A method and device for automatically checking the eccentricity of cable extrusion layer

Abstract

The invention relates to the technical field of cable calibration, and discloses a method and device for automatically calibrating the eccentricity of a cable extrusion layer, wherein the method comprises: performing multi-angle scanning on a target cable to obtain a cable cross-sectional profile data set including a conductor outer diameter and an insulating layer outer diameter; performing digital filtering and curve fitting to obtain the coordinates of the center of a circle; calculating the eccentricity value and thickness change data at each angle position; performing statistical feature calculation to obtain eccentricity distribution features and a maximum eccentricity direction, and performing change feature extraction to obtain thickness change features; inputting the eccentricity distribution features, the maximum eccentricity direction, and the thickness change features into a plurality of autoencoders and an ant colony decision fusion model to perform mold position compensation analysis, and generate mold position adjustment parameters. The method can automatically generate mold position adjustment parameters according to the detection results, thereby realizing closed-loop control of the production process and improving the quality and efficiency of cable manufacturing.

Description

Technical Field

[0001] The invention relates to the technical field of cable calibration, and in particular to a method and a device for automatically calibrating the eccentricity of a cable extrusion layer. Background Art

[0002] During the cable manufacturing process, the eccentricity control of the extruded layer is a key link to ensure product quality. Traditional eccentricity detection methods mainly rely on manual operation and experience judgment, which have problems such as low efficiency and low precision. With the continuous improvement of the automation level of cable production, higher requirements are placed on the accuracy and real-time performance of eccentricity detection. However, the current automatic detection technology still faces challenges such as insufficient sampling accuracy and limited data processing capabilities, and it is difficult to meet the needs of high-precision cable production.

[0003] In addition, there are many influencing factors in the cable extrusion process, such as temperature fluctuations, pressure changes, etc. These factors will cause dynamic changes in eccentricity. Existing detection methods often have difficulty in effectively capturing these dynamic characteristics and cannot adjust and optimize the production process in a timely manner. Summary of the invention

[0004] The present invention provides a method and device for automatically checking the eccentricity of a cable extrusion layer, which are used to automatically generate mold position adjustment parameters according to detection results, realize closed-loop control of the production process, and improve the quality and efficiency of cable manufacturing.

[0005] In a first aspect, the present invention provides a method for automatically checking the eccentricity of a cable extrusion layer, the method comprising:

[0006] Scan the target cable at multiple angles to obtain a cable cross-sectional profile data set including the outer diameter of the conductor and the outer diameter of the insulation layer;

[0007] Performing digital filtering and curve fitting on the cable cross-sectional profile data set to obtain the center coordinates of the conductor outer circle and the insulation layer outer circle;

[0008] Based on the center coordinates, the distance between the outer circle of the conductor and the center of the outer circle of the insulating layer is calculated to obtain the eccentricity value at each angular position, and the thickness variation data of the insulating layer at different angles is calculated;

[0009] Performing statistical characteristic calculation on the eccentricity value to obtain eccentricity distribution characteristics and maximum eccentricity direction, and performing variation feature extraction on the thickness variation data to obtain thickness variation characteristics;

[0010] The eccentricity distribution characteristics, the maximum eccentricity direction and the thickness variation characteristics are input into a preset plurality of autoencoders and an ant colony decision fusion model to perform mold position compensation analysis to generate mold position adjustment parameters.

[0011] In a second aspect, the present invention provides an automatic eccentricity checking device for a cable extrusion layer, the automatic eccentricity checking device for a cable extrusion layer comprising:

[0012] A scanning module is used to scan the target cable at multiple angles to obtain a cable cross-sectional profile data set including the outer diameter of the conductor and the outer diameter of the insulation layer;

[0013] A fitting module, used for performing digital filtering and curve fitting on the cable cross-sectional profile data set to obtain the center coordinates of the conductor outer circle and the insulation layer outer circle;

[0014] A calculation module, for calculating the distance between the outer circle of the conductor and the center of the outer circle of the insulating layer based on the center coordinates, obtaining the eccentricity value at each angular position, and calculating the thickness variation data of the insulating layer at different angles;

[0015] An extraction module is used to perform statistical feature calculation on the eccentricity value to obtain the eccentricity distribution feature and the maximum eccentricity direction, and to perform variation feature extraction on the thickness variation data to obtain the thickness variation feature;

[0016] A generation module is used to input the eccentricity distribution characteristics, the maximum eccentricity direction and the thickness change characteristics into a preset plurality of autoencoders and an ant colony decision fusion model to perform mold position compensation analysis and generate mold position adjustment parameters.

[0017] The third aspect of the present invention provides an automatic calibration device for the eccentricity of a cable extrusion layer, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the automatic calibration device for the eccentricity of a cable extrusion layer executes the above-mentioned automatic calibration method for the eccentricity of a cable extrusion layer.

[0018] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned method for automatically checking the eccentricity of an extruded cable layer.

[0019] In the technical solution provided by the present invention, through multi-angle scanning and digital filtering technology, the method can obtain more accurate cable cross-sectional profile data and greatly improve the accuracy of eccentricity measurement. By adopting efficient data processing algorithms and autoencoder models, fast feature extraction and analysis are achieved, so that eccentricity detection can keep up with the real-time speed of the production line. Combined with multiple autoencoders and ant colony decision fusion models, this method can adapt to the structural characteristics of different types of cables, and through in-depth analysis of eccentricity distribution characteristics and thickness change characteristics, it can effectively capture the dynamic change characteristics during cable extrusion. The ant colony algorithm is used to perform mold position compensation analysis, automatically generate adjustment parameters, reduce human intervention, and improve production efficiency. The RANSAC algorithm and wavelet transform are used to effectively reduce errors in the measurement and calculation process. Through repeated measurements and statistical feature calculations, the reliability and stability of the detection results are enhanced, which is conducive to the realization of automatic control of the entire process.

[0020] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0021] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A schematic diagram of an embodiment of a method for automatically checking the eccentricity of a cable extrusion layer in an embodiment of the present invention;

[0023] Figure 2 A schematic diagram of an embodiment of an automatic eccentricity checking device for a cable extrusion layer in an embodiment of the present invention;

[0024] Figure 3 It is a schematic diagram of an embodiment of an automatic eccentricity calibration device for a cable extrusion layer in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products, or device ends.

[0027] To facilitate understanding of this embodiment, a method for automatically checking the eccentricity of a cable extrusion layer disclosed in an embodiment of the present invention is first described in detail. Figure 1 As shown, the method comprises the following steps:

[0028] 101. Scan the target cable at multiple angles to obtain a cable cross-sectional profile data set including a conductor outer diameter and an insulation layer outer diameter;

[0029] It is understandable that the execution subject of the present invention may be an automatic eccentricity checking device for a cable extrusion layer, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject.

[0030] Specifically, the target cable is fixedly installed to ensure that the cable remains stable and does not move during the entire measurement process, and a cable sample to be tested is obtained. The rotation angle of the cable sample to be tested is calibrated to determine the initial angle position. According to the initial angle position, a precision rotating device is used to rotate the cable sample to be tested at equal intervals, and the cable is subjected to multi-point laser scanning at each rotation angle position. During the laser scanning process, a high-precision laser scanner is used to scan M points at each sampling angle position to obtain the precise coordinate data of these points. The scanning operation of each sampling angle position is repeatedly measured multiple times to form an original measurement data set. The original measurement data set contains data from multiple measurement cycles. By performing mean calculations on these data, random errors are eliminated and the average coordinate value of each sampling angle position is obtained. The cable cross-section contour point cloud is constructed based on the average coordinate value. Point cloud data is a three-dimensional spatial data representation method that can intuitively display the geometric shape and characteristics of the cable cross section. The initial contour model of the cable is generated by processing the point cloud data. The initial contour model contains all important geometric information on the cable cross section. The initial contour model is subjected to edge detection. The edge detection technology can effectively separate the boundaries of the cable contour, which usually represent the outer contour of the conductor and the outer contour of the insulation layer. The detected edges are segmented to separate the outer contour of the conductor and the outer contour of the insulation layer. The radial distance is calculated based on the outer contour of the conductor and the outer contour of the insulation layer. The radial distance refers to the distance from the outer center of the conductor to the outer center of the insulation layer. This distance can reflect the radial size of the cable at different sampling angles. By calculating the radial distances at all sampling angles, the radial size data at each sampling angle is obtained. The radial size data is interpolated. Through mathematical algorithms, the existing data points are used to predict unknown data points, thereby filling in the gaps or noise in the data to form a smoother and more continuous cable contour. After interpolation, a cable cross-sectional profile data set containing the outer diameter of the conductor and the outer diameter of the insulation layer is obtained.

[0031] 102. Perform digital filtering and curve fitting on the cable cross-sectional profile data set to obtain the center coordinates of the conductor outer circle and the insulation layer outer circle;

[0032] Specifically, a wavelet transform is performed on the cable cross-sectional profile data set. Wavelet transform is a tool that can effectively process signal noise. The profile data set is decomposed into multiple scales through wavelet transform to obtain multi-scale decomposition coefficients. The noise threshold is calculated for the multi-scale decomposition coefficients to obtain a denoising threshold to help identify and remove high-frequency noise in the profile data. According to the calculated denoising threshold, the multi-scale decomposition coefficients are soft-threshold processed to obtain denoised wavelet coefficients. Soft threshold processing is a commonly used denoising method that reduces coefficients less than the threshold to zero, thereby effectively reducing the impact of noise. After the processing is completed, the denoised wavelet coefficients are converted back to the time domain through inverse wavelet transform to obtain filtered cable profile data. Edge detection is performed on the filtered profile data to extract the conductor outer contour point set and the insulation layer outer contour point set of the cable. After edge detection is completed, the conductor outer contour point set is circularly fitted by the first RANSAC algorithm to obtain the equation of the conductor outer circle. Similarly, the insulation layer outer contour point set is circularly fitted by the second RANSAC algorithm to obtain the equation of the insulation layer outer circle. The RANSAC algorithm is a robust parameter estimation method that is suitable for fitting geometric models when there are outliers in the data. Through this algorithm, the geometric shapes of the outer circle of the conductor and the outer circle of the insulation layer can be effectively fitted. According to the outer circle equation of the conductor and the outer circle equation of the insulation layer, the center coordinates and radius are extracted to obtain the outer circle parameters of the conductor and the outer circle parameters of the insulation layer respectively. The uncertainty analysis of the outer circle parameters of the conductor and the outer circle parameters of the insulation layer is carried out to evaluate the credibility and accuracy of the measurement results, ensure that the fitting results can accurately reflect the actual cable structure, and obtain the center coordinates of the outer circle of the conductor and the outer circle of the insulation layer.

[0033] 103. Based on the coordinates of the center of the circle, calculate the distance between the center of the outer circle of the conductor and the center of the outer circle of the insulating layer, obtain the eccentricity value at each angle position, and calculate the thickness change data of the insulating layer at different angles;

[0034] Specifically, the coordinate system of the center coordinates of the outer circle of the conductor and the outer circle of the insulation layer is converted. By converting the center coordinates from the Cartesian coordinate system to the polar coordinate system, the position of the center of the circle is more intuitively represented. Based on the center position represented by the polar coordinates, the Euclidean distance between the two centers is calculated to obtain the initial eccentricity value. The Euclidean distance is a standard measure for the distance between two points, reflecting the geometric eccentricity between the conductor and the insulation layer. The initial eccentricity value is normalized, and the eccentricity value is standardized so that the eccentricity data at different angles are comparable, and the relative eccentricity value is obtained. Based on the relative eccentricity value, the cable cross section is angularly divided into N evenly distributed angular positions. At each divided angular position, the radial distance between the outer circle of the conductor and the outer circle of the insulation layer is calculated to obtain the absolute eccentricity value of each angular position. The absolute eccentricity value reflects the degree of eccentricity of the cable at different angular positions. According to the absolute eccentricity value, curve fitting is performed to obtain the eccentricity distribution curve. The curve is a continuous function that describes the distribution of the eccentricity of the cable extrusion layer over the entire cross section. By performing differential operations on the eccentricity distribution curve, the eccentricity change rate curve is obtained. The change rate curve can reveal the trend of how the eccentricity changes with the angle. By analyzing this curve, the eccentricity mutation point, that is, the position where the eccentricity changes significantly, is identified, so as to determine the target eccentricity angle position. Identifying the mutation point helps to adjust the mold position and optimize the cable production process. The insulation layer thickness at the target eccentricity angle position is calculated to obtain the distribution data of the insulation layer thickness, revealing the thickness difference of the cable insulation layer at different angles. According to the thickness distribution data, curve fitting is performed to obtain the thickness change curve of the insulation layer at different angles. This curve is a continuous function that describes the trend of the insulation layer thickness changing with the angle.

[0035] 104. Calculate the statistical characteristics of the eccentricity value to obtain the eccentricity distribution characteristics and the maximum eccentricity direction, and extract the variation characteristics of the thickness variation data to obtain the thickness variation characteristics;

[0036] Specifically, the eccentricity value is Fourier transformed to convert the eccentricity value from the time domain to the frequency domain to reflect its distribution characteristics at different frequencies. By analyzing the frequency domain representation, the target frequency component of the eccentricity is extracted to obtain the harmonic characteristics of the eccentricity. The harmonic characteristics can reflect the periodic changes of the cable eccentricity and help analyze the regularity and periodicity of the eccentricity. The eccentricity value is kernel-density estimated to obtain the probability density function of the eccentricity. The probability density function reflects the distribution of eccentricity in different value ranges. Based on the function, the statistical moment characteristics of the eccentricity, such as the mean, variance, skewness and kurtosis, are calculated. The statistical moment characteristics provide a basic description of the eccentricity distribution. The mean represents the central tendency of the eccentricity, the variance reflects the degree of dispersion of the eccentricity, the skewness describes the symmetry of the distribution, and the kurtosis reveals the sharpness of the distribution. The spatial autocorrelation analysis of the eccentricity value reveals the spatial distribution pattern of the eccentricity and identifies the spatial distribution characteristics of the eccentricity. Based on these characteristics, the maximum value of eccentricity and its corresponding angular position are determined to obtain the direction of maximum eccentricity. The direction of maximum eccentricity is a key parameter in cable quality control, indicating the direction of the most serious eccentricity, helping to locate and adjust problems in the production process. At the same time, the thickness change data is decomposed by wavelet packet decomposition, and the data is decomposed into multiple wavelet coefficients of different scales. Through the multi-scale wavelet coefficients, the energy entropy and corresponding dimension of the thickness change are calculated to obtain the complexity characteristics of the thickness change. The thickness change data is subjected to empirical mode decomposition, and the data is decomposed into multiple intrinsic mode functions. Each intrinsic mode function represents an intrinsic oscillation mode in the data. By analyzing the mode function, the instantaneous frequency and amplitude are extracted to obtain the time-frequency characteristics of the thickness change. The time-frequency characteristics reveal the dynamic behavior of the thickness change, especially the instantaneous changes in the time and frequency domains. The harmonic characteristics, statistical moment characteristics, and spatial distribution characteristics are combined to form a comprehensive distribution characteristic of the eccentricity. At the same time, the complexity characteristics and time-frequency characteristics of the thickness change are combined to obtain a detailed description of the thickness change.

[0037] 105. The eccentricity distribution characteristics, the maximum eccentricity direction and the thickness variation characteristics are input into the preset multiple autoencoders and ant colony decision fusion model to perform mold position compensation analysis and generate mold position adjustment parameters.

[0038] Specifically, the eccentricity distribution features are processed by dimensionality reduction to reduce the dimension of the feature data, thereby reducing the computational complexity and extracting the most representative eccentricity feature vector. The eccentricity feature vector after dimensionality reduction is input into the first autoencoder in the multiple autoencoders and ant colony decision fusion model for processing. The autoencoder consists of a three-layer encoder and a three-layer decoder. The encoder compresses and encodes the input features, while the decoder reconstructs the compressed information into an output similar to the original data. In this way, the key features of the eccentricity distribution are extracted and the data is reconstructed to obtain the potential representation of the target eccentricity. At the same time, the thickness change features are subjected to principal component analysis to extract the most representative principal component feature vector. Principal component analysis is a commonly used data dimensionality reduction technology. Through linear transformation, the original data is converted into unrelated principal components, thereby retaining the most important change trend in the data. The extracted principal component feature vector is input into the second autoencoder for processing. The autoencoder consists of a four-layer encoder and a four-layer decoder, which can more deeply extract the potential representation of the thickness change features and reconstruct the data. In this way, the potential representation of the target thickness change is obtained. The maximum eccentricity direction, target eccentricity potential representation, and target thickness change potential representation are feature fused. Feature information from different sources is combined to form a fused feature vector containing comprehensive information. Adaptive fuzzy c-means clustering is performed on the fused feature vector to identify different feature clusters. These feature clusters represent the eccentricity and thickness change patterns of the cable under different conditions. Based on the feature clusters, the initial pheromone matrix of the ant colony decision layer in the multiple autoencoders and ant colony decision fusion model is constructed. The setting of the pheromone matrix is ​​the core of the ant colony algorithm and determines the tendency of ants to select paths in the search space. At the same time, the number of ants, the number of iterations, and the heuristic factor are set to control the operation process of the ant colony algorithm. During the execution of the ant colony algorithm, based on the roulette selection method and the state transition probability calculation, each ant will choose a path, which represents a possible mold position adjustment scheme. As the iteration proceeds, each ant will update the pheromone matrix according to the fitness value of the selected path. Paths with higher fitness values ​​will leave more pheromones, thereby increasing the probability of other ants choosing this path. During the iteration process, the algorithm will continuously update the global optimal path and the local optimal path to ensure that the optimal die position adjustment solution is finally found. By decoding the global optimal path, the optimal die position adjustment parameters are generated, which will be used for actual die adjustment to achieve precise eccentricity control and thickness adjustment of the cable extrusion layer.

[0039] The eccentricity feature vector after dimensionality reduction is normalized, the feature value is scaled to a standard range, the difference between different feature dimensions is eliminated, and the normalized eccentricity feature vector is obtained. The normalized eccentricity feature vector is input into the first encoding layer of the first autoencoder. The first encoding layer consists of 128 neurons and uses the ReLU activation function for nonlinear transformation. Through this process, the feature vector is mapped to a high-dimensional feature space to generate the first hidden layer representation. The first hidden layer representation is batch normalized. The main function of batch normalization is to stabilize the network training process, avoid the problem of gradient disappearance or explosion, and improve the convergence speed of the network. The normalized first hidden layer representation is input into the second encoding layer of the first autoencoder. The encoding layer consists of 64 neurons and also uses the ReLU activation function. After this layer of feature compression, the feature dimension is further reduced, while retaining the key information in the data to generate the second hidden layer representation. In order to prevent the model from overfitting, the second hidden layer representation is dropout. By randomly discarding a part of the neurons, this sparse process makes the model more robust and avoids overfitting of the training data. The sparsified second hidden layer representation is input to the third encoding layer of the first autoencoder, which consists of 32 neurons and uses the tanh activation function for feature encoding. The tanh function compresses the input values ​​to between -1 and 1, making the feature representation smoother and more adaptable, and obtaining the initial eccentricity potential representation. The initial eccentricity potential representation is input to the first decoding layer of the first autoencoder. The decoding layer performs preliminary decoding through 32 neurons and the ReLU activation function, and reconverts the potential representation back to a higher dimensional data space to obtain the first decoding layer output. The first decoding layer output is subjected to Layer Normalization. Layer Normalization can standardize each sample individually, reduce the change of internal covariates, and improve the generalization ability of the model. The normalized first decoding layer output is input to the second decoding layer. In the second decoding layer, 64 neurons and the ReLU activation function are used for feature reconstruction, which is further expanded and reconstructed to generate the second decoding layer output. In order to enhance the expressive power of the model, the output of the second decoding layer is subjected to a residual connection operation, and the output of the second decoding layer is added to the second hidden layer representation generated during the encoding process to obtain the residual enhanced second decoding layer output. Residual connections help alleviate the gradient vanishing problem in deep networks and ensure that important features are not lost when information is transmitted in the network. The residual enhanced second decoding layer output is input to the third decoding layer of the first autoencoder. This decoding layer performs the final feature reconstruction through 128 neurons and a Sigmoid activation function, mapping the data into a normalized output space to obtain the target eccentricity potential representation.

[0040] The principal component eigenvector is standardized so that each dimension of the eigenvector is on the same scale, thereby eliminating the dimensional differences between different features and obtaining the standardized principal component eigenvector. The standardized principal component eigenvector is input into the first encoding layer of the second autoencoder, which consists of 256 neurons and uses the LeakyReLU activation function for nonlinear transformation. The LeakyReLU activation function allows a certain negative slope when processing negative values, thereby avoiding the problem of neuron death caused by the traditional ReLU when processing negative values, and obtaining the output of the first encoding layer. The output of the first encoding layer is subjected to Instance Normalization. Instance Normalization can adapt to small batch training by independently standardizing each sample, and has a good inhibitory effect on the change of internal covariates during network training. The normalized output of the first encoding layer is input into the second encoding layer of the second autoencoder, which consists of 128 neurons and also uses the LeakyReLU activation function. This layer further compresses the features and reduces redundant information, and obtains the output of the second encoding layer. In order to enhance the generalization ability of the model and prevent overfitting, the Spatial Dropout operation is performed on the output of the second encoding layer. Unlike traditional dropout, Spatial Dropout randomly discards the entire channel of the feature map. This process helps prevent the model from over-relying on local features. The sparsified output of the second encoding layer is input to the third encoding layer, which consists of 64 neurons and uses the LeakyReLU activation function to further compress the feature vector to generate the output of the third encoding layer. In order to increase the diversity of the data and the robustness of the model, the output of the third encoding layer is concatenated with the noise vector to generate a noise-enhanced feature vector. The noise-enhanced feature vector is then input to the fourth encoding layer for encoding. The fourth encoding layer consists of 32 neurons and uses the tanh activation function for encoding to obtain the initial thickness change potential representation. The use of the tanh function compresses the features into the range of -1 to 1, which helps capture complex nonlinear relationships. After completing the encoding stage, the initial thickness change potential representation is input to the first decoding layer of the second autoencoder. The decoding layer consists of 32 neurons and uses the ReLU activation function for preliminary decoding to generate the first decoded output. In order to ensure the stability of the decoding process and the standardization of the data, the first decoding output is conditionally normalized. Conditional normalization can provide adaptive standardization under different decoding conditions, thereby improving the decoding ability of the model. The first decoding output after conditional normalization is then input into the second decoding layer, which consists of 64 neurons and reconstructs the features through the ReLU activation function to generate the second decoding output.In order to enhance the information extraction capability in the decoding process, the Self-Attention mechanism is analyzed for the second decoded output. The Self-Attention mechanism can automatically assign different attention weights to different features during the decoding process, thereby highlighting the most important information for model prediction. The second decoded output after attention weighting is input into the third decoding layer, which consists of 128 neurons, and a deeper feature reconstruction is performed through the ReLU activation function to obtain the third decoded output. In order to enhance the multi-scale representation of the feature, the third decoded output is jump-connected with the second encoding output to form a multi-scale fusion feature. The jump connection can directly transfer the information of the early encoding stage to the later decoding stage, effectively alleviating the gradient vanishing problem in the deep network. The multi-scale fusion feature is then input into the fourth decoding layer, which consists of 256 neurons, and the Sigmoid activation function is used for the final feature reconstruction. Through this process, the potential representation of the target thickness change is obtained.

[0041] The initial pheromone matrix is ​​normalized so that the elements of the pheromone matrix are on the same scale. Based on the normalized pheromone matrix and the Euclidean distance between the feature clusters, the heuristic information matrix is ​​calculated. The heuristic information matrix provides guidance information for path selection, reflects the attraction between different feature clusters, and can help ants choose paths more effectively. According to the normalized pheromone matrix and the heuristic information matrix, the state transition probability matrix of each ant is calculated. The state transition probability matrix reflects the probability distribution of ants moving from one node to another, which is determined by the pheromone concentration and the heuristic information. In order to determine the specific path of each ant, the roulette selection method is used to sample the transition probability distribution. The roulette selection method is a probability-based selection mechanism. Through this method, each ant will select the next node to visit according to the calculated probability until all nodes are visited to form a complete path for each ant. The complete path of each ant is optimized by local search. Local search can further improve the path quality, reduce the path length and optimize the eccentricity change. Through this optimization, a more ideal ant path is obtained. The path length and eccentricity changes of each path are calculated, and these results are substituted into the preset fitness function to calculate the fitness value of each ant path. The fitness value reflects the quality of the path. The higher the fitness, the closer the path is to the ideal mold position adjustment solution. After sorting the fitness values ​​of all ants, the path with the highest fitness is selected as the optimal path of the current iteration, that is, the local optimal path. In order to ensure the update of the global optimal path, the local optimal path of the current iteration is compared with the global optimal path in history, and the global optimal path is updated according to the comparison results. Based on the optimized ant path and fitness value, the pheromone increment of each path is calculated. The pheromone increment reflects the importance of the path. The higher the fitness, the greater the pheromone increment of the path, which can have a greater impact on the subsequent ant path selection. At the same time, the evaporation coefficient is used to attenuate the pheromone matrix to avoid stagnation of path selection due to excessive pheromone concentration. The attenuated pheromone matrix is ​​combined with the new pheromone increment to form an updated pheromone matrix, thereby guiding the path selection in the next iteration. The updated global optimal path is subjected to curve interpolation processing to obtain a smoothed optimal path. Smoothing helps eliminate irregularities in the path and generate an adjustment plan that is more suitable for practical applications. The smoothed optimal path is projected into three-dimensional space, and discretized sampling and coordinate transformation are performed to finally generate the mold position adjustment parameters. These adjustment parameters can be directly applied to the actual mold adjustment operation to achieve precise control of the eccentricity of the cable extrusion layer, thereby improving product quality and production efficiency.

[0042] In the embodiment of the present invention, through multi-angle scanning and digital filtering technology, the method can obtain more accurate cable cross-sectional profile data and greatly improve the accuracy of eccentricity measurement. By adopting efficient data processing algorithms and autoencoder models, fast feature extraction and analysis are achieved, so that eccentricity detection can keep up with the real-time speed of the production line. Combined with multiple autoencoders and ant colony decision fusion models, this method can adapt to the structural characteristics of different types of cables. Through in-depth analysis of eccentricity distribution characteristics and thickness change characteristics, it can effectively capture the dynamic change characteristics during cable extrusion. The ant colony algorithm is used to perform mold position compensation analysis, automatically generate adjustment parameters, reduce human intervention, and improve production efficiency. The RANSAC algorithm and wavelet transform are used to effectively reduce the errors in the measurement and calculation process. Through repeated measurements and statistical feature calculations, the reliability and stability of the detection results are enhanced, which is conducive to the realization of automatic control of the entire process.

[0043] In a specific embodiment, the process of executing step 101 may specifically include the following steps:

[0044] The target cable is fixedly installed to obtain a cable sample to be tested, and the rotation angle of the cable sample to be tested is calibrated to obtain an initial angle position;

[0045] The cable sample to be tested is rotated at equal intervals according to the initial angle position to obtain N sampling angle positions, and multi-point laser scanning is performed on each sampling angle position to obtain coordinate data of M scanning points;

[0046] Repeatedly measure each sampling angle position multiple times according to the coordinate data to obtain an original measurement data set, and perform a mean operation on the original measurement data set to obtain an average coordinate value of each sampling angle position;

[0047] Construct the cable cross-section contour point cloud according to the average coordinate value to obtain the initial contour model, and perform edge detection and contour segmentation on the initial contour model to obtain the conductor outer contour and the insulation layer outer contour;

[0048] The radial distance is calculated according to the outer contour of the conductor and the outer contour of the insulation layer to obtain the radial dimension data of each sampling angle position, and the radial dimension data is interpolated to obtain the cable cross-sectional profile data set including the outer diameter of the conductor and the outer diameter of the insulation layer.

[0049] Specifically, the target cable is fixedly installed to ensure that the cable sample to be tested is firmly installed on the measuring device to obtain the cable sample to be tested. The fixed installation of the cable needs to ensure that the cable does not undergo any displacement or deformation during the measurement process, and also ensure that the axis of the cable is aligned with the rotation axis of the measuring device to prevent deviation during the rotation process. The rotation angle of the cable sample to be tested is calibrated to determine the initial angle position, which will serve as the reference for subsequent multi-angle scanning. For example, the initial angle position is determined by aligning one side of the cable with the initial scanning line of the laser and taking this alignment position as the 0 degree angle. The cable sample to be tested is rotated at equal intervals according to the initial angle position. Assuming that N sampling angle positions are collected, the cable is evenly divided into N angles within a range of 360 degrees. Multi-point laser scanning is performed on the cable sample at each position to obtain the coordinate data of M scanning points at the angle position. The coordinates of each scanning point can be expressed as (x, y), where x and y are the two-dimensional plane coordinates of the scanning point on the cable cross section. In order to improve the accuracy of the measurement data, each sampling angle position is repeatedly measured multiple times, and noise and accidental errors are eliminated through multiple measurements. Assume that K repeated measurements are performed on each angular position to obtain K groups of original measurement data sets. The mean operation is performed on these data, that is, the average value of K measurements is taken for the coordinate value of each scanning point. The contour point cloud of the cable cross section is constructed based on the average coordinate value of all sampled angle positions. A point cloud is a collection of a large number of points used to represent the surface shape of an object. The point cloud reflects the geometric shape of the cable cross section. An initial contour model is formed by connecting the points in sequence. Edge detection and contour segmentation are performed on the initial contour model. The purpose of edge detection is to identify the boundary of the cable contour, which can usually be achieved by detecting areas with large gradient changes in the point cloud. On the cross section of the cable, the boundaries of the outer contour of the conductor and the outer contour of the insulation layer will show obvious gradient changes, and these areas are identified as edges. By segmenting the detected edges, the outer contour of the conductor and the outer contour of the insulation layer are extracted respectively. The radial distance is calculated based on the extracted outer contour of the conductor and the outer contour of the insulation layer. The radial distance refers to the distance from the center of the outer contour of the conductor to each point of the outer contour of the insulation layer. These distances can be used to describe the eccentricity of the cable and the thickness change of the insulation layer. The calculation formula for the radial distance is:

[0050] ;

[0051] in, is the radial distance of the i-th point, is the coordinate of the center of the conductor outer contour, is the center coordinate of the outer contour of the insulation layer, is the coordinate of the i-th point on the outer contour of the insulation layer. After calculating the radial distances of all sampled angle positions, the radial dimension data of each angle position is obtained. In order to generate a continuous cable cross-sectional profile data set, the radial dimension data is interpolated to fill the gaps between angle positions and generate a smooth contour curve. Common interpolation methods include linear interpolation, spline interpolation, etc. After interpolation, a complete cable cross-sectional profile data set including the outer diameter of the conductor and the outer diameter of the insulation layer is obtained.

[0052] In a specific embodiment, the process of executing step 102 may specifically include the following steps:

[0053] Perform wavelet transform on the cable cross-sectional profile data set to obtain multi-scale decomposition coefficients, and perform noise threshold calculation on the multi-scale decomposition coefficients to obtain a denoising threshold;

[0054] According to the denoising threshold, the multi-scale decomposition coefficients are subjected to soft threshold processing to obtain the denoised wavelet coefficients, and the denoised wavelet coefficients are subjected to inverse wavelet transform to obtain the filtered contour data;

[0055] Perform edge detection on the filtered contour data to obtain a conductor outer contour point set and an insulation layer outer contour point set;

[0056] Performing circle fitting on the conductor outer contour point set by a first RANSAC algorithm to obtain the conductor outer circle equation, and performing circle fitting on the insulation layer outer contour point set by a second RANSAC algorithm to obtain the insulation layer outer circle equation;

[0057] Extract the center coordinates and radius according to the conductor outer circle equation to obtain the conductor outer circle parameters, and extract the center coordinates and radius according to the insulation layer outer circle equation to obtain the insulation layer outer circle parameters;

[0058] The uncertainty analysis is performed on the parameters of the outer circle of the conductor and the outer circle of the insulation layer to obtain the center coordinates of the outer circle of the conductor and the outer circle of the insulation layer.

[0059] Specifically, a wavelet transform is performed on the cable cross-sectional profile data set. Wavelet transform is a commonly used signal processing technology, which is suitable for analyzing the local characteristics of non-stationary signals. Through wavelet transform, the cable cross-sectional profile data set is decomposed into different scales to obtain multi-scale decomposition coefficients. These coefficients reflect the characteristics of the signal at different frequencies and help to identify the noise and actual profile characteristics in the data. The noise threshold is calculated for the multi-scale decomposition coefficients. The calculation of the noise threshold is based on an assumption that most of the noise is concentrated in the high-frequency part, while the actual signal is mainly concentrated in the low-frequency part. By analyzing the multi-scale decomposition coefficients, a threshold is determined. The coefficients below this threshold mainly represent the noise component, while the coefficients above this threshold contain useful information of the actual signal. This threshold is called the denoising threshold, which can usually be calculated by methods such as VisuShrink. The calculation formula is as follows:

[0060] ;

[0061] in, is the denoising threshold, represents the noise standard deviation, is the number of data points. The threshold is used to distinguish between signal and noise in the wavelet transform domain. Soft threshold processing is performed on the multi-scale decomposition coefficients according to the denoising threshold. Soft threshold processing is a common denoising method. It reduces the coefficients less than the threshold to zero and reduces the coefficients greater than the threshold by a certain proportion to achieve the purpose of removing noise. After processing, the denoised wavelet coefficients are obtained. The denoised wavelet coefficients are subjected to inverse wavelet transform. The processed coefficients are converted back to the time domain signal to obtain the filtered contour data. Edge detection is performed on the filtered contour data to identify important contour lines in the cable cross section, such as the outer contour of the conductor and the outer contour of the insulation layer. Common edge detection methods include Sobel operator, Canny operator, etc. These methods determine the edge by detecting the area with large grayscale changes in the image. After edge detection, the conductor outer contour point set and the insulation layer outer contour point set are obtained, which represent the key geometric information of the cable cross section. Circle fitting is performed on the conductor outer contour point set to find a best fitting circle so that these contour points can be accurately described by mathematical equations. The RANSAC algorithm is used for circle fitting. The RANSAC (Random Sampling Consensus) algorithm is a robust parameter estimation algorithm that can effectively fit the model when there are outliers in the data. For the conductor outer contour point set, the first RANSAC algorithm is used to fit and obtain the equation of the conductor outer circle, which is in the form of:

[0062] ;

[0063] in, and is the coordinate of the center of the outer circle of the conductor, is the radius of the conductor outer circle. Similarly, the second RANSAC algorithm is used to fit the outer contour point set of the insulating layer to obtain the equation of the outer circle of the insulating layer, and the center coordinates and radius are extracted from the conductor outer circle equation and the insulating layer outer circle equation. The conductor outer circle parameters include the center coordinates of the conductor outer circle and radius , the outer circle parameters of the insulation layer include the center coordinates and radius of the outer circle of the insulation layer. In order to ensure the accuracy of the fitting results, uncertainty analysis is performed on the outer circle parameters of the conductor and the outer circle parameters of the insulation layer to evaluate the measurement accuracy and reliability of these parameters, usually by calculating standard errors and confidence intervals. The standard error can reflect the fluctuation range of the parameter estimate, while the confidence interval provides a range that may contain the true value.

[0064] In a specific embodiment, the process of executing step 103 may specifically include the following steps:

[0065] The coordinate system of the center coordinates of the outer circle of the conductor and the outer circle of the insulating layer is converted to obtain the center position represented by polar coordinates, and the Euclidean distance between the two center positions is calculated based on the center position represented by polar coordinates to obtain the initial eccentricity value;

[0066] The initial eccentricity value is normalized to obtain a relative eccentricity value, and the cable cross section is divided into angles according to the relative eccentricity value to obtain N evenly distributed angle positions;

[0067] The radial distance between the outer circle of the conductor and the outer circle of the insulating layer is calculated for each angular position to obtain the absolute eccentricity value at each angular position, and the eccentricity distribution curve is fitted according to the absolute eccentricity value to obtain a continuous eccentricity function;

[0068] Perform differential operation on the eccentricity function to obtain an eccentricity change rate curve, and identify the eccentricity mutation point according to the eccentricity change rate curve to obtain the target eccentricity angle position;

[0069] The thickness of the insulation layer at the target eccentric angle position is calculated to obtain the insulation layer thickness distribution data, and the thickness variation curve is fitted according to the insulation layer thickness distribution data to obtain the thickness variation data of the insulation layer at different angles.

[0070] Specifically, the coordinates of the center of the outer circle of the conductor and the outer circle of the insulating layer are converted from the rectangular coordinate system to the polar coordinate system. Assume that the coordinates of the center of the outer circle of the conductor are , the center coordinates of the outer circle of the insulation layer are , by converting these two coordinates into polar coordinates, we can get the position of the center of the circle in the polar coordinate system. The polar coordinate representation is ,in, represents the distance from the center of the circle to the origin, Represents the angle between the center of the circle and the polar axis. For the center of the outer circle of the conductor, the polar coordinate is expressed as:

[0071] ;

[0072] Similarly, the polar coordinates of the center of the outer circle of the insulation layer are expressed as:

[0073] ;

[0074] The Euclidean distance between the two circle centers is calculated by polar coordinates. This distance reflects the eccentricity between the outer circle of the conductor and the outer circle of the insulation layer. The calculation formula of the Euclidean distance is:

[0075] ;

[0076] in, Indicates the initial eccentricity value, which directly reflects the concentricity of the conductor and the insulation layer. The larger the value, the greater the eccentricity of the cable. The initial eccentricity value is normalized and scaled to a standard range, usually [0, 1], to obtain the relative eccentricity value. The normalization formula is as follows:

[0077] ;

[0078] in, and are the minimum and maximum eccentricity values ​​in a set of measurement data, respectively. The value of relative eccentricity is between 0 and 1. The closer the value is to 1, the greater the eccentricity. The cable cross section is divided into angles according to the relative eccentricity value. Assume that the cross section is evenly divided into N angle positions. At each angle position, the radial distance between the outer circle of the conductor and the outer circle of the insulation layer is calculated to obtain the absolute eccentricity value of each angle position. The absolute eccentricity value indicates the distance between the centers of the outer circle of the conductor and the outer circle of the insulation layer at a certain angle, reflecting the degree of eccentricity of the cable at this angle position. The eccentricity distribution curve is obtained by fitting the absolute eccentricity values ​​of each angle position. This curve is a continuous function that describes the distribution of eccentricity over the entire cable cross section. Common fitting methods include polynomial fitting and spline fitting. Differentiate the eccentricity distribution curve to obtain the eccentricity change rate curve. The eccentricity change rate curve reflects the rate at which the eccentricity changes with the angle. The extreme value of the change rate usually represents the mutation point of the eccentricity, that is, the eccentricity has changed significantly at this angle position. By identifying the mutation point, the target eccentric angle position is determined. The target position is the angle that requires special attention and adjustment in cable production, usually the position with the largest eccentricity or the most dramatic change. The insulation thickness at the target eccentric angle position is calculated. The insulation thickness is calculated by comparing the radial distance between the outer circle of the conductor and the outer circle of the insulation layer at this angle position. Specifically, the thickness can be expressed as the radial distance of the outer circle of the insulation layer minus the radial distance of the outer circle of the conductor, that is:

[0079] ;

[0080] in, and Respectively represent the radii of the outer circle of the insulation layer and the outer circle of the conductor. By calculating the thickness at different angles, the distribution data of the insulation layer thickness is obtained. According to the distribution data of the insulation layer thickness, a curve fitting is performed to obtain the thickness variation curve of the insulation layer at different angles. This curve describes the distribution of the insulation layer thickness on the cable cross section and reveals the uniformity of the thickness. By analyzing this curve, the thickness variation trend of the insulation layer at different angles can be found, and the areas with uneven thickness can be identified, thereby guiding the adjustment of the mold during the cable production process.

[0081] In a specific embodiment, the process of executing step 104 may specifically include the following steps:

[0082] Performing Fourier transform on the eccentricity value to obtain the frequency domain representation of the eccentricity, and extracting the target frequency component of the eccentricity according to the frequency domain representation to obtain the harmonic characteristics of the eccentricity;

[0083] The kernel density estimation is performed on the eccentricity value to obtain the probability density function of the eccentricity, and the mean, variance, skewness and kurtosis of the eccentricity are calculated according to the probability density function to obtain the statistical moment characteristics of the eccentricity;

[0084] The spatial autocorrelation analysis of the eccentricity values ​​is performed to obtain the spatial distribution characteristics of the eccentricity, and the maximum value of the eccentricity and its corresponding angle are determined according to the spatial distribution characteristics to obtain the maximum eccentricity direction;

[0085] The thickness variation data is decomposed by wavelet packet to obtain multi-scale wavelet coefficients, and the energy entropy and corresponding dimension are calculated based on the multi-scale wavelet coefficients to obtain the complexity characteristics of the thickness variation.

[0086] Perform empirical mode decomposition on the thickness variation data to obtain multiple intrinsic mode functions, and extract the instantaneous frequency and amplitude based on the intrinsic mode functions to obtain the time-frequency characteristics of the thickness variation;

[0087] The harmonic features, statistical moment features and spatial distribution features are combined to obtain the eccentricity distribution features, and the complexity features and time-frequency features are combined to obtain the thickness variation features.

[0088] Specifically, the eccentricity value is converted to the frequency domain for analysis. Fourier transform is a mathematical method that converts time domain signals into frequency domain signals, which can reveal the frequency components in the signal and their corresponding amplitudes. Assume that the eccentricity data is ,in Represents different angular positions, and the frequency domain after Fourier transform is expressed as ,in Represents the frequency component. The formula for Fourier transform is:

[0089] ;

[0090] in, Indicates the eccentricity signal at frequency The complex amplitude at is an imaginary unit. The target frequency components of eccentricity are analyzed and extracted. These frequency components have large amplitudes and represent the most significant periodic changes in the signal. These frequency components are harmonic components and constitute the harmonic characteristics of eccentricity. Harmonic characteristics are used to analyze the source of periodic problems in the cable production process, such as periodic deviation of molds or uneven material flow. Kernel density estimation is performed on the eccentricity value. Kernel density estimation is a nonparametric method used to estimate the probability density function of data. Assume that the eccentricity value is ,in represents different sample points, then the probability density function The estimation formula is:

[0091] ;

[0092] in, is the sample size, is the bandwidth parameter, is a kernel function. Commonly used kernel functions include Gaussian kernel, uniform kernel, etc. The probability density function obtained by kernel density estimation can intuitively show the distribution of eccentricity values ​​in different ranges. Based on the probability density function, the statistical moment characteristics of eccentricity are calculated, including mean, variance, skewness and kurtosis. The mean reflects the central tendency of eccentricity, the variance describes the degree of discreteness of the data, the skewness indicates the symmetry of the data distribution, and the kurtosis reveals the sharpness of the distribution. Perform spatial autocorrelation analysis on the eccentricity values. Spatial autocorrelation analysis is used to measure the spatial similarity of data, usually expressed by Moran's I index. The calculation formula of Moran's I is:

[0093] ;

[0094] in, is the sample size, is the average value of eccentricity, is the spatial weight matrix, is the sum of the weights. The value of Moran's I index is between -1 and 1. A positive value indicates a positive correlation in space, and a negative value indicates a negative correlation in space. The spatial distribution characteristics of eccentricity are determined by spatial autocorrelation analysis, and the maximum value of eccentricity and its corresponding angle are determined based on these characteristics to obtain the direction of maximum eccentricity. The direction of maximum eccentricity usually indicates the direction of the maximum error that may exist in the production process, and is a critical angle that requires special attention and adjustment. Perform wavelet packet decomposition on the thickness variation data. Wavelet packet decomposition is a multi-scale analysis method that can decompose the signal into wavelet coefficients in different frequency bands, thereby revealing the characteristics of the signal at each frequency scale. Assume that the thickness variation data is , wavelet coefficients at different scales are obtained by wavelet packet decomposition ,in represents the level of decomposition, Represents the corresponding frequency band index. The energy entropy and corresponding dimension of the signal are calculated through multi-scale wavelet coefficients. Energy entropy is an indicator to measure the complexity of the signal, which describes the energy distribution of the signal in each frequency component. The calculation formula of energy entropy is:

[0095] ;

[0096] in, Indicates The ratio of the energy of a frequency band to the total energy. The higher the energy entropy, the greater the complexity of the signal. The corresponding dimension reveals the multi-scale characteristics of the signal by analyzing the contribution of each frequency component. These features together constitute the complexity characteristics of thickness change, which are used to describe the fluctuation of the thickness of the cable insulation layer. The thickness change data is subjected to empirical mode decomposition to extract the intrinsic mode functions (IMFs) of the signal. Empirical mode decomposition is a method suitable for nonlinear and non-stationary signal analysis. It decomposes the signal into several IMFs, each of which represents the oscillation mode of the signal on a certain time scale. By analyzing these IMFs, the instantaneous frequency and amplitude are extracted. These parameters constitute the time-frequency characteristics of thickness change. The instantaneous frequency reflects the frequency change of the signal in time, while the amplitude describes the intensity of the signal at different time points. The harmonic characteristics, statistical moment characteristics and spatial distribution characteristics of eccentricity are combined to form the distribution characteristics of eccentricity. This combined feature not only reveals the periodic change of eccentricity, but also includes the central tendency and spatial distribution characteristics of eccentricity. For example, if the spatial autocorrelation of the eccentricity is strong and the statistical moment characteristics show that the kurtosis of the data is high, this may indicate that there are systematic deviations in the cable production and stricter quality control is needed during the production process. Similarly, the complexity characteristics of the thickness variation and the time-frequency characteristics are combined to obtain the comprehensive characteristics of the thickness variation. These characteristics describe the variation pattern of the cable insulation thickness in the time and frequency domains. For example, if the energy entropy is high and the time-frequency characteristics show obvious instantaneous frequency fluctuations, this may indicate that the insulation thickness has changed significantly during the production process and the thickness distribution needs to be stabilized by adjusting the production process.

[0097] In a specific embodiment, the process of executing step 105 may specifically include the following steps:

[0098] Perform dimension reduction processing on the eccentricity distribution characteristics to obtain the eccentricity feature vector after dimension reduction;

[0099] The eccentricity feature vector after dimension reduction is input into the first autoencoder in multiple autoencoders and the ant colony decision fusion model, and feature extraction and reconstruction are performed through three-layer encoders and three-layer decoders to obtain the target eccentricity potential representation;

[0100] Perform principal component analysis on the thickness variation characteristics to obtain the principal component eigenvector;

[0101] The principal component feature vector is input into multiple autoencoders and the second autoencoder in the ant colony decision fusion model, and feature extraction and reconstruction are performed through a four-layer encoder and a four-layer decoder to obtain a potential representation of target thickness change;

[0102] The maximum eccentricity direction, the target eccentricity potential representation and the target thickness change potential representation are feature fused to obtain a fused feature vector;

[0103] Adaptive fuzzy c-means clustering is performed on the fused feature vector to obtain multiple feature clusters, and the initial pheromone matrix of the ant colony decision layer in the ant colony decision fusion model is constructed based on the feature clusters. At the same time, the number of ants, the number of iterations and the heuristic factor are set;

[0104] Based on the roulette selection method and state transition probability calculation, the path of each ant is selected to obtain the complete path of each ant, and the fitness value of each ant path is calculated. The global optimal path and the local optimal path are updated, and the pheromone matrix is ​​updated at the same time, and the global optimal path is decoded into the mold position adjustment parameter.

[0105] Specifically, the high-dimensional eccentricity distribution feature vector is reduced in dimension to simplify the calculation and retain the most important information. Common dimensionality reduction methods include principal component analysis, linear discriminant analysis, etc. Assume that principal component analysis is used to reduce the dimension of the eccentricity distribution feature to obtain the reduced eccentricity feature vector. The reduced eccentricity feature vector is input into the first autoencoder in the multiple autoencoders and ant colony decision fusion model for processing. The first autoencoder consists of three layers of encoders and three layers of decoders. The feature vector is input into the first layer of encoder for preliminary nonlinear transformation, compressing the original features and extracting the main features. As the feature vector passes through the second and third layers of encoders, the features are further compressed to obtain a target eccentricity potential representation containing the core information of the eccentricity. The decoder part decodes the compressed potential representation layer by layer and reconstructs it back to a form similar to the input feature. At the same time, principal component analysis is performed on the thickness change feature to extract the most representative principal component feature vector. Principal component analysis projects the data into a new low-dimensional space by finding the direction with the largest variance in the data set, thereby reducing the dimension of the data and retaining as much information as possible. The principal component feature vector processed by principal component analysis is also input into the second autoencoder in the multiple autoencoders and ant colony decision fusion model. The second autoencoder consists of a four-layer encoder and a four-layer decoder. The feature vector is gradually compressed by the four-layer encoder, and more important and abstract features are extracted in each layer to generate the target thickness change potential representation. The four-layer decoder decodes the potential representation layer by layer and reconstructs it back to the original space to ensure that the potential representation can accurately reflect the original information of the thickness change. The target eccentricity potential representation and the target thickness change potential representation are fused. During the fusion process, the maximum eccentricity direction is considered to reflect the direction with the most serious eccentricity. A fused feature vector is obtained by fusion of the maximum eccentricity direction, the target eccentricity potential representation and the target thickness change potential representation. Adaptive fuzzy C-means clustering is performed on the fused feature vector. Adaptive fuzzy C-means clustering is an improved fuzzy clustering algorithm that can automatically adjust the center of the cluster and the fuzziness of the classification according to the distribution of the data. Through adaptive fuzzy C-means clustering, the fused feature vector is divided into multiple feature clusters, each of which represents a typical eccentricity and thickness change pattern. Based on these feature clusters, the initial pheromone matrix of the ant colony decision layer in the ant colony decision fusion model is constructed. The pheromone matrix is ​​a key parameter in the ant colony algorithm, reflecting the priority of path selection. After setting the number of ants, the number of iterations, and the heuristic factor, the path selection is performed for each ant based on the roulette selection method and the state transition probability calculation. The roulette selection method determines the path node that the ant chooses next based on the current state of the ant and the pheromone concentration. The state transition probability is determined by the pheromone concentration and the heuristic information, which controls the direction of movement of the ant in the search space. After each ant completes the path selection, a complete path is formed.The quality of the path is measured by the fitness value. The higher the fitness value, the better the path. After calculating the fitness value of each ant path, these paths are sorted, and the path with the highest fitness is selected as the optimal path for the current iteration and compared with the global optimal path in the previous iteration. If the fitness value of the current path is higher, the global optimal path is updated. At the same time, the pheromone matrix is ​​updated according to the fitness value of the path. The path with higher fitness will leave more pheromones, thereby guiding more ants to choose this path in the next iteration. The updated global optimal path is decoded into the mold position adjustment parameter. In this way, the entire system can automatically optimize the position of the mold in the cable production process to reduce the unevenness of eccentricity and thickness variation and improve the quality of cable products. For example, if it is found in actual operation that the fitness value of a certain path is much higher than that of other paths, this may mean that the mold position corresponding to the path can significantly reduce the unevenness of eccentricity and thickness variation. By decoding the path, the position of the mold is adjusted to obtain better quality cable products in actual production.

[0106] In a specific embodiment, the execution step inputs the dimension-reduced eccentricity feature vector into the first autoencoder in the multiple autoencoders and the ant colony decision fusion model, extracts and reconstructs the features through the three-layer encoder and the three-layer decoder, and obtains the target eccentricity potential representation. The process may specifically include the following steps:

[0107] The eccentricity feature vector after dimension reduction is normalized to obtain a normalized eccentricity feature vector, and the normalized eccentricity feature vector is input into the first encoding layer of the first autoencoder, and a nonlinear transformation is performed through 128 neurons and a ReLU activation function to obtain the first hidden layer representation;

[0108] Perform batch normalization on the first hidden layer representation to obtain a normalized first hidden layer representation, and input the normalized first hidden layer representation into the second encoding layer of the first autoencoder, perform feature compression through 64 neurons and ReLU activation function, and obtain a second hidden layer representation;

[0109] Perform a dropout operation on the second hidden layer representation to obtain a sparse second hidden layer representation, and input the sparse second hidden layer representation into the third encoding layer of the first autoencoder, perform feature encoding through 32 neurons and a tanh activation function, and obtain an initial eccentricity potential representation;

[0110] The initial eccentricity potential representation is input into the first decoding layer of the first autoencoder, and is initially decoded through 32 neurons and the ReLU activation function to obtain the output of the first decoding layer;

[0111] Perform Layer Normalization on the output of the first decoding layer to obtain a normalized output of the first decoding layer, and input the normalized output of the first decoding layer into the second decoding layer of the first autoencoder, perform feature reconstruction through 64 neurons and ReLU activation function, and obtain the output of the second decoding layer;

[0112] The output of the second decoding layer is residually connected, and the output of the second decoding layer is added to the representation of the second hidden layer to obtain the residual enhanced output of the second decoding layer. The residual enhanced output of the second decoding layer is input into the third decoding layer of the first autoencoder and reconstructed through 128 neurons and the Sigmoid activation function to obtain the potential representation of the target eccentricity.

[0113] Specifically, the eccentricity feature vector after dimensionality reduction is normalized, and the data is scaled to a standard range to eliminate the dimensional differences between different features and ensure that each feature can participate in the training in a balanced manner when training the neural network. The normalization process usually adjusts the data to the range of [0, 1] or [-1, 1]. Through the normalization process, a normalized eccentricity feature vector is obtained. The normalized eccentricity feature vector is input into the first encoding layer of the first autoencoder. The encoding layer consists of 128 neurons and uses the ReLU activation function for nonlinear transformation. The ReLU activation function (Rectified Linear Unit) is an activation function widely used in current deep learning, which can effectively solve the gradient disappearance problem. The ReLU function is defined as:

[0114] ;

[0115] After being processed by the first encoding layer, the feature vector is converted into the first hidden layer representation ,in Contains the output of 128 neurons. In order to improve the stability and training efficiency of the model, the first hidden layer is represented Perform batch normalization. Normalize the output of each batch of data to a standard normal distribution, thereby accelerating the convergence of the network and reducing the dependence on initial parameters. The output after batch normalization is the normalized first hidden layer representation , this output is input to the second encoding layer of the first autoencoder. The second encoding layer consists of 64 neurons and continues to use the ReLU activation function for feature compression. Through this layer, the feature vector is further compressed and more important features are extracted to obtain the second hidden layer representation , contains the output of 64 neurons. In order to prevent the model from overfitting, the second hidden layer is represented Perform dropout operation. Dropout is a technique that randomly discards a certain proportion of neurons during training, which can effectively improve the generalization ability of the model. Assume that the probability of dropout is , then each neuron has The probability of being retained. After the dropout operation, the sparse second hidden layer representation is obtained The sparse second hidden layer is represented as Input to the third encoding layer of the first autoencoder. This layer consists of 32 neurons and uses the tanh activation function for feature encoding. The Tanh activation function (hyperbolic tangent function) is defined as:

[0116] ;

[0117] The Tanh function compresses the input value into the range of [-1, 1], which helps to capture the nonlinear relationship in the input data. After this layer of processing, the initial eccentricity potential representation is obtained. ,in Contains the output of 32 neurons, representing the compressed core features. The initial eccentricity potential representation Input to the first decoding layer of the first autoencoder. This decoding layer consists of 32 neurons and uses the ReLU activation function for preliminary decoding. After processing by the decoding layer, the feature vector is re-expanded to obtain the output of the first decoding layer . Output of the first decoding layer Perform Layer Normalization. Layer Normalization is another normalization technique. Unlike Batch Normalization, it normalizes all features of each sample instead of normalizing each batch of data. The output after Layer Normalization is the normalized output of the first decoding layer. . Output the normalized first decoding layer The input is sent to the second decoding layer, which consists of 64 neurons and uses the ReLU activation function for feature reconstruction. After this layer of processing, the output of the second decoding layer is obtained. In order to enhance the expressive power of the model, the output of the second decoding layer Perform residual connection. Residual connection is a technique that uses skip connections to alleviate the gradient vanishing problem in deep networks. Output the second decoding layer With the second hidden layer represented Add the residual enhanced second decoding layer output . Output the second decoding layer of residual enhancement Input to the third decoding layer of the first autoencoder, which consists of 128 neurons and uses the Sigmoid activation function for final feature reconstruction. The Sigmoid activation function is defined as:

[0118] ;

[0119] The Sigmoid function compresses the output to the range of [0, 1], so that the reconstructed features can be directly used for further analysis and decision-making. Through this layer of processing, the potential representation of the target eccentricity is finally obtained. .

[0120] In a specific embodiment, the execution step inputs the principal component feature vector into a plurality of autoencoders and a second autoencoder in the ant colony decision fusion model, extracts and reconstructs features through a four-layer encoder and a four-layer decoder, and obtains a potential representation of the target thickness change. The process may specifically include the following steps:

[0121] The principal component eigenvector is standardized to obtain a standardized principal component eigenvector;

[0122] The standardized principal component feature vector is input into the first encoding layer of the second autoencoder, and nonlinearly transformed through 256 neurons and the LeakyReLU activation function to obtain the output of the first encoding layer;

[0123] Perform Instance Normalization on the output of the first coding layer to obtain the normalized output of the first coding layer, and input the normalized output of the first coding layer into the second coding layer of the second autoencoder, perform feature compression through 128 neurons and LeakyReLU activation function, and obtain the output of the second coding layer;

[0124] Perform Spatial Dropout on the output of the second coding layer to obtain a sparse output of the second coding layer, and input the sparse output of the second coding layer into the third coding layer of the second autoencoder, compress it through 64 neurons and the LeakyReLU activation function to obtain the output of the third coding layer;

[0125] The output of the third coding layer is concatenated with the noise vector to obtain a noise-enhanced feature vector, and the noise-enhanced feature vector is input into the fourth coding layer and encoded through 32 neurons and the tanh activation function to obtain the initial thickness change potential representation;

[0126] The initial thickness change potential representation is input into the first decoding layer of the second autoencoder, and is initially decoded through 32 neurons and the ReLU activation function to obtain the first decoding output;

[0127] Conditionally normalize the first decoding output to obtain a conditionally normalized first decoding output, and input the conditionally normalized first decoding output into the second decoding layer of the second autoencoder, perform feature reconstruction through 64 neurons and a ReLU activation function, and obtain a second decoding output;

[0128] Perform a Self-Attention mechanism analysis on the second decoded output to obtain an attention-weighted second decoded output, and input the attention-weighted second decoded output into the third decoding layer of the second autoencoder, reconstruct it through 128 neurons and the ReLU activation function to obtain a third decoded output;

[0129] The third decoding output is jump-connected with the second encoding output to obtain multi-scale fusion features, and the multi-scale fusion features are input into the fourth decoding layer of the second autoencoder and reconstructed through 256 neurons and Sigmoid activation function to obtain the potential representation of target thickness change.

[0130] Specifically, the principal component eigenvector is standardized and the data in the eigenvector is scaled to the same range. Usually, standardization adjusts the data to a normal distribution with a mean of 0 and a variance of 1 to obtain a standardized principal component eigenvector. Input to the first encoding layer of the second autoencoder. This encoding layer consists of 256 neurons and uses the LeakyReLU activation function for nonlinear transformation. The LeakyReLU activation function is an improved version of the ReLU function that allows negative values ​​to be passed through a small negative slope to avoid the "neuron death" problem that may be caused by ReLU. The definition of LeakyReLU is as follows:

[0131] ;

[0132] in, is a small constant, usually 0.01. After being processed by the first encoding layer, the feature vector is converted to the output of the first encoding layer . Output of the first coding layer Perform Instance Normalization. Instance Normalization is different from Batch Normalization. It performs normalization on each sample independently and is usually used in image style transfer tasks. Through this operation, the normalized output of the first encoding layer is obtained. , and input it to the second encoding layer of the second autoencoder. The second encoding layer consists of 128 neurons and also uses the LeakyReLU activation function for feature compression. In this layer, the feature vector is further compressed, redundant information is removed, and more representative features are extracted to obtain the output of the second encoding layer In order to enhance the generalization ability of the model and prevent overfitting, the output of the second encoding layer Perform Spatial Dropout. Spatial Dropout is different from traditional Dropout. It is a dropout operation for feature map channels. This operation is suitable for convolutional neural networks. By randomly dropping a part of the feature channels, the model is made more robust to the local features in the input data. After Spatial Dropout processing, the sparse second encoding layer output is obtained. . Output the sparse second encoding layer Input to the third encoding layer, which consists of 64 neurons and continues to use the LeakyReLU activation function for feature compression. Through this step, the feature vector is further compressed and refined to obtain the output of the third encoding layer In order to enhance the expressive power of features, the output of the third encoding layer is with a randomly generated noise vector Splice to get the noise-enhanced feature vector . This operation introduces noise, allowing the model to learn more robust and generalized features. The noise-enhanced feature vector is input to the fourth encoding layer. The fourth encoding layer consists of 32 neurons and is encoded using the tanh activation function. The tanh activation function compresses the input to the range of [-1, 1], which helps capture nonlinear relationships in the input data. After this layer of processing, the initial thickness change potential representation is obtained. , which contains the thickness variation features after noise enhancement and multi-layer compression. The initial thickness variation potential representation Input to the first decoding layer of the second autoencoder, which consists of 32 neurons and uses the ReLU activation function for preliminary decoding. Through this step of decoding, the feature vector begins to be re-expanded from the compressed potential representation to obtain the first decoding output. The first decoding output is conditionally normalized. Conditional normalization is a normalization technique for specific conditions in the decoder that can ensure the stability and consistency of the output under different decoding conditions. The output after conditional normalization is the normalized output of the first decoding layer. The normalized output of the first decoding layer is input to the second decoding layer, which consists of 64 neurons and uses the ReLU activation function for feature reconstruction. After processing this layer, the second decoding output is obtained. The second decoding output is analyzed by the Self-Attention mechanism. The Self-Attention mechanism enables the model to focus on the most important features by calculating the attention weight of each position in the input sequence to other positions. The output after attention weighting is the enhanced output of the second decoding layer. The attention-weighted second decoding output is input to the third decoding layer, which consists of 128 neurons and uses the ReLU activation function for further feature reconstruction to obtain the third decoding output. The third decoded output is jump-connected with the second encoded output to obtain multi-scale fusion features. Jump connection is a technique that enhances the model feature expression capability by directly connecting features between different layers. Multi-scale fusion features integrate features at different levels. The multi-scale fusion features are input to the fourth decoding layer of the second autoencoder, which consists of 256 neurons and uses the Sigmoid activation function for final feature reconstruction. The Sigmoid activation function compresses the output to the range of [0, 1], making the reconstructed features suitable for further analysis and decision-making, and finally obtaining the potential representation of target thickness change.

[0133] In a specific embodiment, the execution step selects a path for each ant based on the roulette selection method and the state transition probability calculation, obtains the complete path of each ant, calculates the fitness value of each ant path, updates the global optimal path and the local optimal path, and updates the pheromone matrix at the same time. The process of decoding the global optimal path into the mold position adjustment parameter can specifically include the following steps:

[0134] The initial pheromone matrix is ​​normalized to obtain a normalized pheromone matrix, and the heuristic information matrix is ​​calculated according to the Euclidean distance between the feature clusters;

[0135] According to the normalized pheromone matrix and the heuristic information matrix, the state transition probability matrix is ​​calculated for each ant to obtain the transition probability distribution of each ant;

[0136] The roulette wheel selection method is used for the transition probability distribution to select the next visited node for each ant, and the single-step path selection result of each ant is obtained until all nodes are visited and the complete path of each ant is obtained;

[0137] Perform local search optimization on the complete path to obtain the optimized ant path, calculate the path length and eccentricity changes based on the optimized ant path, and substitute the path length and eccentricity changes into the preset fitness function to obtain the fitness value of each ant path;

[0138] Sort the fitness values ​​of all ants, select the path with the highest fitness as the optimal path for the current iteration, obtain the local optimal path, and compare the local optimal path with the historical global optimal path to obtain the updated global optimal path;

[0139] According to the optimized ant paths and fitness values, the pheromone increment of each path is calculated, and the pheromone matrix is ​​decayed using the evaporation coefficient to obtain the updated pheromone matrix;

[0140] The updated global optimal path is interpolated by curves to obtain a smoothed optimal path, and the smoothed optimal path is projected into three-dimensional space for discrete sampling and coordinate transformation to obtain mold position adjustment parameters.

[0141] Specifically, the initial pheromone matrix is ​​normalized and the values ​​in the pheromone matrix are adjusted to a standard range to ensure that the pheromone influence of each path can be balanced during the operation of the algorithm. Assume that the initial pheromone matrix is ,in Represents the pheromone concentration from node i to node j. The normalized pheromone matrix The calculation formula is:

[0142] ;

[0143] in, and are the minimum and maximum values ​​in the initial pheromone matrix, respectively. Through normalization, the normalized pheromone matrix is ​​obtained. The heuristic information matrix is ​​calculated based on the Euclidean distance between feature clusters. Euclidean distance is a commonly used distance measurement method that can quantify the similarity or difference between two feature clusters. Assume that the center of feature cluster i is , the center of feature cluster j is , then the Euclidean distance between them It can be expressed as:

[0144] ;

[0145] in, and are the coordinate values ​​of feature cluster i and feature cluster j on the kth dimension respectively. The calculated distance values ​​are used to construct the heuristic information matrix ,in:

[0146] ;

[0147] Represents the heuristic information from node i to node j. The larger the heuristic information, the more attractive this path is in the ant selection process. According to the normalized pheromone matrix and the heuristic information matrix , calculate the state transition probability matrix of each ant. State transition probability Determines the possibility of an ant moving from node i to node j, which is determined by the pheromone concentration and heuristic information. The calculation formula is:

[0148] ;

[0149] in, and is a parameter that controls the importance of pheromone and heuristic information, and allowed represents the set of nodes that the ant has not visited yet. By calculating the state transition probability matrix , and obtain the transition probability distribution of each ant. Perform roulette wheel selection on the transition probability distribution to determine the next visited node for each ant. Roulette wheel selection is a probability-based selection strategy. Each possible choice has a corresponding probability, and the selection is made according to this probability. The ant selects the next node based on the state transition probability until all nodes have been visited, and finally forms the complete path of each ant. Perform local search optimization on the complete path of each ant to improve the quality of the path. Local search optimization usually includes operations such as exchange and reversal. Through these operations, a shorter or better path than the original path is found. The optimized ant path not only has a shorter path length, but also better reflects the change in eccentricity. The path length and eccentricity changes can be calculated by the following formula:

[0150] ;

[0151] in, Represents the distance between consecutive nodes on the path. Substitute the path length and eccentricity changes into the preset fitness function to obtain the fitness value of each ant path. The fitness function may be a weighted sum, or a comprehensive measure of the path length and eccentricity changes. Sort the fitness values ​​of all ants, and select the path with the highest fitness as the optimal path for the current iteration to obtain the local optimal path. Compare the local optimal path with the historical global optimal path to determine whether the global optimal path needs to be updated. By comparison, if the fitness value of the local optimal path is higher, update the global optimal path. According to the optimized ant path and fitness value, calculate the pheromone increment of each path, and use the evaporation coefficient The pheromone matrix is ​​decayed. The calculation of pheromone increment is usually proportional to the fitness of the path, and the decay of pheromone is to avoid infinite accumulation of pheromone, thereby ensuring the exploration ability of the algorithm. Updated pheromone matrix The update formula is:

[0152] ;

[0153] in, is the pheromone increment left by the ant on the path, is the pheromone evaporation coefficient. After updating the pheromone matrix, the global optimal path is subjected to curve interpolation processing to obtain a smoothed optimal path. Curve interpolation can be performed by methods such as spline interpolation to make the path smoother and more natural. The smoothed optimal path is projected into three-dimensional space, and discretized sampling and coordinate transformation are performed to finally obtain the mold position adjustment parameter. This parameter is used for mold adjustment in actual production to optimize the control of eccentricity and thickness during cable production.

[0154] The above describes the automatic eccentricity checking method of the cable extrusion layer in the embodiment of the present invention. The following describes the automatic eccentricity checking device of the cable extrusion layer in the embodiment of the present invention. Figure 2 An embodiment of the automatic eccentricity checking device of the cable extrusion layer in the embodiment of the present invention comprises:

[0155] A scanning module 201 is used to perform multi-angle scanning on the target cable to obtain a cable cross-sectional profile data set including the outer diameter of the conductor and the outer diameter of the insulation layer;

[0156] A fitting module 202 is used to perform digital filtering and curve fitting on the cable cross-sectional profile data set to obtain the center coordinates of the conductor outer circle and the insulation layer outer circle;

[0157] The calculation module 203 is used to calculate the distance between the outer circle of the conductor and the center of the outer circle of the insulating layer based on the coordinates of the center of the circle, obtain the eccentricity value at each angle position, and calculate the thickness change data of the insulating layer at different angles;

[0158] The extraction module 204 is used to perform statistical feature calculation on the eccentricity value to obtain the eccentricity distribution feature and the maximum eccentricity direction, and to extract the variation feature of the thickness variation data to obtain the thickness variation feature;

[0159] The generation module 205 is used to input the eccentricity distribution characteristics, the maximum eccentricity direction and the thickness change characteristics into a preset plurality of autoencoders and an ant colony decision fusion model to perform mold position compensation analysis and generate mold position adjustment parameters.

[0160] Through the collaboration of the above components, multi-angle scanning and digital filtering technology, this method can obtain more accurate cable cross-sectional profile data and greatly improve the accuracy of eccentricity measurement. The use of efficient data processing algorithms and autoencoder models realizes fast feature extraction and analysis, so that eccentricity detection can keep up with the real-time speed of the production line. Combined with multiple autoencoders and ant colony decision fusion models, this method can adapt to the structural characteristics of different types of cables. Through in-depth analysis of eccentricity distribution characteristics and thickness change characteristics, it can effectively capture the dynamic change characteristics of the cable extrusion process. The ant colony algorithm is used for mold position compensation analysis, and adjustment parameters are automatically generated to reduce human intervention and improve production efficiency. The RANSAC algorithm and wavelet transform are used to effectively reduce the errors in the measurement and calculation process. Through repeated measurements and statistical feature calculations, the reliability and stability of the detection results are enhanced, which is conducive to the realization of automatic control of the entire process.

[0161] above Figure 2 The automatic calibration device for the eccentricity of the cable extrusion layer in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The automatic calibration device for the eccentricity of the cable extrusion layer in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0162] Figure 3It is a structural schematic diagram of an automatic eccentricity inspection device for a cable extrusion layer provided by an embodiment of the present invention. The automatic eccentricity inspection device 300 for the cable extrusion layer may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage device terminals) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 can be short-term storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the automatic eccentricity inspection device 300 for the cable extrusion layer. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the automatic eccentricity inspection device 300 for the cable extrusion layer to implement the steps of the automatic eccentricity inspection method for the cable extrusion layer.

[0163] The automatic eccentricity checking device 300 for the cable extrusion layer may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. It will be appreciated by those skilled in the art that Figure 3 The structure of the automatic eccentricity calibration device for the cable extrusion layer shown does not constitute a limitation on the automatic eccentricity calibration device for the cable extrusion layer provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0164] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the method for automatically checking the eccentricity of the cable extrusion layer.

[0165] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0166] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0167] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatically checking the eccentricity of a cable extrusion layer, characterized in that: The method comprises: Scan the target cable at multiple angles to obtain a cable cross-sectional profile data set including the outer diameter of the conductor and the outer diameter of the insulation layer; Performing digital filtering and curve fitting on the cable cross-sectional profile data set to obtain the center coordinates of the conductor outer circle and the insulation layer outer circle; Based on the center coordinates, the distance between the outer circle of the conductor and the center of the outer circle of the insulating layer is calculated to obtain the eccentricity value at each angular position, and the thickness variation data of the insulating layer at different angles is calculated; Performing statistical characteristic calculation on the eccentricity value to obtain eccentricity distribution characteristics and maximum eccentricity direction, and performing variation feature extraction on the thickness variation data to obtain thickness variation characteristics; The eccentricity distribution feature, the maximum eccentricity direction and the thickness variation feature are input into a preset plurality of autoencoders and an ant colony decision fusion model to perform mold position compensation analysis to generate mold position adjustment parameters; specifically comprising: performing dimensionality reduction processing on the eccentricity distribution feature to obtain a reduced eccentricity feature vector; performing normalization processing on the reduced eccentricity feature vector to obtain a normalized eccentricity feature vector, and inputting the normalized eccentricity feature vector into the first encoding layer of the first autoencoder, performing nonlinear transformation through 128 neurons and a ReLU activation function to obtain a first hidden layer representation; performing batch normalization operation on the first hidden layer representation to obtain a normalized first hidden layer representation, and The normalized first hidden layer representation is input into the second encoding layer of the first autoencoder, and feature compression is performed through 64 neurons and ReLU activation function to obtain the second hidden layer representation; the second hidden layer representation is subjected to dropout operation to obtain a sparse second hidden layer representation, and the sparse second hidden layer representation is input into the third encoding layer of the first autoencoder, and feature encoding is performed through 32 neurons and tanh activation function to obtain the initial eccentricity potential representation; the initial eccentricity potential representation is input into the first decoding layer of the first autoencoder, and preliminary decoding is performed through 32 neurons and ReLU activation function to obtain the first decoding layer output; the first decoding layer output is subjected to Layer 2 multiplication. Normalization operation is performed to obtain a normalized output of the first decoding layer, and the normalized output of the first decoding layer is input into the second decoding layer of the first autoencoder, and feature reconstruction is performed through 64 neurons and ReLU activation function to obtain the output of the second decoding layer; residual connection is performed on the output of the second decoding layer, and the output of the second decoding layer is added to the representation of the second hidden layer to obtain a residual enhanced output of the second decoding layer, and the residual enhanced output of the second decoding layer is input into the third decoding layer of the first autoencoder, and reconstructed through 128 neurons and Sigmoid activation function to obtain a potential representation of the target eccentricity; principal component analysis is performed on the thickness variation feature to obtain a principal component feature vector; the principal component feature vector is standardized to obtain a standardized principal component feature vector; the standardized principal component feature vector is input into the first encoding layer of the second autoencoder, and nonlinear transformation is performed through 256 neurons and LeakyReLU activation function to obtain the output of the first encoding layer; the output of the first encoding layer is instanced Normalization operation is performed to obtain a normalized output of the first coding layer, and the normalized output of the first coding layer is input into the second coding layer of the second autoencoder, and feature compression is performed through 128 neurons and a LeakyReLU activation function to obtain a second coding layer output;The output of the second coding layer is subjected to a Spatial Dropout operation to obtain a sparse second coding layer output, and the sparse second coding layer output is input into the third coding layer of the second autoencoder, and compressed through 64 neurons and the LeakyReLU activation function to obtain a third coding layer output; the output of the third coding layer is concatenated with the noise vector to obtain a noise-enhanced feature vector, and the noise-enhanced feature vector is input into the fourth coding layer, and encoded through 32 neurons and the tanh activation function to obtain an initial thickness change potential representation; the initial thickness change potential representation is input into the first decoding layer of the second autoencoder, and preliminary decoding is performed through 32 neurons and the ReLU activation function to obtain a first decoding output; the first decoding output is conditionally normalized to obtain a conditionally normalized first decoding output, and the conditionally normalized first decoding output is input into the second decoding layer of the second autoencoder, and feature reconstruction is performed through 64 neurons and the ReLU activation function to obtain a second decoding output; the second decoding output is subjected to a Self-Attention mechanism analysis to obtain an attention-weighted second decoding output, and the attention-weighted second decoding output is converted into an attention-weighted second decoding output. The output is input into the third decoding layer of the second autoencoder, and reconstructed through 128 neurons and ReLU activation function to obtain a third decoding output; the third decoding output is jump-connected with the second encoding output to obtain a multi-scale fusion feature, and the multi-scale fusion feature is input into the fourth decoding layer of the second autoencoder, and reconstructed through 256 neurons and Sigmoid activation function to obtain a potential representation of target thickness change; the maximum eccentricity direction, the potential representation of target eccentricity and the potential representation of target thickness change are feature fused to obtain a fusion feature vector; adaptive fuzzy c-means clustering is performed on the fusion feature vector to obtain multiple feature clusters, and the initial pheromone matrix of the ant colony decision layer in the multiple autoencoders and ant colony decision fusion model is constructed according to the feature clusters, and at the same time, the number of ants, the number of iterations and the heuristic factor are set; the path selection is performed for each ant based on the roulette selection method and the state transition probability calculation to obtain the complete path of each ant, and the fitness value of each ant path is calculated, and the global optimal path and the local optimal path are updated, and the pheromone matrix is ​​updated at the same time, and the global optimal path is decoded into a mold position adjustment parameter. ; 2. The automatic eccentricity calibration method of the cable extrusion layer according to claim 1 is characterized in that: The target cable is scanned at multiple angles to obtain a cable cross-sectional profile data set including the outer diameter of the conductor and the outer diameter of the insulation layer, including: The target cable is fixedly installed to obtain a cable sample to be tested, and the rotation angle of the cable sample to be tested is calibrated to obtain an initial angle position; Rotating the cable sample to be tested at equal intervals according to the initial angle position to obtain N sampling angle positions, and performing multi-point laser scanning on each sampling angle position to obtain coordinate data of M scanning points; Repeatedly measure each sampling angle position multiple times according to the coordinate data to obtain an original measurement data set, and perform a mean operation on the original measurement data set to obtain an average coordinate value of each sampling angle position; Constructing a cable cross-section contour point cloud according to the average coordinate value to obtain an initial contour model, and performing edge detection and contour segmentation on the initial contour model to obtain a conductor outer contour and an insulation layer outer contour; The radial distance is calculated according to the outer contour of the conductor and the outer contour of the insulation layer to obtain radial dimension data at each sampling angle position, and the radial dimension data is interpolated to obtain a cable cross-sectional profile data set including the outer diameter of the conductor and the outer diameter of the insulation layer.

3. The automatic eccentricity calibration method of the cable extrusion layer according to claim 1 is characterized in that: The step of performing digital filtering and curve fitting on the cable cross-sectional profile data set to obtain the center coordinates of the conductor outer circle and the insulation layer outer circle includes: Performing wavelet transformation on the cable cross-sectional profile data set to obtain multi-scale decomposition coefficients, and performing noise threshold calculation on the multi-scale decomposition coefficients to obtain a denoising threshold; Performing soft threshold processing on the multi-scale decomposition coefficients according to the denoising threshold to obtain denoised wavelet coefficients, and performing inverse wavelet transform on the denoised wavelet coefficients to obtain filtered contour data; Performing edge detection on the filtered contour data to obtain a conductor outer contour point set and an insulation layer outer contour point set; Performing circle fitting on the conductor outer contour point set by a first RANSAC algorithm to obtain a conductor outer circle equation, and performing circle fitting on the insulation layer outer contour point set by a second RANSAC algorithm to obtain an insulation layer outer circle equation; Extracting the center coordinates and radius according to the conductor outer circle equation to obtain the conductor outer circle parameters, and extracting the center coordinates and radius according to the insulation layer outer circle equation to obtain the insulation layer outer circle parameters; Uncertainty analysis is performed on the parameters of the outer circle of the conductor and the outer circle of the insulating layer to obtain the center coordinates of the outer circle of the conductor and the outer circle of the insulating layer.

4. The automatic eccentricity checking method of the cable extrusion layer according to claim 1 is characterized in that: The distance between the outer circle of the conductor and the center of the outer circle of the insulating layer is calculated based on the center coordinates, the eccentricity value at each angle position is obtained, and the thickness change data of the insulating layer at different angles is calculated, including: Performing coordinate system conversion on the center coordinates of the outer circle of the conductor and the outer circle of the insulating layer to obtain the center position represented by polar coordinates, and calculating the Euclidean distance between the two center positions according to the center position represented by polar coordinates to obtain an initial eccentricity value; Normalizing the initial eccentricity value to obtain a relative eccentricity value, and dividing the cable cross section into N evenly distributed angular positions according to the relative eccentricity value; Calculating the radial distance between the outer circle of the conductor and the outer circle of the insulating layer for each angular position to obtain an absolute eccentricity value at each angular position, and fitting an eccentricity distribution curve according to the absolute eccentricity value to obtain a continuous eccentricity function; Performing a differential operation on the eccentricity function to obtain an eccentricity change rate curve, and identifying an eccentricity mutation point according to the eccentricity change rate curve to obtain a target eccentricity angle position; The thickness of the insulation layer at the target eccentric angle position is calculated to obtain insulation layer thickness distribution data, and a thickness variation curve is fitted according to the insulation layer thickness distribution data to obtain thickness variation data of the insulation layer at different angles.

5. The automatic eccentricity checking method of the cable extrusion layer according to claim 1 is characterized in that: The statistical characteristic calculation of the eccentricity value is performed to obtain the eccentricity distribution characteristic and the maximum eccentricity direction, and the variation characteristic extraction is performed on the thickness variation data to obtain the thickness variation characteristic, including: Performing Fourier transformation on the eccentricity value to obtain a frequency domain representation of the eccentricity, and extracting a target frequency component of the eccentricity according to the frequency domain representation to obtain a harmonic feature of the eccentricity; Performing kernel density estimation on the eccentricity value to obtain a probability density function of the eccentricity, and calculating the mean, variance, skewness and kurtosis of the eccentricity according to the probability density function to obtain the statistical moment characteristics of the eccentricity; Performing spatial autocorrelation analysis on the eccentricity values ​​to obtain spatial distribution characteristics of the eccentricity, and determining the maximum value of the eccentricity and its corresponding angle according to the spatial distribution characteristics to obtain the maximum eccentricity direction; Performing wavelet packet decomposition on the thickness variation data to obtain multi-scale wavelet coefficients, and calculating energy entropy and corresponding dimensions according to the multi-scale wavelet coefficients to obtain complexity characteristics of the thickness variation; Performing empirical mode decomposition on the thickness variation data to obtain a plurality of intrinsic mode functions, and extracting instantaneous frequency and amplitude according to the intrinsic mode functions to obtain time-frequency characteristics of the thickness variation; The harmonic feature, the statistical moment feature and the spatial distribution feature are combined to obtain an eccentricity distribution feature, and the complexity feature and the time-frequency feature are combined to obtain a thickness variation feature.

6. The automatic eccentricity checking method of the cable extrusion layer according to claim 1 is characterized in that: The roulette wheel selection method and state transition probability calculation are used to select a path for each ant, obtain a complete path for each ant, calculate the fitness value of each ant path, update the global optimal path and the local optimal path, and update the pheromone matrix at the same time, and decode the global optimal path into a mold position adjustment parameter, including: Normalizing the initial pheromone matrix to obtain a normalized pheromone matrix, and calculating a heuristic information matrix according to the Euclidean distance between feature clusters; Calculating a state transition probability matrix for each ant according to the normalized pheromone matrix and the heuristic information matrix to obtain a transition probability distribution of each ant; A roulette wheel selection method is performed on the transition probability distribution to select the next visited node for each ant, and a single-step path selection result of each ant is obtained, until all nodes are visited and a complete path of each ant is obtained; Performing local search optimization on the complete path to obtain an optimized ant path, and calculating the path length and eccentricity change according to the optimized ant path, and substituting the path length and eccentricity change into a preset fitness function to obtain the fitness value of each ant path; Sort the fitness values ​​of all ants, select the path with the highest fitness as the optimal path of the current iteration, obtain the local optimal path, and compare the local optimal path with the historical global optimal path to obtain the updated global optimal path; According to the optimized ant paths and the fitness values, the pheromone increment of each path is calculated, and the pheromone matrix is ​​decayed using an evaporation coefficient to obtain an updated pheromone matrix; The updated global optimal path is subjected to curve interpolation to obtain a smoothed optimal path, and the smoothed optimal path is projected into a three-dimensional space for discrete sampling and coordinate transformation to obtain mold position adjustment parameters.

7. An automatic eccentricity checking device for a cable extrusion layer, characterized in that: A method for automatically checking the eccentricity of a cable extrusion layer according to any one of claims 1 to 6, the device comprising: A scanning module is used to scan the target cable at multiple angles to obtain a cable cross-sectional profile data set including the outer diameter of the conductor and the outer diameter of the insulation layer; A fitting module, used for performing digital filtering and curve fitting on the cable cross-sectional profile data set to obtain the center coordinates of the conductor outer circle and the insulation layer outer circle; A calculation module, for calculating the distance between the outer circle of the conductor and the center of the outer circle of the insulating layer based on the center coordinates, obtaining the eccentricity value at each angular position, and calculating the thickness variation data of the insulating layer at different angles; An extraction module is used to perform statistical feature calculation on the eccentricity value to obtain the eccentricity distribution feature and the maximum eccentricity direction, and to perform variation feature extraction on the thickness variation data to obtain the thickness variation feature; A generation module is used to input the eccentricity distribution characteristics, the maximum eccentricity direction and the thickness change characteristics into a preset plurality of autoencoders and an ant colony decision fusion model to perform mold position compensation analysis and generate mold position adjustment parameters.

Images