generate a historical disturbance feature profile. This historical disturbance feature profile is then registered and superimposed with the historical mill operating condition profile in feature space to form a historical composite operating condition feature map. Historical multidimensional feature clusters related to historical roll geometry data are extracted from the historical composite operating condition feature map. Using these historical multidimensional feature clusters as input features and historical centerline offset records as labels, a feature importance evaluation network is trained. The feature importance evaluation network outputs initial importance weights by calculating the weights of each feature dimension. A causal graph structure is constructed based on prior knowledge in the field of rolling mill transmission. This structure includes feature nodes, causal edges, and weight parameters. Historical multidimensional feature clusters with initial importance weights are input into the causal graph inference network for training. The network learns the causal relationships between features by activating causal edges and updating weight parameters. Iterative optimization minimizes the error between the model's predicted output and historical centerline offset records, thus completing the construction of the centerline offset trend prediction model.

[0102] In practical implementation, consider an example scenario where the multidimensional feature cluster extracted from the composite working condition feature map contains 15 feature dimensions. For example, feature F1 represents "load mean - fundamental frequency energy coupling," and feature F2 represents "speed trend - third harmonic morphology coupling." Data comparison shows that after training the model using a set of known historical multidimensional feature clusters that cause centerline offset, the model can output a causal strength list for new test data. The list prioritizes the causal strength between the feature "lubricating oil temperature fluctuation - second harmonic energy coupling" and "roll roundness error," which is consistent with the conclusion found in subsequent maintenance records that poor lubrication led to bearing wear, thus affecting the roll system. The centerline offset trend prediction model includes a feature importance evaluation network and a causal graph inference network. The multidimensional feature cluster is input into the feature importance evaluation network, which is a three-layer fully connected neural network that outputs the initial importance weights for each feature dimension. The multidimensional feature cluster with its initial importance weights is input into the causal graph inference network, which, based on a pre-defined causal knowledge graph of the rolling mill transmission, activates and weights the causal edges between feature nodes. The causal graph inference network outputs a set of causal path strengths, with geometric error vectors as result nodes and features from a multi-dimensional feature cluster as cause nodes. This set of causal path strengths is then normalized and sorted to form a causal strength list. Data comparison can be performed by comparing the differences in the causal strength lists output by the model under different historical periods and different fault types. For example, the top three features in the bearing fault case list are vibration-related features, while the top three features in the gear fault case list are load and speed-related features.

[0103] In some embodiments, the construction steps of the centerline offset trend prediction model include collecting historical monitoring data, which includes historical equipment operating parameters, historical transmission component motion data, historical roll geometry data, and historical centerline offset records. The historical equipment operating parameters are analyzed to obtain a historical mill operating condition profile. The historical transmission component motion data are dynamically reconstructed to obtain a historical component motion trajectory spectrum. Periodic disturbance features are stripped from the historical component motion trajectory spectrum to generate a historical disturbance feature profile. The historical disturbance feature profile and the historical mill operating condition profile are registered and superimposed in feature space to form a historical composite operating condition feature map. Historical multidimensional feature clusters related to historical roll geometry data are extracted from the historical composite operating condition feature map. Using the historical multidimensional feature clusters as input features and historical centerline offset records as labels, a feature importance evaluation network is trained. The feature importance evaluation network outputs initial importance weights by calculating the weights of each feature dimension. A causal graph structure is constructed based on prior knowledge in the field of rolling mill transmission. This structure includes feature nodes, causal edges, and weight parameters. Historical multidimensional feature clusters with initial importance weights are input into the causal graph inference network for training. The network learns the causal relationships between features by activating causal edges and updating weight parameters. Iterative optimization minimizes the error between the model's predicted output and historical centerline offset records, thus completing the construction of the centerline offset trend prediction model. Specifically, this involves using historical monitoring data as the training set, historical multidimensional feature clusters as input features, and historical centerline offset records as supervision labels. First, a feature importance evaluation network is trained. This network outputs initial importance weights by calculating the weights of each feature dimension. Then, a causal graph structure is constructed based on prior knowledge of the rolling mill drive domain, containing feature nodes, causal edges, and weight parameters. The historical multidimensional feature clusters with initial importance weights are then input into the causal graph inference network for training. During training, the mean squared error loss function is used to quantify the difference between the model's predicted output and the true label. Causal edges are dynamically activated and weight parameters are updated through a message passing algorithm to learn the causal relationships between features. Iterative optimization continues, with the loss value calculated after each iteration, and network parameters adjusted according to the gradient descent principle until the KL divergence between the causal path strength distribution predicted by the model and the causal relationship distribution marked in the historical records reaches a preset convergence threshold. This completes the model construction, ensuring that the model accurately reflects the causal strength between roll geometric anomalies and various operating condition dimensions. Data comparison can be reflected in the decline curve of the loss function during model training, and the degree of agreement between the causal path strength output by the model and the causal relationship labeled by expert experience when evaluated using an independent validation set.

[0104] Optionally, the feature importance evaluation network structure includes an input layer, a hidden layer with ReLU activation, and an output layer using the Softmax function. The number of neurons in the input layer is equal to the dimension of the multidimensional feature cluster. The number of neurons in the output layer is also equal to the dimension of the multidimensional feature cluster, and the output vector is the initial importance weight, with the sum of all weights being 1. The causal graph inference network adopts a graph neural network structure. The pre-defined causal knowledge graph of the rolling mill drive is defined in the form of an adjacency matrix, where nodes represent features or geometric errors, and edges represent existing causal relationships. During the computation of the causal graph inference network, the causal path strength... The strength of the k-th path from the causal feature node to the result geometric error node is used to represent the strength of the path. Its calculation depends on the initial importance weights of the causal features, the learned weights on the causal edges, and the activation states of the nodes along the path. The formula is expressed as:

[0105]

[0106] in: This represents the strength value of the k-th causal path. This represents the initial importance weight of the causal feature node i. This represents the k-th directed path connecting the cause node i and the result node. Representing a path The learning weight of the causal edge e. Representing a path The activation value of the parent node n, This represents the Sigmoid activation function. Normalized sorting measures the strength of causal paths leading to the same outcome node. Sum the results and calculate the proportion of each path's intensity. Then sort the results by proportion from highest to lowest to form a list of causal intensities.

[0107] In some embodiments, historical centerline offset records are represented as vectors containing the offset direction, offset amount, and occurrence time, serving as labels for supervised learning. When training the feature importance evaluation network, a mean squared error loss function is used, with the optimization objective being to minimize the error between the weighted features of the network output and the true label prediction. Causal edges in the causal graph structure are predefined based on domain knowledge; for example, the "abnormal bearing temperature" node has an edge pointing to the "intensified shaft vibration" node, and the "intensified shaft vibration" node has an edge pointing to the "roll roundness error" node. The training of the causal graph inference network updates the weight parameters of the causal edges through a message-passing algorithm. The model is iteratively optimized until the KL divergence between the distribution of causal path strength predicted by the model and the distribution of causal relationships marked in the historical records is minimized. Refer to Table 1 for data comparison, which shows a comparison between some of the model's outputs on the validation set and the actual situation.

[0108] Table 1: Comparison of Historical Centerline Offset Records and Model Predictions

[0109]

[0110] It is understandable that building a centerline offset trend prediction model involves offline training and online application. The offline training phase requires a large amount of historical monitoring data covering different fault modes and historical states to ensure the model learns broad and robust causal relationships. In the online application phase, the multi-dimensional feature clusters obtained in real-time processing are input into the trained model, allowing for rapid calculation of the causal strength list for the current roll geometric anomaly. Data comparison can be conducted over a long period, periodically using newly generated data with confirmed causes as new samples to incrementally train or retrain the model, continuously improving its accuracy and adaptability. Optionally, the activation of causal edges in the causal graph inference network depends not only on the learned weight parameters. It also depends on whether the value of the corresponding feature node in the input multidimensional feature cluster exceeds its threshold. The activation value of the feature node... The output is obtained by mapping the eigenvalues ​​through a nonlinear function. The final output of the entire causal graph inference network is the set of causal path strengths from all nodes of the multidimensional feature cluster to nodes of the geometric error vector. Data comparison can analyze the impact of different training epochs or different initial causal graph structures on the order of the final output causal strength list under the same set of input features, thereby selecting the optimal model configuration.

[0111] See Figure 4 During model training, the loss function (quantified by mean squared error) shows a significant decreasing trend with each training epoch. Specifically, the training set loss (blue curve) and validation set loss (red curve) decrease from approximately 10 initially. 0 The magnitude of the [something] continued to decline, stabilizing after about 80 rounds of training, and eventually converging to close to 10. -22 The fact that both the model and the loss function maintained a consistent downward trend without significant deviation indicates that the model did not exhibit overfitting or underfitting issues. During training, the decay of the loss function corresponded to the iterative optimization of model parameters (weights of the feature importance evaluation network and weights of the causal edge inference network): the feature importance evaluation network effectively filtered input features by adjusting the initial weights of each feature dimension; the causal graph inference network, based on the causal knowledge graph of the rolling mill drive, activated and updated the causal edge weights through a message passing algorithm, gradually matching the strength of the causal paths learned by the model with the causal relationships of real working conditions. The final loss stabilized near the convergence threshold, indicating that the model had fully learned the patterns in historical data and possessed reliable inference capabilities regarding the causal strength of roll geometric anomalies and working condition dimensions.

[0112] Example 5: In one implementation, based on a causal strength list, reverse tracing and localization processing is performed on the roll geometry data within the spatiotemporal coordinate system defined by the composite operating condition feature map. This process includes selecting the top N causal association pairs with the highest intensity from the causal strength list. Each causal association pair contains a roll geometry anomaly and a composite operating condition feature. On the composite operating condition feature map, all active feature regions that completely correspond to the top N composite operating condition features in time are located. For each active feature region, the corresponding original equipment operating parameters and transmission component motion data are traced back to identify specific physical parameter anomalies and motion state anomaly patterns. The physical parameter anomalies are compared with preset normal operating ranges, and the motion state anomaly patterns are matched with a preset standard motion pattern library, thereby associating the active feature region with specific transmission components and operating condition stages.

[0113] In one implementation, a drive centerline tracking report is generated, containing deviation source identifiers and deviation patterns. This process includes creating a standardized report template that includes fields for deviation source component, deviation condition stage, geometric anomaly type, associated feature intensity, and temporal evolution. The identified specific drive component identifier is entered into the deviation source component field. The identified specific condition stage identifier is entered into the deviation condition stage field. The geometric error pattern code identified from the roll geometry data is entered into the geometric anomaly type field. The corresponding causal intensity value is extracted from the causal intensity list and entered into the associated feature intensity field. The intensity change curve of the active feature region over the entire monitoring period is extracted from the composite condition feature map and entered into the temporal evolution field. All field information is integrated to generate a structured drive centerline tracking report.

[0114] In practical implementation, consider an example scenario where a causal strength list is output from the centerline offset trend prediction model. The list includes entries such as "roll roundness error - feature F7 coupling strength 0.85" and "roll diameter deviation - feature F3 coupling strength 0.76". Data comparison shows that in two centerline deviation events at different times on the same mill, the order of the causal strength list is different: in one instance, "feature F7" has the highest strength, and in the other, "feature F3" has the highest strength, indicating potentially different root causes of the fault. The top N causal pairs with the highest strength are selected from the causal strength list, with N=5. Each causal pair contains a roll geometry anomaly and a composite operating condition feature. On the composite operating condition feature map, all active feature regions that time-completely correspond to the top 5 composite operating condition features are located. For each active feature region, the corresponding original equipment operating parameters and transmission component motion data are traced back to identify specific physical parameter anomalies and motion state anomaly patterns. By comparing abnormal physical parameter values ​​with preset normal operating ranges and matching abnormal motion patterns with a preset standard motion pattern library, the active feature regions are associated with specific transmission components and operating conditions. For example, during the time period corresponding to active feature region A, retrospective analysis reveals that the lubricating oil temperature parameter value consistently exceeds the upper limit of the normal operating range. Simultaneously, the spectral characteristics of the vibration signal match the "poor bearing lubrication" pattern in the standard motion pattern library. Thus, active feature region A is associated with the "lubrication system" transmission component and the "continuous high-temperature operation" operating condition.

[0115] In some embodiments, a drive centerline tracking report containing deviation source identifiers and deviation patterns is generated. This process includes creating a standardized report template containing fields for deviation source component, deviation condition stage, geometric anomaly type, associated feature intensity, and temporal evolution. The identified specific drive component identifier is entered into the deviation source component field, for example, "third stand drive-side tapered roller bearing". The identified specific condition stage identifier is entered into the deviation condition stage field, for example, "high-speed rolling stage (speed > 800 m / min)". The geometric error pattern code identified from the roll geometry data is entered into the geometric anomaly type field, for example, the code "GEO_ERR_002" represents "periodic roundness deviation". The corresponding causal intensity value is extracted from the causal intensity list and entered into the associated feature intensity field, for example, "0.85". The intensity change curve of the active feature region over the entire monitoring period is extracted from the composite condition feature map and entered into the temporal evolution field. All field information is integrated to generate a structured drive centerline tracking report. Data comparison can be seen in the fact that, for the same set of monitoring data, the tracking reports generated using different N values ​​still maintain consistency in identifying the deviation source components and the core conclusions of the operating condition stage. However, the larger the N value, the richer the auxiliary information contained in the report.

[0116] Optionally, in the reverse source tracing and localization process, to locate the active feature region that corresponds completely to the composite working condition features in time, it is necessary to segment the composite working condition feature map. The active feature region is defined as a two-dimensional connected region in the spatiotemporal coordinate system where the feature intensity value continuously exceeds a dynamic threshold in both time and feature dimension. Dynamic threshold The formula is defined as follows: Based on the statistical distribution of specific characteristics throughout the entire monitoring history, the formula is defined as follows:

[0117]

[0118] in: This represents the dynamic threshold used to segment regions with active features. This represents the average value of the feature over a historical normal data period. This represents the standard deviation of the feature over a normal historical data period. This is a sensitivity coefficient set based on the false alarm rate tolerance. In the example scenario, for feature F7, its historical mean and standard deviation for normal periods are calculated, and then set... A dynamic threshold is obtained, and continuous regions on the composite working condition feature map whose intensity exceeds this threshold are identified as active regions of feature F7. Data comparison shows that improving... A lower value will reduce the active regions of identified features, leading to missed detection of minor anomalies; a lower value will reduce the number of active regions of identified features. The more values ​​there are, the more false alarms are likely to occur.

[0119] In some embodiments, the preset normal operating range is stored in the form of intervals, such as "main drive shaft bearing temperature: [45°C, 65°C]". A preset standard motion pattern library is stored in the form of feature vectors, with each pattern corresponding to a label, such as "gear meshing impact", "bearing outer ring failure", and "rotor imbalance". When matching abnormal motion state patterns, the cosine similarity between the feature vector extracted from the motion data of the transmission components within the current time period and each pattern vector in the standard motion pattern library is calculated, and the pattern label with the highest similarity is used as the matching result. Feature active regions are associated with specific transmission components and operating conditions. The output is a mapping list, where each item contains the feature active region ID, time range, associated transmission component name, associated operating condition description, and confidence score. When generating a transmission centerline tracking report, the report template uses Extensible Markup Language or a predefined structured text file. The data in the time-series evolution field is the sequence value of the feature intensity corresponding to the feature active region changing over time, usually stored in array form, used to plot trend curves. Data comparison can compare two tracking reports for the same fault but generated from data at different times. The time evolution field of the earlier report shows that the feature intensity is on the rise, while the report after the fault occurs shows that the feature intensity is in a high plateau period. This helps to determine the development stage of the fault.

[0120] Understandably, reverse tracing and location processing is a crucial step in connecting the causal strength list with specific physical components and stages. The transmission centerline tracing report is the final structured conclusion output by this method, directly usable by maintenance personnel. In the data comparison of the example scenario, the system-generated tracing report can be compared with the subsequent manual inspection report to verify the accuracy of the deviation source component field and the deviation operating condition stage field. For example, the system report indicates that "gearbox input shaft bearing" and "acceleration start-up stage" are the root causes of the problem, while manual inspection confirms that the bearing has early pitting corrosion and the damage characteristics are consistent with the start-up impact.

[0121] Optionally, after generating the drive centerline tracking report, the system can push the report to the equipment maintenance management system, triggering a preventative maintenance work order. The content of the deviation source component field can be directly used for the equipment location description in the work order. Data from the correlation characteristic strength field and the time sequence evolution field can be used by the maintenance management system to assess the urgency of the problem and schedule maintenance priorities. Data comparison can analyze the time taken for the entire process from generating the causal strength list to pushing the tracking report in different historical cases, as well as the relationship between this time and the complexity of the problem, thereby assessing the real-time performance of the system's online diagnostics.

[0122] See Figure 5 In the feature active region identification stage of the automatic tracking and determination method for the centerline of the rolling mill drive, the correlation between the temporal variation of the intensity of feature F7, the dynamic threshold, and the active region is demonstrated. Specifically, the horizontal axis in the figure represents the monitoring period, and the vertical axis represents the intensity value of feature F7; the red curve represents the real-time intensity fluctuation of feature F7, and the black dashed line represents the dynamic threshold calculated based on historical normal data (the threshold of 0.101 is obtained by taking the sensitivity coefficient in the formula). Different colored areas correspond to the active region division of feature F7. In the abnormal stage of 50-70 monitoring cycles (pink and orange areas), the intensity of feature F7 significantly exceeds the dynamic threshold, forming a continuous high-intensity fluctuation range: this range is the active region of feature F7 defined by "two-dimensional connected domain where the feature intensity continuously exceeds the dynamic threshold". Combined with the project method, this region corresponds to the feature segment in the composite working condition feature map that is highly coupled with the geometric anomaly of the roll (such as roundness error). Subsequently, it can be located by reverse tracing and associated with the motion data of the corresponding transmission components and the working condition stage (such as lubrication system anomaly, high-speed rolling stage). In terms of parameter configuration, the determination of active areas is based on dynamic thresholds, and the spatiotemporal matching of feature intensity fluctuations and operating conditions is achieved by monitoring the temporal dimension of the cycle.

[0123] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0124] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for automatically tracking and determining the center line of a rolling mill drive, characterized in that, The method includes: The system receives raw monitoring information from the online monitoring device and historical database of the rolling mill. The raw monitoring information includes equipment operating parameters, motion data of transmission components, and geometric data of the rolls. The operating parameters of the equipment are analyzed using a state profile to obtain a structured rolling mill operating condition profile, including: The operating parameters of the equipment are categorized by parameter type, separating load parameters, speed parameters, temperature parameters, and lubrication parameters; For each type of parameter's time series, trend inflection point detection is performed to identify the stable state phase, state transition phase, and state abnormal phase in the time series. The mean feature vector of the stable state phase, the trend slope feature vector of the transition state phase, and the fluctuation amplitude feature vector of the abnormal state phase are extracted respectively. The mean feature vector, trend slope feature vector, and fluctuation amplitude feature vector are assembled into a three-dimensional feature matrix according to the time sequence and stage category to form the rolling mill working condition profile. Dynamic trajectory reconstruction is performed on the motion data of the transmission component to obtain the component motion trajectory spectrum, wherein the rolling mill working condition profile and the component motion trajectory spectrum correspond to each other in the time dimension; Periodic perturbation feature stripping is performed on the motion trajectory spectrum of the component to generate a perturbation feature profile, including: Spectral analysis is performed on the trajectory line of each motion cycle unit in the motion trajectory spectrum of the component, decomposing it into a fundamental frequency trajectory component and multiple harmonic trajectory components; The fundamental frequency trajectory component is identified and filtered out, and the remaining harmonic trajectory components are synthesized to obtain the periodic disturbance trajectory. Calculate the instantaneous energy density of the periodic perturbation trajectory at each sampling time to form a perturbation energy sequence; Extract the geometric morphological features of the periodic perturbation trajectory in phase space to form a perturbation morphology vector; The disturbance energy sequence and the disturbance morphology vector are merged according to timestamps to form the disturbance feature profile; The disturbance feature profile and the rolling mill operating condition profile are registered and superimposed in feature space to form a composite operating condition feature map. Extract multidimensional feature clusters related to the roll geometry data from the composite working condition feature map, and input the multidimensional feature clusters into the centerline offset trend prediction model. The construction steps of the centerline offset trend prediction model include: Collect historical monitoring data, which includes historical equipment operating parameters, historical transmission component motion data, historical roll geometry data, and historical centerline offset records; The historical equipment operating parameters are analyzed to obtain a historical rolling mill operating condition profile. The motion data of the historical transmission components are dynamically reconstructed to obtain the motion trajectory spectrum of the historical components. Periodic disturbance features are stripped from the historical component motion trajectory spectrum to generate a historical disturbance feature profile. The historical disturbance feature profile is then registered and superimposed with the historical rolling mill operating condition profile to form a historical composite operating condition feature map. Extract historical multidimensional feature clusters related to the historical roll geometry data from the historical composite working condition feature map; Using the historical multidimensional feature cluster as input features and the historical centerline offset record as label, a feature importance evaluation network is trained. The feature importance evaluation network outputs initial importance weights by calculating the weights of each feature dimension. A causal graph structure is constructed based on prior knowledge in the field of rolling mill transmission. The causal graph structure includes feature nodes, causal edges, and weight parameters. Historical multidimensional feature clusters with the initial importance weights are input into the causal graph inference network for training. The causal graph inference network learns the causal relationships between features by activating causal edges and updating weight parameters. By iteratively optimizing the model, the error between the predicted output of the model and the historical centerline offset records is minimized, thus completing the construction of the centerline offset trend prediction model. A list of causal strengths between roll geometric anomalies and various operating conditions was calculated. Based on the causal intensity list, in the spatiotemporal coordinate system defined by the composite working condition feature map, the roll geometry data is subjected to reverse tracing and positioning processing to lock the source working condition stage and transmission component that caused the centerline deviation, and a transmission centerline tracking report containing deviation source identification and deviation mode is generated.

2. The automatic tracking and determination method for the center line of the rolling mill drive according to claim 1, characterized in that, The step of dynamically reconstructing the motion data of the transmission component to obtain the component motion trajectory spectrum includes: An integral transformation is performed on the vibration acceleration signal in the motion data of the transmission component to obtain a velocity signal sequence, and an integral transformation is performed again on the velocity signal sequence to obtain a displacement signal sequence; In the displacement-time coordinate system, the displacement signal sequence, velocity signal sequence, and original vibration acceleration signal are synchronized and aligned. Singularity detection is performed on the aligned multidimensional motion signal to identify discontinuous jump points in the motion trajectory; The signal segments corresponding to the discontinuous jump points are removed, and spline interpolation based on adjacent signals is used to repair the removed positions to generate a smoothed motion signal sequence. The smoothed motion signal sequence is divided into multiple motion cycle units according to a preset trajectory segmentation rule, and each motion cycle unit is mapped to a trajectory line in a high-dimensional phase space. The set of trajectory lines of all motion cycle units constitutes the motion trajectory spectrum of the component.

3. The automatic tracking and determination method for the center line of the rolling mill drive according to claim 1, characterized in that, The step of registering and superimposing the disturbance feature profile with the rolling mill operating condition profile in feature space to form a composite operating condition feature map includes: A unified time grid based on the time axis is established, and the data points of the disturbance feature profile and the rolling mill condition profile are resampled to the nodes of the unified time grid. Calculate the coupling coefficient matrix between each working condition feature vector in the mill working condition profile and the disturbance feature vector at the corresponding time in the disturbance feature profile; Based on the coupling coefficient matrix, tensor product operation is performed on the working condition feature vector and the disturbance feature vector to generate a high-order coupling feature tensor; The higher-order coupling feature tensor is expanded along the time dimension to form a two-dimensional feature image with time as the horizontal axis and coupling features as the vertical axis. The two-dimensional feature image is the composite working condition feature map.

4. The automatic tracking and determination method for the center line of the rolling mill drive according to claim 3, characterized in that, The extraction of multidimensional feature clusters related to the roll geometry data from the composite working condition feature map includes: Geometric error pattern recognition is performed on the geometric data of the rolls, and the diameter deviation, roundness error, and cylindricity error of the rolls are quantified into geometric error vectors; On the composite working condition feature map, slide a time window that matches the length of the geometric error vector; At each time window location, calculate the local feature map block of the composite working condition feature map within the time window; Calculate the projection correlation score between the geometric error vector and each local feature patch, and filter out local feature patches whose projection correlation scores exceed a threshold; The features corresponding to all the selected local feature patches are aggregated and redundancy removed to form the multidimensional feature cluster.

5. The automatic tracking and determination method for the center line of the rolling mill drive according to claim 4, characterized in that, The process involves inputting the multidimensional feature clusters into the centerline offset trend prediction model to calculate a list of causal strengths between roll geometric anomalies and various operating conditions, including: The centerline offset trend prediction model includes a feature importance evaluation network and a causal graph inference network; The multidimensional feature clusters are input into the feature importance evaluation network, which outputs the initial importance weights for each feature dimension. The multidimensional feature clusters with the initial importance weights are input into the causal graph inference network. The causal graph inference network is based on a preset rolling mill transmission causal knowledge graph and activates and calculates the weights of the causal edges between feature nodes. The causal graph inference network outputs a set of causal path strengths with the geometric error vector as the result node and each feature in the multidimensional feature cluster as the cause node. The causal path strength set is normalized and sorted to form the causal strength list.

6. The automatic tracking and determination method for the center line of the rolling mill drive according to claim 1, characterized in that, The reverse tracing and localization processing of the roll geometry data in the spatiotemporal coordinate system defined by the composite working condition feature map, based on the causal strength list, includes: Select the top N causal association pairs with the highest strength from the causal strength list. Each causal association pair contains a roll geometry anomaly term and a composite working condition feature. On the composite operating condition feature map, locate all active feature regions that are completely time-corresponding to the first N composite operating condition features; For each active feature region, trace back its corresponding original equipment operating parameters and transmission component motion data to identify specific abnormal physical parameter values ​​and abnormal motion state patterns; The abnormal values ​​of the physical parameters are compared with the preset normal operating range, and the abnormal motion state patterns are matched with the preset standard motion pattern library, thereby associating the active feature area with specific transmission components and operating conditions.

7. The automatic tracking and determination method for the center line of the rolling mill drive according to claim 6, characterized in that, The generation of the drive centerline tracking report, which includes deviation source identification and deviation pattern, includes: Create a standardized report template, which includes fields for deviation source component, deviation condition stage, geometric anomaly type, correlation feature strength, and time series evolution. Fill the identified specific transmission component identifier into the deviation source component field; Fill the identified specific operating condition stage identifier into the deviation operating condition stage field; Fill the geometric error pattern code identified in the roll geometry data into the geometric anomaly type field; Extract the corresponding causal strength value from the causal strength list and fill it into the association feature strength field; Extract the intensity change curve of the active feature region during the entire monitoring period from the composite working condition feature map and fill it into the time series evolution field; Integrate all field information to generate a structured drive centerline tracking report.

Images