Intelligent Correction Method for Die Core Position Based on Cable Insulation Extrusion Eccentricity Detection

Abstract

This invention relates to an intelligent correction method for die core position based on eccentricity detection in cable insulation extrusion, aiming to address the shortcomings of existing systems in identifying and accurately correcting anomalies. It primarily acquires multimodal monitoring data, including high-precision X-ray tomography, die temperature, screw speed, traction wheel phase, and melt pressure, to establish an error source fingerprint template. Feature extraction, real-time feature stream encoding, and similarity comparison are then performed. Combined with Bayesian inference and dynamic screening of candidate causes based on operating parameters, a gating mechanism automatically determines and triggers servo / temperature zone correction, forming a data closed-loop iterative optimization of the fingerprint template. This achieves beneficial effects such as accurate cause classification, timely response, and system self-evolutionary correction, improving the long-term stability and intelligent level of cable extrusion quality.

Description

Technical Field

[0001] This invention relates to the field of cable manufacturing process control and intelligent detection technology, and in particular to an intelligent correction method for the die core position based on the detection of cable insulation layer extrusion eccentricity. Background Technology

[0002] Eccentricity detection and correction during cable insulation extrusion is a core aspect of ensuring product quality and stable production line operation. Currently, on-site control of cable extrusion eccentricity primarily employs automated control systems with a core link of "real-time detection—error attribution—die core positioning correction." Mainstream technologies typically fall into two categories: one is online thickness detection and spatial center deviation extraction based on high-precision X-ray tomography and other image processing methods, using edge recognition and statistical fitting to determine the eccentricity amount; the other combines process parameter acquisition (such as temperature distribution, screw speed, traction speed, melt pressure, etc.) with expert experience rules, employing fixed models or data-driven execution to output correction commands for the die core, temperature zone, or servo actuators. These systems are widely used in the manufacturing of high-end cables, communication optical cables, and power cables, playing a positive role in improving automation levels, reducing manual intervention, and shortening stabilization time.

[0003] However, existing eccentricity correction systems generally suffer from the following problems in the attribution and control command generation process: First, attribution mechanisms are predominantly deterministic, relying on low-confidence or non-quantitative results from the attribution model output. They typically classify detected eccentricities into a single specific cause (such as mold core misalignment, melt fluctuation, mechanical vibration, etc.), and directly generate strong intervention correction commands based on this. Traditional methods lack the quantification of dynamic uncertainties related to various mixed causes, operating condition drift, sensor anomalies, etc., and only consider protection mechanisms against extreme fluctuations in initial system optimization or emergency fault-tolerant scenarios.

[0004] Secondly, to reduce control complexity, technical solutions often employ fixed thresholds or attribution-response mapping logic based on expert experience, which is difficult to adapt to the variable characteristics of real-world operating conditions such as complex disturbances, multi-source fluctuations, and equipment aging in production environments. For example, when detecting eccentricity anomalies, some systems automatically invoke full-amplitude correction algorithms or directly execute large-amplitude core movements, temperature zone increases and decreases, etc., causing overcorrection, frequent adjustments, and residual oscillations, which may lead to equipment damage or product quality fluctuations in severe cases.

[0005] Furthermore, existing process control systems do not adequately consider the uncertainty and confidence level of detection attribution results, and lack scientific data fusion and decision gating mechanisms. Specifically, regardless of the detection signal-to-noise ratio, acquisition synchronization, or parameter fluctuation, the system uniformly performs corrections with the same weight, rarely dynamically adjusting the response intensity based on real-time operating condition health and historical attribution consistency. This results in poor sensitivity to erroneous attributions or sensor anomalies during actual operation, and a lack of collaborative robust optimization capabilities.

[0006] Furthermore, although some studies have drawn on theories such as causal reasoning and variable fusion to attempt to construct multi-model integration or introduce complex statistical methods to improve attribution reliability, these methods often suffer from bloated models, poor on-site interpretability, and long inference chains, making it difficult to meet the cable manufacturing industry's stringent requirements for real-time performance, low computational power, and transparent controllability. On the other hand, some systems do not even consider the reverse correction of the attribution model by the feedback of correction command execution, resulting in the actual error attribution library failing to adapt to the natural performance drift of equipment and processes over a long period of time, leading to a significant decline in effectiveness after long-term operation.

[0007] The main technical challenges faced by existing technologies are reflected in the following aspects: The complex and varied causes of eccentricity due to multiple process disturbances and equipment state drift make it difficult for existing attribution methods to accurately characterize the mixed causes and their dynamic changes.

[0008] The lack of confidence modeling and gating in biased attribution results leads to the system still generating strong intervention instructions for low-confidence or abnormal attributions, which affects product quality and equipment safety.

[0009] The correction execution and feedback information did not form a closed loop with the attribution input, making it difficult for the fingerprint database and control decision model to self-evolve and compensate for changes under real-world operating conditions in a timely manner.

[0010] Existing multi-model ensemble and reinforcement learning strategies lack real-time performance and interpretability in on-site inference optimization, and industrial control systems find it difficult to accept overly complex inference chains.

[0011] Therefore, there is an urgent need for a closed-loop decision-making method that not only possesses multimodal feature recognition capabilities but also dynamically quantifies attribution uncertainty, enables targeted control response intensity adjustment, and self-evolves with changing operating conditions. It should have adaptive capabilities driven by rapid screening, robust matching, and feedback of correction effects to maximize the safety and accuracy of eccentricity correction, meeting the inherent requirements of high robustness, high real-time performance, and high interpretability in cable insulation extrusion processes. These issues are precisely the real-world pain points that this patent attempts to fundamentally overcome. Summary of the Invention

[0012] This application provides an intelligent correction method for the die core position based on the detection of cable insulation extrusion eccentricity, aiming to solve one of the problems or issues of the prior art mentioned in the background.

[0013] The intelligent correction method for die core position based on cable insulation extrusion eccentricity detection provided in this application specifically includes: S1: Acquire multimodal monitoring data during the cable insulation extrusion process. The multimodal monitoring data includes X-ray tomographic images, temperature distribution sequences, rotational speed fluctuation spectra, phase jitter signals, and pressure time series data.

[0014] S2: Based on the historical eccentricity sample annotation information, feature extraction is performed on the multimodal monitoring data to generate an error fingerprint template set that includes spatial asymmetry index, main frequency energy concentration range, phase lag feature and thermo-coupling gradient feature.

[0015] S3: Input the multimodal monitoring data into a multi-branch feature encoder and output a real-time multidimensional feature stream containing spatial, frequency domain, and time-series feature vectors.

[0016] S4: Perform a similarity comparison operation based on the real-time multidimensional feature stream and the error source fingerprint template set to generate a preliminary matching score sequence corresponding to each candidate cause of core offset, material melt fluctuation, traction jitter and temperature gradient mismatch.

[0017] S5: Using the operating condition stability parameters, sensor health indicators and historical matching consistency statistics as prior validation evidence, perform Bayesian posterior probability update calculation on the preliminary matching score sequence to generate a set of candidate causal gating coefficients with values ​​ranging from 0 to 1.

[0018] S6: Determine whether there are any gate coefficients in the candidate cause gating coefficient set that are higher than the preset dynamic threshold. If there are, activate the corresponding correction sub-strategy and generate a target correction instruction. If there are no, generate a micro-amplitude same-direction compensation instruction with an amplitude not exceeding 15% of the current eccentricity as the target correction instruction.

[0019] S7: Drive the core displacement servo mechanism or temperature zone PID adjustment unit to perform physical actions according to the target correction command, and collect a closed-loop feedback evidence set containing the corrected eccentric residual, execution delay time and servo current fluctuation data after the action is completed.

[0020] S8: Based on the closed-loop feedback evidence set, perform an update operation on the prior distribution of the confidence of the corresponding template in the error source fingerprint template set to generate an updated error source fingerprint template set to complete the self-evolution correction of the control decision model.

[0021] The intelligent correction method for die core position based on cable insulation extrusion eccentricity detection provided in this application has the following beneficial effects: (1) By constructing a resilient decision flow of “detection-fingerprint matching-confidence assessment-gated response”, this scheme effectively overcomes the technical defects of the rigid link of “detection → attribution → execution” in traditional cable eccentricity control, which is prone to misjudgment, response lag and poor robustness under complex working conditions. Existing methods usually rely on static threshold judgment or fixed rule mapping, which is difficult to accurately separate the dominant cause when facing multi-source coupled disturbances (such as the coexistence of thermal field drift and mechanical vibration), resulting in mismatch between the correction action and the real error source, and even causing system oscillation. This invention establishes a structured error source fingerprint database, characterizing each typical cause of eccentricity (such as core offset, melt fluctuation, and traction jitter) as a quantifiable combination of multimodal features—including spatial asymmetry, dominant frequency energy distribution, cross-sensor phase hysteresis, and thermo-mechanical gradient curvature—achieving refined differentiation of complex fault modes. In the online stage, a lightweight multi-branch encoder is used to extract the spatial morphology, frequency domain stability, and temporal evolution features of the current eccentricity state in parallel, and perform rapid similarity comparison with the fingerprint template, significantly improving the accuracy and response speed of cause identification. Especially under cold start or transitional conditions, it can still maintain good matching performance, fundamentally improving the problem of ineffective intervention caused by fuzzy attribution in traditional methods.

[0022] (2) By introducing a Bayesian confidence gating mechanism, the dynamic conditionalization and risk controllability of control command activation are realized, achieving for the first time at the industrial field level a collaborative decoupling of error attribution uncertainty and execution robustness. Unlike the aggressive strategy of "execute as soon as it matches" in existing technologies, this solution innovatively designs a gating unit that comprehensively considers contextual factors such as the stability of the current operating conditions (e.g., temperature fluctuations, traction CV value), sensor health status (signal-to-noise ratio attenuation, lost pulse count), and historical matching consistency, dynamically generating the posterior confidence weight of each candidate cause. The corresponding correction sub-strategy is triggered only when the confidence level is higher than the adaptive threshold, thereby avoiding erroneous actions during periods of high data noise, unstable environment, or sensor anomalies. For low confidence scenarios, the system automatically switches to "steady-state maintenance mode" and outputs a limited-amplitude, same-direction compensation command (≤15% of the current eccentricity), ensuring continuous operation of the production line while preventing over-adjustment from introducing new disturbances. This mechanism significantly enhances the system's tolerance to uncertainty in real manufacturing environments, greatly reduces quality fluctuations and equipment wear caused by misjudgments, and maintains the simplicity of the algorithm structure and the feasibility of deployment without the need to introduce complex causal graph modeling, latent variable inference, or multi-model integration frameworks.

[0023] The synergistic effect of these two aspects enables this solution to not only possess high-precision attribution capabilities and robust anti-interference characteristics, but also to construct a closed-loop control system with self-awareness and continuous evolution capabilities: the feedback data after each execution (residual changes, delays, current fluctuations) is used to update the confidence prior distribution of the fingerprint template online, achieving incremental optimization of the knowledge base. Without relying on highly complex models such as graph neural networks, variational autoencoders, reinforcement learning, or meta-learning, the entire system, using a lightweight paradigm of "fingerprint recognition + confidence gating," completes the leap from passive response to proactive and prudent decision-making. It possesses excellent interpretability, real-time performance, and engineering portability, making it particularly suitable for cable insulation extrusion processes with extremely high requirements for safety and continuity. It provides a new control paradigm for high-end cable manufacturing that combines intelligence and reliability. Attached Figure Description

[0024] Figure 1 This is the main flowchart of a method for intelligent correction of die core position based on the detection of eccentricity in cable insulation extrusion.

[0025] Figure 2 This is a sub-flowchart of a method for intelligent correction of die core position based on the detection of eccentricity in cable insulation extrusion.

[0026] Figure 3 This is another sub-flowchart of the intelligent correction method for the die core position based on the detection of eccentricity in cable insulation extrusion. Detailed Implementation

[0027] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0028] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0029] like Figure 1 As shown, this application provides an intelligent correction method for the die core position based on the detection of cable insulation extrusion eccentricity, specifically including: S1: Acquire multimodal monitoring data during the cable insulation extrusion process. The multimodal monitoring data includes X-ray tomographic images, temperature distribution sequences, rotational speed fluctuation spectra, phase jitter signals, and pressure time series data.

[0030] S2: Based on the historical eccentricity sample annotation information, feature extraction is performed on the multimodal monitoring data to generate an error fingerprint template set that includes spatial asymmetry index, main frequency energy concentration range, phase lag feature and thermo-coupling gradient feature.

[0031] S3: Input the multimodal monitoring data into a multi-branch feature encoder and output a real-time multidimensional feature stream containing spatial, frequency domain, and time-series feature vectors.

[0032] S4: Perform a similarity comparison operation based on the real-time multidimensional feature stream and the error source fingerprint template set to generate a preliminary matching score sequence corresponding to each candidate cause of core offset, material melt fluctuation, traction jitter and temperature gradient mismatch.

[0033] S5: Using the operating condition stability parameters, sensor health indicators and historical matching consistency statistics as prior validation evidence, perform Bayesian posterior probability update calculation on the preliminary matching score sequence to generate a set of candidate causal gating coefficients with values ​​ranging from 0 to 1.

[0034] S6: Determine whether there are any gate coefficients in the candidate cause gating coefficient set that are higher than the preset dynamic threshold. If there are, activate the corresponding correction sub-strategy and generate a target correction instruction. If there are no, generate a micro-amplitude same-direction compensation instruction with an amplitude not exceeding 15% of the current eccentricity as the target correction instruction.

[0035] S7: Drive the core displacement servo mechanism or temperature zone PID adjustment unit to perform physical actions according to the target correction command, and collect a closed-loop feedback evidence set containing the corrected eccentric residual, execution delay time and servo current fluctuation data after the action is completed.

[0036] S8: Based on the closed-loop feedback evidence set, perform an update operation on the prior distribution of the confidence of the corresponding template in the error source fingerprint template set to generate an updated error source fingerprint template set to complete the self-evolution correction of the control decision model.

[0037] Step S1: Acquire multimodal monitoring data during the cable insulation extrusion process. This multimodal monitoring data includes X-ray tomography images, temperature distribution sequences, rotational speed fluctuation spectra, phase jitter signals, and pressure time-series data. Specifically, it includes: S1.1: The Compton scattering suppression processing is performed on the penetrating X-ray beam emitted by the high-precision X-ray tomography device to obtain a high signal-to-noise ratio projection raw data sequence characterizing the radial density distribution of the cable insulation layer, and the high signal-to-noise ratio projection raw data sequence is used as the basic input source for constructing the spatial morphological feature vector.

[0038] It should be noted that the pressure timing data is continuously acquired by a melt pressure sensor installed in the die head area at a fixed sampling frequency, and is used to characterize the dynamic fluctuation characteristics of melt pressure during the extrusion process.

[0039] For the penetrating X-ray beam emitted by the high-precision X-ray tomography device, based on the known geometric calibration parameters and detector response curve, the upper and lower limit boundary values ​​of the X-ray energy spectrum window are set, and the scattering components that exceed the energy spectrum range are identified and marked for subsequent rejection.

[0040] In the detector imaging plane, an energy discrimination operation is performed on the original charge accumulation signal of each pixel. By setting the Compton scattering characteristic energy threshold, the low-energy scattering component is distinguished from the direct penetration component, thus achieving the initial separation of the scattering signal.

[0041] After initial separation, an adaptive background estimation process based on polynomial fitting is performed on the low-energy scattering components. The estimated background components are subtracted from the original pixel signal pixel by pixel to eliminate the non-uniform attenuation effect introduced by scattering.

[0042] For the direct-penetration signal sequence after background subtraction, a two-dimensional Gaussian filter operator is used for spatial domain smoothing to suppress random fluctuations in detector noise and scattering residue, resulting in a projection matrix with stable signal amplitude.

[0043] The signal-to-noise ratio (SNR) is calculated on the projection matrix using the following formula: in, To reach the average value of the penetrating signal, To determine the standard deviation of the direct-penetrating signal, the signal-to-noise ratio is used to judge the data quality and select high-quality frame sequences.

[0044] The output high signal-to-noise ratio projected raw data sequence serves as the basic input source for constructing spatial morphological feature vectors, enabling accurate representation of radial density distribution in subsequent feature extraction processes.

[0045] For example, in a medium- and high-voltage cable extrusion production line, a high-precision X-ray tomography device with an energy resolution of 2keV is used. The energy spectrum range of the penetrating X-ray beam is set to 60keV to 120keV, and the Compton scattering discrimination threshold is set to 65keV. Through energy discrimination of detector pixels, the proportion of low-energy scattering components in each frame of the image is reduced to less than 5% after removal. Background estimation uses a third-order polynomial fitting window of 7×7 pixels. After removing the background, the signal mean is increased to 1.2 times the original value. The two-dimensional Gaussian filter radius is 2 pixels, and the variance is 1.5. After smoothing, the signal-to-noise ratio is increased from the original 18 to over 40. In the signal-to-noise ratio calculation, the mean S is taken from the global mean of the projection matrix, and the standard deviation N is taken from the global standard deviation of the projection matrix. The calculated signal-to-noise ratio value is significantly improved, ensuring that the projection data can accurately reflect the radial density distribution of the insulation layer in the subsequent spatial morphology feature vector construction process, effectively supporting the generation of core offset state detection and correction strategies.

[0046] S1.2: Based on the high signal-to-noise ratio projection original data sequence, a filtered back-projection reconstruction algorithm is executed to generate a two-dimensional grayscale tomographic image containing the geometric center coordinate deviation information of the mold core, and the two-dimensional grayscale tomographic image is used as the direct basis for calculating the spatial asymmetry index.

[0047] For the high signal-to-noise ratio (SNR) projection raw data sequence input, the preprocessing module of the filtered back-projection reconstruction algorithm is invoked to perform convolution filtering on the ray attenuation data of each projection viewpoint in the sequence. The frequency response characteristics of the filter kernel are configured according to the preset bandpass enhancement and low-frequency suppression principles to reduce low-frequency drift and high-frequency noise distortion in the projection data. Geometric back-projection interpolation is performed on the convolutionally filtered data using a set of projection geometric parameters (including detector unit center coordinates, ray beam emission angle sequence, source-detector distance, etc.). The filtering results of each projection viewpoint are then superimposed onto the voxel array of the reconstruction plane according to the corresponding ray path to form a preliminary reconstructed grayscale matrix. Iterative back-projection residual correction is performed on the grayscale matrix to calculate the residual matrix between the reconstructed result and the predicted value of the original projection data. Based on this, Laplacian constraint pixel-by-pixel correction is performed in the voxel grayscale space to improve the reconstruction accuracy of the geometric center coordinate details of the model core. Following the principle of matching spatial resolution with grayscale dynamic range, the matrix after residual correction is subjected to two-dimensional grayscale normalization and mean filtering smoothing to obtain a two-dimensional tomographic image with uniform grayscale distribution. This image contains the coordinate deviation information of the geometric center of the core, serving as a direct basis for calculating the spatial asymmetry index. Through filtered back projection and residual iterative correction, the high signal-to-noise ratio projection raw data sequence output from the previous step S1.1 is transformed into a two-dimensional grayscale tomographic image that meets the requirements for spatial morphological feature extraction, achieving high-precision visualization of the eccentricity state of the cable insulation layer cross-section.

[0048] S1.3: Perform cold-end compensation and linearization correction on the analog voltage signal output by the mold temperature sensor array to generate a temperature distribution sequence that reflects the thermal field distribution of the melt, and use the temperature distribution sequence as a key variable for extracting the thermo-coupling gradient characteristics.

[0049] S1.4: The current waveform of the screw drive motor and the pulse signal of the traction wheel encoder are subjected to spectral decomposition processing using the Fast Fourier Transform algorithm to generate a speed fluctuation spectrum characterizing plasticization stability and a phase jitter signal characterizing mechanical transmission smoothness, respectively. The speed fluctuation spectrum and the phase jitter signal are used together as the constituent elements of the frequency domain stability feature vector.

[0050] S1.5: Based on the unified timestamp protocol, the generated two-dimensional grayscale tomographic scan image, temperature distribution sequence, rotation speed fluctuation spectrum, phase jitter signal, and time series data output by the melt pressure sensor are subjected to multi-source spatiotemporal alignment and fusion processing to generate a multimodal monitoring data packet with strict causal relationship, and the multimodal monitoring data packet is used as the standard input object of the subsequent multi-branch feature encoder.

[0051] For the generated 2D grayscale tomographic images, temperature distribution sequences, rotational speed fluctuation spectra, phase jitter signals, and time-series data output from the melt pressure sensor, the unified timestamp protocol parsing unit is invoked to extract the original acquisition timestamps of each data file. These timestamps are then converted to a unified high-precision time base format to eliminate the heterogeneous effects of cross-device sampling delays. Time base drift compensation is performed on the converted timestamps. Based on the offset between the reference clock source and the local clocks of each sensor, a linear interpolation correction method is used to synchronize the time series. The adjusted timestamps and corresponding data are indexed and reconstructed. The absolute acquisition time of each frame in the 2D grayscale tomographic images is located. The global time coordinates of each temperature sample are calibrated in the temperature distribution sequence. Frequency domain components are correlated with the global time in the rotational speed fluctuation spectrum and phase jitter signals. A mapping relationship between pressure values ​​and a unified time axis is established in the pressure time-series data. A time-series registration algorithm is used to align the above multi-source data frame by frame along the unified time axis, ensuring the consistency of the physical state of each data source at the same time node. The aligned multi-source data undergoes spatial dimension fusion processing. Using a multimodal data fusion operator based on feature matrix concatenation, the spatial morphological features of the two-dimensional grayscale tomographic image, the thermal field features of the die temperature distribution, the frequency domain stability features of the rotational speed fluctuation spectrum and phase jitter signal, and the pressure waveform features of the melt pressure time series are combined into a unified multimodal monitoring data package. Through this multi-source spatiotemporal alignment and fusion processing, the dispersed data sources from the previous step are transformed into multimodal monitoring data packages with strict causal relationships, enabling the construction of the standard input object for the subsequent multi-branch feature encoder.

[0052] For example, on a medium- and high-voltage cable production line, the sampling rate of the two-dimensional grayscale tomographic scan image is 30 frames per second, the sampling period of the die temperature sensor array is 200ms, the rotational speed fluctuation spectrum is sampled at 512 points per second based on the current signal, the phase jitter signal sampling period is 50ms, and the sampling frequency of the melt pressure sensor timing data is 10Hz. A unified timestamp protocol converts the local time stamps of each data format into nanosecond-level global time, and corrects the maximum offset of 13ms between the reference clock and the sensor clock through time base drift compensation. After interpolation correction, a mapping is established on the global time axis to ensure that each group of temperature samples is precisely aligned with the corresponding scanned image frame and pressure data. The timing registration algorithm uses a nearest-neighbor mapping method to match all data sources on a unified time axis, ensuring the consistency of multi-source data at the same nanosecond-level time point. During spatial dimension fusion, the cross-sectional contour matrix of the scanned image and the thermal field matrix of the temperature distribution sequence are concatenated column-wise and merged with the frequency domain feature vector and pressure waveform features to form a fused feature matrix of size (30×N), where N is the total feature dimension after fusion. This multimodal monitoring data packet is fed as standard input into the multi-branch feature encoder, resulting in significantly improved eccentricity detection response speed and accuracy in subsequent runs.

[0053] Step S2: Based on historical eccentricity sample annotation information, feature extraction is performed on the multimodal monitoring data to generate an error fingerprint template set containing spatial asymmetry indicators, dominant frequency energy concentration intervals, phase lag characteristics, and thermo-coupling gradient characteristics. Specifically, this includes: S2.1: Based on the X-ray tomographic image sequence in the historical eccentricity sample annotation information with known historical causes, the point cloud data of the insulation layer cross-sectional contour edge is processed by least squares circle fitting and centroid deviation calculation to extract the geometric center coordinate deviation, and the geometric center coordinate deviation is used as the basic input variable for generating the spatial asymmetry index.

[0054] It should be noted that the historical eccentricity sample annotation information refers to the multimodal monitoring data collected in historical production batches, in which process experts or offline detection systems annotate the dominant eccentricity cause type of each sample frame. The cause types include core offset, material melt fluctuation, traction jitter and temperature gradient mismatch, and the corresponding eccentricity level (light, medium and heavy) and the qualified correction effect mark are also annotated.

[0055] Based on high-precision X-ray tomographic image sequences from historically known eccentricity sample annotation information, a contour extraction operator is used to separate the edge pixel set of the insulating layer cross-section from each frame of the image, obtaining an edge point cloud data matrix representing the radial morphology. Noise removal and coordinate normalization are performed on the edge point cloud data matrix to remove isolated points and abnormal deviations, and the remaining point cloud is mapped to a normalized Cartesian coordinate system to ensure the stability and comparability of the fitting operation. A least-squares circle fitting algorithm is used to calculate the center coordinates and radius of the best-fit circle in the normalized point cloud coordinate system, and an objective function is constructed. in For the number of point clouds, The coordinates of the edge points, To fit the coordinates of the circle center, Given the distance from a point to the center of a circle, iteratively minimize... The fitting parameters are obtained. Based on the coordinates of the fitted circle's center and the overall centroid coordinates of the image, the deviation vector is calculated using the formula... in For the centroid coordinate components, To fit the coordinate components of the circle center, the geometric center coordinate deviation is output. Temporal statistical analysis is performed on this deviation, calculating the multi-frame average and standard deviation to form a stability correction coefficient to suppress the influence of transient noise. The stability-corrected deviation is used as the basic input variable for generating a spatial asymmetry index, achieving a quantitative characterization of the mechanical offset of the mold core.

[0056] The above processing method converts the two-dimensional grayscale tomographic image from the previous step into geometric center coordinate deviation data, thereby achieving a high-precision quantitative description of the core center offset state.

[0057] For example, in a certain high-voltage cable production batch, the historical eccentricity sample annotation information includes 150 frames of high-precision X-ray tomographic images, each with a resolution of 1024×1024 pixels. The edge extraction threshold is set to 0.75 gray-level normalization units, and outliers with a radius deviation exceeding 2% are removed. The least-squares circle fitting iteration count is set to 500 steps, with a convergence tolerance of 1×10⁻⁶. 6The average center coordinates of the circle are (512.4, 511.8), and the centroid coordinates of the image are (515.0, 512.5). The deviation is calculated using the above formula, yielding 2.65 pixels. After multi-frame statistical correction, the deviation stabilizes at 2.60 pixels, corresponding to a spatial asymmetry index quantization value of 0.00255 (normalized relative to the radius). This quantization value is assigned to the fingerprint template of the core mechanical offset error source, providing a high-confidence reference standard for subsequent similarity comparison.

[0058] S2.2: The Fast Fourier Transform algorithm is used to perform spectral energy density integration on the rotational speed fluctuation spectrum and pressure time series data in the historical samples to identify and extract the dominant frequency components and their bandwidth ranges that characterize the instability of material flow, thereby obtaining well-defined dominant frequency energy concentration interval parameters, and using the dominant frequency energy concentration interval parameters as the core elements for constructing the frequency domain stability feature vector.

[0059] S2.3: Based on the edge point cloud data and phase jitter signal of the X-ray tomographic scan image synchronously acquired in the historical samples, the peak position of the time lag between the two is calculated by the cross-correlation analysis algorithm, so as to quantify the cross-sensor phase lag time value reflecting the mechanical transmission coupling effect, and the cross-sensor phase lag time value is used as the key intermediate variable for deriving the cross-sensor phase lag characteristics.

[0060] Time series preprocessing is performed on the edge point cloud data and phase jitter signal of the X-ray tomographic scan image synchronously acquired in the annotation information of historical eccentricity samples with known causes. The raw data output by different sensors are aligned under a unified time reference to eliminate sampling clock differences.

[0061] The radial thickness sequence of aligned edge point clouds and the phase jitter sequence of traction wheels are normalized to ensure that the dimensions of each input quantity are consistent during the cross-correlation calculation and to avoid the bias of the correlation coefficient caused by differences in spatial scale or voltage amplitude.

[0062] The core function of the cross-correlation analysis algorithm is called, with the normalized edge point cloud sequence as the reference signal input and the normalized traction wheel phase jitter sequence as the target signal input. The correlation coefficient curve of the two sequences is calculated point by point through a sliding window.

[0063] Perform a maximum value search on the correlation coefficient curve, mark the time offset corresponding to the peak value of the correlation coefficient on the curve, and calculate the position of the time lag peak using the following formula: in, Represents the edge point cloud sequence. This indicates the traction wheel phase jitter sequence. For cross-correlation function, This represents the time lag.

[0064] The calculated peak time lag position is converted from the sampling unit to the actual time unit millisecond, and the result is defined as the cross-sensor phase lag time value.

[0065] Through the above processing method, the signal characteristics of the previous step are transformed into quantitative cross-sensor phase lag time characteristic data, thereby realizing the explicit quantification of the mechanical transmission coupling effect.

[0066] For example, in a sample set of a high-voltage cable extrusion production line, the sampling frequency of the X-ray tomography system is set to 200Hz, and the sampling frequency of the traction wheel encoder signal is set to 200Hz. After alignment with a unified time reference, the length of both sequences is 4000 points. The edge point cloud thickness value is normalized to the range of 0 to 1, and the traction wheel phase jitter signal is normalized to... Within the range of 1 to 1, the peak value of the correlation coefficient curve obtained by the cross-correlation operation is 0.92, corresponding to a sampling point offset of 15 points. The peak position of the time lag is calculated using the formula, and the cross-sensor phase lag time value is obtained by multiplying the sampling point offset by the sampling period of 0.005 seconds, resulting in a value of 0.075 seconds. This result, used as a fingerprint template set calibration process for cross-sensor phase lag feature input error sources, significantly improves the accurate description of the mechanical transmission coupling effect and demonstrates a high degree of consistency in lag time values ​​across multiple batches of samples in repeated testing.

[0067] S2.4: Construct a thermo-coupled gradient matrix by combining the temperature distribution sequence and phase jitter signal in historical samples. Perform second-order partial derivative curvature estimation processing on the matrix to derive the thermo-coupled gradient curvature scalar that characterizes the intensity of process environment disturbance. Use the thermo-coupled gradient curvature scalar as a specific fingerprint feature to characterize the cause of temperature gradient mismatch.

[0068] S2.5: Integrate the spatial asymmetry index, the main frequency energy concentration interval parameter, the cross-sensor phase lag characteristics, and the thermo-coupling gradient curvature scalar. Perform multi-dimensional feature vector encapsulation and standardized mapping processing according to four known cause labels: core offset, material melt fluctuation, traction jitter, and temperature gradient mismatch, to generate a structured error source fingerprint template set. The error source fingerprint template set is then used as the sole standard reference library for subsequent real-time multi-dimensional feature flow similarity comparison.

[0069] The spatial asymmetry index, main frequency energy concentration interval parameter, cross-sensor phase lag characteristics and thermo-coupling gradient curvature scalar from steps S2.1 to S2.4 are subjected to feature type grouping processing. The various indices are established into corresponding feature sets according to four known cause labels: core offset, material melt fluctuation, traction jitter and temperature gradient mismatch.

[0070] For each type of feature set, a multidimensional vector encapsulation operation is used, placing the spatial asymmetry index in the first dimension, the main frequency energy concentration interval parameter in the second dimension, the cross-sensor phase lag feature in the third dimension, and the thermo-coupling gradient curvature scalar in the fourth dimension, forming a multidimensional feature vector with a fixed dimensional order.

[0071] The encapsulated multidimensional feature vectors are subjected to standardization mapping. The mean and standard deviation of each feature dimension under the historical sample benchmark are calculated, and numerical mapping is performed using the following standardization formula: in, These are the original eigenvalues. This is the historical average. The historical standard deviation ensures that features of different dimensions are comparable on the same scale.

[0072] Label binding is performed on the multidimensional feature vectors after normalization mapping, generating key-value pair structures with the corresponding causal category labels, and stored in the data table of the error source fingerprint template set.

[0073] Perform a consistency check operation on the set of error source fingerprint templates to verify the consistency of dimension order and storage format of each template under the same label, so as to ensure that the matching operation during subsequent real-time feature stream similarity comparison can run under a unified benchmark.

[0074] Through the above grouping, encapsulation, standardization mapping and label binding processing methods, the results of the previous step are transformed into a structured set of error source fingerprint templates with a uniform scale, realizing a standard reference for multi-causal feature patterns.

[0075] like Figure 2 As shown, step S3 involves inputting the multimodal monitoring data into a multi-branch feature encoder and outputting a real-time multidimensional feature stream containing spatial, frequency domain, and temporal feature vectors. Specifically, this includes: S3.1: Perform edge contour point cloud extraction processing based on X-ray tomographic images to obtain a thermal map of the radial thickness deviation of the insulation layer as the basic data object for spatial morphology analysis.

[0076] It should be noted that the multi-branch feature encoder refers to a parallel encoding structure that includes a convolutional neural network branch, a frequency domain transform branch, and a time-series convolutional network branch, which respectively perform feature extraction on the image, spectrum, and sequence data in the multimodal monitoring data, and fuse the outputs of each branch through an attention mechanism.

[0077] S3.2: The fast Fourier transform algorithm is used to perform spectral decomposition on the thermal map of the radial thickness deviation of the insulating layer to generate a vector of the main frequency energy concentration interval that characterizes the energy distribution of the eccentric fluctuation.

[0078] Two-dimensional discrete resampling processing is performed on the extracted radial thickness deviation heatmap of the insulation layer. The original image matrix is ​​uniformly quantized in both the radial and angular directions to ensure that the input data for the spectral decomposition operation has uniform resolution and phase consistency.

[0079] A window function is applied to the thickness deviation matrix after resampling. The Hanning window or Blackman window is selected to suppress spectral leakage. The window length is equal to the number of radial sampling points and the sliding step size is set according to the process acquisition cycle to achieve stable extraction of frequency domain components.

[0080] A Fast Fourier Transform (FFT) operation is performed radially on the matrix after window function processing. Using an FFT algorithm based on the Cooley-Tukey splitting strategy, the spatial distribution pattern of the thickness deviation is converted into a frequency domain energy spectrum, as shown in the following formula: in This is a radial thickness deviation sequence. For frequency index, The total number of sampling points. For the first Complex values ​​of frequency components.

[0081] Perform amplitude calculations on the obtained complex spectrum to convert each frequency component into an energy value. The energy of each frequency component is calculated and normalized to the interval [0,1] to eliminate the influence of absolute amplitude differences.

[0082] The normalized energy spectrum is processed by extracting the dominant frequency energy concentration interval. Based on the cumulative integral curve of the energy spectrum, the start and end frequency indices of the cumulative energy value reaching the preset ratio are determined, forming a dominant frequency energy concentration interval vector that characterizes the eccentric fluctuation energy distribution.

[0083] By using the above-mentioned FFT spectrum decomposition, energy normalization and interval extraction processing methods, the radial thickness deviation heatmap of the insulation layer in the previous step is converted into a main frequency energy concentration interval vector that reflects the eccentric fluctuation characteristics, thereby realizing the quantitative expression of the frequency domain characteristics of the eccentric fluctuation.

[0084] For example, in a 110kV medium-voltage cable extrusion production line, 256-point radial resampling and 128-point angular resampling were performed on the thermal map of the radial thickness deviation of the insulation layer with a collection period of 0.5 seconds. The Hanning window function was selected, and the spectral leakage suppression factor was set to 0.5. The FFT transformation adopted the divide-and-conquer method, with a transformation scale of 256 points and a frequency resolution of 1.95Hz. After the spectral amplitude was squared to obtain the energy spectrum, it was normalized, and the cumulative energy ratio threshold was set to 0.75. The start and end frequencies when the cumulative energy reached 75% were extracted as 12Hz and 24Hz, respectively, forming the main frequency energy concentration interval vector [12,24]. This interval vector was directly used to construct the frequency domain stability features of the real-time multidimensional feature flow. During the test, the detection effect of the eccentric fluctuation energy distribution was significantly improved, and the frequency interval positioning was accurate and stable, supporting the efficient execution of subsequent similarity comparison and Bayesian confidence gating.

[0085] S3.3: Perform cross-correlation time delay estimation calculation based on the rotation speed fluctuation spectrum and pressure time series data to output a cross-sensor phase hysteresis characteristic sequence reflecting the material flow stability.

[0086] For the synchronous signal input consisting of the speed fluctuation spectrum and pressure time series data, the cross-correlation function calculation module is called to perform dual-signal normalization preprocessing, uniformly mapping the original amplitude range to the [-1,1] interval to eliminate the interference of dimensional differences on time delay estimation. Based on the normalized signal sequence, time window vectors of consistent length are constructed, and convolutional multiplication is performed on each pair of corresponding time windows to obtain the cross-correlation cumulative value sequence that changes with time shift. The cross-correlation cumulative value sequence is subjected to maximum value search processing, and the time shift corresponding to the peak position is extracted as the preliminary time delay estimate. The cross-correlation cumulative value curve near the peak is locally smoothed using the interpolation polynomial fitting method, and the corresponding time shift is read at the maximum value of the fitted curve to obtain the optimized high-precision time delay value. Based on the optimized time delay value, a cross-sensor phase lag feature sequence is constructed, where each feature element represents the phase difference between the material flow characteristics under the current operating conditions and the screw drive and melt pressure response. The cross-correlation function is calculated using the following mathematical expression: in, The time window length, Let be the amplitude of the rotational speed fluctuation spectrum at time t. For pressure time series data at time amplitude, These are the means of the two sequences, respectively. By using cross-correlation smoothing and peak localization, the frequency and time domain signal processing results from the previous step are transformed into quantitative cross-sensor phase hysteresis characteristics, thereby achieving accurate characterization of material flow stability.

[0087] For example, in a medium- and high-voltage cable extrusion production line, the sampling frequency of the rotation speed fluctuation spectrum is set to 200Hz, the sampling frequency of the melt pressure sensor is 200Hz and synchronized with a unified clock, and cross-correlation calculation is performed on a normalized signal sequence of 1000 points. The peak position is found to be at the 46th sampling point, and the time delay after interpolation fitting optimization is 0.225s. The cross-sensor phase lag characteristic sequence corresponding to this time delay maintains a stable fluctuation range of ±0.005s within the current batch production cycle. When mapped to the material flow stability index, it exhibits a high stability state. Combined with the known stability threshold in the fingerprint database, a high-confidence identification of the material flow cause is achieved, significantly improving the response accuracy of the control decision module under this operating condition.

[0088] S3.4: Combine the temperature distribution sequence and the phase jitter signal to construct a thermo-coupled gradient matrix, and derive the thermo-coupled gradient characteristic scalar that characterizes the intensity of process environment disturbance.

[0089] The temperature distribution sequence and the phase jitter signal are resampled at the same time scale to ensure that the two types of data can be matrix-combined under the same sampling frequency and time alignment.

[0090] The normalized temperature distribution sequence is used as the thermal field data component, and the same normalized phase jitter signal is used as the mechanical disturbance component. They are assembled in a row-indexed manner to construct a two-dimensional matrix containing the thermo-mechanical coupling state at multiple time points.

[0091] After constructing the matrix, gradient calculations are performed for each spatial location and time segment to obtain the thermal field gradient and force field gradient, and then the two are concatenated in a column vector manner to form a gradient coupling matrix.

[0092] Perform second-order partial derivative operations on each element of the gradient coupling matrix to obtain the curvature distribution matrix, where the curvature calculation formula is: in These are the element values ​​in the thermal coupling matrix. The coordinates are the spatial location coordinates. The curvature distribution matrix is ​​aggregated along the spatial location dimension to obtain a single scalar value as the thermally coupled gradient curvature feature scalar.

[0093] Through the matrix construction, gradient extraction and curvature aggregation processing methods described above, the die temperature distribution and phase jitter signal from the previous step are transformed into thermo-coupling gradient curvature feature scalars that characterize the intensity of process environment disturbances, thereby realizing real-time plotting of the causes of temperature gradient mismatch.

[0094] For example, the sampling frequency of the temperature distribution sequence is set to 200Hz, the sampling frequency of the phase jitter signal is resampled to 200Hz via interpolation, and the normalization range is set to []. [1,1]. The two-dimensional thermo-coupling matrix has 1000 rows. Gradient calculation is performed on each matrix element. The thermal field gradient uses the central difference method, and the force field gradient uses the backward difference method. The curvature distribution matrix performs second-order partial derivatives on each row, and the resulting curvature values ​​fluctuate between 0.05 and 0.15. Finally, after spatial mean aggregation, the thermo-coupling gradient curvature feature scalar is obtained as 0.092. This scalar is input into the subsequent multi-source data fusion and encapsulation module, which effectively improves the temperature gradient mismatch identification capability in similarity comparison and significantly improves the posterior probability calculation results of this cause in the Bayesian confidence gating module, thereby improving the targeting of the control decision strategy.

[0095] S3.5: Integrate the spatial morphological feature vector, frequency domain stability feature vector, temporal evolution feature vector, and thermally coupled gradient feature scalar to perform multi-source data fusion and encapsulation operations to generate a real-time multidimensional feature stream for similarity comparison calculations.

[0096] It should be noted that the real-time multidimensional feature stream is a structured feature vector sequence with fixed dimensions. Each frame's feature vector can be concatenated by spatial morphology feature sub-vectors, frequency domain stability feature sub-vectors, and temporal evolution feature sub-vectors according to the channel dimension. The spatial morphology feature sub-vector has a length of M, the frequency domain stability feature sub-vector has a length of N, and the temporal evolution feature sub-vector has a length of L, for a total dimension of M+N+L. This feature stream is output frame by frame at fixed time intervals (e.g., 0.5 seconds) as input for subsequent similarity comparisons. The spatial morphology feature vector in the real-time multidimensional feature stream includes a "current eccentricity" field. This field is directly obtained by processing X-ray tomographic images using a geometric morphology analysis algorithm. Its value is equal to the Euclidean distance between the geometric center of the insulating layer cross-section and the geometric center of the mold core, expressed in millimeters or pixels.

[0097] like Figure 3 As shown, step S4 involves performing a similarity comparison operation based on the real-time multidimensional feature stream and the error source fingerprint template set to generate preliminary matching score sequences corresponding to each candidate cause, including core offset, material melt fluctuation, traction jitter, and temperature gradient mismatch. Specifically, this includes: S4.1: Based on the spatial morphological feature vector in the real-time multidimensional feature stream, call the pre-constructed spatial asymmetry mapping operator to perform geometric topological projection processing to generate a real-time spatial asymmetry index sequence that characterizes the degree of distortion of the current insulation layer cross-sectional profile.

[0098] S4.2: Using the frequency domain stability feature vector in the real-time multidimensional feature flow, and combining the fast Fourier transform algorithm to perform spectral energy density integration on the rotational speed fluctuation spectrum and pressure time series data, in order to extract the real-time main frequency energy concentration interval parameter characterizing the material flow instability.

[0099] Data parsing is performed on the frequency domain stability feature vector in the real-time multidimensional feature stream to extract the frequency domain components containing the speed fluctuation spectrum and pressure time series data as input objects.

[0100] The Fast Fourier Transform algorithm is applied to the rotational speed fluctuation spectrum to map the sampled time-domain signal to the frequency-domain spectrum, thereby achieving the separation of the amplitude and phase of each frequency component.

[0101] The fast Fourier transform algorithm is performed on the pressure time series data to generate a frequency domain distribution curve that reflects the dynamic pressure changes inside the material melt, while maintaining the same frequency resolution as the rotation speed fluctuation spectrum.

[0102] The rotational speed fluctuation spectrum and the melt pressure frequency spectrum are registered along the frequency axis to construct a matching matrix containing the correspondence between the amplitudes of the two frequency domain signals, so as to perform synchronization in the integral calculation.

[0103] Perform energy density integration on the registered frequency domain matching matrix to calculate the energy concentration of material flow instability in the main frequency region, where the integration formula is: in Energy concentration This is the amplitude function of the rotational speed fluctuation spectrum. This is a function of the frequency spectrum amplitude of the melt pressure. and It is the boundary of the main frequency integration range.

[0104] Based on the energy concentration results, parameters of the real-time main frequency energy concentration range are extracted. These parameters include the main frequency value, bandwidth range, and corresponding average energy density.

[0105] By using the spectral energy density integral processing method, the frequency domain stability feature vector from the previous step is transformed into a real-time dominant frequency energy concentration interval parameter that can directly characterize the instability of material flow, thereby enabling quantitative analysis of the material flow state.

[0106] For example, in the extrusion production process of a high-voltage cable insulation layer, the sampling frequency of the rotational speed fluctuation spectrum is set to 500Hz, and the sampling frequency of the pressure time sequence data is set to 500Hz, both with a frequency resolution of 0.5Hz. After performing a Fast Fourier Transform on the screw drive motor current sampling signal, the dominant frequency peak appears in the range of 8Hz to 12Hz, with an average amplitude of 2.3A. After performing a Fast Fourier Transform on the melt pressure sensor sampling signal, the dominant frequency peak of the pressure fluctuation spectrum is in the range of 8Hz to 12Hz, with an average amplitude of 1.8MPa. A matching matrix is ​​constructed through frequency axis registration. For 8Hz, Perform integration calculations over the 12Hz interval, where A(f) and B(f) are the amplitudes at the corresponding frequencies, and the integration result E is: The integrated value, after normalization, yielded an energy concentration of 2.07, corresponding to the real-time dominant frequency energy concentration interval parameters (dominant frequency = 10Hz, bandwidth = 4Hz, mean energy density = 2.07). This result was used for subsequent multidimensional feature similarity comparison, significantly improving the accuracy of attributing material flow instability.

[0107] S4.3: Based on the temporal evolution feature vector in the real-time multidimensional feature stream, the cross-correlation analysis algorithm is used to calculate the time lag between the edge point cloud of the X-ray tomographic image and the phase jitter signal, so as to output the real-time cross-sensor phase lag feature value that characterizes the mechanical transmission coupling effect.

[0108] Based on the temporal evolution feature vector in the real-time multidimensional feature stream, the edge point cloud sequence and phase jitter signal of the high-precision X-ray tomography image are indexed and matched under a unified time reference to ensure accurate correspondence of timestamps output by different sensors. Interpolation resampling is performed on the matched edge point cloud sequence to ensure that its sampling frequency is numerically identical to that of the phase jitter signal, thereby eliminating time axis deviation caused by sampling rate differences. A bandpass filtering algorithm is used to preprocess the two sets of signals, filtering out low-frequency drift and high-frequency noise components unrelated to mechanical transmission errors, making the time delay characteristics reflected by the cross-correlation analysis more stable. Normalization is performed on the preprocessed signals, mapping their amplitude range to the [-1, 1] interval to eliminate the influence of dynamic response differences between different sensors on the correlation calculation. The cross-correlation analysis algorithm is called to calculate the full-range cross-correlation function of the two sets of signals, and the x-coordinate position of the point with the maximum absolute value of the cross-correlation function is identified as the time lag, where the formula for calculating the lag is: in, Input signal and The cross-correlation function, The index of the point with the largest absolute value. To standardize the sampling period, the time lag is used as a real-time cross-sensor phase lag characteristic value to characterize the mechanical transmission coupling effect. This characteristic value is then encapsulated into a comprehensive feature set for subsequent multi-dimensional spatial proximity calculations. Through cross-correlation analysis and time lag extraction, the results of the previous step are transformed into quantitative indicators that reflect the dynamic coupling relationship between the traction system and the extrusion system, enabling real-time identification and location of mechanical transmission error sources.

[0109] For example, on a medium-voltage cable extrusion production line, the sampling frequency of the edge point cloud signal generated by the X-ray tomography unit is set to 200Hz, and the sampling frequency of the phase jitter signal is adjusted to 200Hz through interpolation resampling to ensure consistency between the two. A 1Hz to 50Hz bandpass filter is applied to the edge point cloud signal to filter out thermal field drift below 1Hz and X-ray noise above 50Hz. The same filter is performed on the encoder jitter signal. After normalization, the maximum value of the cross-correlation function appears at index position 12, and the sampling period Ts is 0.005 seconds, according to the formula... The calculated time lag was 0.06 seconds. This lag was encapsulated into the real-time integrated feature set and compared with the historical mechanical transmission coupled fingerprint template in subsequent matching. The output proximity was significantly improved, indicating that the accuracy and stability of the real-time identification of the mechanical transmission error source were greatly improved.

[0110] S4.4: For the comprehensive feature set consisting of the real-time spatial asymmetry index sequence, the real-time main frequency energy concentration interval parameter, and the real-time cross-sensor phase lag characteristic value, the weighted Euclidean distance metric algorithm is used to perform multi-dimensional spatial proximity calculation with the four types of templates in the error source fingerprint template set, namely core offset, material melt fluctuation, traction jitter and temperature gradient mismatch, to generate a preliminary matching score sequence corresponding to each candidate cause.

[0111] Based on the real-time spatial asymmetry index sequence, real-time main frequency energy concentration interval parameters, and real-time cross-sensor phase lag characteristic values ​​obtained through the preceding sub-steps S4.1 to S4.3, a comprehensive feature set containing three dimensions is constructed as the input object for similarity calculation.

[0112] Feature mapping relationships are established for the four types of causal templates in the comprehensive feature set and the error source fingerprint template set, respectively, to ensure that the feature vectors of each type of template are completely consistent with the comprehensive feature set in terms of dimensional order and unit dimension, so as to eliminate scale differences.

[0113] Weighted processing is performed on the feature vectors of each type of template. The weight of the spatial asymmetry index is set based on its contribution to the historical attribution accuracy statistics. The weight of the main frequency energy concentration interval is set based on the sensitivity of material flow stability discrimination. The weight of the cross-sensor phase lag feature value is set based on the importance of mechanical transmission coupling effect identification.

[0114] A weighted Euclidean distance metric algorithm is used to perform multidimensional spatial proximity calculations between the comprehensive feature set and the feature vectors of each template. The distance calculation for each type of template is performed according to the following formula: in For the first Weights of dimensional features For the first comprehensive feature set 3D real-time feature values, For the template vector corresponding to the first Dimensional eigenvalues.

[0115] The weighted Euclidean distances calculated for all templates are normalized and mapped using the inverse of the distance value to obtain a preliminary matching score sequence for each candidate cause. This inverse mapping is used to emphasize the proportion of templates with smaller distances (higher similarity) in the score.

[0116] By using weighted Euclidean distance metric and normalized score processing, the real-time feature stream from the previous step is transformed into preliminary matching score data that can be used for Bayesian posterior probability updates, thereby achieving a quantitative expression of the similarity of each candidate cause.

[0117] For example, in the operating condition of a medium-voltage cable extrusion production line, the real-time spatial asymmetry index is quantized to 0.35, the real-time main frequency energy concentration interval parameter is quantized to 125Hz, and the real-time cross-sensor phase lag characteristic value is quantized to 0.012s. The weights are configured as follows: spatial asymmetry index weight 0.5, main frequency energy concentration interval weight 0.3, and cross-sensor phase lag characteristic weight 0.2. For the die core offset template, its spatial asymmetry index value is 0.32, the main frequency energy concentration interval value is 120Hz, and the cross-sensor phase lag value is 0.010s. Substituting these values ​​into the formula: The calculation results show that the template has a high proximity to the comprehensive feature set in the multidimensional space. Through distance inverse normalization, this factor accounts for a significant proportion in the initial matching score, which verifies the effectiveness of weighted Euclidean distance in the quantification of multidimensional feature similarity.

[0118] It should be noted that the similarity comparison operation can use any of the following measurement methods: weighted Euclidean distance, Mahalanobis distance, or cosine similarity. In this embodiment, weighted Euclidean distance is used as an example for explanation.

[0119] Step S5: Using operating condition stability parameters, sensor health indicators, and historical matching consistency statistics as prior validation evidence, perform Bayesian posterior probability update calculation on the preliminary matching score sequence to generate a set of candidate causal gating coefficients with values ​​ranging from 0 to 1. Specifically, this includes: S5.1: Obtain the standard deviation of temperature fluctuation and the coefficient of variation of traction speed from the real-time multimodal monitoring data as the raw data of operating condition stability, and perform normalization mapping processing on the raw data of operating condition stability to generate a quantitative vector of operating condition stability parameters that characterizes the current smoothness of the extrusion process.

[0120] It should be noted that the standard deviation of temperature fluctuation can be calculated as follows: for the temperature values ​​of each measuring point of the mold temperature sensor array within a preset time window (such as 60 seconds), first calculate the standard deviation of each measuring point, and then take the average of the standard deviations of all measuring points; the coefficient of variation of traction speed is calculated as follows: the standard deviation of the pulse frequency of the traction wheel encoder divided by its mean.

[0121] S5.2: Based on the quantization vector of the working condition stability parameter, combined with the X-ray detector signal-to-noise ratio attenuation rate and encoder pulse loss count as the original indicators of sensor health, a weighted fusion algorithm is used to comprehensively evaluate the original indicators of sensor health and generate a comprehensive evaluation value of sensor health.

[0122] Based on the quantization vector of operating condition stability parameters, and combining the signal-to-noise ratio attenuation rate data from the X-ray detector and the pulse loss count data from the traction wheel encoder as the original indicators of sensor health, a corresponding weight coefficient table is established for the physical characteristics of different sensors and their impact on the eccentricity detection link. The original sensor health indicators are then normalized: the signal-to-noise ratio attenuation rate values ​​are mapped to the zero-to-one range for a unified scale, and the pulse loss count is scaled proportionally to a preset maximum number of lost pulses to eliminate dimensional differences. Based on the weight coefficient table, a weighted fusion algorithm is used to calculate the weighted composite value of the two types of sensor health indicators. The core of the weighted fusion algorithm is: in, This is a comprehensive evaluation value for the sensor's health. For X-ray detector weighting coefficients, These are the encoder weight coefficients. This is the normalized signal-to-noise ratio attenuation rate value. To normalize the lost pulse count values, a quantization vector of the operating condition stability parameter is added as a third weight reference during the fusion process, and dynamically adjusted... and To adapt to changes in current operating conditions, this ensures that the weights of corresponding sensors are prioritized when there are large fluctuations in the thermal field or unstable mechanical speeds, thereby enhancing the scenario adaptability of health assessment. The calculated... Boundary truncation is performed to ensure that the comprehensive evaluation value is strictly limited to the range of zero to one, serving as a confidence reference for subsequent historical matching consistency correction calculations. Through weighted normalization fusion, the quantized vector of the operating condition stability parameters from the previous step and the original sensor health indicators are transformed into a comprehensive sensor health evaluation value characterizing the reliability of data acquisition, thus achieving input accuracy and robustness for Bayesian prior dynamic control.

[0123] For example, in the quantification vector of stability parameters for a cable extrusion production line, the standard deviation of temperature fluctuation is 0.12, the coefficient of variation of traction speed is 0.05, the normalized value of the X-ray detector signal-to-noise ratio attenuation rate is 0.08, and the normalized value of the encoder lost pulse count is 0.03. Based on field experience, the weighting coefficient of the X-ray detector is set to 0.6, and the weighting coefficient of the encoder is set to 0.4. Substituting these values ​​into the formula, the calculated comprehensive evaluation value of sensor health is 0.062. This comprehensive evaluation value, ranging from zero to within one level, falls within a relatively high reliability range. It will be given a higher confidence weight in subsequent historical matching consistency correction, thereby significantly improving the stability and anti-interference capability of the gating coefficient during Bayesian inference.

[0124] S5.3: Extract the standard deviation of fingerprint matching results from multiple consecutive frames as the original statistic of historical matching consistency, and perform confidence correction processing on the original statistic of historical matching consistency based on the comprehensive evaluation value of sensor health to generate the corrected statistic of historical matching consistency.

[0125] It should be noted that the "continuous multiple frames" refers to the most recent N matching results. The value of N can be set according to the process stability requirements. In this embodiment, N is an integer value between 10 and 20. The standard deviation is calculated as follows: for each cause score value in the matching score sequence of N consecutive frames, calculate its standard deviation, and then take the average of the standard deviations of the four causes as the original statistic of historical matching consistency.

[0126] S5.4: Construct a dynamic prior probability distribution model using the quantization vector of the working condition stability parameter, the comprehensive evaluation value of the sensor health, and the historical matching consistency correction statistic. Input the preliminary matching score sequence as likelihood evidence into the dynamic prior probability distribution model to perform Bayesian inference operation and generate an intermediate confidence sequence.

[0127] A dynamic prior probability distribution model is constructed based on the quantized vector of operating condition stability parameters, the comprehensive evaluation value of sensor health, and the historical matching consistency correction statistics. A multivariate Gaussian prior structure is selected to cover the correlation characteristics between different candidate causes. The covariance matrix of the three types of prior evidence is solved, and the correction amount of operating condition stability to the prior mean and the adjustment amount of sensor health to the prior variance are fused element-wise to form a prior parameter set with temporal smoothness. In the dynamic prior probability distribution model, a prior probability value is defined for each candidate cause and its corresponding joint probability function is established. The preliminary matching score sequence is used as the likelihood evidence input, and the posterior probability is calculated using the Bayesian inference formula.

[0128] in, For candidate causal events, To match score vectors in real time, For posterior probability, The likelihood function value, For prior probability, This serves as evidence. During the calculation, the weights of the likelihood function are adjusted using the prior covariance matrix to suppress the bias of low-health sensor data on the posterior output. The posterior probability values ​​of all candidate causal events are encapsulated into an intermediate confidence sequence in causal index order, ensuring that the numerical range of this sequence is consistent with the model definition.

[0129] By fusing Bayesian inference with prior models, the preliminary matching score sequence from the previous step is transformed into an intermediate confidence sequence that covers the trade-off between operational stability and data reliability, thereby achieving dynamic quantification of the posterior probability of each candidate cause.

[0130] For example, on a medium- and high-voltage cable extrusion production line, the quantization vector of the working condition stability parameter is [0.82, 0.76], representing the normalized value of the standard deviation of temperature fluctuation and the normalized value of the coefficient of variation of traction speed, respectively; the comprehensive evaluation value of sensor health is 0.91, indicating that the overall sensor condition is good; the historical matching consistency correction statistic is 0.14, which is derived from the result of the standard deviation of matching results of 20 consecutive frames after health correction. A two-dimensional Gaussian prior distribution is constructed, with the mean vector set to [0.5, 0.5]. The covariance matrix is ​​scaled according to the health value, with a scaling factor of 0.91. The preliminary matching score sequence [0.78, 0.33, 0.56, 0.41] is used as the likelihood evidence input, and the likelihood value of the cause of the core offset is calculated to be 0.79. The corresponding posterior probability is calculated using Bayes' formula, and the result is 0.705, which is filled into the corresponding position of the intermediate confidence sequence. The confidence level was verified to be activated in the subsequent gating decision. Based on this high confidence level, the control module invoked the core displacement servo adjustment strategy to produce a target instruction that significantly improves the correction accuracy.

[0131] S5.5: Based on the intermediate confidence sequence, perform linear scaling and boundary truncation processing to map the posterior probability values ​​of each candidate cause to an open interval of 0 to 1, thereby generating a set of gating coefficients for each candidate cause.

[0132] Based on the posterior probability values ​​corresponding to each candidate cause in the intermediate confidence sequence, the linear scaling operator is invoked to perform numerical range standardization mapping on different probability values, converting the original probability values ​​into intermediate coefficients after interval normalization, wherein the scaling coefficients are obtained by back-calculation from the mean of the current working condition stability parameter quantization vector.

[0133] Boundary truncation is performed on the intermediate coefficients after linear scaling. By setting hard thresholds for the lower and upper limits, noise probabilities below the lower threshold and abnormal peaks above the upper threshold are eliminated to ensure that each intermediate coefficient is within the open interval of zero to one.

[0134] For the truncated intermediate coefficient set, the numerical smoothing filter module is called to perform local mean updates on the coefficients of adjacent candidate causes, so as to suppress the jump in probability values ​​within the same working condition cycle and enhance the stability of subsequent gating decisions.

[0135] The smoothed coefficient set is repackaged into a structured gated coefficient set according to the candidate cause label order, so that each cause corresponds to a unique coefficient in the open interval of zero to one, and is accompanied by the time label of the current decision period and the working condition association index.

[0136] By employing the aforementioned linear scaling, boundary truncation, and smoothing methods, the posterior probability value output from the previous step is transformed into a standardized and stable set of gating coefficients that satisfy the gating judgment of the control strategy, thereby achieving controllable adaptation to the quantification results of error attribution uncertainty.

[0137] For example, for an intermediate confidence sequence containing four causes—die core misalignment, melt flow fluctuation, traction jitter, and temperature gradient mismatch—the values ​​are 0.82, 0.46, 0.33, and 0.15 under a certain extrusion condition. A normalization mapping is performed using a linear scaling operator with a scaling factor of 0.95, yielding normalized values ​​of 0.779, 0.437, 0.314, and 0.142. Boundary truncation is then performed on these normalized values, setting a lower threshold of 0.05 and an upper threshold of 0.95. Values ​​exceeding the boundaries are removed, and the results are retained. Compared to the original values, no out-of-limit conditions are observed. Then, through three-point local mean smoothing filtering, the mean is calculated using 0.779 for core offset and 0.437 for melt fluctuation, and their respective coefficients are updated to 0.608. This is combined with 0.314 for traction jitter and 0.142 for temperature gradient mismatch to calculate the mean and update it to 0.228. This results in a new set of gating coefficients {0.608 (core offset), 0.608 (melt fluctuation), 0.228 (traction jitter), 0.228 (temperature gradient mismatch)}, with an attached time tag T20240625 and a stability index S0.93. Using this set as input to the dynamic threshold determination module in step S6 significantly improves the stability of the gating state and reduces false triggering of low-confidence strategies.

[0138] Step S6: Determine whether there are any gate coefficients in the candidate cause gating coefficient set that are higher than a preset dynamic threshold. If so, activate the corresponding correction sub-strategy and generate a target correction instruction. If not, generate a micro-amplitude same-direction compensation instruction with an amplitude not exceeding 15% of the current eccentricity as the target correction instruction. Specifically, this includes: S6.1: Perform dynamic threshold comparison processing on each numerical element in the set of candidate cause gating coefficients to generate a gating state determination sequence containing high-confidence activation markers and low-confidence suppression markers, wherein the dynamic threshold is calculated in real time based on the stability parameters of the current operating condition.

[0139] For each numerical element in the set of candidate cause gating coefficients, a dynamic threshold generation operator is invoked. The quantized vector of the working condition stability parameter output from the previous step S5 is used as the real-time input signal, and a dynamic threshold for the current moment is generated according to the set floating threshold calculation model. The difference between each gating coefficient and the aforementioned dynamic threshold is calculated and a sign determination is performed, resulting in two types of intermediate determination results: determination flags above the threshold and determination flags below the threshold. The intermediate determination results are input into the state labeling mapping module, and state labels of "high confidence activation" or "low confidence suppression" are assigned according to the flag category. The state label sequence is used for index sorting to ensure that the labeling order of different candidate causes is consistent with the index order of the gating coefficient set. Through the binding operation of state labels and index order, the comparison determination results are converted into a gating state determination sequence with a clear execution priority, realizing the optional labeling of different candidate cause correction paths.

[0140] By using dynamic threshold comparison processing, the set of working condition stability parameters and gating coefficients from the previous step is transformed into a gating state judgment sequence, thereby achieving real-time judgment of high-confidence activation or low-confidence suppression for different candidate causes.

[0141] For example, in the scenario of detecting eccentricity during extrusion of medium and high voltage cable insulation, the S5 outputs a core offset gating coefficient of 0.78, a melt flow gating coefficient of 0.65, a traction jitter gating coefficient of 0.52, a temperature gradient mismatch gating coefficient of 0.41, a temperature fluctuation standard deviation of 0.012, and a traction speed variation coefficient of 0.008 in the quantification vector of operating condition stability parameters. The temperature fluctuation standard deviation and the traction speed variation coefficient are then input into the dynamic threshold calculation model. in For dynamic thresholds, As the baseline threshold, This represents the normalized standard deviation of the temperature range fluctuation. This is the normalized value of the coefficient of variation of traction speed.

[0142] The fractional part of the above formula represents the mean of the stability parameters under the current operating condition, and the calculated threshold is 0.506. The gate coefficients are then successively compared with this threshold using difference calculations. Differences greater than zero are marked as high-confidence activation, and differences less than zero are marked as low-confidence suppression. This results in three causes: core offset, material melt fluctuation, and traction jitter, labeled as high-confidence activation, and temperature gradient mismatch as low-confidence suppression. The state determination sequence generated after indexing and sorting significantly improves the rationality of the control strategy. In the activation channel, the servo displacement or PID control unit can quickly respond to the source of eccentricity, while the suppression channel avoids introducing additional disturbances due to uncertainties, thus improving the stability and response accuracy of the system.

[0143] S6.2: Based on whether there is a high-confidence activation marker in the gated state determination sequence, a logical branch judgment is made. If it exists, the specific error source fingerprint template identifier associated with the corresponding high-confidence activation marker is extracted. If it does not exist, a steady-state maintenance mode trigger signal is directly generated.

[0144] Based on the numerical tag set of the gated state determination sequence, the Boolean logic operation unit performs an existence check on the high-confidence activation tag bits of each element in the sequence to output a high-confidence conditional Boolean variable. Based on the determination result of this Boolean variable, the index position of each element in the sequence is conditionally screened, and the index positions containing the high-confidence activation tag bits are extracted and mapped to the corresponding error source fingerprint template identifier index value. For the extracted index values, the structured index table in the error source fingerprint template set is accessed, parsed, and the specific error source fingerprint template identifier matching each index value is output as the sole input object for subsequent correction strategy matching processing. If the Boolean variable determination result is empty, the steady-state mode trigger construction unit is called to generate a steady-state maintenance mode trigger signal, and this signal is marked as a conservative strategy preset state for use in the next sub-step scheduling. The results of the above extraction or triggering are output to the input buffer of the control strategy execution module, realizing accurate extraction of target causes and selection of strategy branches based on the gated state determination sequence.

[0145] By determining the high-confidence activation flag and extracting the error source fingerprint template identifier or generating the steady-state maintenance mode trigger signal, the gating state determination sequence of the previous step is transformed into control strategy input data with a clear execution path, thereby realizing the activation of the dedicated correction strategy under high-confidence attribution and the switching of the conservative compensation mode under low-confidence attribution.

[0146] For example, in the extrusion production of a high-voltage cable, the gated state judgment sequence length is 4, corresponding to the cause categories of core displacement, material melt fluctuation, traction jitter, and temperature gradient mismatch. The dynamic threshold is set to 0.72, and the sequence values ​​are 0.85, 0.68, 0.74, and 0.53 respectively. During the verification process, the Boolean logic unit determines that the first and third bits meet the high confidence condition, with index positions 1 and 3 respectively. The error source fingerprint template set index table is accessed, with the first bit mapped to the identifier "MCPD" and the third bit mapped to the identifier "TRPH". "MCPD" and "TRPH" are output to the control strategy execution module, corresponding to the pre-calibrated core displacement servo drive mapping curve and traction system phase adjustment mapping curve, respectively. Under another operating condition, the sequence values ​​are all below 0.72, the Boolean variable judgment result is empty, and the steady-state mode trigger construction unit generates the trigger signal "STABLE_RUN". This signal will be used in the next step to generate a unidirectional compensation command with an amplitude limited to within 15% of the eccentricity. Application results show that in the former case, the system can quickly enter the targeted correction path, and the deviation of the mold core axis is reduced to less than 0.05mm after correction, and the traction phase drift is significantly improved; in the latter case, the system maintains a stable operating state, and the eccentricity fluctuation amplitude is maintained at an extremely low level, realizing precise diversion of control decisions and risk avoidance.

[0147] S6.3: For the extracted specific error source fingerprint template identifier, the pre-calibrated multidimensional mapping relationship library is called to perform correction strategy matching processing to generate a special correction instruction set containing the servo drive voltage increment sequence or temperature zone PID parameter adaptive offset. The multidimensional mapping relationship library stores the nonlinear coupling curves between different error source fingerprint template identifiers and physical actuator action parameters.

[0148] For the specific error source fingerprint template identifier output by step S6.2, the pre-calibrated multidimensional mapping relationship library is called as the unique policy matching execution object.

[0149] A hash matching search is performed on the template identifier and the causal classification index in the mapping relationship library to locate the corresponding nonlinear coupling curve parameter set.

[0150] Based on the retrieved set of coupling curve parameters, a multidimensional control function model is loaded, which includes two types of action mappings: servo drive voltage increment and temperature zone PID parameter adaptive offset.

[0151] The spatial asymmetry index, the main frequency energy concentration interval parameter, and the thermo-coupling gradient curvature scalar obtained from the previous steps are input into the multidimensional control function model. Multidimensional variable substitution operation is performed to calculate the original output value of the control function.

[0152] Boundary constraints and smoothing filtering are used to correct the original output value of the control function to ensure that the servo drive voltage increment and the temperature zone PID parameter offset both meet the physical execution safety range and response stability requirements.

[0153] The corrected servo drive voltage increment data and temperature zone PID parameter offset data are combined through format encapsulation to generate a dedicated correction instruction set, which serves as the input instruction for subsequent execution modules.

[0154] Through the above-mentioned multi-dimensional mapping and safety correction processing, the template identifier of the previous step is transformed into a structured control instruction that accurately matches the action parameters of the actuator, thereby realizing the output of customized correction strategies for different causes of eccentricity.

[0155] For example, in the scenario of detecting the cause of core misalignment in a medium- and high-voltage cable production line, the fingerprint template of a specific error source is identified as "MC_OFF," corresponding to the nonlinear coupling curve form stored in the multidimensional mapping relationship library as the servo drive voltage increment function and the temperature zone PID parameter offset function. The servo drive voltage increment function is: ,in The displacement sensitivity coefficient, The quantification value is the spatial asymmetry index; the temperature zone PID parameter offset function is: ,in The thermal field sensitivity coefficient, The thermo-mechanical coupling gradient curvature scalar is used. In practical applications, the displacement sensitivity coefficient is set to 2.5V / mm, the spatial asymmetry index is quantized to 0.8mm, and the calculated servo drive voltage increment is 2.0V; the thermal field sensitivity coefficient is set to 1.2 PID units / degree, the thermo-mechanical coupling gradient curvature scalar is 0.5 degrees, and the calculated temperature zone PID parameter offset is 0.6 PID units. After boundary constraint correction, the servo drive voltage increment remains at 2.0V, and the temperature zone PID offset is corrected to 0.58 PID units. The final generated dedicated correction instruction set drives the servo mechanism to precisely move the mold core in the execution module, significantly reducing the eccentricity residual and significantly improving the uniformity of the insulation layer.

[0156] S6.4: Perform conservative intervention amplitude limiting operation on the steady-state maintenance mode trigger signal to generate a micro-amplitude same-direction compensation command with the amplitude strictly constrained to within 15% of the current eccentricity, wherein the current eccentricity is obtained by quantification of the spatial asymmetry index in the real-time multidimensional feature stream output by the previous step.

[0157] For the steady-state maintenance mode trigger signal generated by S6.2, the spatial asymmetry index of the real-time multidimensional feature flow in the preceding steps is used as the quantization basis value of the current eccentricity to establish the initial input set for the intervention amplitude calculation. The spatial asymmetry index undergoes unit consistency verification and physical scale normalization to ensure dimensional consistency of the eccentricity value in subsequent amplitude scaling calculations. The normalized current eccentricity is input into the amplitude scaling coefficient calculation unit, and an amplitude constraint formula is constructed based on fixed scaling constraints. in, Indicates the maximum permissible level of intervention. This represents the normalized value of the current eccentricity. Based on the calculation result of the amplitude constraint formula, and combined with the intervention direction determination logic, the same-direction compensation strategy mapping table is called to generate direction vector parameters to calibrate the compensation direction. The maximum intervention amplitude A and the direction vector parameters are input into the intervention amplitude limit filter to perform amplitude boundary truncation and signal smoothing processing to avoid transient overshoot when the command amplitude approaches the limit value. The filtered amplitude value and the direction parameters are vector synthesized to generate a micro-amplitude same-direction compensation command parameter set with the amplitude strictly constrained within 15% of the current eccentricity. This set is then encapsulated into a control signal that can be directly parsed by the actuator through the command formatting module. Through amplitude ratio constraint, direction determination, and filtering truncation processing, the result of the previous step is transformed into a compensation command that meets the conservative intervention requirements under steady-state maintenance mode, achieving the expected technical effect of maintaining system eccentricity stability and avoiding over-adjustment even under low-confidence attribution conditions.

[0158] For example, in a medium-voltage cable extrusion production site, the spatial asymmetry index of the real-time multidimensional characteristic flow is detected at 0.32 mm, and the normalized P value is 0.32. Substituting this value into the amplitude constraint formula, the calculated result A is 0.048 mm. Based on the direction determination logic of the current eccentricity, the compensation direction is clockwise. 0.048 mm and the direction parameter are input into the intervention amplitude limiting filter. The filter is set with a boundary cutoff value of 0.05 mm and a smoothing coefficient of 0.8, resulting in an output amplitude of 0.046 mm and a clockwise direction. The compensation command parameter set is obtained through vector synthesis: amplitude = 0.046 mm, direction = clockwise. This parameter set is encapsulated into a standard control signal by the command formatting module and transmitted to the core displacement servo mechanism for execution. In this scenario, the compensation amplitude is significantly lower than the eccentricity and the direction is consistent. The execution results show that the insulation layer thickness deviation remains stable within the range of 0.32 mm to 0.33 mm, the system operates smoothly, and there are no additional temperature fluctuations or abnormal mechanical loads.

[0159] S6.5: Perform final format encapsulation and priority arbitration processing on the dedicated correction instruction set or the micro-amplitude unidirectional compensation instruction to output a target correction instruction with unique execution effect, wherein the target correction instruction carries a corresponding decision confidence label for subsequent closed-loop feedback evidence set acquisition module to trace and verify.

[0160] Based on the dedicated correction instruction set or micro-amplitude co-directional compensation instruction output from the previous steps, the control instruction structure definition table is called to perform field-level parsing of the input instructions, identify core parameters such as servo drive voltage increment, temperature zone PID parameter offset, and decision confidence label, and match a unique identifier code for each type of parameter.

[0161] The parsed instruction parameters are mapped to the execution interface protocol specifications. A one-to-one parameter correspondence constraint check is used to ensure that the mapping results are within the acceptable range of the physical execution mechanism, and boundary value replacement is performed on parameters that do not meet the constraints.

[0162] Perform format encapsulation operations on the instruction data that has been mapped to protocol parameters, and generate a binary frame structure containing instruction type identifier, parameter combination and execution priority fields according to the byte order requirements of the underlying driver control signals.

[0163] Based on the execution priority arbitration rules, the priority calculation operator is invoked to compare the priority fields of multiple candidate instructions. The maximum value selection strategy is used to determine the unique target instruction with execution effect, and non-selected instructions are marked as pending cancellation.

[0164] The decision confidence label field output by the preceding steps is appended to the end of the encapsulated data of the target instruction, and an integrity check is performed. The check value generated by CRC16 check is appended to the end of the data frame to ensure the data reliability of the encapsulated instruction during transmission.

[0165] Through the above-mentioned format encapsulation and priority arbitration processing, the result of the previous step is transformed into a target correction instruction with unique execution effect and carrying a decision confidence label, thereby realizing unambiguous data transmission and traceability verification capabilities between the control decision module and the execution feedback module.

[0166] For example, in a scenario where the gating coefficient for the cause of mechanical misalignment in the mold core exceeds a threshold, the input dedicated correction instruction set includes a servo drive voltage increment of 0.45V, a temperature zone PID parameter adaptive offset of 0.02, and a decision confidence label of 0.88. After field-level parsing, the unique codes for each parameter are E01, P02, and C88, respectively. Protocol mapping checks show that the 0.45V drive voltage increment conforms to the allowable 0~1V range for the actuator, and the offset of 0.02 conforms to the allowable 0~0.05V range. The instruction type identifier for the frame structure generated during encapsulation is T01, and the priority field is set to 5. During arbitration, this frame has a higher priority than the priority field values ​​4 and 3 of other instructions in the same batch, and is therefore selected as the sole execution instruction. An additional decision confidence tag C88 is added, and the CRC16 checksum value 0x3F92 is calculated and appended to the frame tail, forming the final transmission format [Frame Header|T01|E01|P02|Priority 5|C88|CRC160x3F92]. In practical applications, after the target correction command is sent to the servo interface, the core displacement adjustment accuracy is significantly improved. Subsequent closed-loop feedback verification shows that the eccentric residual decreases to the millimeter level, and the execution latency remains at a low level, verifying the effectiveness of the encapsulation and arbitration mechanism.

[0167] Step S7: Drive the core displacement servo mechanism or temperature zone PID control unit to perform physical actions according to the target correction command, and collect a closed-loop feedback evidence set containing the corrected eccentricity residual, execution delay time, and servo current fluctuation data after the action is completed. Specifically, this includes: S7.1: Perform protocol parsing and signal conversion processing on the target correction command to generate a digital pulse sequence adapted to the drive interface of the mold core displacement servo mechanism or an analog voltage setting value adapted to the temperature zone PID adjustment unit, thereby obtaining a low-level drive control signal with a clear execution timing.

[0168] S7.2: Based on the underlying drive control signal, perform physical drive operation on the core displacement servo mechanism or temperature zone PID adjustment unit to induce the cable extrusion production line to generate corresponding mechanical displacement compensation or heating power adjustment, thereby obtaining an execution action completion flag that reflects the actual physical response state.

[0169] S7.3: The X-ray tomography image acquisition unit performs a secondary scan of the cable insulation layer cross-section after the execution action completion flag is triggered, in order to extract the corrected insulation layer thickness distribution data and calculate the geometric center deviation value, thereby obtaining the corrected eccentricity residual data characterizing the correction effect.

[0170] S7.4: Calculate and process the time difference between the timestamp of the underlying drive control signal and the timestamp of the completion flag of the execution action to quantify the time interval from the instruction to the physical response stabilization, thereby obtaining execution delay time data that characterizes the dynamic response characteristics of the system.

[0171] S7.5: Perform spectrum analysis and fluctuation amplitude extraction processing on the current sensor sampling data of the core displacement servo mechanism or temperature zone PID adjustment unit during the execution process to identify the abnormal current oscillation characteristics caused by load changes, thereby obtaining servo current fluctuation data characterizing the load stability of the actuator, and integrate the corrected eccentric residual data, the execution delay time data and the servo current fluctuation data into a closed-loop feedback evidence set.

[0172] Step S8: Based on the closed-loop feedback evidence set, perform an update operation on the prior distribution of the confidence scores of the corresponding templates in the error source fingerprint template set to generate an updated error source fingerprint template set, thereby completing the self-evolutionary correction of the control decision model. Specifically, this includes: S8.1: Obtain the corrected eccentric residual data, execution delay time and servo current fluctuation data from the closed-loop feedback evidence set, and use the multi-dimensional state mapping algorithm to convert the above physical quantities into likelihood function parameters that characterize the effectiveness of this correction action, and generate a correction performance evaluation vector that includes residual convergence rate index, response hysteresis coefficient and energy consumption anomaly index.

[0173] S8.2: Based on the residual convergence rate index in the correction effectiveness evaluation vector, combined with the currently activated candidate cause type, the support strength of the current observation evidence for different error source fingerprint templates is derived using the Bayesian likelihood calculation rule, and a real-time likelihood score sequence corresponding to each cause such as core offset, material melt fluctuation, traction jitter and temperature gradient mismatch is generated.

[0174] S8.3: Extract the confidence prior distribution parameters of the target fingerprint template that matches the activated corrector strategy from the error source fingerprint template set, and use the Bayesian posterior update formula to fuse the real-time likelihood score sequence with the confidence prior distribution parameters to generate an intermediate posterior probability distribution set that reflects the latest working condition characteristics.

[0175] S8.4: Perform confidence smoothing on the intermediate posterior probability distribution set, introduce a time decay factor to suppress drastic probability jumps caused by a single abnormal feedback, calculate the final confidence weight value with time stability, and generate the updated confidence prior distribution parameters after smoothing correction.

[0176] S8.5: Replace the original corresponding template prior data in the error source fingerprint template set with the updated confidence prior distribution parameters, reconstruct the error source fingerprint template set containing the latest statistical features, and generate an updated error source fingerprint template set with self-evolution capability to support fingerprint matching and gating judgment in the next round of control decision cycle.

[0177] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0178] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0179] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A smart correction method for die core position based on cable insulation extrusion eccentricity detection, specifically including: S1: Acquire multimodal monitoring data during the cable insulation extrusion process, including X-ray tomographic images, temperature distribution sequences, rotational speed fluctuation spectra, phase jitter signals, and pressure time series data; S2: Based on historical eccentricity sample annotation information, feature extraction is performed on the multimodal monitoring data to generate an error fingerprint template set containing spatial asymmetry index, main frequency energy concentration range, phase lag feature and thermo-coupling gradient feature; S3: Input the multimodal monitoring data into a multi-branch feature encoder and output a real-time multidimensional feature stream containing spatial, frequency domain, and temporal feature vectors; S4: Perform a similarity comparison operation based on the real-time multidimensional feature stream and the error source fingerprint template set to generate a preliminary matching score sequence corresponding to each candidate cause of core offset, material melt fluctuation, traction jitter and temperature gradient mismatch; S5: Using the operating condition stability parameters, sensor health indicators and historical matching consistency statistics as prior validation evidence, perform Bayesian posterior probability update calculation on the preliminary matching score sequence to generate a set of candidate causal gating coefficients with values ​​ranging from 0 to 1. S6: Determine whether there are any gate coefficients in the candidate cause gating coefficient set that are higher than the preset dynamic threshold. If there are, activate the corresponding correction sub-strategy and generate a target correction instruction. If there are no, generate a micro-amplitude same-direction compensation instruction with an amplitude not exceeding 15% of the current eccentricity as the target correction instruction.

2. The intelligent correction method for die core position based on cable insulation extrusion eccentricity detection according to claim 1, characterized in that, The steps S6 are followed by S7-S8, which specifically include: S7: Drive the core displacement servo mechanism or temperature zone PID adjustment unit to perform physical actions according to the target correction command, and collect a closed-loop feedback evidence set containing the corrected eccentricity residual, execution delay time and servo current fluctuation data after the action is completed. S8: Based on the closed-loop feedback evidence set, perform an update operation on the prior distribution of the confidence of the corresponding template in the error source fingerprint template set to generate an updated error source fingerprint template set to complete the self-evolution correction of the control decision model.

3. The intelligent correction method for die core position based on cable insulation extrusion eccentricity detection according to claim 1, characterized in that, Step S2 specifically includes: Based on the X-ray tomographic image sequence in the historical eccentricity sample annotation information with known historical causes, the least squares circle fitting and centroid deviation calculation processing are performed on the point cloud data of the insulation layer cross-section contour edge to extract the geometric center coordinate deviation, and the geometric center coordinate deviation is used as the basic input variable for generating the spatial asymmetry index. The Fast Fourier Transform algorithm is used to perform spectral energy density integration on the rotational speed fluctuation spectrum and pressure time series data in historical samples to identify and extract the dominant frequency components and their bandwidth ranges that characterize the instability of material flow, thereby obtaining well-defined dominant frequency energy concentration interval parameters, and using the dominant frequency energy concentration interval parameters as the core elements for constructing frequency domain stability feature vectors; Based on the edge point cloud data and phase jitter signal of X-ray tomographic images synchronously acquired in historical samples, the peak position of the time lag between the two is calculated by cross-correlation analysis algorithm to quantify the cross-sensor phase lag time value reflecting the mechanical transmission coupling effect, and the cross-sensor phase lag time value is used as a key intermediate variable for deriving the cross-sensor phase lag characteristics. A thermo-coupled gradient matrix is ​​constructed by combining the temperature distribution sequence and phase jitter signal in historical samples. The second-order partial derivative curvature estimation process is performed on the matrix to derive the thermo-coupled gradient curvature scalar that characterizes the intensity of process environment disturbance. The thermo-coupled gradient curvature scalar is then used as a specific fingerprint feature to characterize the cause of temperature gradient mismatch. The spatial asymmetry index, the main frequency energy concentration interval parameter, the cross-sensor phase lag characteristics, and the thermo-coupling gradient curvature scalar are integrated. Multidimensional feature vector encapsulation and standardized mapping are performed according to four known cause labels: core offset, material melt fluctuation, traction jitter, and temperature gradient mismatch, to generate a structured error source fingerprint template set. The error source fingerprint template set is then used as the sole standard reference library for subsequent real-time multidimensional feature flow similarity comparison.

4. The intelligent correction method for die core position based on cable insulation extrusion eccentricity detection according to claim 1, characterized in that, Step S5 specifically includes: The standard deviation of temperature fluctuation and coefficient of variation of traction speed in real-time multimodal monitoring data are obtained as raw data of working condition stability. The raw data of working condition stability is normalized and mapped to generate a quantitative vector of working condition stability parameters that characterizes the current smoothness of extrusion process operation. Based on the quantization vector of the working condition stability parameter, combined with the signal-to-noise ratio attenuation rate of the X-ray detector and the encoder pulse loss count as the original indicators of sensor health, a weighted fusion algorithm is used to comprehensively evaluate the original indicators of sensor health and generate a comprehensive evaluation value of sensor health. The standard deviation of fingerprint matching results from multiple consecutive frames is extracted as the original statistic of historical matching consistency. The original statistic of historical matching consistency is then subjected to confidence correction processing based on the comprehensive evaluation value of sensor health to generate the corrected statistic of historical matching consistency. A dynamic prior probability distribution model is constructed using the quantization vector of the working condition stability parameter, the comprehensive evaluation value of the sensor health, and the historical matching consistency correction statistic. The preliminary matching score sequence is then used as likelihood evidence to input the dynamic prior probability distribution model to perform Bayesian inference operations and generate an intermediate confidence sequence. Based on the intermediate confidence sequence, linear scaling and boundary truncation are performed to map the posterior probability values ​​of each candidate cause to an open interval of 0 to 1, thereby generating a set of gating coefficients for each candidate cause.

5. The method according to claim 1, characterized in that, When extracting the spatial asymmetry index, a geometric morphology analysis algorithm is used to calculate the thickness variance and eccentricity characteristics of the cross-sectional profile of the insulation layer along the circumferential direction. The thickness variance and eccentricity characteristics are then normalized and added to the error fingerprint template set as spatial asymmetry indices.

6. The method according to claim 1, characterized in that, The extraction of phase lag features specifically includes: performing cross-correlation analysis on the speed fluctuation spectrum and pressure time series data in the multimodal monitoring data, extracting phase lag values, and adding the phase lag values ​​as phase lag features to the error fingerprint template set.

7. The method according to claim 1, characterized in that, The multi-branch feature encoder includes a convolutional neural network branch, a frequency domain transform branch, and a time-series convolutional network branch. These branches respectively extract features from the image, spectrum, and sequence data in the multimodal monitoring data. The outputs of each branch are then fused through an attention mechanism to generate the real-time multidimensional feature stream.

8. The method as described in claim 1, characterized in that, When normalizing the set of gating coefficients, the process includes linear scaling and boundary constraints on the Bayesian posterior probability, limiting the non-zero gating coefficients to between 0.05 and 0.95, maintaining the stability of the results and preventing extreme probability triggering.

9. The method according to claim 1, characterized in that, The correction sub-strategies include horizontal translation of the mold core, vertical translation of the mold core, dynamic adjustment of screw speed, and adjustment of power ratio in the temperature zone. Each correction sub-strategy corresponds to a specific gate coefficient threshold and execution parameters.

10. The method as described in claim 1, characterized in that, The calculation of the thermo-coupled gradient curvature characteristics is performed by taking the temperature distribution sequence as the thermal field and the traction wheel phase jitter signal as the force field. After resampling at a unified sampling frequency, a two-dimensional thermo-matrix is ​​constructed and its second-order partial derivative is calculated and the mean is aggregated.

Images