A method for on-line detection and rectification of the gap and thickness of the insulating tape of a pre-branching cable

Abstract

The present application relates to a kind of pre-branch cable insulation tape layer lapping gap and thickness online detection and deviation correction method.It solves the problem of high-precision detection and fast correction of dynamic deviation in lapping process.Core scheme includes: multi-source signal time sequence binding to generate synchronous motion data set with displacement and angle mark;Based on image texture and real-time motion parameter, the lapping phase angle is inversed;Through automatic clustering and statistical analysis, standard phase fingerprint atlas and deviation classification are established;According to the deviation type, adaptive compensation model is selected and real-time output correction instruction is outputted, to directly drive actuator to adjust lapping state;At the same time, dynamic weight updating optimization reference fingerprint atlas is used to adapt to equipment slow drift.The scheme significantly improves the dynamic response speed, correction accuracy and process adaptability of lapping process, effectively guarantees product quality and production efficiency.

Description

Technical Field

[0001] This invention relates to the field of online detection and closed-loop control technology in cable manufacturing process, and in particular to a method for online detection and correction of the gap and thickness of the insulation tape layering in pre-branched cables. Background Technology

[0002] Currently, online detection and correction systems for insulation tape lamination accuracy in cable manufacturing generally employ closed-loop control strategies based on error feedback. Mainstream solutions typically combine image-based visual inspection with pre-set machine learning models. This involves acquiring real-time image information on the insulation tape lamination gap and thickness, and using image processing techniques such as edge detection, template matching, or pitch measurement to calculate the geometric deviation of the tape. The result serves as input to traditional proportional-integral-derivative (PID) control or adaptive control algorithms, dynamically adjusting the lateral correction action of the lamination actuator. More advanced systems integrate multi-sensor information fusion technology, superimposing physical quantities such as temperature, tension, and guide wheel displacement to estimate and compensate for errors, further enhancing the system's correction capability and robustness to complex operating conditions. With the increasing demand for high-speed, high-consistency cable production, some enterprises and research institutions have introduced fuzzy control, neural network self-optimization algorithms, or offline modeling-online iterative identification methods to alleviate the bottlenecks of traditional feedback loops caused by lag and error accumulation due to sudden changes in process parameters.

[0003] Currently, the industry generally focuses on the dynamic response speed and control accuracy of cable wrapping correction systems. For example, some manufacturers are improving the reliability of the system under dynamic scenarios such as wrapping speed fluctuations and tension disturbances by enhancing the frame rate and accuracy of visual acquisition and improving the response rate of displacement actuators. Industry trends indicate that future online detection and correction control of cable wrapping is developing towards multi-source information fusion, high-speed, low-latency dynamic compensation, and online model adaptive collaborative evolution, striving to break through the high dependence of traditional control loops on historical error integration and static model identification, and achieve precise real-time adjustment under dynamic changes in multiple process parameters.

[0004] Current representative technologies largely rely on vision-based error detection, combined with known geometric models and feedback loops to achieve deviation correction. For example, online visual inspection systems for the layered wrapping of pre-branched cable insulation tape primarily use high-resolution cameras to periodically acquire wrapping images, employing edge detection and template matching algorithms to extract gap and thickness deviations. The feedback loop then uses fixed proportional-integral-differential parameters to automatically adjust the deviation correction action. This system is suitable for applications with stable wrapping rhythms, high equipment structural rigidity, and moderate production speeds. In situations with increased production speeds or rapid tension changes, errors are often reduced by adding external tension sensors and offline model correction methods. However, the system still fundamentally relies on periodic error feedback and static model parameter settings.

[0005] However, existing technologies still have significant shortcomings under dynamic process parameter variations, especially when key process parameters such as traction speed and tension change rapidly during the wrapping process. Traditional closed-loop control methods based on error accumulation and feedback adjustment struggle to respond to these changes in a timely manner, often resulting in correction delays, accuracy fluctuations, and even closed-loop oscillations. On one hand, static gap positioning using visual inspection is sensitive to minute changes in the on-site mechanical structure, minor wear of guide wheels, or thermal expansion, causing accumulated errors that cannot effectively distinguish between mechanical factors and process disturbances, thus affecting the system's adaptive capability. On the other hand, existing online adaptive algorithms largely rely on historical error integration and parameter iterative identification, making them highly susceptible to abnormal disturbances, signal delays, and noise accumulation. Their robustness and stability are limited, making it difficult to guarantee the sustained accuracy of correction control under high-speed, continuous operation.

[0006] Existing technologies lack dynamic modeling and utilization methods that fully consider the kinematic state of the insulation tape wrapping process. There is a lack of a deviation decoupling method that can reflect the intrinsic mapping relationship between the wrapping motion phase and the geometric changes of the image in real time. As a result, when the multi-source process parameters fluctuate synchronously and drastically, the system's correction response is always limited by the external error feedback mechanism and static model, and it is impossible to achieve effective feedforward compensation for dynamic disturbances and essential structural noise isolation.

[0007] Therefore, there is an urgent need to provide a new paradigm for motion phase recognition and deviation decoupling in pre-branched cable insulation tape layered wrapping systems. This paradigm can synchronize multiple process parameters such as traction speed, tension, and image timing in real time, accurately extract the phase fingerprint in the wrapping motion, and achieve dynamic decoupling of deviation in the time and spatial domains. By combining statistics, signal processing, and mechanism compensation, the system's sensitivity and feedforward capability to dynamic process disturbances can be improved. Ultimately, the goal of deviation correction control no longer relies on historical error integration and online model iteration can be achieved, thereby significantly improving the dynamic response speed, control accuracy, and long-term operational stability of the wrapping process. Summary of the Invention

[0008] This application provides a method for online detection and correction of the gap and thickness of the insulation tape layering in pre-branched cables, aiming to solve one of the problems or issues of the prior art mentioned in the background.

[0009] This application provides a method for online detection and correction of the gap and thickness of the insulation tape lamination in pre-branched cables, specifically including: S1: Acquire the traction speed, tension parameters and image acquisition timing signals during the wrapping process of the pre-branched cable insulation tape as multi-source signals, and bind the multi-source signals with timestamps to generate a synchronous motion raw dataset; S2: Obtain the wrapping pitch and guide wheel diameter parameters, and perform reverse calculation based on the periodic structure of the texture of continuous multi-frame images in the original synchronous motion dataset, the wrapping pitch and guide wheel diameter parameters to generate the real-time wrapping phase angle; S3: Utilize the initial light-load no-load running process during the start-up of the production shift to collect standard-length cable image sequences and multi-source motion parameters. Through automatic clustering, construct a benchmark phase fingerprint cluster containing confidence weights from the standard-length cable image sequences and the multi-source motion parameters to generate a benchmark map. S4: Map the real-time wrapping phase angle to the matching reference fingerprint sub-region in the reference map, perform local cross-correlation matching, and extract the residual offset; S5: Calculate the statistical moment features of the residual offset, including skewness, kurtosis and phase span, and determine the deviation type label based on the distribution pattern of the statistical moment features to generate process status classification results; S6: Select the corresponding compensation model according to the process state classification result. If it is determined to be guide wheel axial movement, call the harmonic elimination model and generate a correction command that has been filtered by mechanism matching. S7: Inject the correction command into the actuator drive unit to complete the dynamic correction action on the gap and thickness of the insulation tape layer wrapping, and generate the corrected wrapping physical state. S8: Monitor the original dataset of the new synchronous motion under the corrected wrapping physical state and update the confidence weights in the baseline graph.

[0010] The method for online detection and correction of the gap and thickness of the insulation tape layering in pre-branched cables provided in this application has the following advantages: (1) To address the technical shortcomings of traditional pre-branched cable insulation tape layering wrapping process, where deviation calculation generally relies on static geometric error detection or simple feedback control, resulting in delayed correction response and susceptibility to inherent mechanical disturbances, this solution proposes a new deviation decoupling paradigm based on time-varying kinematic phase modeling. By reconstructing the wrapping process into a dynamic system with definite phase characteristics, and utilizing the time-space synchronization mechanism of a high-precision encoder and image acquisition module, each frame of visual information is strictly bound to the cable axial displacement, rotation angle, and traction speed, thereby establishing a joint characterization space of "phase angle-pixel offset". Based on this, the current phase state is inverted by combining the known wrapping pitch and guide wheel parameters, and an initial phase fingerprint spectrum that drifts with the working conditions is constructed. This allows strongly correlated disturbances such as mechanical eccentricity, lens distortion, and pitch fluctuations to be naturally incorporated into the phase periodic structure, significantly improving the separability of background interference. Compared to traditional methods that directly threshold or differentiate image gaps, this approach effectively avoids misjudging deterministic periodic disturbances as process anomalies, significantly reducing false alarm rates and overcorrection risks, and substantially improving the authenticity and robustness of deviation identification.

[0011] (2) Furthermore, to address the problems of existing correction systems being unable to distinguish between different types of disturbance sources, having a single control strategy, and relying on manual experience for parameter tuning, this solution introduces a residual deviation classification decision mechanism based on phase fingerprint statistical characteristics. Without the need for additional vibration sensors or tension detection devices, it can automatically identify typical fault modes such as tension abrupt changes, axial movement, and imaging degradation simply by analyzing the distribution of non-periodic offset components after local cross-correlation matching in the phase domain. Specifically, the system completes the disturbance type discrimination based on higher-order statistical moments such as skewness, kurtosis, and phase span: single-peak concentrated type triggers an exponential decay model to cope with material elastic deformation; double-peak symmetrical type activates a harmonic elimination algorithm to match the guide wheel vibration mode library; and broadband diffuse type drives a self-diagnostic process to check for optical link anomalies. This mechanism achieves a leap from "unified error processing" to "mechanism-driven differential response", enabling the correction command to have physical interpretability and process adaptability, significantly improving the response sensitivity and decision reliability of the control system to complex operating conditions, while eliminating the dependence on external sensor upgrades and complicated calibration processes, demonstrating excellent hardware lightweighting and deployment convenience.

[0012] The aforementioned technical approaches collectively construct an intelligent error correction system that is process-inherent, semantically clear, and closed-loop autonomous. Its core advantage lies in transforming traditional passive error correction into active feedforward control based on an understanding of the nature of motion. The final output net error correction command undergoes triple filtering—phase decoupling, disturbance classification, and mechanism matching—before being directly injected into the actuator as an open-loop feedforward. It does not participate in model parameter updates or rely on historical integrals, fundamentally avoiding common problems in adaptive control such as integral saturation, oscillation divergence, and convergence delay. This approach not only significantly improves the system's dynamic stability and real-time performance but also enables efficient deployment on low-computing-power edge devices. It is suitable for multi-specification, small-batch, and fast-change production scenarios, providing a novel technical solution for high-precision cable manufacturing that combines accuracy assurance with cost control. Attached Figure Description

[0013] Figure 1 This is the main flowchart of a method for online detection and correction of the gap and thickness of the insulation tape layering in pre-branched cables.

[0014] Figure 2 This is a sub-flowchart of a method for online detection and correction of the gap and thickness of the insulation tape layering in pre-branched cables.

[0015] Figure 3 This is another sub-flowchart of a method for online detection and correction of the gap and thickness of the insulation tape layering in pre-branched cables. Detailed Implementation

[0016] 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.

[0017] 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.

[0018] like Figure 1 As shown, this application provides an online detection and correction method for the gap and thickness of the insulation tape lamination of pre-branched cables, specifically including: S1: Acquire the traction speed, tension parameters and image acquisition timing signals during the wrapping process of the pre-branched cable insulation tape as multi-source signals, and bind the multi-source signals with timestamps to generate a synchronous motion raw dataset; S2: Obtain the wrapping pitch and guide wheel diameter parameters, and perform reverse calculation based on the periodic structure of the texture of continuous multi-frame images in the original synchronous motion dataset, the wrapping pitch and guide wheel diameter parameters to generate the real-time wrapping phase angle; S3: Utilize the initial light-load no-load running process during the start-up of the production shift to collect standard-length cable image sequences and multi-source motion parameters. Through automatic clustering, construct a benchmark phase fingerprint cluster containing confidence weights from the standard-length cable image sequences and the multi-source motion parameters to generate a benchmark map. S4: Map the real-time wrapping phase angle to the matching reference fingerprint sub-region in the reference map, perform local cross-correlation matching, and extract the residual offset; S5: Calculate the statistical moment features of the residual offset, including skewness, kurtosis and phase span, and determine the deviation type label based on the distribution pattern of the statistical moment features to generate process status classification results; S6: Select the corresponding compensation model according to the process state classification result. If it is determined to be guide wheel axial movement, call the harmonic elimination model and generate a correction command that has been filtered by mechanism matching. S7: Inject the correction command into the actuator drive unit to complete the dynamic correction action on the gap and thickness of the insulation tape layer wrapping, and generate the corrected wrapping physical state. S8: Monitor the original dataset of the new synchronous motion under the corrected wrapping physical state and update the confidence weights in the baseline graph.

[0019] Step S1: Obtain the traction speed, tension parameters, and image acquisition timing signals during the wrapping process of the pre-branched cable insulation tape as multi-source signals, and bind the multi-source signals with timestamps to generate a synchronous motion raw dataset. Specifically, this includes: S1.1: Obtain the pulse sequence signal output by the encoder and the analog voltage signal output by the tension sensor on the pre-branched cable insulation tape wrapping production line. Use a high-frequency sampling clock to count and accumulate the pulse sequence signal and perform analog-to-digital conversion on the analog voltage signal to generate a discrete traction displacement raw data stream and a discrete tension fluctuation raw data stream with a unified sampling reference.

[0020] The input conditions include a pulse sequence signal output from a high-resolution pulse encoder installed on the pre-branched cable insulation tape wrapping production line and a continuous analog voltage signal output from a tension sensor, with the sampling reference time base derived from a high-frequency clock signal.

[0021] High-frequency sampling triggers are performed on the encoder pulse sequence signal, and precise counting and accumulation operations are performed based on the number of pulses in each sampling period to obtain a discretized cumulative sequence of axial displacement.

[0022] A pulse resolution coefficient is introduced during the counting and accumulation process to map the accumulated pulse count to a physical displacement value. The traction displacement is then calculated using the following formula: in, For traction displacement, This represents the cumulative number of pulses. This represents the physical displacement corresponding to each pulse.

[0023] The analog voltage signal output by the tension sensor is subjected to analog-to-digital conversion processing, and the voltage value is quantized into a digital signal under a unified sampling reference, forming a discrete tension data point sequence corresponding to the sampling time.

[0024] A synchronous index based on the sampling clock is used to align the discretized axial displacement sequence and the discretized tension sequence in the time domain, ensuring that both belong to the same sampling reference system.

[0025] Through the above processing method, the signal acquisition results of the previous step are transformed into a discrete traction displacement raw data stream and a discrete tension fluctuation raw data stream with a unified sampling benchmark, so as to realize the accurate input for subsequent timestamp binding and motion state derivation.

[0026] For example, on a pre-branched cable insulation tape wrapping production line, the encoder resolution is set to output 500 pulses per millimeter of displacement, the tension sensor range is 0 to 50 N, corresponding to an analog voltage of 0 to 10 V, and the high-frequency sampling clock is 200 kHz. In a certain sampling period, the cumulative count is 1250 pulses. According to the formula R = 1 / 500 mm, multiplying 1250 by 0.002 mm yields a displacement of 2.5 mm. In the same period, the tension sensor output voltage is 4 V, which, after A / D conversion, yields a value of 1024, corresponding to a tension value of 20 N. During continuous sampling periods, the displacement data stream and tension data stream are time-aligned using a uniform 200 kHz clock index, ensuring microsecond-level synchronization. The final output discrete traction displacement raw data stream has millimeter-level resolution, and the discrete tension fluctuation raw data stream has 0.1 N-level resolution, which can be used as high-precision input for subsequent image acquisition timestamp binding and spatial position state vector generation.

[0027] S1.2: Obtain the hardware interrupt signal triggered by the industrial camera at the moment of exposure, extract the image acquisition timestamp accurate to the microsecond level based on the hardware interrupt signal, and use the image acquisition timestamp as the master synchronization index key to generate an image frame trigger event sequence with absolute time stamp.

[0028] It receives the hardware interrupt signal generated by the industrial camera at the moment of exposure as input. The signal comes from the camera's built-in trigger interface and is acquired through a high-speed digital input port.

[0029] The hardware interrupt signal is phase-locked by a high-precision time reference module to ensure that the time reference is synchronized with the main control clock of the production line, thereby eliminating cross-device clock drift.

[0030] The locked hardware interrupt signal is timestamped, and a microsecond-level resolution timer is used to record the absolute moment of the signal occurrence, forming an event record entry containing the time value.

[0031] Noise suppression filtering is applied to outlier detection and removal in the timestamp sequence. Time records exceeding a preset deviation threshold are removed to ensure the stability and accuracy of the timestamps.

[0032] The filtered timestamp is used as the primary synchronization index key and associated with the image frame buffer to ensure that each time record corresponds to a unique image frame number and to form a preliminary image frame trigger event mapping table.

[0033] By combining the mapping table to perform structured data encapsulation, timestamps, frame numbers, camera working status metadata, etc. are combined into data objects in a unified format to form a sequence of image frame trigger events with absolute time stamps.

[0034] By using the above processing method, a precise correlation is established between the sampling benchmark of the discretized traction displacement and tension data from the previous step and the time domain of image acquisition, thus achieving the expected technical effect of cross-domain data alignment.

[0035] For example, on a pre-branched cable insulation tape wrapping production line, an industrial camera (model XG-5000) is selected. Its trigger interface outputs a TTL level pulse with a width of 10 microseconds, which is captured by a high-speed digital acquisition card. The production line's main control system clock is 100MHz, and the time reference module uses a GPS-synchronized clock to achieve microsecond-level timing accuracy. The acquired pulse sequence is time-calibrated, and the absolute moment of each exposure trigger is recorded. The unit is microseconds. An algorithm with a median filter window length of 5 is used to remove outliers with time drift exceeding 2 microseconds to ensure output stability. In this scenario, each event in the generated image frame trigger event sequence contains a structured object with "timestamp=12451238μs", "frame number=F147", and "camera status=normal". This sequence is used for subsequent linear interpolation resampling processing in S1.3 to achieve accurate temporal matching between the traction velocity vector and tension value pairs, significantly improving the temporal consistency and control response accuracy of the original synchronous motion dataset.

[0036] S1.3: Based on the absolute time stamp in the image frame trigger event sequence, perform linear interpolation resampling processing on the discretized traction displacement raw data stream and the discretized tension fluctuation raw data stream to generate an instantaneous traction velocity vector and instantaneous tension value pair that strictly correspond to the time of each frame image.

[0037] S1.4: Using the known cable guide wheel diameter parameters and wrapping pitch constant, perform kinematic integral transformation on the instantaneous traction velocity vector to generate a spatial position state vector characterizing the cumulative axial displacement and cumulative rotational angle of the cable.

[0038] Based on the generated instantaneous traction velocity vector, the wrapping pitch constant and guide wheel diameter parameters are used as physical geometric inputs to establish a correlation model between axial displacement and rotation angle in the kinematic calculation module. For the instantaneous traction velocity vector, time-domain integration is performed, integrating the velocity function over the image acquisition period to accumulate the axial displacement. A fixed-step numerical integration method is used to ensure consistency with the image frame timestamp. For the same velocity vector, a circular motion conversion is performed using the known guide wheel diameter parameters, converting the axial displacement into a rotation angle using a scaling factor defined based on the ratio of the guide wheel circumference to the axial travel. The wrapping pitch constant is used to perform phase normalization on the rotation angle, mapping the angle to the wrapping phase space for subsequent image matching. The accumulated axial displacement and accumulated rotation angle are combined into a vector to construct a spatial position state vector containing two components. A time series marker is appended to the state vector to maintain synchronization with other sensor information.

[0039] S1.5: The spatial position state vector and the instantaneous tension value are structurally encapsulated into the original image data block corresponding to the image frame trigger event sequence to generate a synchronous motion raw dataset containing axial displacement markers, rotation angle markers, and multi-source sensor information.

[0040] Step S2: Obtain the wrapping pitch and guide wheel diameter parameters, and perform inverse calculation based on the periodic structure of the texture of consecutive multi-frame images in the original synchronous motion dataset, the wrapping pitch, and the guide wheel diameter parameters to generate the real-time wrapping phase angle. Specifically, this includes: S2.1: Obtain the original image data blocks from the synchronous motion original dataset, and use the gray-level co-occurrence matrix algorithm to perform spatial statistical feature extraction on the edge texture of the insulating strip to generate a two-dimensional texture energy distribution map that represents the periodic repetition pattern of the texture.

[0041] S2.2: Based on the high-frequency energy peak points in the two-dimensional texture energy distribution map, apply the Fast Fourier Transform algorithm to perform frequency domain spectrum analysis to generate a texture space frequency vector sequence containing fundamental frequency components and harmonic components.

[0042] Based on the two-dimensional texture energy distribution map already generated in the original synchronous motion dataset, the input object of frequency domain analysis is set as the matrix element of the distribution map. The matrix size is consistent with the resolution of the original image, and the spatial arrangement order of gray intensity values ​​is maintained to ensure the spatial consistency of spectrum analysis.

[0043] Frequency domain sampling preprocessing is performed on the two-dimensional texture energy distribution map. A window function is used to suppress boundary effects. The Hanning window is selected as the window type to smoothly truncate the edge signal. The window size is close to the spatial period corresponding to the wrapping pitch to reduce peak shift caused by spectral leakage.

[0044] The windowed two-dimensional texture energy distribution map is input into the Fast Fourier Transform (FFT) operation module. The one-dimensional FFT is then performed in the row and column directions using the split two-dimensional FFT algorithm to resynthesize the spectrum, and a two-dimensional spectrum matrix containing amplitude and phase information is calculated.

[0045] Amplitude spectrum extraction is performed on the two-dimensional spectrum matrix. The complex value of each frequency component is moduloed to obtain the spectrum amplitude distribution. High-frequency energy peak points are searched in the spectrum amplitude distribution. The peak selection adopts a local peak detection algorithm. The detection radius is set to be related to the guide wheel diameter to adapt to the spatial resolution.

[0046] For the detected high-frequency energy peak points, a texture space frequency vector sequence is constructed. Each vector element contains the horizontal frequency component, vertical frequency component and corresponding amplitude parameter of the peak point, and is sorted by amplitude to distinguish the fundamental frequency component and harmonic components.

[0047] The two-dimensional FFT amplitude spectrum is calculated using the following formula: in, , These represent the row and column lengths of the two-dimensional texture energy distribution map, respectively. , For frequency index, This is a grayscale value function. It is a natural exponential function. It is the imaginary unit.

[0048] By combining FFT operation with peak detection, the two-dimensional texture energy distribution map is transformed into a texture spatial frequency vector sequence containing fundamental and harmonic components. This enables real-time extraction of the periodic and non-periodic components of the insulating strip edge texture in the spatial domain, providing accurate frequency basis data for subsequent normalized phase shift calculation.

[0049] For example, in a 2D energy distribution map of an insulating tape texture with a resolution of 1024×1024 pixels, Hanning window preprocessing is applied, with the window size set to 128 pixels, corresponding to the spatial period of the wrapping pitch. After performing a 2D FFT, a spectrum matrix is ​​obtained. A peak with a horizontal frequency component of 4 periods / pixel and a vertical frequency component of 2 periods / pixel is detected in the amplitude spectrum as the fundamental frequency component, with an amplitude of 3.6 units; simultaneously, a harmonic component with a horizontal frequency of 8 periods / pixel is detected, with an amplitude of 1.2 units. Using the above formula, the corresponding complex spectrum values ​​can be extracted and their magnitudes verified under the conditions of M=1024 and N=1024. In the output texture space frequency vector sequence, the first vector element (4,2,3.6) corresponds to the fundamental frequency component, and the second vector element (8,2,1.2) corresponds to the harmonic component. Under this condition, the frequency data successfully captures the spatial periodic characteristics of the insulating tape wrapping, significantly improving the stability and accuracy of subsequent phase displacement calculations.

[0050] S2.3: Using the reciprocal of the fundamental frequency component in the texture space frequency vector sequence, combined with the known wrapping pitch constant, perform spatial wavelength normalization operation to generate a normalized phase displacement scalar characterizing the relative position offset of the insulating strip within a single frame image.

[0051] S2.4: Based on the normalized phase displacement scalar and the cumulative rotation angle in the original synchronous motion dataset, perform coordinate transformation processing from axial displacement to angle through trigonometric function mapping relationship to generate an initial geometric phase angle value characterizing the instantaneous position of the cable spiral wrapping trajectory.

[0052] S2.5: Based on the initial geometric phase angle value and the known guide wheel diameter parameters, a mechanical transmission clearance compensation coefficient is introduced to perform nonlinear error correction processing, so as to finally generate a real-time wrapping phase angle that eliminates the influence of mechanical hysteresis and accurately represents the current frame wrapping geometry.

[0053] Based on the initial geometric phase angle value and the guide wheel diameter parameter, the range of values ​​for the mechanical transmission clearance compensation coefficient is set, and it is used as the input benchmark for the error correction model.

[0054] The initial geometric phase angle values ​​are normalized by the guide wheel circumference, mapping the angle values ​​to a unit phase domain with a period of one rotation of the guide wheel, so that the subsequent compensation coefficients can operate under a unified dimension.

[0055] Based on the standard and actual measured values ​​of the guide wheel diameter, the angular hysteresis introduced by the mechanical transmission clearance is calculated, and the hysteresis correction variable is calculated using the following formula: in, This is the angle lag correction amount. The difference between the actual diameter and the standard diameter of the guide wheel. This is the standard diameter.

[0056] The above-mentioned hysteresis correction amount is multiplied by the preset mechanical transmission clearance compensation coefficient to obtain the nonlinear correction increment, which is calculated using the following formula: in, To compensate for the increase, This is the mechanical transmission clearance compensation coefficient.

[0057] The compensation increment is added to the initial geometric phase angle value, and the accumulated error is corrected twice by a nonlinear mapping function to obtain a high-precision real-time wrapping phase angle that represents the wrapping geometry of the current frame.

[0058] Through the above compensation mechanism, the initial geometric phase angle result of the previous step is transformed into a real-time wrapping phase angle that eliminates the influence of mechanical hysteresis and conforms to the dynamic characteristics of the process, thereby significantly improving the phase calculation accuracy and dynamic response capability.

[0059] For example, in a pre-branched cable production line, the standard diameter d of the guide roller is set to 120.00 mm, the actual measured diameter is 120.25 mm, Δd is 0.25 mm, and the compensation coefficient is set to 0.85. The lag correction is calculated according to the formula. : The result is 0.01309 radians. Compensation increment. The calculated result is 0.011126 radians. The initial geometric phase angle was assumed to be 1.5708 radians, and after compensation, the real-time wrapping phase angle is 1.581926 radians. This compensated phase angle shows improved stability in subsequent texture feature extraction and phase fingerprint matching, significantly improved accuracy in determining process deviation types, and a marked improvement in the dynamic response lag problem.

[0060] like Figure 2 As shown, step S3 involves: utilizing the initial light-load no-load running process during the production shift start-up to collect standard-length cable image sequences and multi-source motion parameters. Through automatic clustering, a baseline phase fingerprint cluster containing confidence weights is constructed from the standard-length cable image sequences and the multi-source motion parameters, generating a baseline map. Specifically, this includes: S3.1: Obtain the standard length cable image sequence and corresponding multi-source motion parameters synchronously collected during the light-load no-load running process at the beginning of the production shift. Perform texture periodic structure extraction operation on the standard length cable image sequence, and calculate and generate an initial wrapping phase angle sequence in combination with the known wrapping pitch and guide wheel diameter parameters to form an original phase data set with geometric state markings.

[0061] It should be noted that automatic clustering refers to unsupervised learning methods that do not require pre-specifying the number of categories. These include prototype-based K-means algorithms, learned vector quantization (LVQ), Gaussian mixture models, density-based DBSCAN, and hierarchical clustering algorithms such as AGNES. This method employs an improved density peak clustering algorithm, which automatically identifies and groups samples with similar kinematic phase characteristics by calculating the local density of each sample and its minimum distance to higher-density samples, thereby constructing a baseline phase fingerprint cluster.

[0062] The input objects include a sequence of standard-length cable images and corresponding multi-source motion parameters collected synchronously during the light-load no-load running process at the beginning of the production shift. The multi-source motion parameters include traction speed, tension value, and cumulative rotation angle.

[0063] Edge detection processing is performed on the standard length cable image sequence to extract continuous pixel strips in the edge region of the insulation tape as input data blocks for texture analysis.

[0064] Based on the pixel stripes, the gray-level co-occurrence matrix is ​​applied to calculate the gray-level pair frequency distribution of each image block in the horizontal and vertical directions, and a two-dimensional texture energy distribution map is generated.

[0065] In the two-dimensional texture energy distribution map, the fundamental frequency component and the main harmonic components are extracted by fast Fourier transform, and the image texture repetition period length is calculated based on the reciprocal of the fundamental frequency component.

[0066] The period length is normalized using a known wrapping pitch constant to obtain a normalized phase displacement scalar.

[0067] By using the normalized phase displacement scalar and the cumulative rotation angle in the multi-source motion parameters, the initial geometric phase angle value is calculated through the trigonometric mapping function. All initial geometric phase angle values ​​are then encapsulated together with the corresponding image frame index, traction velocity, and tension value to form a set of original phase data with geometric state labels.

[0068] This processing method transforms the real-time wrapping phase angle results from the previous step into a batch of standardized raw phase data sets, enabling stable input for subsequent phase feature vector construction.

[0069] S3.2: Based on the initial wrapping phase angle sequence in the original phase data set, the offset of the insulating strip edge texture pixels in multiple consecutive frames of images is mapped to the phase space, and multi-dimensional feature vector construction processing is performed to generate a high-dimensional phase feature vector set that characterizes the wrapping geometric distortion characteristics of the current frame.

[0070] S3.3: Using the high-dimensional phase feature vector set as input, an improved density peak clustering algorithm is applied to perform automatic clustering analysis to identify sample clusters with similar kinematic phase features, so as to generate a preliminary set of candidate reference phase fingerprint clusters.

[0071] The high-dimensional phase feature vector set is input into the core module of the preset improved density peak clustering algorithm, and the sample similarity calculation operator based on the dual measures of multidimensional Euclidean distance and local density is called to generate an initial feature index matrix containing the local density values ​​of each feature vector and the neighborhood distance of the phase angle.

[0072] Normalization is performed on the initial feature index matrix, and the contribution ratio of local density values ​​in different feature dimensions is adjusted by the weighting coefficient of phase span and texture distortion amplitude to obtain the normalized feature index matrix after the distribution features are balanced.

[0073] Based on the normalized feature index matrix, the density peak criterion is used to calculate the candidate score of the cluster center for each sample. The calculation formula is as follows: in For local density, The product score is the minimum distance from a sample to a high-density sample, and it is used to determine potential cluster centers.

[0074] Samples with scores greater than the preset center threshold are marked as candidate cluster centers, and the minimum phase difference between cluster centers is determined to eliminate redundant centers with too close phase angles.

[0075] Using the remaining candidate cluster centers as the initial clustering cores, and employing the shortest path search of phase angle and the decreasing texture distortion criterion, non-center samples are sequentially assigned to the nearest cluster cores with decreasing local density, thus completing the initial division of candidate clusters.

[0076] Through the above processing method, the high-dimensional phase feature vector set of the previous step is transformed into a candidate benchmark phase fingerprint cluster set containing cluster center identifiers and member distributions, thereby realizing the structured grouping of samples with similar kinematic phase features.

[0077] For example, during the initial light-load no-load operation at the start of a production shift, a sequence of standard-length cable images is collected and processed by S3.2 to generate a high-dimensional phase feature vector set of 500 samples. The texture distortion amplitude dimension ranges from 0 to 2.5 pixels, and the phase span dimension ranges from 0 to 15 degrees. This vector is then input into the improved density peak clustering algorithm module to calculate the local density. With the Gaussian kernel bandwidth set to 1.2, the minimum distance is calculated. The phase angle weight was set to 0.7, and the texture distortion amplitude weight was set to 0.3. Using the formula, the distribution of all sample scores ranged from 0.2 to 4.8. A cluster center score threshold of 3.5 was set, and 12 candidate cluster centers were selected. Redundant centers with a phase difference less than 1.5 degrees were removed based on the minimum phase difference criterion, ultimately retaining 9 cluster centers. The remaining 491 samples were assigned to these 9 clusters according to the shortest phase angle path and local density decrease criteria, generating a candidate benchmark phase fingerprint cluster set. In this set, the number of members and phase angle distribution of each cluster met the compactness and separation requirements for subsequent confidence calculations, significantly improving the accuracy and stability of subsequent benchmark map construction.

[0078] S3.4: For each cluster in the candidate benchmark phase fingerprint cluster set, calculate the distribution compactness and boundary separation index of the high-dimensional phase feature vector set within the cluster, and perform weight assignment operation based on the statistical confidence evaluation model to generate a weighted benchmark phase fingerprint cluster containing quantized confidence weights.

[0079] S3.5: Integrate all weighted benchmark phase fingerprint clusters containing quantized confidence weights, perform spatiotemporal index reconstruction processing in ascending order of phase angle, establish a multidimensional mapping relationship between phase angle and pixel offset, and finally generate a benchmark map for subsequent comparison.

[0080] The weighted benchmark phase fingerprint cluster set is subjected to unified data structuring processing, and the phase angle sequence within the cluster is bound to the corresponding pixel offset set as a multi-dimensional index item.

[0081] All weighted baseline phase fingerprint clusters are sorted in ascending order of phase angle values, and a spatiotemporal index linked list is established according to the sorting order to form a one-to-one correspondence between phase angle, timestamp, and pixel offset.

[0082] During the index reconstruction process, a phase angle interpolation algorithm is introduced to perform linear or spline interpolation to fill in the missing phase points between adjacent clusters, ensuring the continuity and queryability of the mapping relationship.

[0083] The phase angle and pixel offset mapping matrix after interpolation is arranged into a data table by rows, and a cluster confidence weight value is added to each row to form a mapping dataset containing three-dimensional information.

[0084] Normalization is used to map the phase angle, pixel offset, and confidence weight to a unified dimension, which makes it easier for subsequent matching algorithms to be directly called under different working conditions.

[0085] Through the above processing method, the weighted reference phase fingerprint cluster from the previous step is transformed into a searchable reference map, achieving a high-precision mapping effect from the phase angle to the pixel offset in real time.

[0086] For example, on a cable production line with a guide wheel diameter of 200mm and a wrapping pitch of 15mm, the reference data collected during the light-load no-load running phase includes 50 weighted reference phase fingerprint clusters. Each cluster contains a phase angle sequence with a resolution of 0.5° and a corresponding pixel offset range of ±3 pixels. During spatiotemporal index reconstruction, the phase angle interpolation step size is set to 0.25°, and cubic spline interpolation is used to complete missing points, forming a mapping matrix of size 720×2 (phase angle and pixel offset), with an additional confidence weight column to form a 720×3 mapping data table. For the interpolation completion process, the pixel offset value is calculated using the following spline interpolation formula: in, The coefficients are cubic spline coefficients. For phase angle interpolation points, This corresponds to the pixel offset. Through this interpolation and index reconstruction, the output reference map achieves a significant improvement in phase angle mapping accuracy in actual production, ensuring the accuracy of deterministic perturbation removal during subsequent real-time matching.

[0087] like Figure 3 As shown, step S4 involves mapping the real-time wrapping phase angle to the matched reference fingerprint sub-region in the reference map, performing local cross-correlation matching, and extracting the residual offset. Specifically, this includes: S4.1: Obtain the real-time wrapping phase angle and reference map, perform hash index lookup processing based on the phase angle value range, locate and output the reference fingerprint sub-region data block that strictly corresponds to the geometric state of the current frame.

[0088] The input conditions include the real-time wrapping phase angle value calculated in step S2 and the complete mapping structure of the baseline map constructed in step S3. Based on this phase angle value, interval positioning operations are performed within a preset phase domain, mapping the value to the corresponding phase index key to determine its corresponding phase interval. Using a pre-established hash index table, the phase interval is used as the input parameter of the hash function to calculate the hash value and search for the matching baseline fingerprint cluster identifier in the index table. Based on the found baseline cluster identifier, the physical storage address of the corresponding sub-region in the map data structure is located, and the data access interface is called to read the two-dimensional pixel offset matrix and associated statistical features of the sub-region. Integrity verification is performed on the read sub-region data block to confirm that its confidence weight meets the preset threshold and that the data structure is not damaged. Through the combined processing of hash index lookup and data block reading, the real-time wrapping phase angle is transformed into a baseline fingerprint sub-region data block that strictly corresponds to the geometric state of the current frame, realizing fast positioning and accurate output of the baseline region.

[0089] For example, in a pre-branched cable production line, during a single wrapping process, the real-time wrapping phase angle is 47.36°. The phase domain range of the reference spectrum is set to 0° to 360°, the phase interval resolution is 0.5°, and the hash index table contains 720 index keys. This real-time phase angle corresponds to the 94th interval after interval positioning calculation, and is input into the hash function. Get hash value The reference cluster identifier B217 was located in the index table. Based on identifier B217, the sub-region storage address 0x04F2_7A in the map data structure was located. The two-dimensional pixel offset matrix size was read as 128×128, with auxiliary statistical features including a mean offset of 0.12mm and a variance of 0.03mm². A confidence weight check was performed, and the value was 0.86, which is higher than the preset threshold of 0.80. The data structure check field passed CRC verification and was found to be correct. The output sub-region data block was used for the subsequent sliding window grayscale matrix alignment preprocessing in S4.2, effectively improving matching accuracy and computation speed, and ensuring a significant improvement in the removal effect of deterministic phase perturbation during the offset component extraction process.

[0090] S4.2: Receive the reference fingerprint sub-region data block and the original image texture data of the current frame, and perform pixel-level grayscale matrix alignment preprocessing based on the sliding window mechanism to generate a sequence of matching image pairs that eliminates axial displacement timing errors.

[0091] S4.3: Using the sequence of image pairs to be matched, perform a two-dimensional spatial domain convolution operation based on the normalized cross-correlation algorithm to construct a peak response surface of correlation coefficients that characterizes the distribution of image similarity.

[0092] The system receives a sequence of image pairs to be matched, pre-processed using a sliding window mechanism, as input. A correspondence matrix is ​​established along the row and column dimensions of the pixel grayscale matrix to define the matching search range. A matching response function for each pair of image blocks in the two-dimensional spatial domain is constructed based on a normalized cross-correlation algorithm. The normalization process ensures the consistency of the correlation coefficient calculation scale by normalizing the pixel grayscale values ​​according to a range standard. A two-dimensional convolution operation is performed on each pair of image blocks to calculate the normalized cross-correlation coefficient for the local region.

[0093] The normalized cross-correlation coefficient calculation results are arranged in a two-dimensional manner according to horizontal and vertical displacements to construct a correlation coefficient peak response surface data matrix. Matrix smoothing is applied to this response surface to reduce the influence of high-frequency noise, thereby enhancing the discernibility of peak positions. Local peak enhancement filtering is performed on the response surface to suppress mismatches that may be caused by non-global extrema, thus ensuring accurate positioning in subsequent peak coordinate extraction. Through a collaborative processing method of normalized cross-correlation algorithm and two-dimensional convolution operation, the sequence of image pairs to be matched is transformed into a correlation coefficient peak response surface representing the image similarity distribution, realizing the construction of quantitative basic data for deterministic phase perturbation.

[0094] S4.4: Analyze the extreme point coordinates of the peak response surface of the correlation coefficient, and perform fine reconstruction of the peak position based on the sub-pixel interpolation algorithm to calculate the comprehensive displacement vector field that characterizes the total amount of deterministic phase perturbation.

[0095] Receive the correlation coefficient peak response surface data output by the normalized cross-correlation algorithm, establish the extreme point coordinate detection matrix, and use it to extract the initial integer pixel coordinate pairs of the two-dimensional peak position.

[0096] Two-dimensional gradient operators are performed on the initial integer pixel coordinate pairs in the detection matrix to obtain the directional gradient distribution and amplitude gradient distribution of the peak neighborhood, providing gradient constraints for subsequent sub-pixel interpolation fitting.

[0097] A biquadratic surface fitting model is introduced to map the correlation coefficient values ​​of the peak neighborhood to the fitting function. The coefficients of the fitting function are solved based on the least squares method to achieve sub-pixel level peak position reconstruction.

[0098] By utilizing the analytical extremum condition of the subpixel fitting function in two-dimensional space, the precise row and column coordinates of the reconstructed peak are calculated, and the peak offset vector is generated by the difference between the coordinate difference and the original integer coordinates.

[0099] The peak offset vector and the peak response amplitude are jointly encoded into vector field elements. All peak positions are traversed to generate a comprehensive displacement vector field containing row offset, column offset and response amplitude, which serves as a quantization result characterizing the total amount of deterministic phase perturbation.

[0100] By using subpixel interpolation and surface fitting, the normalized cross-correlation response data is transformed into a spatially continuous comprehensive displacement vector field, enabling high-precision quantitative positioning of deterministic phase perturbations.

[0101] S4.5: Read the integrated displacement vector field and the original image edge coordinate data, and perform a deterministic component removal operation based on vector difference operation to generate the residual offset after removing periodic interference.

[0102] Step S5: Calculate the statistical moment features of the residual offset, including skewness, kurtosis, and phase span, and determine the deviation type label based on the distribution pattern of the statistical moment features to generate a process status classification result. Specifically, this includes: S5.1: Obtain the residual offset sequence extracted by local cross-correlation matching, and perform basic operations of the first and second moments on the residual offset sequence using the sliding window integration algorithm to generate a set of basic statistics including the mean center location and variance dispersion.

[0103] For the residual offset sequence extracted through local cross-correlation matching, continuous data segments are extracted point-by-point in the time domain using a fixed-length sliding window to form a batched window sample set. The data points in each window sample set are summed to obtain the sum of the values ​​within that window, which is then divided by the number of data points in the window to generate the corresponding first-order moment value, representing the mean center position of the deviation signal segment. For each data point in each window sample set, the mean center position of that window is subtracted, and the difference is squared. All squared differences are summed and divided by the number of data points in the window minus one to generate the corresponding second-order moment value, representing the variance dispersion of the deviation signal segment. The second-order moment values ​​in the calculation step are calculated using the following variance formula: in The number of data points in the window. For a single data point value, The mean center location is defined within the window. The first and second moment values ​​are stored as the mean and variance indices of the basic statistics set, respectively, for use in subsequent feature calculations such as the third and fourth normalized moment transformations. A set of basic statistics containing the mean center location and variance dispersion is generated using a sliding window integral algorithm. This transforms the residual offset sequence from the previous step into quantifiable statistical indicators characterizing the data distribution center and dispersion, thus achieving basic quantitative preparation before determining the type of bias.

[0104] For example, on a pre-branched cable insulation tape production line, the residual offset sequence after removing deterministic disturbances through cross-correlation matching is processed using a sliding window integral algorithm with a window length of 50 frames and a window sliding step size of 10 frames. For a certain window sample set, the data point sequence is [0.12, 0.15, 0.10, 0.09, 0.14, a total of 50 points]. The sum of these points is 6.40, which, divided by the number of data points (50), gives the mean center position as 0.128. Then, each data point is subtracted from the mean of 0.128, squared, and summed to obtain the sum of squared differences, which is 0.0263. Dividing this by 49 gives the variance dispersion as 0.000537. These mean and variance values ​​are recorded in the basic statistics set for subsequent skewness and kurtosis calculations. In another window, the mean center shifted to 0.145, and the variance dispersion increased to 0.000812, indicating a significant increase in the volatility of the deviation signal within this window. Through full-sequence sliding processing, the set of basic statistics covers the distribution characteristics of deviation signals across all segments of the entire production cycle, providing high-precision statistical support for deviation type determination and ensuring the accuracy of adaptive adjustment of dynamic correction parameters.

[0105] S5.2: Based on the mean center position in the set of basic statistics, perform a third-order normalized moment transformation on the residual offset sequence to quantify the degree of asymmetry of the data distribution relative to the mean and generate skewness characteristic parameters.

[0106] S5.3: Using the variance dispersion in the set of basic statistics, perform a fourth-order normalized moment transformation on the residual offset sequence to quantify the thickness of the tail of the data distribution and the sharpness of the peak and generate kurtosis characteristic parameters.

[0107] Based on the variance dispersion input object in the set of basic statistics, each residual offset in the sequence is selected as a calculation unit and centered to eliminate the interference of mean offset on the calculation of higher-order moments.

[0108] Perform a fourth-power operation on each centered offset component to construct the cumulative value of the fourth-order original moments and retain the sequence length for normalization processing.

[0109] By using a normalized denominator of the squared variance, the cumulative value of the fourth-order raw moments is divided by the cumulative value of the squared variance to obtain the kurtosis statistical index, which describes the thickness of the tail and the sharpness of the peak of the distribution.

[0110] The following standardized kurtosis calculation formula is introduced: Where n is the number of samples, and x is the value of the offset component. This is the mean of the offset components.

[0111] The calculation results are compared with the preset kurtosis threshold range. If the kurtosis value is higher than the upper limit, the data distribution is determined to have a sharp peak shape and may correspond to transient change-type deviation. If it is lower than the lower limit, it is determined to be a smooth distribution and may correspond to dispersed interference.

[0112] Through this fourth-order normalized moment transformation, the variance dispersion of the previous step is transformed into kurtosis characteristic parameters that quantitatively describe the tail thickness and peak sharpness, thereby achieving high-precision classification input for deviation morphology.

[0113] For example, on an insulation tape wrapping production line, a residual offset sequence of 1000 sampling points is collected, with a mean μ of 0.12 mm and a variance σ² of 0.005 mm². Subtracting the mean from each sample and raising the result to the fourth power, the summation yields 0.0048 mm. 4 With a total sample size n of 1000, 0.0048 is divided by the squared variance of 0.000025, and the kurtosis is calculated using the above formula, resulting in 192. This value is significantly higher than the preset sharpness threshold of 50. Based on this, the system classifies this condition as a "sharp peak abrupt change" deviation type. Subsequently, in S5.4, the deviation shape is further confirmed by combining skewness characteristics, and in S6, the tension abrupt change compensation model is invoked to implement feedforward correction.

[0114] S5.4: Combining the skewness characteristic parameter and the kurtosis characteristic parameter, calculate the effective support interval width of the residual offset sequence in the phase domain to characterize the dispersion range of the deviation signal in the phase space and generate the phase span characteristic parameter.

[0115] Joint analysis is performed on the kurtosis feature parameters generated by S5.3 and the skewness feature parameters generated by S5.2, and the residual offset sequence is selected as the operation object.

[0116] The residual offset sequence is normalized in the phase domain to ensure that the data value range is within a uniform range of [-1,1], so as to eliminate the interference of the difference in amplitude dimensions under different working conditions on the calculation of the support interval width.

[0117] A validity threshold is set for the normalized offset component sequence. The threshold amplitude is determined according to the matching judgment rules of the skewness feature parameter and the kurtosis feature parameter. For example, when the absolute value of skewness is large and the kurtosis is higher than the preset sharpness threshold, a wider threshold is taken, and vice versa.

[0118] Scan the normalized offset component sequence in the phase domain, mark all phase position index sets that exceed the validity threshold, and aggregate the phase position index sets in segments according to continuity.

[0119] The total span of the phase position set of each segment is calculated, and the phase span feature parameters are calculated. The phase span feature parameters of all segment sets are weighted and averaged. The weights are set according to the frequency or stability statistics of the corresponding segment in the sequence to form the global phase span feature parameters.

[0120] By using the above processing method, the skewness and kurtosis results from the previous step are transformed into quantitative indicators that characterize the dispersion range of the deviation signal in the phase space, thereby achieving accurate quantification of the deviation distribution width.

[0121] S5.5: Based on the multidimensional feature vector composed of the skewness feature parameter, kurtosis feature parameter and phase span feature parameter, perform a distribution pattern matching operation based on the rule engine to determine the deviation type label of single-peak concentrated type, double-peak symmetrical type or broadband diffuse type and generate process state classification results that characterize tension change, guide wheel movement or image anomaly.

[0122] The skewness feature parameters, kurtosis feature parameters, and phase span feature parameters generated from S5.2 to S5.4 are obtained and used to construct a three-dimensional feature vector as the input object of the rule engine.

[0123] The three-dimensional feature vector is input into a predefined set of deviation distribution shape matching rules, which contains quantitative judgment conditions for the combination relationship of skewness, kurtosis, and phase span.

[0124] Based on the skewness and absolute value range, the first type of judgment condition regarding the symmetry and centrality of the distribution in the matching rule set is used to identify the corresponding label of the unimodal centralized deviation distribution.

[0125] Based on the numerical value of kurtosis, the second type of judgment condition regarding tail thickness and peak sharpness in the matching rule set is used to identify the label corresponding to the bimodal symmetrical deviation distribution.

[0126] Based on the numerical range of the phase span, the third type of judgment condition in the matching rule set regarding the degree of signal dispersion in the phase space is used to identify the corresponding label of the broadband dispersion type deviation distribution.

[0127] By matching the combination relationships of skewness, kurtosis, and phase span with the morphological judgment conditions in the rule set one by one, the statistical moment features of the previous step are transformed into deviation type labels, thereby achieving an accurate mapping from deviation type to process state classification results.

[0128] For example, in the operation of a pre-branched cable production line with a traction speed of 1.2 m / s and an average tension of 320 N, the residual offset, after being statistically analyzed using a sliding window, yields a skewness of -0.15, a kurtosis of 4.2, and a phase span of 0.12 rad. The rule engine threshold condition is set as follows: an absolute skewness less than 0.3, a kurtosis greater than 3.8, and a phase span less than 0.15 rad, classifying it as a single-peak concentrated type. Inputting these characteristic parameters into the rule engine satisfies the first type of judgment condition, and the output deviation type label is "single-peak concentrated type," corresponding to the process state classification result of "tension abrupt change." The corresponding formula is calculated as follows: in These are the skewness characteristic parameter values; in This refers to the kurtosis characteristic parameter value; in This represents the phase span characteristic parameter value. After rule matching, all three conditions are met, and the deviation type label output by the rule engine is unimodal. The classification result will directly trigger the exponential decay model in step S6, significantly improving the dynamic response performance of the deviation correction control.

[0129] Step S6: Select the corresponding compensation model based on the process state classification results. If the problem is determined to be guide wheel axial movement, call the harmonic rejection model to generate a correction command filtered by mechanism matching. Specifically, this includes: S6.1: Obtain the process state classification result and the corresponding residual offset sequence generated in the previous step, and execute the conditional branch judgment logic based on the deviation type label in the process state classification result to generate the target compensation model identifier to be activated.

[0130] It should be noted that the harmonic elimination model is a compensation model used to handle process anomalies caused by guide wheel swaying. This model first calls the pre-stored guide wheel vibration mode database to extract the fundamental frequency and harmonic parameters that match the current wrapping phase angle. Then, it performs a fast Fourier transform on the residual offset sequence to construct an adaptive filter in the frequency domain, accurately eliminating the periodic harmonic components caused by the mechanical swaying of the guide wheel, thereby obtaining the guide wheel period compensation amount.

[0131] The process state classification results generated in the previous step are received as the main decision input, and the corresponding residual offset sequence is loaded as the core data carrier for the compensation model operation. The process state classification results are structured and parsed, and the deviation type label field is extracted and encoded to ensure the uniqueness and comparability of the type labels in subsequent conditional branch judgments. A conditional matching matrix is ​​constructed based on the deviation type label field, establishing a one-to-one mapping relationship between each type label and the identifier of the pre-set compensation model, and eliminating multi-value conflicts in the matrix to maintain the determinism of the matching logic. Matrix matching retrieval is performed on the deviation type labels, outputting the corresponding target compensation model identifier, and recording the identifier's calling priority and execution version information for downstream model loading and parameter matching. Through the above processing method based on label parsing and matrix matching, the static classification results of the previous step are transformed into callable dynamic compensation model identifiers, realizing the automation of the compensation model selection logic and the complete closure of the decision-making chain.

[0132] For example, in a startup scenario of a pre-branched cable running under light load, the process status classification result field shows the deviation type label code as "TC01", representing the tension mutation type; the residual offset sequence is a set of 100-millimeter-level lateral displacement measurements. The type label "TC01" is input into the conditional matching matrix, where the first row and first column correspond to the tension mutation model identifier "M_EXP_DECAY", the second row corresponds to the guide wheel axial movement model identifier "M_HARM_REMOVAL", and the third row corresponds to the image anomaly model identifier "M_IMG_DIAG". The matching result locates "M_EXP_DECAY", and the identifier field is appended with a call priority of 1 and an execution version number of V2.3. In this scenario, the target compensation model identifier "M_EXP_DECAY" will be called in sub-step S6.2, using the material stress relaxation time constant τ combined with the residual offset to perform transient response fitting. Assuming the time constant is set to 0.85 seconds, the compensation model will then call... Weight attenuation is performed, where t is the timestamp of the offset occurrence. The entire processing chain maintains logical consistency under different operating conditions. The guide wheel malfunction label "RW02" will match "M_HARM_REMOVAL", and the offset component will be cleared using the harmonic component removal formula; the image anomaly label "IM03" will match "M_IMG_DIAG", triggering the anomaly analysis submodule. Ultimately, this decision-making step significantly improved the accuracy of the compensation model in testing and ensured the consistency and traceability of inputs in subsequent compensation calculation stages.

[0133] S6.2: When the target compensation model identifier indicates a tension mutation type, the preset exponential decay model is invoked, and the residual offset sequence is subjected to transient response fitting using the material stress relaxation time constant to generate a tension transient compensation amount characterizing the elastic deformation recovery trend.

[0134] The system receives the target compensation model identifier output from the conditional branch decision logic, confirms the deviation type label as tension mutation, and extracts the residual offset sequence as the input object for response fitting. The exponential decay model is a compensation model used to handle tension mutation-type process anomalies.

[0135] The material stress relaxation time constant is read and used as the core parameter of the exponential decay model to construct the initial mathematical framework for transient response fitting.

[0136] The residual offset sequence is segmented along the time axis to ensure that the sampling timing of each offset component unit is complete and corresponds to the instantaneous changes in the process state.

[0137] The least squares fitting method is applied to match the offset component sequence with the exponential decay function in the time domain. The fitting function has the following form: Where A is the initial offset amplitude, t is the time after the offset occurs, and τ is the material stress relaxation time constant.

[0138] Based on the fitting results, the transient amplitude of the decay curve at each sampling time is extracted to form time series data characterizing the elastic deformation recovery trend.

[0139] The transient amplitude sequence and the residual offset sequence are differentially processed to obtain the tension transient compensation sequence.

[0140] By using the above-mentioned exponential decay fitting and differential calculation, the tension abrupt change deviation identified by the conditional branch is transformed into a quantifiable tension transient compensation amount, thereby realizing the dynamic prediction and correction of the elastic deformation recovery trend.

[0141] For example, in the production process of pre-branched cable insulation tape wrapping, the deviation type is determined to be a tension mutation, the residual offset sequence length is 120 points, the sampling interval is 0.05 seconds, and the material stress relaxation time constant τ is set to 1.2 seconds. The initial offset amplitude A is statistically calculated to be 0.45 mm. This sequence is input into an exponential decay fitting model, and the transient amplitude is calculated using the formula at each sampling time t. For example, at t=0.25 seconds, the transient amplitude is 0.45 mm. e^( (0.25 / 1.2) = 0.364 mm. Throughout the sequence, the transient amplitude exhibits a smooth decreasing trend. The tension transient compensation sequence generated after differential operation shows significant values ​​in the first 15 sampling points, reflecting the rapid recovery segment of elastic deformation. Executing this compensation amount injects a feedforward command into the correction control, stabilizing the wrapping gap, significantly reducing the image detection gap deviation, and significantly improving the system's dynamic response capability.

[0142] S6.3: When the target compensation model identifier indicates the guide wheel sway type, the pre-stored guide wheel vibration mode database is called to extract the fundamental frequency and harmonic parameters that match the current real-time wrapping phase angle. Adaptive harmonic component elimination operation is performed on the residual offset sequence to generate the guide wheel period compensation amount that characterizes the mechanical sway suppression effect.

[0143] Receive the target compensation model identifier generated by the conditional branch judgment logic and confirm that its indication is the guide wheel axial movement type, and load the pre-stored guide wheel vibration mode database as a reference for spectrum elimination.

[0144] The current real-time wrapping phase angle value is analyzed and used as an index key to locate the corresponding vibration mode record entry in the guide wheel vibration mode database. The fundamental frequency parameter and harmonic parameters contained in the record are then extracted.

[0145] The residual offset sequence is input into the frequency domain transformation module to perform a fast Fourier transform, converting the displacement signal from the time domain to the frequency domain so that it can be compared with the extracted fundamental and harmonic parameters.

[0146] An adaptive harmonic rejection filter is constructed based on the spectrum comparison results. The amplitude and phase information of the fundamental frequency and harmonics are introduced in the filter coefficient design process to ensure that the rejection operation only acts on the mechanical oscillation harmonic components.

[0147] The constructed adaptive harmonic elimination filter is used to perform filtering operations on the residual offset frequency domain signal, and the filtered signal is reconstructed back into the time domain displacement sequence through inverse fast Fourier transform.

[0148] The mechanical oscillation suppression effect index is calculated for the reconstructed displacement sequence, and the guide wheel period compensation amount characterizing the mechanical oscillation suppression effect is generated and output to the subsequent dimension normalization processing stage.

[0149] By calling the pre-stored guide wheel vibration mode database and performing adaptive harmonic component elimination calculation, the residual offset sequence from the previous step is transformed into the guide wheel period compensation amount, thereby achieving a targeted suppression effect on the guide wheel axial movement type deviation.

[0150] For example, on a pre-branched cable insulation tape layering production line, the guide wheel diameter is set to 120mm, the wrapping pitch constant is 15mm, and the guide wheel vibration mode database pre-stores the fundamental frequency of this specification guide wheel at a working condition phase angle of 45° as 12Hz and the second harmonic as 24Hz. The residual offset sequence length is set to 2048 points and the sampling frequency is set to 256Hz. The data is input into a Fast Fourier Transform module to obtain the frequency domain spectrum peaks corresponding to 12Hz and 24Hz, with amplitudes of 0.35mm and 0.12mm respectively, which are matched to the database fundamental and harmonic parameters. A band-stop filter is constructed, with the band-stop center frequencies set to 12Hz and 24Hz respectively, and a bandwidth of 0.5Hz. The filter coefficients are adjusted according to the peak amplitude and phase difference. The filtered frequency domain signal is transformed by inverse fast Fourier transform to obtain the time domain compensated displacement curve. The amplitude fluctuation is reduced to 0.05mm, and the mechanical oscillation suppression effect is significantly reduced from the original 0.21mm to 0.06mm. The output guide wheel period compensation effectively suppresses the periodic oscillation caused by the guide wheel movement.

[0151] S6.4: Based on the generated tension transient compensation amount or guide wheel period compensation amount, combined with the current traction speed and insulation tape wrapping pitch parameters, perform dimensional normalization and time domain alignment processing to generate an initial correction command feedforward amount with physical execution meaning.

[0152] S6.5: Apply actuator saturation limiting constraint and safety rate of change filtering to the feedforward amount of the initial correction command to eliminate high-frequency noise interference and prevent actuator overshoot, so as to finally generate a correction command that has been filtered by mechanism matching and can directly drive the execution unit.

[0153] Step S7: The correction command is injected into the actuator drive unit to complete the dynamic correction of the insulation tape layering gap and thickness, generating the corrected wrapping physical state. Specifically, this includes: S7.1: Obtain the correction command after mechanism matching and filtering, and use the digital-to-analog conversion module to perform quantization mapping processing on the correction command from discrete signal to continuous analog voltage signal to generate a correction drive analog voltage signal with an industrial control standard amplitude range.

[0154] S7.2: Receive the correction drive analog voltage signal, and perform decoupling operation on the torque component and excitation component of the correction drive analog voltage signal based on the servo motor vector control algorithm to generate a servo motor stator current reference vector characterizing the instantaneous electromagnetic thrust requirement.

[0155] The system receives the analog voltage signal for correction drive after it has been filtered by mechanism matching, and calls the vector control module of the servo drive controller to perform signal amplitude and phase analysis on the analog voltage signal to extract the instantaneous voltage component and phase offset in the drive voltage waveform.

[0156] The instantaneous voltage components obtained from the analysis are input into the Clarke coordinate transformation function to convert the voltage components in the three-phase stationary coordinate system into voltage vectors in the two-phase orthogonal stationary coordinate system for subsequent excitation and torque component control.

[0157] In a two-phase orthogonal stationary coordinate system, a Park coordinate transformation is performed on the voltage vector to generate voltage components in a synchronous rotating coordinate system, where the direct axis component represents the excitation component and the cross axis component represents the torque component.

[0158] Based on the electromagnetic parameter model of the servo motor, and using the voltage components and motor speed in the synchronous rotating coordinate system, the required stator current components are calculated using the motor state equation, as shown in the following formula: in, For voltage components, The back electromotive force component, For stator resistance, For inductance, This is the electrical frequency factor.

[0159] The direct-axis excitation current component and the cross-axis torque current component are merged into a unified stator current reference vector, and the amplitude is normalized and the phase is synchronized to match the input specifications of the drive control system.

[0160] Through the above processing method, the analog voltage signal of the correction drive in the previous step is converted into a servo motor stator current reference vector that can characterize the instantaneous electromagnetic thrust demand, thereby improving the instantaneous response capability of the drive control system to the mechanical output.

[0161] For example, on a pre-branched cable production line, the analog voltage signal for correction drive generated by S7.1 has an amplitude of 2.5V and a phase deviation of 15°. This signal undergoes Clarke transformation to obtain an α-axis voltage of 1.8V and a β-axis voltage of 1.6V. Then, through Park transformation, at a motor speed of 1500rpm, a direct axis voltage of 1.7V and a cross axis voltage of 1.5V are generated in the synchronous rotating coordinate system. Combining the motor stator resistance of 0.5Ω, inductance of 0.01H, electrical frequency factor of 314rad / s, and back electromotive force of 0.3V, the direct axis current is calculated to be 2.8A and the cross axis current to be 3.1A using the above formulas. This forms a stator current reference vector with a normalized amplitude of 3.5A and a phase of 12°. After this vector is input to the drive controller, the electromagnetic thrust response delay of the actuator is significantly shortened, and the lateral correction mechanical displacement accuracy of the insulation tape is improved to the micrometer level, ensuring stable control of the wrapping gap and thickness under dynamic operating conditions.

[0162] S7.3: Input the servo motor stator current reference vector, and perform switching duty cycle timing arrangement processing on the servo motor stator current reference vector through space vector modulation to generate a high-frequency pulse width modulation wave sequence for driving the power bridge arm.

[0163] The servo motor stator current reference vector output from the previous sub-step S7.2 is received as the input object for space vector pulse width modulation timing arrangement.

[0164] The stator current reference vector of the servo motor is decomposed into three symmetrical components and Clark transformation is performed to convert it into a stationary α–β coordinate system vector so as to uniformly handle duty cycle calculation in a two-dimensional plane.

[0165] The Park transformation is applied to the transformed α–β components to map them to the rotating dq coordinate system, and the target sector number and vector magnitude are determined by combining the real-time rotor position feedback.

[0166] Based on the target sector number and vector amplitude, the two-phase conduction time and zero vector time are determined using the spatial vector geometry construction method, and the switching duty cycle of each bridge arm is calculated using the following mathematical formula: Where D is the duty cycle, V is the target sector voltage component, and T is the sampling period. This is the DC bus voltage.

[0167] Quantization resolution alignment is applied to the duty cycle calculation results to map floating-point values ​​to the integer bit width range recognizable by the fixed-point PWM counter.

[0168] The timing arrangement module arranges the fixed duty cycle of each bridge arm according to the switching action sequence within the sampling period to generate a high-frequency pulse width modulation wave sequence, which is then output to the power drive port to achieve effective control of the bridge arm.

[0169] Through the aforementioned duty cycle geometry construction, resolution alignment, and timing arrangement processing, the servo motor stator current reference vector from the previous step is transformed into a high-frequency pulse width modulation wave sequence that can directly drive the power bridge arm, thereby achieving accurate vector voltage synthesis and high-speed response control of the actuator.

[0170] For example, in a pre-branched cable wrapping production line, the servo motor stator current reference vector received in this step is (2.8A, 1.6A, ...). 4.4A). After performing the Clark transform, the α component is 2.3A and the β component is 2.0A. Based on this, performing the Park transform and combining it with the rotor position of 72°, the d component in the dq coordinate system is 3.1A and the q component is 0.8A. According to the spatial vector geometry construction method, the target sector is determined to be the 3rd sector. The calculated two-phase conduction times are 150μs and 120μs, respectively, and the zero vector time is 30μs. Combining the DC bus voltage of 680V and the sampling period of 300μs, according to the formula... The three-phase duty cycles were calculated to be 0.53, 0.48, and 0.45, respectively. After fixed-point mapping, these values ​​were converted into integer values ​​within the bit width of a PWM counter and arranged in sector order to generate a high-frequency pulse width modulation waveform sequence. This waveform sequence, when output to the servo drive arm, achieved a mechanical axial displacement speed of 2.4 mm / ms at the linear actuator end. This verified the high-speed response characteristics of the combination of duty cycle arrangement and space vector modulation, which can significantly improve the real-time performance and accuracy of dynamic correction.

[0171] S7.4: Apply the high-frequency pulse width modulation wave sequence to the linear actuator drive unit, and use the principle of electromagnetic induction to convert the high-frequency pulse width modulation wave sequence into a mechanical axial displacement output, so as to generate a mechanical displacement of the insulating tape with micron-level positioning accuracy for lateral correction.

[0172] The high-frequency pulse width modulation wave sequence generated by S7.3 is received as an input signal and sent to the power bridge arm control interface of the linear actuator drive unit to trigger the turn-on and turn-off timing actions of the drive power device.

[0173] Inside the drive unit, the duty cycle of the high-frequency pulse width modulation wave sequence is analyzed in real time using the principle of electromagnetic induction, the instantaneous current change rate in the stator winding is calculated, and the corresponding electromagnetic force component is derived.

[0174] The analyzed electromagnetic force components are coupled with the motion parameter model of the actuator, using formulas. Calculate the linear thrust acting on the mover, where B is the magnetic flux density, I is the winding current, and L is the effective conductor length.

[0175] Based on the calculated instantaneous thrust value, and combining the friction coefficient of the linear actuator guide rail and the mass parameters of the mover, the formula of Newton's second law is applied. in, For instantaneous acceleration, For linear thrust, The mass of the mover is calculated by integrating the instantaneous acceleration over the sampling period to obtain the instantaneous velocity, and then performing a second integration to obtain the mechanical displacement, ensuring that the displacement output is strictly synchronized with the high-frequency modulation characteristics of the drive signal.

[0176] Through the above processing method, the high-frequency pulse width modulation wave sequence of the previous step is converted into a mechanical displacement of the insulating tape with micron-level positioning accuracy for lateral correction, thereby realizing the instantaneous physical adjustment of the actuator during the wrapping process.

[0177] For example, in a pre-branched cable production line, the mass of the linear actuator mover is set to 2.5 kg, the guide rail friction coefficient is measured to be 0.015, the magnetic flux density B is 0.8 T, the effective conductor length L is 0.12 m, and the PWM wave sequence duty cycle varies from 45% to 55%, corresponding to a winding current I variation range of 3.2 A to 3.9 A. Substituting B = 0.8 T, I = 3.5 A, and L = 0.12 m into the formula... The thrust F = 0.336 N was obtained. Substituting F = 0.336 N and the mover mass m = 2.5 kg into Newton's second law, the instantaneous acceleration a = 0.1344 m / s² was obtained. Integrating the acceleration once with a sampling period of 2 ms yielded an instantaneous velocity of 0.0002688 m / s, and a second integration yielded a displacement of approximately 0.5376 μm, directly driving the synchronous correction of the lateral position of the insulation tape. Under the action of a periodic pulse width modulation wave sequence, the cumulative displacement achieved micron-level adjustment, significantly improving the accuracy of wrapping gap and thickness control, and eliminating dynamic deviations caused by sudden tension changes or guide wheel movement.

[0178] S7.5: Monitor the wrapping site under the action of the lateral correction mechanical displacement of the insulation tape, and integrate the lateral correction mechanical displacement of the insulation tape into the running insulation tape layer wrapping trajectory in real time through geometric superposition effect, so as to generate the corrected wrapping physical state after eliminating dynamic deviation.

[0179] The lateral mechanical displacement data of the insulation tape generated by the preceding sub-step S7.4 is obtained. Based on the real-time observation system formed by the combination of high-precision displacement sensors and industrial cameras installed at the wrapping site, the position and geometric status information of the insulation tape during the operation process are collected.

[0180] The lateral mechanical displacement of the insulating tape is matched with the real-time collected data of the stacked and wrapped trajectory of the insulating tape. A multidimensional homogeneous transformation matrix is ​​used to unify the coordinate system of the two types of data to eliminate sensor installation errors and imaging parallax effects.

[0181] Within the registered coordinate system, the wrap-around trajectory geometric superposition algorithm is invoked to add the tangential and normal components of the lateral correction mechanical displacement on the trajectory curve to the original trajectory vector field, thereby forming an instantaneous trajectory vector set of fused correction displacement.

[0182] A continuous curvature smoothing operation is performed on the instantaneous trajectory vector set, and a third-order B-spline interpolation function is used to fit the trajectory curve to ensure a smooth transition in geometric continuity and tension balance in the overlapping and wrapping action.

[0183] The gap and thickness parameters of the smoothed and fitted trajectory within each pitch cell are calculated to verify whether the corrected trajectory meets the preset process control tolerance range. By comparing the offset of the trajectory before and after correction in the phase domain, it is confirmed that the dynamic deviation has been effectively eliminated.

[0184] By using the above-mentioned geometric superposition and smoothing methods, the lateral correction mechanical displacement of the previous step is transformed into corrected trajectory data, thereby realizing dynamic correction of the gap and thickness of the insulation tape layer wrapping.

[0185] For example, in a pre-branched cable wrapping production line operating at a traction speed of 2.5 m / s, a wrapping pitch of 15 mm, and a guide wheel diameter of 240 mm, a lateral correction mechanical displacement of 0.12 mm was collected. Trajectory data was simultaneously acquired using a 0.01 mm resolution laser displacement sensor and a 500 fps industrial camera installed at the wrapping site. After homogeneous transformation matrix registration, the displacement data was mapped to a trajectory vector field, with a tangential component of 0.08 mm and a normal component of 0.07 mm. A geometric superposition algorithm was used to fuse the displacement data, and a third-order B-spline smoothing curve was applied. The gap of the corrected trajectory was recalculated in units of pitch, reducing it from 0.85 mm to 0.73 mm, and the thickness fluctuation from 0.05 mm to 0.02 mm. Comparing the offset before and after correction in the phase domain, the change in the offset curve across the entire phase range was reduced to approximately one-third of its original value. Using this method, the wrapping stability of the production line under the above operating conditions is significantly improved. Field verification shows that the gap and thickness of the insulation tape are stable within the control threshold, meeting the accuracy and stability requirements for long-term operation.

[0186] Step S8: Monitor the original dataset of the new synchronous motion under the corrected wrapping physical state, and update the confidence weights in the baseline spectrum. Specifically, this includes: S8.1: Obtain the newly acquired synchronous motion raw dataset under the corrected wrapping physical state, and perform time series sliding window truncation processing on the real-time wrapping phase angle in the synchronous motion raw dataset to generate a phase angle time series segment containing the current moment and a historical preset duration window.

[0187] S8.2: Based on the phase angle time segment, perform a neighbor cluster search operation in the reference map, and call the Gaussian kernel density estimation algorithm to calculate the Mahalanobis distance between the real-time wrapping phase angle and the center of the existing reference phase fingerprint cluster, so as to generate a dynamic drift quantification index characterizing the degree of deviation of the current working condition.

[0188] S8.3: Perform exponentially weighted moving average filtering based on the dynamic drift quantification index, and construct an adaptive forgetting factor function by combining the preset equipment temperature rise rate threshold and the strip elastic creep time constant to generate dynamic weight update coefficients for adjusting the contribution rate of historical data.

[0189] S8.4: Utilize the dynamic weight update coefficients to perform a recursive Bayesian update operation on the original confidence weights of the matching benchmark fingerprint sub-regions in the benchmark map, so as to generate updated confidence weights that reflect the latest state characteristics of the device after the drift of slowly changing factors.

[0190] S8.5: Based on the updated confidence weights, perform renormalization on the overall topology of the reference map and remove outdated reference phase fingerprint clusters with confidence levels below a preset validity threshold to generate a new version of the reference map that has been dynamically adapted.

[0191] 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.

[0192] 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.

[0193] 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 method for online detection and correction of the gap and thickness of the insulation tape layering in pre-branched cables, specifically comprising: S1: Acquire the traction speed, tension parameters and image acquisition timing signals during the wrapping process of the pre-branched cable insulation tape as multi-source signals, and bind the multi-source signals with timestamps to generate a synchronous motion raw dataset; S2: Obtain the wrapping pitch and guide wheel diameter parameters, and perform reverse calculation based on the periodic structure of the texture of continuous multi-frame images in the original synchronous motion dataset, the wrapping pitch and guide wheel diameter parameters to generate the real-time wrapping phase angle; S3: Utilize the initial light-load no-load running process during the start-up of the production shift to collect standard-length cable image sequences and multi-source motion parameters. Through automatic clustering, construct a benchmark phase fingerprint cluster containing confidence weights from the standard-length cable image sequences and the multi-source motion parameters to generate a benchmark map. S4: Map the real-time wrapping phase angle to the matching reference fingerprint sub-region in the reference map, perform local cross-correlation matching, and extract the residual offset; S5: Calculate the statistical moment features of the residual offset, including skewness, kurtosis and phase span, and determine the deviation type label based on the distribution pattern of the statistical moment features to generate process status classification results; S6: Select the corresponding compensation model based on the process state classification results. If it is determined to be guide wheel axial movement, call the harmonic rejection model to generate a correction command that has been filtered by mechanism matching.

2. The method for online detection and correction of the gap and thickness of the insulation tape layering in a pre-branched cable according to claim 1, characterized in that, The process of adding steps S7-S8 after step S6 includes: S7: Inject the correction command into the actuator drive unit to complete the dynamic correction action on the gap and thickness of the insulation tape layer wrapping, and generate the corrected wrapping physical state. S8: Monitor the original dataset of the new synchronous motion under the corrected wrapping physical state and update the confidence weights in the baseline graph.

3. The method for online detection and correction of the gap and thickness of the insulation tape layering in a pre-branched cable according to claim 1, characterized in that, Step S4 specifically includes: The real-time wrapping phase angle and the reference map are obtained. Based on the phase angle value range, a hash index lookup process is performed to locate and output the reference fingerprint sub-region data block that strictly corresponds to the geometric state of the current frame. Receive the reference fingerprint sub-region data block and the original image texture data of the current frame, perform pixel-level grayscale matrix alignment preprocessing based on the sliding window mechanism, and generate a sequence of image pairs to be matched; Using the sequence of image pairs to be matched, a two-dimensional spatial domain convolution operation is performed based on the normalized cross-correlation algorithm to construct a peak response surface of the correlation coefficient; The extreme point coordinates of the peak response surface of the correlation coefficient are analyzed, and the peak position is finely reconstructed based on the sub-pixel interpolation algorithm to obtain the comprehensive displacement vector field. Obtain the original image edge coordinate data. Based on the comprehensive displacement vector field and the original image edge coordinate data, perform a deterministic component removal operation through vector difference operation to generate the residual offset.

4. The method for online detection and correction of the gap and thickness of the insulation tape layering in a pre-branched cable according to claim 1, characterized in that, Step S5 specifically includes: The residual offset is obtained, and the first and second moments are performed on the residual offset sequence using the sliding window integration algorithm to generate a set of basic statistics. Based on the mean center position in the set of basic statistics, the residual offset sequence is subjected to a third-order normalized moment transformation to generate skewness characteristic parameters. Using the variance dispersion in the set of basic statistics, a fourth-order normalized moment transformation is performed on the residual offset sequence to generate kurtosis characteristic parameters. By combining the skewness characteristic parameter and the kurtosis characteristic parameter, the effective support interval width of the residual offset is calculated to generate the phase span characteristic parameter; Based on the skewness feature parameter, the kurtosis feature parameter, and the phase span feature parameter, a multidimensional feature vector is constructed. A distribution pattern matching operation based on a rule engine is performed on the multidimensional feature vector to determine the deviation type label of single-peak concentrated type, double-peak symmetrical type, or broadband diffuse type, and the process state classification result is generated.

5. The method for online detection and correction of the gap and thickness of the insulation tape layering in a pre-branched cable as described in claim 1, characterized in that, The acquisition of the multi-source signals includes: acquiring the cable traction displacement pulse sequence through an encoder, acquiring the tension analog voltage signal through a tension sensor, and achieving microsecond-level data alignment by using a high-frequency clock to unify the sampling reference.

6. The method for online detection and correction of the gap and thickness of the insulation tape layering in a pre-branched cable as described in claim 1, characterized in that: By interpolating and matching the timestamps of exposure interruptions in industrial cameras with displacement and tension data, instantaneous kinematic state extraction that is strictly synchronized with each frame of image data is achieved. The image frame data is processed by a gray-level co-occurrence matrix algorithm to extract edge textures, and then by a fast Fourier transform to obtain the fundamental frequency and harmonic components.

7. The method for online detection and correction of the gap and thickness of the insulation tape layering in a pre-branched cable as described in claim 1, characterized in that: Based on the hash index lookup of the real-time phase angle in the reference map, the sliding window grayscale matrix alignment and normalized cross-correlation algorithm are introduced in the region matching to support the extraction of displacement at the sub-pixel level and further remove deterministic periodic disturbances caused by guide wheel eccentricity, lens distortion, etc., and output the residual offset.

8. The method for online detection and correction of the gap and thickness of the insulation tape layering in a pre-branched cable as described in claim 1, characterized in that, The calculation of statistical moment features specifically includes: performing first-, second-, third-, and fourth-order normalized moment transformations on the residual offset sequence using a sliding window to realize the full calculation of statistical features such as the mean, variance, skewness, and kurtosis of the sequence; and combining the phase domain normalization method to obtain the diffusion width of the deviation signal in the phase space to determine the deviation category such as single-peak, double-peak, or broadband diffusion type.

9. The method for online detection and correction of the gap and thickness of the insulation tape layering in a pre-branched cable as described in claim 1, characterized in that: The rule engine automatically matches a preset deviation pattern template based on statistical features, maps multidimensional feature vectors to process states such as tension mutation, guide wheel movement or image anomaly, and outputs category labels to drive the compensation module to enter different correction branches.

10. The method for online detection and correction of the gap and thickness of the insulation tape layering in a pre-branched cable as described in claim 1, characterized in that: For the sudden tension condition, an exponential decay model of the material stress relaxation time constant parameter is adopted. For the guide wheel axial movement condition, the built-in vibration mode database is called to suppress mechanical periodic disturbances in real time through an adaptive harmonic elimination algorithm. The compensated correction command is then input into the servo or linear motor to realize the physical action.

Images