An online defect detection and self-adaptive deviation correction method and system for aluminum enameled flat wire

Abstract

The application provides an aluminum enameled flat wire online defect detection and self-adaptive correction method and system, relates to the enameled wire defect detection technical field, and comprises the following steps: immersing the aluminum enameled flat wire into a fluid with a set viscosity, collecting a reference fluid resistance signal and a reference electromagnetic signal; real-time detecting the fluid resistance signal and the electromagnetic induction signal; positioning the defect position of the surface defect of the enameled wire; identifying the type and distribution of the internal defect of the aluminum enameled flat wire; distinguishing the defect type through a machine learning algorithm; outputting a corresponding correction scheme through a pre-trained neural network model; and outputting a correction control signal to a correction module for correction operation. Through synchronous collection and analysis of the fluid resistance signal and the electromagnetic induction signal, the application realizes synchronous detection of surface defects and internal defects, solves the limitations of traditional single detection means, forms a high-precision three-stage classification system, generates an interpretable report, and significantly improves the classification accuracy of composite defects.

Description

Technical Field

[0001] This invention relates to the field of enameled wire defect detection technology, and in particular to an online defect detection and adaptive correction method and system for aluminum enameled flat wire. Background Technology

[0002] In the field of enameled wire manufacturing, enameled wire serves as a core conductive component in electrical equipment such as motors and transformers, and the quality of its surface insulation coating directly affects the safety and reliability of equipment operation. The production process of enameled wire involves multiple steps, including wire drawing, annealing, coating, and baking. The insulation coating formed after the coating has cured may contain microscopic defects such as bubbles, pinholes, impurities, uneven thickness, or film damage. These defects are typically on the order of micrometers to sub-millimeters and are randomly distributed. Traditional manual visual inspection methods are inefficient and subjective, making them unsuitable for large-scale production. With the improvement of industrial automation, machine vision-based defect detection technology for enameled wire has gradually become a research focus.

[0003] In the field of aluminum enameled flat wire manufacturing, with the increasing demand for lightweight and integrated conductor materials in high-end equipment such as new energy vehicle drive motors and high-efficiency transformers, aluminum enameled flat wire is gradually replacing traditional copper enameled wire as a key conductive component due to its light weight, excellent conductivity, and cost advantages. The production process of aluminum enameled flat wire involves core steps such as drawing, annealing and softening, multi-layer insulating varnish coating, and high-temperature curing. Its flat cross-sectional structure and the characteristics of the aluminum substrate make the varnish film susceptible to mechanical stress, thermal field distribution, and differences in metal ductility during curing, resulting in microscopic defects such as scratches, bubbles, missed coatings, edge burrs, or uneven varnish film thickness. To address these issues, online defect detection and adaptive correction technology have become core components of automated aluminum enameled flat wire production.

[0004] Conventional online defect detection and correction methods for enameled aluminum flat wire typically require separate detection of surface and internal defects, resulting in low efficiency and a tendency to miss complex defects, making it difficult to meet the demands of high-precision quality control. Existing correction schemes often employ fixed parameter adjustments, failing to consider the differences in wire defect types and dynamic deformations. This leads to delayed or over-adjusted correction actions, which can easily cause secondary damage or correction failure, making it difficult to meet the real-time accuracy requirements of high-speed production scenarios.

[0005] To address the shortcomings of the existing technology, this technical solution proposes an online defect detection and adaptive correction method and system for aluminum enameled flat wire. Summary of the Invention

[0006] This invention provides an online defect detection and adaptive correction method and system for aluminum enameled flat wire, which addresses the shortcomings of existing technologies.

[0007] On one hand, the present invention provides an online defect detection and adaptive correction method for aluminum enameled flat wire, comprising:

[0008] S1: Immerse the aluminum enameled flat wire into a fluid of a set viscosity and apply a high-frequency alternating current to collect the reference fluid resistance signal and the reference electromagnetic signal;

[0009] S2: Based on the reference fluid resistance signal and the reference electromagnetic signal, the aluminum enameled flat wire is passed through the detection area at a constant speed, and the fluid resistance signal and electromagnetic induction signal are detected in real time.

[0010] S3: Based on the fluid resistance signal and combined with the reference fluid resistance signal, locate the defect position of the surface defect of the enameled wire and output the surface defect data; based on the defect position and combined with the reference electromagnetic signal, analyze the spectral characteristics of the electromagnetic induction signal, identify the type and distribution of internal defects of the aluminum enameled flat wire, and output the internal defect data.

[0011] S4: Perform spatiotemporal matching of surface defect data and internal defect data, distinguish defect types using machine learning algorithms, establish a defect feature matrix based on defect types, and output corresponding correction schemes through a pre-trained neural network model;

[0012] S5: Based on the correction scheme, output the correction control signal to the correction module to perform the correction operation.

[0013] According to the present invention, an online defect detection and adaptive correction method for aluminum enameled flat wire is provided. Step S1, the step of acquiring a reference fluid resistance signal and a reference electromagnetic signal, includes:

[0014] S1.1: The aluminum enameled flat wire is vertically immersed into the detection fluid in the constant temperature control box through the guide wheel, and the reference fluid resistance signal at each axial position is collected by a distributed fluid pressure sensor array;

[0015] S1.2: Apply a standard frequency alternating current through a high-frequency power generator. The current intensity is automatically matched to preset parameters based on the cross-sectional area of ​​the wire.

[0016] S1.3: Based on preset parameters, a reference electromagnetic signal is collected in the stationary state by a ring electromagnetic induction coil group.

[0017] According to the present invention, an online defect detection and adaptive correction method for aluminum enameled flat wire is provided. Step S2, the step of real-time detection of fluid resistance signals and electromagnetic induction signals, includes:

[0018] S2.1: Control the precision traction mechanism to move the enameled wire at a constant linear speed, record the dynamic resistance signal of the fluid pressure sensor in real time during the movement, and use a sliding window algorithm to eliminate mechanical vibration noise;

[0019] S2.2: Synchronously acquire the time-varying induced signal of the electromagnetic induction coil under motion conditions;

[0020] S2.3: Perform time-domain synchronous marking processing on dynamic resistance signals and time-varying induction signals to establish a precise mapping relationship with the physical position of the wire, and output fluid resistance signals and electromagnetic induction signals.

[0021] According to the online defect detection and adaptive correction method for aluminum enameled flat wire provided by the present invention, step S3, the step of performing surface defect detection includes:

[0022] S3.1: Multi-level digital filtering is used to eliminate vibration noise of the traction mechanism, and an effective signal segment is extracted through an adaptive threshold algorithm to establish a resistance position mapping model;

[0023] S3.2: Based on the fluid dynamics model and the resistance location mapping model, the fluid resistance signal is converted into a three-dimensional surface profile; the ideal surface model is reconstructed using the moving least squares method, the deviation matrix between the three-dimensional surface profile and the ideal surface model is calculated, and anomalies are identified through curvature analysis to obtain the abnormal regions;

[0024] S3.3: Based on the abnormal region, calculate the defect feature parameters and establish the defect feature vector;

[0025] S3.4: Using a pre-trained random forest model, the defect feature vectors are mapped to specific defect types, and a list of defects with spatial coordinates is output.

[0026] According to the present invention, an online defect detection and adaptive correction method for aluminum enameled flat wire is provided, wherein step S3 includes the following steps for internal defect detection:

[0027] S3.5: Perform synchronous compression transformation on the electromagnetic induction signal, separate the skin effect and defect response in the time and frequency domain, use singular value decomposition to eliminate baseline drift, and extract transient characteristic components that are synchronized with the wire movement.

[0028] S3.6: Establish an electromagnetic-thermal-mechanical coupling model, and decompose the transient characteristic components into conductor defect response and insulation defect response;

[0029] S3.7: Construct a forward model based on Maxwell's equations. According to the conductor defect response and insulation defect response, use the Monte Carlo optimization algorithm to invert the conductor cross-sectional area change and insulation layer thickness fluctuation parameters, and output internal defect data.

[0030] According to the online defect detection and adaptive correction method for aluminum enameled flat wire provided by the present invention, step S4, the step of performing spatiotemporal matching includes:

[0031] S4.1: Spatial registration is performed between the coordinate data in the surface defect data and the distribution cloud map in the internal defect data to establish a high-precision mapping relationship;

[0032] S4.2: Analyze the internal region below the coordinate data, compare the correlation between electromagnetic features and surface morphology, and identify associated defects; when the distance between the coordinate data and the distribution cloud map is less than a set threshold, it is determined to be an associated composite defect.

[0033] According to the present invention, an online defect detection and adaptive correction method for aluminum enameled flat wire is provided. In step S4, the step of distinguishing defect types using a machine learning algorithm includes:

[0034] S4.3: Based on the coordinate data and distribution cloud map, construct the spatiotemporal correlation feature matrix and generate the fused feature vector;

[0035] S4.4: Load the pre-trained three-stage classification model based on machine learning algorithm, input the fused feature vector into the three-stage classification model to classify defect types, and output the classification result;

[0036] S4.5: Based on surface defect data and internal defect data, construct a three-dimensional feature matrix according to the defect type based on the classification results;

[0037] S4.6: Use SHAP values ​​to evaluate the contribution of each feature and dynamically weight key features; perform feature importance analysis on the three-dimensional feature matrix and generate an interpretability analysis report.

[0038] According to the online defect detection and adaptive correction method for aluminum enameled flat wire provided by the present invention, step S4.4, the step of outputting the classification result includes:

[0039] S4.4.1: Load the pre-trained random forest model, perform coarse-grained classification on the fused feature vectors, and output low-confidence samples;

[0040] S4.4.2: Call the 1D-CNN sub-model to perform dynamic feature enhancement on low-confidence samples and output the corrected probability distribution;

[0041] S4.4.3: Based on the modified probability distribution, a graph neural network is used to process the correlation defects of low-confidence samples and output the classification results.

[0042] According to the present invention, an online defect detection and adaptive correction method for aluminum enameled flat wire is provided. Step S5, the step of performing the correction operation, includes:

[0043] S5.1: Calculate the required correction action parameters based on the defect type and defect location;

[0044] S5.2: Based on the correction action parameters, the action parameters of the multi-axis actuator are adjusted by the PID controller to drive the radial pressure roller to adjust the wire posture; the torsion mechanism is controlled to compensate for the axial deformation of the wire in real time and dynamically track the deflection change of the wire.

[0045] S5.3: Re-introduce the corrected cable into the testing area for verification.

[0046] This invention also provides an online defect detection and adaptive correction system for aluminum enameled flat wire, comprising:

[0047] The fluid detection module is used to immerse the aluminum enameled flat wire into a fluid of a set viscosity and collect a reference fluid resistance signal; based on the reference fluid resistance signal, the aluminum enameled flat wire is passed through the detection area at a constant speed to detect the fluid resistance signal in real time.

[0048] The electromagnetic detection module is used to apply a high-frequency alternating current to the aluminum enameled flat wire in the fluid and collect a reference electromagnetic signal; based on the reference electromagnetic signal, the electromagnetic induction signal is detected in real time.

[0049] The defect location module is used to locate the surface defects of enameled wire based on fluid resistance signals and reference fluid resistance signals; and based on the defect location and reference electromagnetic signals, analyze the spectral characteristics of electromagnetic induction signals to identify the type and distribution of internal defects in aluminum enameled flat wires.

[0050] The defect identification module is used to perform spatiotemporal matching of surface defect data and internal defect data, and to distinguish defect types through machine learning algorithms;

[0051] The error correction decision module is used to establish a defect feature matrix based on the defect type and output the corresponding error correction scheme through a pre-trained neural network model.

[0052] The correction execution module is used to output correction control signals to the correction module for correction operations based on the correction scheme.

[0053] This invention provides an online defect detection and adaptive correction method and system for enameled aluminum flat wire. By synchronously acquiring and analyzing fluid resistance signals and electromagnetic induction signals, it achieves simultaneous detection of surface and internal defects, overcoming the limitations of traditional single detection methods. An ideal surface model is reconstructed using the moving least squares method, combined with curvature analysis to identify abnormal regions. Electromagnetic signals undergo synchronous compression transformation and singular value decomposition to separate noise, significantly improving the detection sensitivity of minute defects. A random forest model enables rapid coarse-grained classification, 1D-CNN dynamically enhances the features of low-confidence samples, and a graph neural network processes correlated defects, forming a high-precision three-stage classification system. Combined with SHAP value feature importance analysis, an interpretable report is generated, significantly improving the classification accuracy of complex defects. By generating a correction scheme based on defect type, location, and spatial mapping relationships, a PID controller dynamically adjusts the multi-axis actuator to compensate for axial deformation and deflection changes in the wire in real time. A post-correction re-inspection mechanism is introduced to form a closed-loop control, ensuring the effectiveness of defect correction. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0055] Figure 1 This is a flowchart of an online defect detection and adaptive correction method for aluminum enameled flat wire provided in an embodiment of the present invention;

[0056] Figure 2 This is a step-by-step diagram of how machine learning algorithms distinguish defect types;

[0057] Figure 3 This is a schematic diagram of an online defect detection and adaptive correction system for aluminum enameled flat wire provided in an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0059] Example 1:

[0060] The following is combined with Figures 1-3This invention describes an online defect detection and adaptive correction method and system for aluminum enameled flat wire.

[0061] like Figures 1-2 As shown in the figure, an online defect detection and adaptive correction method for aluminum enameled flat wire provided by an embodiment of the present invention includes:

[0062] S1: Immerse the aluminum enameled flat wire in a fluid of a set viscosity and apply a high-frequency alternating current to collect reference fluid resistance signals and reference electromagnetic signals. The steps include:

[0063] S1.1: The aluminum enameled flat wire is vertically immersed in the detection fluid within the temperature-controlled chamber via guide wheels. These guide wheels are typically high-precision ceramic wheels with a polished surface roughness less than a set threshold, ensuring no lateral displacement of the wire during vertical immersion. The concentricity error of the guide wheels is controlled to the micrometer level using laser calibration to prevent wire twisting due to wheel deformation. A distributed fluid pressure sensor array is used to collect reference fluid resistance signals at each axial position. During fluid resistance signal acquisition, the fluid resistance on the surface of the aluminum enameled flat wire is calculated and expressed as:

[0064]

[0065] Among them, F f The fluid resistance per unit length of the wire is given by μ, and the fluid viscosity is given by A. c The effective contact area between the wire and the fluid. Let v0 be the average velocity gradient of the fluid on the surface of the flat wire, v0 represent the velocity of the fluid, and y represent the coordinate direction perpendicular to the surface of the aluminum enameled flat wire.

[0066] S1.2: A standard frequency alternating current is applied via a high-frequency power generator, and the current intensity is automatically matched to preset parameters based on the wire cross-sectional area. The power generator automatically adjusts the current frequency and amplitude based on the real-time cross-sectional area of ​​the wire (measured online via a laser diameter gauge) to ensure that the current density is within the effective detection range, while avoiding the risk of paint film breakdown.

[0067] S1.3: Based on preset parameters, a reference electromagnetic signal in a stationary state is acquired using a ring-shaped electromagnetic induction coil group. The coil group employs a multi-turn tightly wound structure, with the coil spacing matching the wire width and the center concentric with the wire's movement trajectory to reduce edge effects. The coil is protected by a high-permeability shield to reduce external electromagnetic interference. The formula for acquiring the reference electromagnetic signal in a stationary state using the ring-shaped electromagnetic induction coil group is expressed as:

[0068]

[0069] Where ε is the induced electromotive force of a single-turn coil, N is the number of turns in the coil winding, B is the magnetic induction intensity generated by the coil, L is the effective detection length of the wire, and v1 is the constant moving speed of the wire.

[0070] S2: Based on the reference fluid resistance signal and the reference electromagnetic signal, the aluminum enameled flat wire is passed through the detection area at a constant speed, and the fluid resistance signal and electromagnetic induction signal are detected in real time. The steps include:

[0071] S2.1: The precision traction mechanism moves the enameled wire at a constant linear velocity. The traction mechanism employs a servo motor and ball screw closed-loop control. An encoder provides real-time feedback of the linear velocity, and a PID algorithm dynamically adjusts the rotational speed to ensure speed fluctuations are less than the set accuracy. The dynamic resistance signal from the fluid pressure sensor is recorded in real-time during the movement, and a sliding window algorithm is used to eliminate mechanical vibration noise. The PID algorithm for the servo motor and ball screw closed-loop control of the traction mechanism is expressed as follows:

[0072]

[0073] Where u(t) is the servo motor control voltage, e(t) is the error between the actual speed and the set speed, reflecting the deviation between the current operating state and the target state of the system, and is the input signal for PID control. t is the current operating time of the system, and K... p K i K d These are the proportional, integral, and derivative coefficients of the PID controller, respectively. It is used to control the accumulation of historical errors, eliminate steady-state errors through integration, and improve system accuracy. Used to control the rate of change of error, reflect the trend of error change, suppress overshoot (reduce the error growth trend in advance), improve system damping, and reduce oscillation.

[0074] S2.2: Synchronously acquire the time-varying induction signal of the electromagnetic induction coil under motion. The induction signal is connected to the data acquisition card via a preamplifier and the position coordinates of the wire are recorded synchronously (obtained through the absolute encoder of the traction mechanism) to ensure that the signal and position correspond accurately.

[0075] S2.3: Perform time-domain synchronization marking processing on dynamic resistance signals and time-varying induction signals. Use hardware triggering to synchronously trigger the acquisition of two types of sensors when the traction mechanism pulse signal (corresponding to the fixed distance the wire moves). Realize microsecond-level synchronization through timestamp recording, establish a precise mapping relationship with the physical position of the wire, and output fluid resistance signals and electromagnetic induction signals.

[0076] S3: Based on the fluid resistance signal and combined with the reference fluid resistance signal, locate the surface defects of the enameled wire and output surface defect data. Based on the defect location and combined with the reference electromagnetic signal, analyze the spectral characteristics of the electromagnetic induction signal to identify the type and distribution of internal defects in the aluminum enameled flat wire and output internal defect data.

[0077] The steps for surface defect detection include:

[0078] S3.1: Multi-level digital filtering is used to eliminate vibration noise of the traction mechanism. The three-level filtering structure consists of low-pass filtering (removing high-frequency noise), band-pass filtering (extracting defect feature frequencies), and adaptive notch filtering (suppressing power frequency interference) to gradually improve the signal-to-noise ratio. An adaptive threshold algorithm is used to extract effective signal segments. The threshold is dynamically adjusted based on the mean and standard deviation of the signal within the sliding window. Signal segments that continuously exceed the threshold are determined to be effective defect segments. A resistance position mapping model is established.

[0079] S3.2: Based on the fluid dynamics model and the drag location mapping model, the fluid drag signal is converted into a three-dimensional surface profile. The fluid dynamics model is based on the laminar flow assumption and simplifies the relationship between drag and surface morphology using the Navier-Stokes equations. The model parameters are calibrated using standard samples. The ideal surface model is reconstructed using the moving least squares method. The moving least squares method employs a variable window and an adaptive polynomial order to balance the fitting accuracy between the overall profile and local details. The deviation matrix between the three-dimensional surface profile and the ideal surface model is calculated, and anomalies are identified through curvature analysis. Curvature analysis calculates Gaussian curvature and average curvature. Curvature anomaly regions that exceed the deviation threshold of the ideal surface and are continuously distributed are identified as defects, thus obtaining the anomaly regions. The three-dimensional surface profile reconstruction model based on fluid dynamics is expressed as follows:

[0080]

[0081] in, Let be the local resistance at coordinate (x,s), μ be the fluid viscosity, v1 be the constant moving speed of the wire, and h(x,s) be the surface height deviation function. , These represent the gradients of the surface profile in the x and s directions, respectively. n x1 represents the upper and lower limits of integration in the x-direction, and s m s1 represents the upper and lower limits of the integral in the s direction. dxds is an area element in two-dimensional Cartesian coordinates, serving as a cumulative volume from local to global.

[0082] S3.3: Based on the abnormal region, calculate the defect feature parameters (such as the maximum value of the contour deviation, standard deviation, number of curvature anomalies and distribution density), and establish the defect feature vector.

[0083] S3.4: Map the defect feature vectors into specific defect types through a pre-trained random forest model, and output a defect list with spatial coordinates. The model input is a multi-dimensional feature vector, and the classification boundary is trained and optimized through historical defect samples (scratches, pits, protrusions, etc.).

[0084] The steps for internal defect detection include.

[0085] S3.5: Perform synchronous compression transform on the electromagnetic induction signal. The synchronous compression transform adjusts the time-frequency resolution, separates the skin effect (high-frequency linear correlation) and defect response (low-frequency non-linearity) signals. When separating the skin effect and defect response in the time-frequency domain, singular value decomposition is used to eliminate baseline drift. The signal time-varying matrix is decomposed into a low-rank (baseline drift) and a sparse (defect transient) component. Retain the low-rank part to remove noise and extract the transient feature component synchronized with the wire movement. The formula for time-frequency domain signal separation is as follows:

[0086]

[0087] where is the wavelet coefficient,[[]]END]] is the original electromagnetic induction signal, t1 is the time window, ψ(t1) is the mother wavelet function, which is the basic function of wavelet transform. It has characteristics such as oscillation, fast decay, and compact support or approximate compact support in both the time domain and the frequency domain. represents the complex conjugate of the mother wavelet function. a is the scale factor, which is a positive real number used to scale the mother wavelet function. When a > 1, the mother wavelet function is stretched, the width in the time domain increases, and the width in the frequency domain decreases; when 0 < a < 1, the mother wavelet function is compressed, the width in the time domain decreases, and the width in the frequency domain increases. b is the translation factor, which is a real number used to translate the scaled mother wavelet function along the time axis. It determines the sliding position of the wavelet function on the original signal f(t1). By changing the value of b, local analysis of the original signal can be performed at different time points, thereby obtaining the time-frequency characteristics of the signal at different moments.

[0088] S3.6: Establish an electromagnetic-thermal-mechanical coupling model. The model is based on the coupling of multiple fields including Maxwell's equations (electromagnetic field), heat conduction equation (eddy current loss heat generation), and elastic mechanics equation (stress distribution). Solve the temperature field and stress field at the defect location through finite element simulation, and decompose the transient feature component into conductor defect response and insulation defect response.

[0089] S3.7: A forward model is constructed based on Maxwell's equations. The forward model uses finite element simulation to establish the conductor / insulator geometric model, inputting parameters such as cross-sectional area variation and insulation thickness fluctuation to simulate the electromagnetic induction signal response. Based on the conductor defect response and insulation defect response, the Monte Carlo optimization algorithm is used to inversely retrieve the conductor cross-sectional area variation and insulation thickness fluctuation parameters, outputting internal defect data. Monte Carlo optimization obtains the parameter solution closest to the actual signal by randomly generating parameter combinations and calculating errors. The model can be simplified using Maxwell's equations as follows:

[0090]

[0091] in, Let γ be the first partial derivative of the electric field strength E(j) with respect to position j, where E(j) is the electric field strength at defect position j, γ is the magnetic permeability of the conductor, and σ is the electrical conductivity of the conductor. Let t2 be the function of the change in conductor cross-sectional area caused by the defect, and t2 be the duration. The function of conductor cross-sectional area change The first-order partial derivative with respect to time t2.

[0092] S4: Spatiotemporally match surface defect data and internal defect data, and distinguish defect types using machine learning algorithms. Establish a defect feature matrix based on the defect type, and output the corresponding correction scheme through a pre-trained neural network model.

[0093] The steps for spatiotemporal matching include:

[0094] S4.1: Spatial registration is performed between the coordinate data in the surface defect data and the distribution cloud map in the internal defect data. A feature point-based registration method is used to extract the defect center point (surface) and centroid (internal cloud map). A high-precision mapping relationship is established through affine transformation (translation, rotation, scaling).

[0095] S4.2: Analyze the internal region below the coordinate data, compare the correlation between electromagnetic characteristics and surface morphology, quantify the degree of correlation using cross-correlation functions and mutual information indices, and identify associated defects when the correlation exceeds a threshold. When the distance between the coordinate data and the distribution cloud map is less than a set threshold, it is identified as a correlated composite defect.

[0096] The steps involved in distinguishing defect types using machine learning algorithms include:

[0097] S4.3: Based on the coordinate data and distribution cloud map, construct the spatiotemporal correlation feature matrix and generate the fusion feature vector. The fusion feature vector contains multi-dimensional information such as the spatial coordinates of surface defects, contour deviation, curvature anomaly, internal defect cloud map, and electromagnetic features. The dimensions are eliminated after normalization.

[0098] S4.4: Load the pre-trained three-stage classification model based on machine learning algorithm, input the fused feature vector into the three-stage classification model to classify defect types, and output the classification results.

[0099] The steps for outputting classification results include:

[0100] S4.4.1: Load the pre-trained random forest model, perform coarse-grained classification on the fused feature vectors, and output low-confidence samples, represented as:

[0101]

[0102] Where Gini is the node Gini index, C is the total number of defect types, such as scratches, dents, etc., c is the defect sample index, and P is the defect sample index. c Let be the probability of the c-th type defect sample in the node.

[0103] S4.4.2: The 1D-CNN submodel is invoked to dynamically enhance the features of low-confidence samples. The 1D-CNN extracts and enhances temporal features (such as the correlation of defect development time), correcting the probability distribution. The 1D-CNN submodel (One-Dimensional Convolutional Neural Network Submodel) is a deep learning model specifically designed for processing one-dimensional temporal signals (such as electromagnetic induction signals, defect feature sequences, etc.). Its main function is to dynamically enhance the features of low-confidence samples to improve the accuracy of defect classification.

[0104] S4.4.3: Based on the modified probability distribution, a graph neural network is used to process the low-confidence samples for correlation defects. The graph neural network models the correlation between defect types (such as the probability of surface scratches and insulation detachment occurring together) and outputs the classification results.

[0105] S4.5: Based on surface defect data and internal defect data, construct a three-dimensional feature matrix according to the defect type based on the classification results.

[0106] S4.6: SHAP values ​​are used to evaluate the contribution of each feature, with dynamic weighting for key features. Feature importance analysis is performed on the three-dimensional feature matrix, generating an interpretability analysis report. SHAP values ​​quantify the contribution of features to classification, an importance ranking chart is drawn and the direction of influence is marked, and textual explanations are generated in conjunction with expert knowledge. SHAP values ​​(SHapley Additive Explanations) are an interpretability analysis method used to quantify the contribution of each feature in a machine learning model to the prediction results. Based on the concept of Shapley values ​​in game theory, it can fairly allocate the "contribution" of each feature to the model output, thereby helping to understand the model's decision-making logic. It calculates the influence of a feature under different feature subsets, ultimately obtaining the global and local interpretation of the feature's contribution to the model output. Shapley values ​​are a mathematical method derived from game theory, used to fairly allocate cooperative payoffs or quantify the contribution of individuals to collective outcomes. In the field of machine learning, it is used to interpret model predictions and measure the degree of contribution of each feature to the final result.

[0107] S5: Based on the correction scheme, output a correction control signal to the correction module to perform the correction operation. The steps for performing the correction operation include:

[0108] S5.1: Calculate the required correction action parameters based on the defect type and defect location. The parameters include the radial pressure roller pressure adjustment amount (lateral deviation calculation), the torsion mechanism torque adjustment amount (axial deformation calculation), and the multi-axis actuator movement speed (deflection rate adjustment).

[0109] S5.2: Based on the correction action parameters, the action parameters of the multi-axis actuator are adjusted by a PID controller to drive the radial pressure roller for wire posture adjustment. The control torsion mechanism compensates for the axial deformation of the wire in real time and dynamically tracks the deflection changes of the wire. The PID parameters are tuned using the Ziegler-Nichols method, combined with the dynamic characteristics of the actuator to optimize response speed and stability. The Ziegler-Nichols method is a classic PID controller parameter tuning method used to quickly determine the initial values ​​of the proportional, integral, and derivative control parameters, enabling the correction system to achieve good dynamic response performance. The critical gain and critical oscillation period of the system are tested experimentally, and then the PID parameters are calculated according to a specific formula.

[0110] S5.3: The corrected wire is re-introduced into the inspection area for verification. The same fluid and electromagnetic detection modules are used for verification. After the signal is collected, the effect is verified through the defect location and identification module. If the standard is not met, the parameters are adjusted and the verification is repeated.

[0111] In summary, this invention provides an online defect detection and adaptive correction method and system for enameled aluminum flat wire. By synchronously acquiring and analyzing fluid resistance signals and electromagnetic induction signals, it achieves simultaneous detection of surface and internal defects, overcoming the limitations of traditional single detection methods. The ideal surface model is reconstructed using the moving least squares method, combined with curvature analysis to identify abnormal regions. Noise in the electromagnetic signals is separated through synchronous compression transformation and singular value decomposition, significantly improving the detection sensitivity of minute defects. A random forest model is used for rapid coarse-grained classification, 1D-CNN dynamically enhances the features of low-confidence samples, and a graph neural network processes correlated defects, forming a high-precision three-stage classification system. Combined with SHAP value feature importance analysis, an interpretable report is generated, significantly improving the classification accuracy of complex defects. By generating a correction scheme based on defect type, location, and spatial mapping relationships, and dynamically adjusting multi-axis actuators using a PID controller to compensate for axial deformation and deflection changes in real time, a post-correction re-inspection mechanism is introduced to form closed-loop control, ensuring the effectiveness of defect correction.

[0112] like Figure 3 As shown, the present invention also provides an online defect detection and adaptive correction system for aluminum enameled flat wire, comprising: a fluid detection module, an electromagnetic detection module, a defect location module, a defect identification module, a correction decision module, and a correction execution module.

[0113] The fluid detection module is used to immerse the enameled aluminum flat wire in a fluid of a set viscosity and collect a reference fluid resistance signal. Based on the reference fluid resistance signal, the aluminum enameled flat wire is passed through the detection area at a constant speed, and the fluid resistance signal is detected in real time. The fluid detection module includes a fluid circulation and filtration unit (pump body + filter to maintain cleanliness) and Teflon-coated guide wheels (to reduce the coefficient of friction).

[0114] The electromagnetic detection module applies a high-frequency alternating current to aluminum enameled flat wire in a fluid to acquire a reference electromagnetic signal. Based on the reference electromagnetic signal, it detects the electromagnetic induction signal in real time. The electromagnetic detection module is equipped with an electromagnetic shield (made of high-permeability material) and a coaxial cable (to reduce transmission loss).

[0115] The defect location module is used to locate surface defects in enameled wires based on fluid resistance signals, combined with reference fluid resistance signals. Based on the defect location, and combined with reference electromagnetic signals, it analyzes the spectral characteristics of the electromagnetic induction signal to identify the type and distribution of internal defects in the aluminum enameled flat wire. The reference signals are stored in an industrial-grade database (supporting version management and backup), and the location mapping model is continuously optimized using historical data.

[0116] The defect identification module is used to perform spatiotemporal matching of surface defect data and internal defect data, and to distinguish defect types through machine learning algorithms.

[0117] The error correction decision module is used to establish a defect feature matrix based on the defect type and output the corresponding error correction scheme through a pre-trained neural network model.

[0118] The correction execution module is used to output correction control signals to the correction module for correction operations based on the correction scheme. The correction module includes a radial pressure roller group (servo motor driven), an axial torsion mechanism (stepper motor driven), and a laser displacement sensor (to monitor deflection changes).

[0119] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for online defect detection and adaptive correction of aluminum enameled flat wire, characterized in that, include: S1: Immerse the aluminum enameled flat wire into a fluid of a set viscosity and apply a high-frequency alternating current to collect the reference fluid resistance signal and the reference electromagnetic signal; S2: Based on the reference fluid resistance signal and the reference electromagnetic signal, the aluminum enameled flat wire is passed through the detection area at a constant speed, and the fluid resistance signal and electromagnetic induction signal are detected in real time. S3: Based on the fluid resistance signal and the reference fluid resistance signal, locate the defect position of the surface defect of the enameled wire and output the surface defect data; Based on the defect location and the reference electromagnetic signal, the spectral characteristics of the electromagnetic induction signal are analyzed to identify the type and distribution of internal defects in the aluminum enameled flat wire, and internal defect data is output. S4: Perform spatiotemporal matching of the surface defect data and the internal defect data, and distinguish the defect types using machine learning algorithms; establish a defect feature matrix based on the defect types, and output the corresponding correction scheme through a pre-trained neural network model; S5: Based on the aforementioned correction scheme, output a correction control signal to the correction module to perform correction operations.

2. The method for online defect detection and adaptive correction of aluminum enameled flat wire according to claim 1, characterized in that, Step S1, which involves acquiring the reference fluid resistance signal and the reference electromagnetic signal, includes: S1.1: The aluminum enameled flat wire is vertically immersed into the detection fluid in the constant temperature control box through the guide wheel, and the reference fluid resistance signal at each axial position is collected by a distributed fluid pressure sensor array; S1.2: Apply a standard frequency alternating current through a high-frequency power generator. The current intensity is automatically matched to preset parameters based on the cross-sectional area of ​​the wire. S1.3: Based on the preset parameters, the reference electromagnetic signal in the stationary state is collected by a ring electromagnetic induction coil group.

3. The method for online defect detection and adaptive correction of aluminum enameled flat wire according to claim 1, characterized in that, Step S2, the steps of real-time detection of fluid resistance signals and electromagnetic induction signals include: S2.1: Control the precision traction mechanism to move the enameled wire at a constant linear speed, record the dynamic resistance signal of the fluid pressure sensor in real time during the movement, and use a sliding window algorithm to eliminate mechanical vibration noise; S2.2: Synchronously acquire the time-varying induced signal of the electromagnetic induction coil under motion conditions; S2.3: Perform time-domain synchronous marking processing on the dynamic resistance signal and the time-varying induction signal to establish a precise mapping relationship with the physical position of the wire, and output the fluid resistance signal and electromagnetic induction signal.

4. The method for online defect detection and adaptive correction of aluminum enameled flat wire according to claim 1, characterized in that, Step S3, the steps for surface defect detection include: S3.1: Multi-level digital filtering is used to eliminate vibration noise of the traction mechanism, and an effective signal segment is extracted through an adaptive threshold algorithm to establish a resistance position mapping model; S3.2: Based on the fluid dynamics model and the resistance position mapping model, the fluid resistance signal is converted into a three-dimensional surface profile; the ideal surface model is reconstructed using the moving least squares method, the deviation matrix between the three-dimensional surface profile and the ideal surface model is calculated, and anomalies are identified through curvature analysis to obtain the abnormal regions; S3.3: Based on the abnormal region, calculate the defect feature parameters and establish the defect feature vector; S3.4: Using a pre-trained random forest model, the defect feature vectors are mapped to specific defect types, and a list of defects with spatial coordinates is output.

5. The method for online defect detection and adaptive correction of aluminum enameled flat wire according to claim 1, characterized in that, Step S3 includes the following steps for internal defect detection: S3.5: Perform synchronous compression transformation on the electromagnetic induction signal, separate the skin effect and defect response in the time and frequency domain, eliminate baseline drift by singular value decomposition, and extract transient characteristic components that are synchronized with the wire movement. S3.6: Establish an electromagnetic-thermal-mechanical coupling model to decompose the transient characteristic components into conductor defect response and insulation defect response; S3.7: Construct a forward model based on Maxwell's equations, and use the Monte Carlo optimization algorithm to invert the conductor cross-sectional area change and insulation layer thickness fluctuation parameters according to the conductor defect response and insulation defect response, and output the internal defect data.

6. The online defect detection and adaptive correction method for aluminum enameled flat wire according to claim 1, characterized in that, Step S4, the spatiotemporal matching steps include: S4.1: Spatial registration is performed between the coordinate data in the surface defect data and the distribution cloud map in the internal defect data to establish a high-precision mapping relationship; S4.2: Analyze the internal region below the coordinate data, compare the correlation between electromagnetic features and surface morphology, and identify associated defects; when the distance between the coordinate data and the distribution cloud map is less than a set threshold, it is determined to be an associated composite defect.

7. The online defect detection and adaptive correction method for aluminum enameled flat wire according to claim 6, characterized in that, Step S4, the step of distinguishing defect types using machine learning algorithms, includes: S4.3: Based on the coordinate data and the distribution cloud map, construct a spatiotemporal correlation feature matrix and generate a fused feature vector; S4.4: Load the pre-trained three-stage classification model based on machine learning algorithm, and input the fused feature vector into the three-stage classification model to classify the defect type, and output the classification result; S4.5: Based on the surface defect data and the internal defect data, construct a three-dimensional feature matrix according to the defect type based on the classification results; S4.6: Use SHAP values ​​to evaluate the contribution of each feature and dynamically weight key features; perform feature importance analysis on the three-dimensional feature matrix and generate an interpretability analysis report.

8. The online defect detection and adaptive correction method for aluminum enameled flat wire according to claim 7, characterized in that, Step S4.4, the steps for outputting the classification results, include: S4.4.1: Load the pre-trained random forest model, perform coarse-grained classification on the fused feature vector, and output low-confidence samples; S4.4.2: Call the 1D-CNN sub-model to perform dynamic feature enhancement on the low-confidence samples and output the corrected probability distribution; S4.4.3: Based on the corrected probability distribution, the low-confidence samples are processed for correlation defects using a graph neural network, and the classification result is output.

9. The method for online defect detection and adaptive correction of aluminum enameled flat wire according to claim 1, characterized in that, Step S5 includes the following steps for performing the correction operation: S5.1: Calculate the required correction action parameters based on the defect type and defect location; S5.2: Based on the aforementioned correction action parameters, the action parameters of the multi-axis actuator are adjusted by the PID controller to drive the radial pressure roller to adjust the wire posture; the torsion mechanism is controlled to compensate for the axial deformation of the wire in real time and dynamically track the deflection change of the wire. S5.3: The corrected cable is re-introduced into the detection area for verification.

10. An online defect detection and adaptive correction system for aluminum enameled flat wire, comprising the online defect detection and adaptive correction method for aluminum enameled flat wire as described in any one of claims 1 to 9, characterized in that, include: The fluid detection module is used to immerse the aluminum enameled flat wire in a fluid of a set viscosity and collect a reference fluid resistance signal. Based on the reference fluid resistance signal, the aluminum enameled flat wire is passed through the detection area at a constant speed, and the fluid resistance signal is detected in real time. The electromagnetic detection module is used to apply a high-frequency alternating current to the aluminum enameled flat wire in the fluid and collect reference electromagnetic signals. Based on a reference electromagnetic signal, electromagnetic induction signals are detected in real time. The defect location module is used to locate the defect position on the surface of the enameled wire based on the fluid resistance signal and the reference fluid resistance signal. Based on the location of the defect and in conjunction with the reference electromagnetic signal, the spectral characteristics of the electromagnetic induction signal are analyzed to identify the type and distribution of internal defects in the aluminum enameled flat wire. The defect identification module is used to perform spatiotemporal matching of the surface defect data and the internal defect data, and to distinguish the defect type through machine learning algorithms. The error correction decision module is used to establish a defect feature matrix based on the defect type and output the corresponding error correction scheme through a pre-trained neural network model. The correction execution module is used to output a correction control signal to the correction module to perform correction operations based on the correction scheme.

Images