An Adaptive Production Method and System for Data Cables Based on Image Analysis

Abstract

This invention discloses an adaptive production method and system for data cables based on image analysis. The method specifically includes: simultaneously acquiring three-modal image data (visible light, infrared, and polarization) in the gas injection, extrusion, and forming sections of the data cable using a multispectral imaging unit; spatially aligning the three-modal image data using a sub-pixel registration algorithm to obtain standard image data; based on the standard image data, jointly diagnosing the cable's gas injection structure, insulation quality, and surface defects using a multi-task analysis engine to obtain defect diagnosis results; and based on the defect diagnosis results, dynamically adjusting the traction speed, extrusion temperature, and gas injection pressure using a reinforcement learning-optimized fuzzy PID controller to form online closed-loop control of process parameters. This invention effectively overcomes the shortcomings of traditional data cable production and inspection methods, such as limited functionality, static detection, and strong data dependence, providing a more efficient and intelligent solution for data cable production.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and in particular to an adaptive production method and system for data cables based on image analysis. Background Technology

[0002] In the manufacturing of data cables, ensuring product quality and production efficiency are core elements for the sustainable development of enterprises. As a crucial carrier of information transmission, the quality of data cables directly affects the stability and reliability of data transmission, thus impacting the normal operation of various electronic devices. However, traditional data cable production and testing methods have gradually revealed numerous shortcomings in addressing increasingly complex and diverse production demands. These shortcomings severely restrict the level of automation in the production process and the improvement of product quality.

[0003] First, traditional testing methods have significant functional limitations. These methods often only test a specific aspect of the data cable, such as focusing solely on the integrity of the cable structure or checking only for surface defects like scratches and cracks. This single-function testing approach makes it difficult to comprehensively and accurately assess the overall quality of the cable during the production process. The inability to promptly detect and correct other potential quality issues, such as uneven insulation thickness or internal air bubbles, significantly increases the risk of product defects. Once these problems surface during subsequent use, they not only damage the company's reputation but may also lead to customer complaints and returns, resulting in substantial economic losses.

[0004] Secondly, many traditional testing methods employ static testing, which involves offline testing after cable production is complete. This method has significant drawbacks. On one hand, offline testing requires additional testing equipment and space, increasing production costs and space usage. On the other hand, since testing is conducted after production, if quality problems are discovered, rework or scrapping of the entire batch is often necessary, leading to low production efficiency and significant resource waste. More importantly, static testing cannot reflect quality fluctuations during the production process in real time, making it difficult for companies to adjust production parameters and processes promptly, thus failing to effectively prevent quality problems from occurring.

[0005] Finally, with the rapid development of intelligent manufacturing and the Industrial Internet, some data analysis-based inspection methods are gradually being applied to the data cable manufacturing industry. These methods train models by collecting and analyzing large amounts of historical data to achieve intelligent inspection of cable quality. However, in actual production, especially when production lines are being upgraded or new products are being introduced, companies often face the problem of insufficient sample data. Due to the limited production data for new specifications or products, it is difficult to train models with sufficient generalization ability, leading to a decrease in inspection accuracy and making it difficult to meet the needs of high-precision inspection. In addition, even with sufficient historical data, data differences between different production lines and different equipment can lead to difficulties in model transfer, further limiting the application scope of data analysis-based inspection methods.

[0006] In summary, traditional data cable production and testing methods have many shortcomings in terms of functionality, efficiency, and data dependence, making it difficult to meet the demands of modern production for high quality, high efficiency, and high flexibility. Summary of the Invention

[0007] The purpose of this invention is to provide an adaptive production method and system for data cables based on image analysis, which effectively overcomes the shortcomings of traditional data cable production and testing methods, such as single function, static testing, and strong data dependence, and provides a more efficient and intelligent solution for data cable production, thereby solving at least one of the aforementioned problems in the prior art.

[0008] In a first aspect, the present invention provides an adaptive production method for data cables based on image analysis, the method specifically comprising:

[0009] The multispectral imaging unit simultaneously acquires three-modal image data, including visible light, infrared and polarization, in the gas injection section, extrusion section and forming section of the data cable, and uses a sub-pixel registration algorithm to spatially align the three-modal image data to obtain standard image data.

[0010] Based on standard image data, a multi-task analysis engine is used to jointly diagnose cable air injection structure, insulation quality and surface defects to obtain defect diagnosis results.

[0011] Based on the defect diagnosis results, a fuzzy PID controller optimized by reinforcement learning is used to dynamically adjust the traction speed, extrusion temperature and gas injection pressure, forming an online closed-loop control of process parameters.

[0012] During production line changeovers, historical process knowledge is transferred through cross-production line federated learning and model distillation technology to perform rapid specification switching under limited sample conditions.

[0013] Secondly, the present invention provides an adaptive data cable production system based on image analysis, the system specifically comprising:

[0014] The first production module is used to simultaneously acquire three-modal image data containing visible light, infrared and polarization in the gas injection section, extrusion section and molding section of the data cable through a multispectral imaging unit, and to spatially align the three-modal image data using a sub-pixel registration algorithm to obtain standard image data.

[0015] The second production module is used to perform joint diagnosis of cable air injection structure, insulation quality and surface defects based on standard image data and through a multi-task analysis engine to obtain defect diagnosis results.

[0016] The third production module is used to dynamically adjust the traction speed, extrusion temperature and gas injection pressure based on the defect diagnosis results and using a fuzzy PID controller optimized by reinforcement learning, forming an online closed-loop control of process parameters.

[0017] The fourth production module is used to transfer historical process knowledge through cross-production line federated learning and model distillation technology during production line changeovers, enabling rapid specification switching under limited sample conditions.

[0018] Thirdly, the present invention provides a computer device, comprising: a memory and a processor, and a computer program stored in the memory, wherein when the computer program is executed on the processor, it implements the image analysis-based adaptive production method for data cables as described in any of the above methods.

[0019] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the image analysis-based adaptive production method for data cables as described in any of the above methods.

[0020] Compared with the prior art, the present invention has at least one of the following technical effects:

[0021] 1. This invention effectively overcomes the shortcomings of traditional data cable production and testing methods, such as limited functionality, static testing, and strong data dependence, and provides a more efficient and intelligent solution for data cable production.

[0022] 2. This invention uses a multispectral imaging unit to simultaneously acquire three-modal image data and a multi-task analysis engine to perform joint diagnosis, which can comprehensively and accurately assess the quality of cables and promptly identify and address potential problems.

[0023] 3. This invention utilizes a fuzzy PID controller optimized through reinforcement learning to dynamically adjust the traction speed, extrusion temperature, and injection pressure based on defect diagnosis results, thereby achieving real-time optimization of production parameters and improving production efficiency and product quality.

[0024] 4. This invention utilizes cross-production line federated learning and model distillation techniques to transfer historical process knowledge and combines it with the FE-GAN generator to expand the few-sample defect data, thereby achieving high-precision defect detection under few-sample conditions and reducing the dependence on new sample data.

[0025] 5. This invention acquires three-modal image data simultaneously through a multispectral imaging unit and uses a sub-pixel registration algorithm for spatial alignment, thereby realizing multi-dimensional and high-precision image data acquisition of the gas injection section, extrusion section, and molding section of the data cable, providing a reliable data foundation for subsequent joint diagnosis and process parameter adjustment.

[0026] 6. This invention achieves synchronous acquisition of three-modal image data and adaptive exposure time adjustment through the collaborative work of a rotary encoder, FPGA and three-modal camera. Combined with sub-pixel registration algorithm, it ensures the spatial alignment accuracy of image data and improves the accuracy of subsequent image analysis.

[0027] 7. Based on standard image data, this invention uses a multi-task analysis engine to jointly diagnose cable air injection structure, insulation quality, and surface defects, achieving comprehensive and real-time monitoring of data cable quality and improving the efficiency and accuracy of defect diagnosis.

[0028] 8. This invention detects the air core distribution by fusing grayscale-gradient-texture entropy feature vectors and corrects it by combining the stress distribution of polarization images, thus achieving accurate detection of the air core distribution and providing strong support for the evaluation of cable air injection structures.

[0029] 9. This invention performs spatiotemporal wavelet decomposition on infrared image sequences and establishes a thickness-thermal conduction regression model by combining laser thickness measurement data, thereby realizing accurate measurement of insulation layer thickness and analysis of thermal conduction characteristics, providing a scientific basis for the evaluation of insulation layer quality.

[0030] 10. This invention deploys an attention network with infrared temperature correction on visible light images and uses an FE-GAN generator to expand the few-sample defect data, achieving high-precision detection of surface defects, especially maintaining high detection performance under few-sample conditions.

[0031] 11. Based on defect diagnosis results, this invention utilizes a reinforcement learning-optimized fuzzy PID controller to dynamically adjust traction speed, extrusion temperature, and injection pressure, forming an online closed-loop control of process parameters. This enables adaptive adjustment of the production process, improving production efficiency and product quality.

[0032] 12. When changing production lines, this invention transfers historical process knowledge through cross-production line federated learning and model distillation technology, enabling rapid specification switching under limited sample conditions. This reduces the cost and time of deploying new specification production lines and improves the flexibility and efficiency of the production line. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a flowchart illustrating an embodiment of the adaptive production method for data cables based on image analysis provided by the present invention.

[0035] Figure 2 This is a schematic diagram of the structure of an image analysis-based adaptive data cable production system provided in an embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0037] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0038] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0039] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0040] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0041] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0042] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0043] In the embodiments of this application, Figure 1 A flowchart illustrating an embodiment of the adaptive data cable production method based on image analysis disclosed in this invention is shown below:

[0044] S101 uses a multispectral imaging unit to simultaneously acquire three-modal image data, including visible light, infrared, and polarization, in the gas injection section, extrusion section, and forming section of the data cable. It then uses a sub-pixel registration algorithm to spatially align the three-modal image data to obtain standard image data.

[0045] In this embodiment, a three-modal imaging system comprising a visible light camera, an infrared camera, and a polarization camera is configured to ensure that each camera has high resolution, high sensitivity, and synchronous acquisition capability. The multispectral imaging unit is installed above the gas injection section, extrusion section, and forming section of the data cable production line to ensure coverage of the entire cable production process and clear, unobstructed imaging.

[0046] A rotary encoder is configured to acquire cable movement position signals in real time, ensuring that the acquisition by the multispectral imaging unit is synchronized with cable production. An FPGA (Field Programmable Gate Array) is configured as a synchronization trigger controller. When the cable enters the detection area, the FPGA generates a synchronization trigger pulse signal to control the three-modal camera to adaptively adjust the exposure time and acquire data synchronously.

[0047] Configure a high-performance computer or server to receive, process, and store the three-modal image data acquired by the multispectral imaging unit. Install image processing software and a subpixel registration algorithm library for subsequent spatial alignment processing.

[0048] The multispectral imaging unit and synchronous triggering and control system are activated and initialized, including camera parameter adjustments and synchronous trigger pulse signal generation. The cable production line is ensured to operate normally, with the cable sequentially passing through the gas injection section, extrusion section, and forming section. When the cable enters the detection area, the rotary encoder acquires the cable's movement position signal in real time and transmits it to the FPGA. The FPGA generates a synchronous trigger pulse signal based on the cable's movement position signal, controlling the three-modal camera to adaptively adjust the exposure time, ensuring that visible light, infrared, and polarization images are acquired at the same time point. The three-modal camera transmits the acquired image data to the data processing and storage device.

[0049] Preprocessing of the acquired visible light, infrared, and polarization images, including noise reduction and contrast enhancement, is performed to improve the accuracy of subsequent registration. Feature points, such as SURF (Accelerated Robust Features) points, are extracted from the visible light, infrared, and polarization images respectively. The extracted feature points are matched to establish the correspondence between feature points in the visible light, infrared, and polarization images. Taylor expansion correction is applied to the matched feature points to obtain sub-pixel-level position corrections, improving registration accuracy. Based on the feature point correspondence and sub-pixel-level position corrections, coordinate mapping relationships are calculated using an affine transformation model. Normalized mutual information and other methods are used to verify the registration results to ensure accuracy. The registered visible light, infrared, and polarization images are then fused to generate spatially aligned standard image data.

[0050] S102, based on standard image data, uses a multi-task analysis engine to jointly diagnose cable air injection structure, insulation quality, and surface defects to obtain defect diagnosis results.

[0051] In this embodiment, a feature extractor based on a convolutional neural network (CNN) is used to extract features at multiple scales from standard image data. Specifically, a network structure containing multiple convolutional and pooling layers is constructed, using convolutional kernels of different sizes to capture local and global features in the image. For example, 3×3, 5×5, and 7×7 convolutional kernels are used to extract feature maps at different scales, and the multi-scale features are fused through a concatenation operation to form a more discriminative feature representation.

[0052] The data cable quality inspection task is decomposed into three sub-tasks: cable air injection structure inspection, insulation layer quality inspection, and surface defect inspection. Each sub-task corresponds to an independent inspection branch, but they share the previous feature extraction part to improve computational efficiency and model generalization ability.

[0053] In the cable air-filled structure inspection branch, extracted feature maps are used to measure air-filled structure features such as air core distribution characteristics (e.g., air core diameter, air core spacing) and airtightness-related structural parameters (e.g., air core wall thickness uniformity, air core-insulation layer bonding tightness). Specialized convolution kernels and pooling layers are designed to enhance and extract the cable air-filled structure features. Simultaneously, a classifier is used to determine whether the cable air-filled structure conforms to specifications, outputting the cable air-filled structure inspection results.

[0054] The insulation layer quality inspection branch focuses on detecting defects such as the uniformity of insulation layer thickness and internal bubbles. It determines whether there are quality problems in the insulation layer by analyzing the grayscale distribution and texture features in the feature map. For example, it uses the Gray-Level Co-occurrence Matrix (GLCM) to extract the texture features of the insulation layer and then uses thresholding or a classifier to determine whether the insulation layer is up to standard.

[0055] The surface defect detection branch is responsible for detecting defects such as scratches and cracks on the surface of data cables. Specialized defect detection algorithms, such as defect recognition methods based on edge detection and morphological processing, are designed to further analyze the feature maps. Specifically, the Canny edge detection algorithm is used to extract edge information from the image, and morphological opening and closing operations are used to remove noise and fine edges, ultimately yielding the surface defect detection results.

[0056] The features extracted from the three detection branches are fused to form a comprehensive feature representation. The fusion method can be a simple concatenation operation or a more complex attention mechanism that dynamically adjusts the feature weights according to the importance of each task.

[0057] Based on the fused features, a joint decision-making mechanism is employed to evaluate the overall quality of the data cable. Specifically, a multi-task classifier is designed to comprehensively judge the detection results of the cable's air injection structure, insulation quality, and surface defects. The classifier can be a multi-output neural network, with each output node corresponding to the result of a detection task. By training this classifier, it can accurately determine whether the overall quality of the data cable is up to standard based on the input features.

[0058] Finally, the results of the joint diagnostics will be output in the form of structured data, including whether the cable air injection structure meets specifications, whether there are quality problems in the insulation layer, whether there are surface defects, and an overall quality assessment of the data cable. These results will be used for subsequent process parameter adjustments and production process optimization.

[0059] In this embodiment, simultaneous detection of the air injection structure, insulation quality, and surface defects of data cables can be achieved, effectively improving detection efficiency and accuracy. Furthermore, because this method employs multispectral imaging technology and a multi-task analysis engine, it can capture various quality characteristics during the data cable manufacturing process, providing strong support for optimizing the production process.

[0060] S103, based on the defect diagnosis results, uses a reinforcement learning-optimized fuzzy PID controller to dynamically adjust the traction speed, extrusion temperature and injection pressure, forming an online closed-loop control of process parameters.

[0061] In this embodiment, by combining defect diagnosis results with a fuzzy PID controller optimized by reinforcement learning, dynamic adjustment of process parameters is achieved, effectively improving the automation level of the production process and product quality.

[0062] Specifically, the defect diagnosis results output by the multi-task analysis engine are analyzed to identify specific defect types, such as abnormal cable air injection structure, uneven insulation thickness, and surface scratches. Each defect type corresponds to different quality problems and possible causes. For each identified defect type, its severity is further assessed. This can be done by setting thresholds or using fuzzy logic to classify defect severity into different levels, such as minor, moderate, and severe. The severity assessment results will serve as an important basis for subsequent process parameter adjustments.

[0063] A fuzzy PID controller is constructed, which includes three control elements: proportional (P), integral (I), and derivative (D). Fuzzy logic is used to dynamically adjust the PID parameters. The fuzzy PID controller can automatically adjust the PID parameters according to changes in the input signal to adapt to different production conditions.

[0064] To further improve the performance of the fuzzy PID controller, a reinforcement learning algorithm is introduced for optimization. The reinforcement learning algorithm learns the optimal control strategy through continuous trial and error by interacting with the production environment. Specifically, the process parameter adjustment process is modeled as a Markov decision process (MDP), where the state is the current defect diagnosis result and production parameters, the action is the adjustment amount of the process parameters, and the reward function is designed based on the adjusted product quality and production efficiency. The reinforcement learning algorithm learns the optimal process parameter adjustment strategy by maximizing the cumulative reward.

[0065] Based on the defect diagnosis results, especially those related to the cable air injection structure, the traction speed is dynamically adjusted. For example, when an abnormality in the cable air injection structure is detected, the traction speed can be appropriately reduced to decrease stress during the cable air injection process and improve the quality of the cable air injection structure. The adjustment strategy is implemented through a fuzzy PID controller, which automatically calculates and outputs the adjustment amount of the traction speed based on the severity of the defect and the current traction speed.

[0066] Extrusion temperature has a significant impact on the thickness and uniformity of the insulation layer. When defect diagnosis results indicate problems such as uneven insulation layer thickness or internal air bubbles, the extrusion temperature is dynamically adjusted using a fuzzy PID controller. For example, appropriately increasing the extrusion temperature can improve the flowability of the insulation material and help reduce the formation of internal air bubbles. The adjustment strategy is also calculated based on the severity of the defects and the current extrusion temperature.

[0067] Injection pressure is a key factor affecting the stability of the cable's air-injection structure. When abnormalities such as looseness or excessive tightness are detected in the cable's air-injection structure, the injection pressure is dynamically adjusted using a fuzzy PID controller. The adjustment strategy takes into account the characteristics of the cable's air-injection structure material, the cable's air-injection speed, and the current injection pressure value to ensure the stability and consistency of the cable's air-injection structure.

[0068] During production, sensors collect real-time data on process parameters such as traction speed, extrusion temperature, and injection pressure, as well as image data acquired by a multispectral imaging unit. This data serves as input to a fuzzy PID controller, which adjusts the process parameters in real time.

[0069] The fuzzy PID controller calculates the optimal process parameter adjustment based on the input defect diagnosis results and current process parameters, and adjusts the traction speed, extrusion temperature, and injection pressure in real time through actuators (such as frequency converters, heaters, etc.). The adjusted process parameters are then fed back into the production process, forming a closed-loop control.

[0070] Reinforcement learning algorithms continuously learn and optimize the parameters of the fuzzy PID controller during closed-loop control to adapt to different production conditions and defect types. Through continuous trial and error and adjustments, optimal control of process parameters is ultimately achieved, improving the automation level of the production process and product quality.

[0071] In this embodiment, key process parameters such as traction speed, extrusion temperature, and injection pressure can be dynamically adjusted in real time based on defect diagnosis results, effectively improving the automation level of the production process and product quality. Meanwhile, the reinforcement learning-optimized fuzzy PID controller exhibits stronger adaptability and robustness, enabling it to handle different production conditions and defect types.

[0072] S104, during production line changeover, transfers historical process knowledge through cross-production line federated learning and model distillation technology to perform rapid specification switching under limited sample conditions.

[0073] In this embodiment, production line changeovers and new product introductions are common production needs in the data cable manufacturing process. However, traditional methods often require the re-collection of large amounts of sample data to train new detection or control models when facing production line changeovers. This is not only time-consuming and labor-intensive but may also lead to production interruptions. Furthermore, data differences between different production lines and equipment can make model migration difficult, limiting the flexibility and efficiency of the production line. The cross-production line federated learning and model distillation technology proposed in this embodiment can effectively solve the above problems and achieve rapid specification switching under conditions of limited samples.

[0074] Specifically, federated learning nodes are deployed on each production line, with each node responsible for collecting and processing production data for its respective line. This data includes tri-modal image data acquired by the multispectral imaging unit, process parameter records, defect diagnosis results, etc. A secure and reliable communication protocol is established to ensure data exchange and model updates between the federated learning nodes. The communication protocol should employ encryption technology to protect the security and privacy of data transmission. A base model is initialized on each federated learning node; this model can be a pre-trained convolutional neural network (CNN) or other models suitable for image analysis and quality inspection. The structure and parameters of the base model remain consistent across all nodes.

[0075] Periodically (e.g., daily or after each batch of production), local model updates from each federated learning node are aggregated on the central server. The aggregation process can employ weighted averaging or other federated learning algorithms to synthesize the production experience and knowledge from various production lines. The aggregated global model is then distributed back to each federated learning node as the starting point for the next round of local training. In this way, different production lines can share the production experience and knowledge of others, accelerating model convergence and optimization. During federated learning, feature representations of historical data are extracted. These features may include characteristics such as cable air injection structure, insulation quality, and surface defects. The feature extraction process should be consistent to facilitate transfer and comparison between different production lines.

[0076] When changing production lines, a well-trained and high-performing global model is selected as the teacher model. The teacher model contains rich historical process knowledge and production experience. For the new production line or new product, a student model is initialized. The student model's structure can be similar to the teacher model, but its parameters need to be reinitialized. Knowledge distillation is performed on the student model using the teacher model. Specifically, the teacher model's predictions on historical data (such as classification probabilities, regression values, etc.) are used as soft labels to guide the training of the student model. Simultaneously, fine-tuning can be performed using a small amount of new sample data (few-sample condition) to adapt to new production requirements. The student model is evaluated to check its performance on the new sample data. If the performance does not meet requirements, the knowledge distillation and fine-tuning process can be iterated until the student model reaches a satisfactory performance level.

[0077] Based on new production requirements, corresponding specifications are configured, such as core pitch, insulation layer thickness, and surface quality requirements. These parameters will serve as targets for model prediction and control. The trained student model is deployed on the production line, and data from the multispectral imaging unit is collected in real time for defect diagnosis and process parameter prediction. Based on the prediction results, a reinforcement learning-optimized fuzzy PID controller is used to dynamically adjust process parameters such as traction speed, extrusion temperature, and injection pressure to achieve rapid specification switching. After specification switching, quality fluctuations and process parameter changes during production are continuously monitored. Based on the monitoring results, model parameters or control strategies are adjusted promptly to ensure the stability of the production process and the consistency of product quality.

[0078] In this embodiment, historical process knowledge can be effectively transferred during production line changeovers, enabling rapid specification switching under limited sample conditions. This not only shortens production downtime and improves the flexibility and efficiency of the production line, but also reduces the cost of recollecting sample data and training models.

[0079] In some embodiments, in step S101 above, the simultaneous acquisition of three-modal image data including visible light, infrared, and polarization by the multispectral imaging unit at the gas injection section, extrusion section, and forming section of the data cable, and the spatial alignment of the three-modal image data using a sub-pixel registration algorithm to obtain standard image data, specifically includes:

[0080] The cable movement position signal is acquired in real time by a rotary encoder. When the cable enters the detection area, the FPGA generates a synchronous trigger pulse signal to control the three-mode camera to adaptively adjust the exposure time and simultaneously acquire visible light image, infrared image and polarized image.

[0081] The first SURF feature point set, the second SURF feature point set, and the third SURF feature point set were extracted from the visible light image, the infrared image, and the polarization image, respectively.

[0082] Based on the Hessian matrix response value, key feature regions are selected from the first SURF feature point set;

[0083] Taylor expansion correction is performed on the second and third SURF feature point sets to obtain sub-pixel level position correction.

[0084] Based on key feature regions and sub-pixel level position corrections, coordinate mapping relationships are calculated using an affine transformation model, and the registration results are verified using normalized mutual information. The spatially aligned standard image data is then output.

[0085] In this embodiment, quality inspection of data cables is crucial during the manufacturing process. Traditional inspection methods often only acquire image data in a single modality, making it difficult to comprehensively reflect the internal structure and surface quality of the cable. Multispectral imaging technology can simultaneously acquire image data in multiple modalities, including visible light, infrared, and polarization, providing richer information for cable quality inspection. However, spatial discrepancies exist between image data from different modalities, requiring precise spatial alignment to ensure the accuracy of subsequent analysis. The method proposed in this embodiment effectively solves this problem, achieving simultaneous acquisition and spatial alignment of three-modal image data.

[0086] Specifically, rotary encoders are installed on the data cable production line to acquire real-time position signals of the cables. Connected to the cable's drive mechanism, the rotary encoder accurately measures the cable's movement distance and speed. When the cable enters a pre-defined detection area, the rotary encoder transmits a position signal to a field-programmable gate array (FPGA). Based on the received position signal, the FPGA generates a synchronous trigger pulse signal. This pulse signal features precise timing control, ensuring that the three-modal cameras expose simultaneously.

[0087] The tri-modal camera system comprises a visible light camera, an infrared camera, and a polarization camera. Synchronous trigger pulse signals generated by the FPGA simultaneously control the exposure time of these three cameras. The camera exposure time is adaptively adjusted based on the imaging characteristics of different modal images and the cable movement speed to obtain clear and accurate image data. For example, for infrared images, due to the weaker infrared radiation, a longer exposure time may be required; while for visible light images, the exposure time can be relatively shorter.

[0088] Under the control of a synchronous trigger pulse signal, the three-modal camera simultaneously exposes the gas injection section, extrusion section, and forming section of the data cable, acquiring visible light images, infrared images, and polarization images respectively. The acquired image data is then transmitted to a subsequent image processing system for further processing.

[0089] The acquired visible light images are processed to extract the first SURF (Speeded Up Robust Features) feature point set. SURF feature points are local feature descriptors with scale invariance and rotation invariance, which can effectively represent key information in the image. In the extraction process, the scale space of the image is first constructed, then feature points are detected at different scales, and the principal orientation and feature descriptor of each feature point are calculated.

[0090] Similarly, the infrared image is processed to extract the second SURF feature point set. Because infrared images differ from visible light images in imaging principles and characteristics, appropriate parameter adjustments are needed when extracting SURF feature points to ensure accurate and stable feature point extraction.

[0091] SURF feature point extraction is performed on the polarization image to obtain a third SURF feature point set. The polarization image contains polarization information of the cable surface, which can reflect the surface texture and material properties of the cable. When extracting feature points, the special characteristics of the polarization image need to be fully considered to improve the quality and reliability of the feature points.

[0092] For the first SURF feature point set (visible light image feature points), the Hessian matrix response value of each feature point is calculated. The Hessian matrix response value reflects the local structural information of the region where the feature point is located; the larger the response value, the more prominent the feature of the region. Based on the calculated Hessian matrix response values, key feature regions are selected from the first SURF feature point set. A response value threshold is set, and the regions where feature points with response values ​​greater than this threshold are identified as key feature regions. These key feature regions play an important role in subsequent image registration, improving the accuracy and stability of the registration process.

[0093] Taylor expansion correction is applied to the second SURF feature point set (infrared image feature points). Taylor expansion can locally approximate the image near the feature points. By calculating the coefficients of the Taylor expansion, the sub-pixel-level position correction of the feature points can be obtained. The calculation of the correction takes into account the local gradient information of the image, enabling more accurate determination of the feature point positions.

[0094] Similarly, a Taylor expansion correction is performed on the third SURF feature point set (polarization image feature points) to obtain its sub-pixel-level position correction. By performing sub-pixel-level correction on the feature points of the infrared and polarization images, the matching accuracy of feature points between different modal images can be improved.

[0095] Based on the selected key feature regions and sub-pixel-level positional corrections, the coordinate mapping relationship between the three modal images is calculated using an affine transformation model. The affine transformation model can describe linear transformations such as translation, rotation, and scaling between images. By solving the affine transformation matrix, image data from different modalities can be mapped to the same coordinate system.

[0096] Normalized mutual information (NMI) is used as the verification metric for the registration results. NMI measures the statistical correlation between two images; a higher value indicates better registration. The registration results are evaluated and verified by calculating the NMI between the images before and after registration. If the NMI does not meet the preset threshold, the registration parameters can be adjusted, and the registration calculation can be performed again.

[0097] After spatial alignment and registration result verification, spatially aligned standard image data is output. These standard image data achieve precise spatial alignment, providing an accurate and reliable basis for subsequent cable quality inspection and analysis.

[0098] In this embodiment, the multispectral imaging unit can simultaneously acquire and spatially align three-modal image data of the data cable. Simultaneous acquisition ensures temporal consistency of the image data across different modalities, while spatial alignment eliminates spatial discrepancies between the images, improving the accuracy and usability of the image data. This method provides more comprehensive and accurate image information for data cable quality inspection, contributing to improved accuracy and efficiency.

[0099] In some embodiments, in step S102 above, the step of jointly diagnosing the cable air injection structure, insulation quality, and surface defects based on standard image data using a multi-task analysis engine to obtain defect diagnosis results specifically includes:

[0100] Based on standard image data, the air core distribution is detected by fusing gray-level-gradient-texture entropy feature vectors, and the air core distribution detection result is determined.

[0101] Based on standard image data, the sequence of infrared images is decomposed in the spatiotemporal domain using wavelet decomposition, and a thickness-heat conduction regression model is established by combining laser thickness measurement data.

[0102] Based on standard image data, an attention network with infrared temperature correction is deployed on visible light images, and the FE-GAN generator is used to expand the few-sample defect data to obtain feature maps.

[0103] Based on the core distribution detection results, thickness-heat conduction regression model, and feature map, the output includes defect diagnosis results containing defect type, location, and severity level.

[0104] In this embodiment, grayscale features, gradient features, and texture entropy features of the image are extracted separately. Grayscale features reflect the brightness distribution of the image; gradient features describe the changes in edges and contours in the image; and texture entropy features reflect the texture complexity of local areas of the image. The grayscale features, gradient features, and texture entropy features are fused to construct a grayscale-gradient-texture entropy fused feature vector. The fusion process can employ weighted averaging or other feature fusion methods to ensure that the fused feature vector comprehensively reflects multiple feature information of the air core distribution.

[0105] Based on the constructed fused feature vector, image segmentation algorithms are used to segment the cable core distribution area. For example, a threshold-based segmentation method can be used to determine an appropriate threshold based on the distribution of each feature value in the fused feature vector, distinguishing the core area from other main parts of the cable (such as insulation and conductors). Alternatively, a segmentation method based on region growing or edge detection can be used to further accurately determine the boundary contour of the core. Post-processing of the segmentation results is performed, such as removing noise areas (eliminating non-core mis-segmented areas caused by imaging interference, etc.) and filling small holes (filling in small unsegmented areas inside the core caused by imaging defects, etc.), to obtain accurate core distribution detection results.

[0106] Infrared image sequences were obtained from standard image data, reflecting the thermal distribution of the data cable at different times. Spatiotemporal wavelet decomposition was performed on the infrared image sequences. In the time domain, the image sequences were decomposed chronologically to extract features at different time scales; in the spatial domain, wavelet transform was applied to each frame to decompose the image into sub-bands of different frequencies. Through spatiotemporal wavelet decomposition, both temporal and spatial feature information of the image sequences can be obtained simultaneously, providing a more comprehensive description of the thermal conductivity characteristics of the insulation layer.

[0107] In the data cable manufacturing process, a laser thickness gauge is used to measure the thickness of the insulation layer in real time. The laser thickness gauge features high precision and high resolution, enabling accurate acquisition of insulation layer thickness information. The laser thickness measurement data is then correlated with the wavelet decomposition results of the infrared image sequence. By using time synchronization, the laser thickness measurement data and infrared image data at the same moment can be correlated, establishing a correspondence between the two.

[0108] Based on the correlated laser thickness measurement data and the wavelet decomposition results of infrared images, a thickness-heat conduction regression model is established. The regression model can employ linear regression, support vector regression, or other suitable regression algorithms. During model training, the wavelet decomposition results of the infrared images are used as input features, and the laser thickness measurement data is used as the output target. Through training with a large amount of sample data, the model learns the relationship between the insulation layer thickness and its heat conduction characteristics.

[0109] An attention network is deployed on a visible light image. This network automatically learns important regions in the image and assigns more attention weights to them, thereby improving the detection capability of surface defects. An infrared temperature correction mechanism is introduced. Since infrared images can reflect the temperature distribution of the cable surface, and temperature changes may be related to surface defects, the temperature information in the infrared image is used as a correction factor to adjust the output of the attention network. For example, the attention weights can be adjusted according to the temperature, making the network pay more attention to areas with abnormal temperatures, which may contain surface defects.

[0110] In actual production, there may be situations where sample data for certain defect types is scarce. To address this issue, an FE-GAN (Feature Enhancement Generative Adversarial Network) generator is used to expand the limited sample defect data. The FE-GAN generator learns from and analyzes existing, limited defect samples to extract defect feature information. Then, the generator produces new samples with similar features to real defect samples. During the generation of new samples, features can be enhanced and modified to increase sample diversity.

[0111] An infrared temperature-corrected attention network is used to process a visible light image to obtain a preliminary feature map. This feature map contains important feature information related to surface defects in the image. The attention network is then further trained and optimized using expanded few-sample defect data, enabling it to better identify various types of surface defects. The optimized network is then processed again to obtain a more accurate and richer feature map.

[0112] Based on the air core distribution segmentation results, analyze the air core structural parameters of the cable, such as the diameter and uniformity of the air core distribution. If the air core diameter exceeds the normal range (too large or too small) or the air core distribution is uneven, it may affect the cable's inflation effect, electrical performance (e.g., changes in local electric field distribution due to air core abnormalities), and mechanical strength (e.g., uneven air core distribution causing uneven local stress on the cable). This will be identified as a defect type, and the location information of the defect will be recorded.

[0113] A thickness-thermal conduction regression model is used to evaluate the thickness and thermal conductivity characteristics of the insulation layer. Based on the model's predictions, the presence of defects such as uneven thickness or abnormal thermal conduction in the insulation layer is determined. For example, if the thickness of a certain region of the insulation layer is significantly lower than normal, or if the thermal conduction rate is abnormal, then that region is considered to have a defect, and the severity level of the defect is determined.

[0114] Based on the obtained feature maps, surface defects in visible light images are identified. By analyzing the feature distribution and intensity in the feature maps, the type of surface defect, such as scratches, cracks, and bubbles, is determined, and the location of the defect is accurately marked. Simultaneously, the severity level of the defect is assessed based on the prominence of its features and its potential impact on cable quality.

[0115] The system integrates the results of cable air injection structure analysis, insulation quality assessment, and surface defect identification to output defect diagnosis results that include defect type, location, and severity level. The diagnosis results can be presented in the form of reports or a visual interface, allowing production personnel to promptly understand the cable's quality status and take appropriate measures.

[0116] In this embodiment, a joint diagnosis of the air-filled structure, insulation quality, and surface defects of data cables can be achieved. Accurate air core distribution segmentation clearly reflects the structural condition of the cable's air-filled structure; the thickness-thermal conduction regression model accurately assesses the insulation quality; and the combination of an attention network with infrared temperature correction and an FE-GAN generator improves the accuracy of surface defect detection and its adaptability to defects in small sample sizes. The comprehensive diagnostic results provide production personnel with comprehensive and accurate cable quality information, helping to promptly identify and resolve quality issues in the production process, thereby improving the production quality and efficiency of data cables.

[0117] Furthermore, the air core distribution is detected based on standard image data using a grayscale-gradient-texture entropy fusion feature vector. The air core distribution detection results specifically include:

[0118] Based on standard image data, the gray-level mean of visible light images is calculated through a sliding window, and the gradient magnitude is calculated using an improved Scharr operator. The texture entropy is then calculated using the gray-level co-occurrence matrix.

[0119] The grayscale mean, gradient magnitude, and texture entropy are fused to form a three-dimensional feature vector;

[0120] The three-dimensional feature vectors are input into a radial basis kernel support vector machine for classification to obtain the initial detection results of the air core distribution;

[0121] Based on the initial detection results, the stress distribution is calculated using the Stokes parameters extracted from the polarization image. The initial detection results are then corrected using the stress distribution to obtain the core distribution detection results.

[0122] In this embodiment, a visible light image is obtained from standard image data. To extract the grayscale features of the image, a sliding window approach is used to process the visible light image. The sliding window moves across the image with a certain step size, covering various regions of the image. For each pixel within the sliding window, its mean grayscale value is calculated. The mean grayscale value reflects the average brightness level of the pixels in that region, and can reflect the difference in brightness between the gas core distribution and the main part of the gas injection section. For example, the mean grayscale value of the gas core distribution may be significantly different from that of the main part of the gas injection section due to light reflection and other reasons.

[0123] An improved Scharr operator is used to calculate the gradient magnitude of visible light images. The Scharr operator is effective for edge detection, and the improved version better adapts to the characteristics of data cable images, improving edge detection accuracy. The improved Scharr operator performs convolution operations in both the horizontal and vertical directions of the image, obtaining gradient components in both directions. Then, the gradient magnitude of each pixel is calculated based on these two gradient components. The gradient magnitude reflects the degree of change in pixel values ​​in the image; pixel values ​​at the core distribution area typically change more drastically, resulting in larger gradient magnitudes.

[0124] Texture entropy of visible light images is calculated using the gray-level co-occurrence matrix (GLCM). The GLCM describes the spatial dependencies between gray levels in an image and is constructed by statistically analyzing the frequency of different gray level pairs in the image. When constructing the GLCM, parameters such as distance and orientation between pixel pairs need to be considered. Appropriate parameter combinations are selected based on the specific application scenario and data characteristics. Then, texture entropy is calculated based on the GLCM. Texture entropy reflects the textural complexity of local image regions. The texture features of the gas core distribution and the main part of the gas injection section may differ; texture entropy helps distinguish between these two regions.

[0125] The extracted grayscale mean, gradient magnitude, and texture entropy are fused. Specifically, the grayscale mean, gradient magnitude, and texture entropy corresponding to each pixel are combined to form a three-dimensional feature vector. This three-dimensional feature vector integrates the grayscale, gradient, and texture information of the image, and can more comprehensively describe the characteristics of the pixel.

[0126] The resulting 3D feature vectors are input into a Radial Basis Function (RBF-SVM) for classification. SVM is a commonly used classification algorithm; the radial basis function maps input features to a high-dimensional space, making it easier to find an optimal classification hyperplane in that space. Before classification, training samples are prepared. These samples include the 3D feature vectors and class labels corresponding to pixels in the core distribution region and the main injection section region. The RBF-SVM is trained using a large number of training samples, allowing it to learn the classification boundaries of the core distribution and the main injection section in the 3D feature space. The 3D feature vectors of all pixels in the image to be segmented are input into the trained RBF-SVM, and the initial detection results for the core distribution are obtained based on the classification results. The initial detection results classify the pixels in the image into two categories: the core distribution and the main injection section.

[0127] Polarization images are obtained from standard image data. These images contain polarization information about the cable surface. By calculating the Stokes parameters of the polarization image, the stress distribution on the cable surface can be further analyzed. Stokes parameters are a set of parameters describing the polarization state of light. By performing polarization analysis on the polarization image, the Stokes parameters corresponding to each pixel can be obtained. Based on the relationship between the Stokes parameters and stress, the stress distribution on the cable surface is calculated. Generally, areas of stress concentration may cause changes in the optical properties of the cable surface, which are reflected in the Stokes parameters of the polarization image.

[0128] The relationship between stress distribution and the initial detection results of air core distribution is analyzed. In some cases, factors such as stress concentration or material inhomogeneity on the cable surface may lead to errors in the initial detection results. For example, stress concentration areas may alter the grayscale, gradient, or texture features of the image, resulting in misclassification as air core distribution or the main part of the air injection section. The initial detection results are corrected based on stress distribution information. For instance, if a region is initially identified as an air core distribution, but stress distribution analysis suggests that the region may be affected by stress, leading to abnormal features, the segmentation result for that region can be adjusted, reclassifying it as the main part of the air injection section, or vice versa. In this way, more accurate air core distribution detection results are obtained.

[0129] In this embodiment, the air core distribution in data cable images can be effectively segmented. The fusion of grayscale, gradient, and texture entropy into a feature vector improves the comprehensiveness and accuracy of feature description, while the radial basis function (RBF) support vector machine classification provides fast and accurate initial detection results. Utilizing stress distribution information from polarization images to correct the initial detection results further enhances the accuracy of segmentation and reduces segmentation errors caused by environmental interference and image feature complexity.

[0130] Furthermore, based on standard image data, the sequence of infrared images undergoes spatiotemporal wavelet decomposition, and a thickness-heat conduction regression model is established by combining laser thickness measurement data. Specifically, this includes:

[0131] Based on standard image data, a four-level three-dimensional wavelet decomposition was performed on the sequence of infrared images to extract the time-domain average temperature change rate and spatial temperature gradient.

[0132] The real-time data of cable thickness, time-domain average temperature change rate, and spatial temperature gradient obtained by the laser thickness gauge are merged to form a training sample set;

[0133] Based on the training sample set, the regression coefficients are solved by the least squares method to establish a thickness-heat conduction mapping model.

[0134] In this embodiment, infrared image sequences are obtained from standard image data. These image sequences are continuously acquired using an infrared thermal imager during the data cable manufacturing process, reflecting the thermal distribution of the cable at different times. The acquired infrared image sequences undergo preprocessing operations, including denoising to eliminate random noise and interference signals in the images; and image registration to ensure that images acquired at different times are spatially aligned, facilitating subsequent feature extraction and analysis.

[0135] A four-layer three-dimensional wavelet decomposition method was used to process the preprocessed infrared image sequence. Three-dimensional wavelet decomposition considers information in both the time and spatial domains of the image sequence, decomposing it into sub-bands of different frequencies and scales. In the time domain, wavelet decomposition decomposes the image sequence chronologically, extracting temperature change information at different time scales. In the spatial domain, wavelet transform is applied to each frame, decomposing the image into approximate and detail sub-bands. The approximate sub-band reflects the general outline information of the image, while the detail sub-band contains detailed information such as edges and textures. After four layers of decomposition, rich time-frequency feature information can be obtained.

[0136] For each pixel, its temperature change over different time periods is calculated based on the time-domain sub-band information obtained from wavelet decomposition. Specifically, the average rate of temperature change of each pixel within a certain time window is statistically analyzed; this rate of change reflects the trend of cable surface temperature change over time. For example, if the temperature in a certain area rises rapidly in a short period of time, it may indicate abnormal heat conduction in that area.

[0137] Using the spatial domain sub-band information obtained from wavelet decomposition, the temperature gradient of the region surrounding each pixel is calculated. The temperature gradient describes the degree and direction of temperature variation in space and can reflect the non-uniformity of temperature distribution on the cable surface. By calculating the spatial temperature gradient, regions with drastic temperature changes can be identified; these regions may be related to variations in insulation thickness or defects.

[0138] A laser thickness gauge is installed on the data cable production line to acquire real-time data on the thickness of the cable insulation layer. The laser thickness gauge features high precision and high resolution, enabling accurate measurement of thickness changes during the production process. It is crucial to ensure that the laser thickness gauge's measurement position corresponds to the acquisition area of ​​the infrared thermal imager to correlate the thickness data with temperature information from the infrared image sequence.

[0139] The real-time cable thickness data acquired by the laser thickness gauge, the extracted time-domain average temperature change rate, and the spatial temperature gradient are combined. For each time point or each image frame, the corresponding thickness value, time-domain average temperature change rate, and spatial temperature gradient are combined to form a data sample.

[0140] Collect a large number of data samples from different time points or locations to construct a training sample set. The training sample set should cover various normal and abnormal situations in the cable production process to ensure that the established regression model has good generalization ability. Preprocess the training sample set, such as data normalization, to transform data of different dimensions to the same numerical range, avoiding the impact of differences in data dimensions on the regression model.

[0141] This embodiment selects a linear regression model based on the least squares method to establish the thickness-heat conduction mapping relationship. The linear regression model is simple, easy to understand and implement, and suitable for describing the linear relationship between two or more variables. It is assumed that there is a linear relationship between thickness T and the time-domain average temperature change rate R and the spatial temperature gradient G, i.e., T = aR + bG + c, where a, b, and c are regression coefficients. Based on the training sample set, the regression coefficients a, b, and c are solved using the least squares method. The goal of the least squares method is to minimize the sum of squared errors between the predicted and actual values ​​of all samples in the training sample set. During the solution process, statistical analysis is performed on the training sample set to calculate the estimated values ​​of the regression coefficients. For example, the optimal solution for the regression coefficients can be obtained by calculating statistics such as the mean and covariance of the sample data.

[0142] The established thickness-heat conduction mapping model was validated using an independent validation sample set. The time-domain average temperature change rate and spatial temperature gradient from the validation sample set were input into the model to obtain the predicted thickness value, which was then compared with the actual thickness value to calculate the prediction error. If the prediction error is large, it indicates that the model's accuracy needs improvement. In this case, the model can be optimized, such as by increasing the number of training samples, adjusting the model's structure or parameters, or using a more complex regression algorithm. Through multiple iterations and optimizations, the model's prediction accuracy was improved until it met practical requirements.

[0143] In this embodiment, a regression model between the insulation thickness and thermal conductivity characteristics of data cables can be effectively established. Four-layer three-dimensional wavelet decomposition can fully extract the temporal and spatial features from the infrared image sequence, and the temporal average temperature change rate and spatial temperature gradient can accurately reflect the thermal conductivity of the cable. Combined with laser thickness measurement data, the established thickness-thermal conductivity regression model can accurately predict the insulation thickness, providing a reliable basis for cable quality inspection. Furthermore, this method has high flexibility and scalability, and can be adjusted and optimized according to actual production needs.

[0144] Furthermore, the process of deploying an attention network with infrared temperature correction on visible light images based on standard image data, and using an FE-GAN generator to expand the few-sample defect data to obtain feature maps, specifically includes:

[0145] Based on standard image data, a temperature difference map is calculated through the infrared temperature field and linearly fused with visible light image features to generate a spatial attention weight map.

[0146] In the FE-GAN generator, the style vector and defect encoding are concatenated and used as input to generate defect samples with diverse textures;

[0147] Based on defect samples, the features of visible light images are dynamically enhanced according to the spatial attention weight map to obtain infrared temperature-corrected feature maps.

[0148] In this embodiment, during the data cable manufacturing process, an infrared thermal imager is used to collect infrared temperature field data of the cable surface. The infrared thermal imager can acquire temperature information at various points on the cable surface in real time, forming an infrared temperature image. Based on the collected infrared temperature field data, a temperature difference map is calculated. Specifically, a reference temperature value is selected; this reference temperature value can be the average temperature under normal cable operation or a threshold temperature set empirically. Then, the difference between the temperature of each pixel and the reference temperature value is calculated to obtain the temperature difference map. The temperature difference map can highlight areas of abnormal temperature on the cable surface, which may be related to defects.

[0149] Visible light images, reflecting the surface appearance of cables, are obtained from standard image data. Feature extraction methods such as Convolutional Neural Networks (CNNs) are used to process these images and extract their features. CNNs can automatically learn local and global features in the image, such as edges, textures, and shapes, which are crucial for describing the appearance of the cable surface.

[0150] The calculated temperature difference map is linearly fused with the extracted visible light image features. This linear fusion can employ a weighted average method, assigning different weights to the temperature difference map and visible light image features based on their importance. For example, if temperature difference information is more critical for defect detection, the weight of the temperature difference map can be appropriately increased.

[0151] The fused features are processed by an attention mechanism module to generate a spatial attention weight map. This module assigns a weight value to each location within the fused features based on the importance of information at different locations. A larger weight value indicates that the feature at that location is more important for defect detection. The spatial attention weight map highlights feature regions in the visible light image that are related to temperature anomalies, providing guidance for subsequent feature enhancement.

[0152] Style information is extracted from a large number of normal cable images to construct style vectors. Style vectors can reflect the overall characteristics of the cable surface, such as texture and color distribution. For example, methods such as Principal Component Analysis (PCA) can be used to reduce the dimensionality of the features in normal cable images, obtaining the main components representing the image style; these main components form the style vector.

[0153] The existing small number of defect samples are analyzed to extract the feature information of the defects and encode it into a defect code. The defect code can include features such as the shape, size, and location of the defect. For example, image segmentation techniques can be used to separate the defect region from the visible light image, and then the features of the defect region can be quantified and encoded.

[0154] In the FE-GAN generator, the acquired style vector and defect encoding are concatenated and used as input. The concatenated input information contains both the style features of normal cables and the feature information of defects. The FE-GAN generator generates defect samples with diverse textures based on the input information. By learning the style features of normal cables and the feature information of defects, the generator can generate defect images similar to real defect samples but with different textures. For example, the generator can change the texture details of the defect region while keeping the basic shape and size of the defect unchanged, thereby generating diverse defect samples and expanding the number of defect samples.

[0155] The generated defect samples are associated with the corresponding visible light images. Since the defect samples are generated based on the defect features and style information of the visible light images, the defect samples can be matched with the corresponding regions in the visible light images using information such as the location and size of the defects.

[0156] Based on the generated spatial attention weight map, the features of the visible light image are dynamically enhanced. Regions with larger weight values ​​in the spatial attention weight map represent feature regions related to defects or temperature anomalies; for these regions, their contribution to the final feature map can be increased. Specifically, the features of the visible light image are multiplied by the spatial attention weight map, strengthening the features of regions with larger weight values ​​and suppressing the features of regions with smaller weight values. Simultaneously, the enhanced visible light image features are further refined by incorporating feature information from generated defect samples. For example, the features of defect samples can be fused with the enhanced visible light image features to obtain an infrared temperature-corrected feature map. This feature map includes both the appearance features of the visible light image and the infrared temperature information and expanded defect sample features, enabling a more accurate description of defects on the cable surface.

[0157] In this embodiment, the accuracy and robustness of surface defect detection in data cables can be effectively improved. The spatial attention weight map generated by combining infrared temperature information and visible light image features highlights feature regions related to defects. The FE-GAN generator is used to expand the defect samples, solving the problem of insufficient model training data in cases with few samples. Dynamic enhancement of visible light image features based on the spatial attention weight map yields an infrared temperature-corrected feature map, which more comprehensively reflects the defect information on the cable surface, helping to improve the recall and accuracy of defect detection.

[0158] In some embodiments, step S103 above, which involves dynamically adjusting the traction speed, extrusion temperature, and injection pressure using a reinforcement learning-optimized fuzzy PID controller based on the defect diagnosis results to form an online closed-loop control of process parameters, specifically includes:

[0159] The gas core distribution variance, thickness deviation, and defect density are used as state vectors;

[0160] The optimal PID parameter adjustment is calculated using the Double DQN algorithm, and the optimal PID parameter adjustment is used as the action vector.

[0161] The reward function is to minimize the vector of defect density, energy consumption per unit length, and the magnitude of change in control quantity.

[0162] Based on state vectors, action vectors, and reward functions, the control quantities of traction speed, extrusion temperature, and injection pressure are calculated through reinforcement learning and then sent to the PLC actuator.

[0163] The reward function value is calculated based on the control effect of the PLC actuator, and the Q network parameters are updated once every N meters of cable produced.

[0164] In this embodiment, the variance of the air core distribution along a certain length of cable is calculated. Variance reflects the dispersion of the air core distribution; a large variance indicates unstable air core distribution, which may affect the air injection quality and electrical performance of the cable. Incorporating the air core distribution variance as part of the state vector helps the controller understand the current fluctuations in the air core distribution.

[0165] Online inspection equipment such as laser thickness gauges is used to measure the thickness of the data cable insulation layer in real time. The measured thickness value is compared with the preset target thickness to calculate the thickness deviation. The thickness deviation directly reflects the accuracy of the cable insulation layer thickness and is one of the important indicators for measuring cable quality. By incorporating the thickness deviation into the state vector, the controller can adjust process parameters such as extrusion temperature in a timely manner based on the magnitude and trend of the thickness deviation to ensure that the cable insulation layer thickness meets the requirements.

[0166] Based on previous defect diagnosis results, defect detection is performed on the produced cables. The number of defects on a certain length of cable is counted, and the defect density is calculated. Defect density directly reflects the production quality of the cable; a higher defect density indicates poorer cable quality. Using defect density as an important component of the state vector allows the controller to dynamically adjust parameters such as traction speed, extrusion temperature, and injection pressure according to the cable's quality, thereby reducing defect generation.

[0167] The Double DQN (Double Deep Q-Network) algorithm is a deep reinforcement learning-based algorithm that estimates the state-action value function using two neural networks (a target network and an evaluation network). It effectively addresses the overestimation problem inherent in traditional DQN algorithms, improving stability and convergence. In this embodiment, the Double DQN algorithm is chosen to calculate the optimal PID parameter adjustment. Because the data cable manufacturing process is a complex dynamic system with nonlinear and time-varying characteristics, the Double DQN algorithm can continuously learn and optimize through interaction with the environment, finding the PID parameter adjustment strategy that optimizes system performance under different states.

[0168] The state vector, composed of core distribution variance, thickness deviation, and defect density, is input into the evaluation network of the Double DQN algorithm. Based on the current state vector, the evaluation network outputs the Q-value corresponding to each possible PID parameter adjustment. The PID parameter adjustment with the largest Q-value is selected as the optimal PID parameter adjustment, and this adjustment is used to form the action vector. The optimal PID parameter adjustment includes adjustments to the proportional, integral, and derivative coefficients in the PID controller. By adjusting these parameters, the control performance of the fuzzy PID controller can be altered, making it better suited to the current production conditions.

[0169] Defect density is a key indicator of cable quality, and minimizing defect density is an important component of the reward function. When defect density decreases, the controller is positively rewarded; conversely, when defect density increases, the controller is negatively penalized. This guides the controller to adjust process parameters to reduce cable defects and improve cable quality.

[0170] Energy consumption is a significant cost factor in the production of data cables. To reduce production costs, it is necessary to minimize energy consumption per unit length of cable. By incorporating energy consumption per unit length into a reward function, the controller is rewarded when energy consumption is reduced. This incentivizes the controller to optimize process parameters to further reduce energy consumption while maintaining cable quality.

[0171] Frequent changes in process parameters (traction speed, extrusion temperature, and injection pressure) can adversely affect production equipment and cable quality. To maintain production stability, the vector of control variable variation is incorporated as part of the reward function. When the control variable variation is small, the controller is positively rewarded; when the variation is large, the controller is negatively penalized. This guides the controller to maintain stability when adjusting process parameters, avoiding large fluctuations.

[0172] Based on defined state vectors, action vectors, and reward functions, the reinforcement learning system (with the Double DQN algorithm at its core) continuously interacts with the production environment. Within each control cycle, the optimal PID parameter adjustment (action vector) is selected based on the current state vector. Then, according to the principle of the fuzzy PID controller, the control quantities for traction speed, extrusion temperature, and injection pressure are calculated. The fuzzy PID controller combines the advantages of fuzzy control and PID control, enabling real-time adjustment of PID parameters based on system error and error rate of change through fuzzy inference rules, thereby achieving precise control of process parameters. The calculated control quantities for traction speed, extrusion temperature, and injection pressure are sent to the PLC (Programmable Logic Controller) actuator via a communication interface. The PLC actuator can then precisely control the traction equipment, extrusion equipment, and injection equipment on the production line based on the received control quantities, achieving real-time adjustment of process parameters.

[0173] Based on the effect of the PLC actuator on adjusting the process parameters, data such as the variance of the gas core distribution, thickness deviation, and defect density are collected again. Then, according to the design of the reward function, the reward function value for this control cycle is calculated. For example, if the defect density decreases, energy consumption decreases, and the change in control quantity is small after adjusting the process parameters, the reward function value will be a large positive value; conversely, if the adjustment effect is poor, the reward function value will be a small value or even a negative value.

[0174] For every N meters of cable produced, the collected state vectors, action vectors, and calculated reward function values ​​are used to form training samples to update the Q-network parameters in the Double DQN algorithm. By continuously updating the Q-network parameters, the reinforcement learning system can gradually learn the optimal PID parameter adjustment strategy under different production conditions, thereby continuously improving the control accuracy of traction speed, extrusion temperature, and injection pressure, and realizing online closed-loop control of process parameters in the data cable production process.

[0175] In this embodiment, the production quality and efficiency of data cables can be significantly improved. The reinforcement learning-optimized fuzzy PID controller can dynamically adjust process parameters such as traction speed, extrusion temperature, and injection pressure based on defect diagnosis results and real-time production status, effectively reducing cable defects, lowering energy consumption per unit length, and maintaining production stability. By continuously updating the Q-network parameters, the controller can gradually adapt to various changes in the production process, improving the adaptive capability and intelligence level of the control system.

[0176] In some embodiments, step S104 above, which involves transferring historical process knowledge through cross-production line federated learning and model distillation techniques during production line changeover to perform rapid specification switching under limited sample conditions, specifically includes:

[0177] When changing production line models, calculate the cosine-Euclidean mixture similarity between the new specifications and parameters and each specification in the historical configuration library, and select the top k configurations as initial parameters.

[0178] Based on the initial parameters, the process models of multiple production lines are aggregated through a federated learning framework. Each production line updates the process models of multiple production lines using local historical data and a small number of new samples.

[0179] The knowledge distillation technique is used to compress the multi-production line process model into a lightweight model, and key process knowledge is preserved by KL divergence and feature alignment loss.

[0180] Channel pruning is performed on the lightweight model. When the average channel weight is lower than the preset channel weight threshold, redundant parameters are removed to complete the deployment of the new specification production line.

[0181] In this embodiment, production line changeover is a common operation in industrial production. For example, data cable manufacturers need to produce cables of different specifications according to market demand. However, each production line changeover faces the problem of insufficient sample data for the new specifications. Traditional model training methods require a large amount of sample data to achieve good performance, which is difficult to achieve in the early stages of a changeover. At the same time, different production lines accumulate rich historical process knowledge, but this knowledge is often scattered across various production lines, making it difficult to effectively integrate and utilize. The method proposed in this embodiment, through cross-production line federated learning and model distillation techniques, can transfer historical process knowledge, achieve rapid specification switching under conditions of limited samples, and improve production efficiency and product quality.

[0182] When changing production lines, the new specifications and parameters, along with those in the historical configuration library, are represented in a structured manner. For example, for data cable production, specifications may include cable diameter, insulation thickness, and air injection method. Each specification is represented as a multi-dimensional vector, with each dimension corresponding to a feature of the parameter.

[0183] Calculate the cosine-Euclidean mixture similarity between the new specification parameters and each specification parameter in the historical configuration library. The mixture similarity calculation can combine the advantages of both cosine similarity and Euclidean distance; for example, each can be assigned a certain weight, and then a weighted sum can be performed. Based on the mixture similarity calculation, the top k historical configurations most similar to the new specification parameters are selected as initial parameters. These initial parameters contain production experience similar to the new specification, providing a good starting point for subsequent model training.

[0184] A cross-production-line federated learning framework is built, capable of connecting the computing and data storage devices of multiple production lines. In federated learning, each production line is treated as a client, possessing its own local historical data. For example, different production lines may have produced data cables of similar specifications but different models, and these lines have accumulated their own production data and process models.

[0185] Based on the selected top k configurations as initial parameters, a multi-production-line process model is aggregated using a federated learning framework. During federated learning, each production line updates its multi-production-line process model using local historical data and a small number of new samples. Each production line trains and updates its model on its local data, then uploads the updated model parameters to a central server. The central server aggregates these parameters to obtain a more robust multi-production-line process model. This distributed learning approach fully utilizes the historical data of each production line while protecting the data privacy of each line. Each production line updates its multi-production-line process model using local historical data and a small number of new samples. The local historical data contains rich information about the production process in the past, while the small number of new samples reflects some characteristics under the new specifications. By combining these two types of data, the model can adapt well to new specifications even with limited samples.

[0186] Knowledge distillation is used to compress a multi-production-line process model into a lightweight model. Knowledge distillation is a technique that transfers knowledge from a complex model (teacher model) to a simpler model (student model). In this embodiment, the multi-production-line process model serves as the teacher model, containing rich process knowledge; the lightweight model serves as the student model, needing to learn the knowledge from the teacher model.

[0187] Key process knowledge is preserved through KL divergence and feature alignment loss. KL divergence measures the difference between two probability distributions, and here it is used to measure the similarity of the output distributions of the student model and the teacher model; feature alignment loss ensures that the student model can extract key features similar to those of the teacher model during the learning process. For example, in data cable production, key features might include the effect of extrusion temperature on the cable insulation thickness, and the effect of traction speed on the cable air injection quality.

[0188] Channel pruning is performed on the lightweight model. Channel pruning is a model compression technique that reduces the number of parameters and computational cost by removing unimportant channels. When the average channel weight is below a preset channel weight threshold, the channel is considered to contribute little to the model and can be removed, thereby reducing redundant parameters. After channel pruning, the lightweight model has the ability to be quickly deployed to new production lines. This model retains the key knowledge of multi-production line process models while having low computational and storage requirements, making it suitable for real-time production line control.

[0189] In this embodiment, historical process knowledge can be quickly utilized during production line changeovers to achieve specification switching under limited sample conditions. Cross-production line federated learning breaks down data silos, enabling the sharing of process experience across different production lines; model distillation technology compresses complex models into lightweight models, improving deployment efficiency and response speed. In practical applications, this method can significantly shorten the debugging time for production line changeovers, improve production efficiency, and ensure the quality stability of new product specifications.

[0190] Reference Figure 2 An embodiment of the present invention provides an adaptive data cable production system 2 based on image analysis, wherein the system 2 specifically includes:

[0191] The first production module 201 is used to simultaneously acquire three-modal image data containing visible light, infrared and polarization in the gas injection section, extrusion section and molding section of the data cable through a multispectral imaging unit, and to spatially align the three-modal image data using a sub-pixel registration algorithm to obtain standard image data.

[0192] The second production module 202 is used to perform joint diagnosis of cable air injection structure, insulation quality and surface defects based on standard image data and through a multi-task analysis engine to obtain defect diagnosis results.

[0193] The third production module 203 is used to dynamically adjust the traction speed, extrusion temperature and gas injection pressure based on the defect diagnosis results and using a fuzzy PID controller optimized by reinforcement learning, forming an online closed-loop control of process parameters.

[0194] The fourth production module 204 is used to transfer historical process knowledge through cross-production line federated learning and model distillation technology during production line changeover, and to perform rapid specification switching under few sample conditions.

[0195] It is understandable that, such as Figure 1 The content shown in the embodiment of the image analysis-based adaptive production method for data cables is applicable to the embodiment of this image analysis-based adaptive production system for data cables. The specific functions implemented in this embodiment of the image analysis-based adaptive production system for data cables are the same as those shown in the image analysis-based adaptive production system for data cables. Figure 1 The illustrated embodiment of the adaptive data cable production method based on image analysis is the same, and the beneficial effects achieved are the same as those shown. Figure 1 The beneficial effects achieved by the illustrated embodiment of the adaptive production method for data cables based on image analysis are also the same.

[0196] It should be noted that the information interaction and execution process between the above systems are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0197] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0198] Reference Figure 3 The present invention also provides a computer device 3, including: a memory 302 and a processor 301, and a computer program 303 stored in the memory 302. When the computer program 303 is executed on the processor 301, it implements the image analysis-based adaptive production method for data cables as described in any of the above methods.

[0199] The computer device 3 may be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device 3 may include, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will understand that... Figure 3 The computer device 3 is merely an example and does not constitute a limitation on the computer device 3. It may include more or fewer components than shown in the figure, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0200] The processor 301 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0201] In some embodiments, the memory 302 may be an internal storage unit of the computer device 3, such as a hard disk or memory of the computer device 3. In other embodiments, the memory 302 may be an external storage device of the computer device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 3. Furthermore, the memory 302 may include both internal and external storage units of the computer device 3. The memory 302 is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of the computer program. The memory 302 can also be used to temporarily store data that has been output or will be output.

[0202] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the image analysis-based adaptive production method for data cables as described in any of the above methods.

[0203] In this embodiment, if the integrated control unit is implemented as a software functional unit and sold or used as an independent product for machine tool control, it can be stored in a computer-readable storage medium specifically designed for machine tools. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by instructing the relevant hardware of the machine tool through a specific computer program. This computer program can be stored in a machine tool-specific computer-readable storage medium. When executed by the machine tool's processor, this computer program can implement the application steps of the various method embodiments in machine tool control. The computer program includes computer program code for machine tool control, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: a recording medium capable of carrying the computer program code for machine tool control to any entity or device in the machine tool equipment, a machine tool computer memory, a read-only memory (ROM), a random access memory (RAM), and other media suitable for machine tool software distribution, such as a dedicated machine tool control card, a machine tool data storage card, etc.

[0204] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0205] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0206] In the embodiments disclosed in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0207] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A data cable self-adapting production method based on image analysis, characterized in that, The method specifically comprises: Synchronously collecting three-modal image data containing visible light, infrared and polarization by a multi-spectral imaging unit at the gas injection section, extrusion section and forming section of the data cable, and using a sub-pixel registration algorithm to spatially align the three-modal image data to obtain standard image data; Based on the standard image data, jointly diagnosing the cable gas injection structure, insulation layer quality and surface defects by a multi-task analysis engine to obtain defect diagnosis results; Based on the defect diagnosis results, dynamically adjusting the pulling speed, extrusion temperature and gas injection pressure using a fuzzy PID controller optimized by reinforcement learning to form an online process parameter closed-loop control; When the production line is changed, migrating historical process knowledge through cross-line federated learning and model distillation technology to perform rapid specification switching under a few-shot condition; The method specifically comprises: Real-time acquisition of cable movement position signals by a rotary encoder, generation of a synchronous trigger pulse signal by the FPGA when the cable enters the detection area, adaptive adjustment of the exposure time of the three-modal camera, and synchronous acquisition of visible light images, infrared images and polarization images; Extracting first, second and third SURF feature point sets from the visible light images, infrared images and polarization images, respectively; Based on the Hessian matrix response value, screening key feature regions from the first SURF feature point set; Taylor expansion correction of the second and third SURF feature point sets to obtain sub-pixel level position correction amounts; Based on the key feature regions and sub-pixel level position correction amounts, calculating coordinate mapping relationships through an affine transformation model, verifying the registration results based on normalized mutual information, and outputting the spatially aligned standard image data.

2. The method of claim 1, wherein, The method specifically comprises: Based on the standard image data, detecting the air core distribution by a gray-scale-gradient-texture entropy fusion feature vector to determine an air core distribution detection result; Based on the standard image data, performing spatiotemporal wavelet decomposition on the sequence of infrared images, and establishing a thickness-thermal conductivity regression model in combination with laser thickness measurement data; Based on the standard image data, deploying an attention network with infrared temperature correction on the visible light images, and using an FE-GAN generator to expand few-shot defect data to obtain feature maps; Based on the air core distribution detection result, thickness-thermal conductivity regression model and feature maps, outputting defect diagnosis results containing defect types, locations and severity levels.

3. The method of claim 2, wherein, The method specifically comprises: Based on the standard image data, the gray mean value of the visible light image is calculated by a sliding window, the gradient amplitude is calculated by an improved Scharr operator, and the texture entropy is calculated by a gray level co-occurrence matrix; The gray mean value, gradient amplitude and texture entropy are fused to form a three-dimensional feature vector; The three-dimensional feature vector is input into a radial basis kernel support vector machine for classification to obtain an initial detection result of the air core distribution; Based on the initial detection result, the stress distribution is calculated according to the Stokes parameters extracted from the polarization image, and the initial detection result is corrected by the stress distribution to obtain the air core distribution detection result.

4. The method of claim 2, wherein, Based on the standard image data, the sequence of infrared images is subjected to spatiotemporal wavelet decomposition, and a thickness-heat conduction regression model is established in combination with the laser thickness measurement data, specifically including: Based on the standard image data, the sequence of infrared images is subjected to four-layer three-dimensional wavelet decomposition, and the temporal average temperature change rate and spatial temperature gradient are extracted; The data cable thickness data obtained by the laser thickness gauge in real time, the temporal average temperature change rate and the spatial temperature gradient are combined to form a training sample set; Based on the training sample set, the regression coefficients are solved by the least square method to establish a thickness-heat conduction mapping model.

5. The method of claim 2, wherein, Based on the standard image data, an attention network with infrared temperature correction is deployed on the visible light image, and a FE-GAN generator is used to expand the few-sample defect data to obtain a feature map, specifically including: Based on the standard image data, a temperature difference map is calculated from the infrared temperature field, and is linearly fused with the visible light image features to generate a spatial attention weight map; In the FE-GAN generator, the style vector and the defect code are spliced as input to generate a defect sample with diverse textures; Based on the defect sample, the features of the visible light image are dynamically enhanced according to the spatial attention weight map to obtain an infrared temperature corrected feature map.

6. The method according to any one of claims 1 to 5, characterized in that, Based on the defect diagnosis result, a fuzzy PID controller optimized by reinforcement learning is used to dynamically adjust the traction speed, extrusion temperature and gas injection pressure to form an online process parameter closed-loop control, specifically including: The air core distribution variance, thickness deviation and defect density are taken as the state vector; The optimal PID parameter adjustment amount is calculated by the Double DQN algorithm, and the optimal PID parameter adjustment amount is taken as the action vector; The minimum defect density, unit length energy consumption and control amount change amplitude vector are taken as the reward function; Based on the state vector, action vector and reward function, the control amount of the traction speed, extrusion temperature and gas injection pressure is calculated by reinforcement learning and is sent to the PLC actuator; Based on the control effect of the PLC actuator, the reward function value is calculated, and the Q network parameters are updated every N meters of cable produced.

7. The method according to any one of claims 1 to 5, characterized in that, When the production line is changed, historical process knowledge is transferred by cross-line federated learning and model distillation technology to perform rapid specification switching under few-sample conditions, specifically including: When the production line is changed, the cosine-Euclidean hybrid similarity of the new specification parameters and the specification parameters in the historical configuration library is calculated, and the top k configurations are selected as the initial parameters; Based on the initial parameters, the multi-line process model is aggregated through a federated learning framework, and each line updates the multi-line process model using local historical data and a small number of new samples; The knowledge distillation technology is used to compress the multi-line process model into a lightweight model, and the key process knowledge is retained through the KL divergence and feature alignment loss; The channel pruning is implemented on the lightweight model, and when the mean value of the channel weight is lower than the preset channel weight threshold, the redundant parameters are removed, and the deployment of the new specification line is completed.

8. An image analysis based data cable adaptive production system, characterized by, The system specifically comprises: The first production module is configured to acquire three-modal image data containing visible light, infrared and polarization through a multi-spectral imaging unit at the gas injection section, extrusion section and forming section of the data cable, and perform spatial alignment on the three-modal image data through a sub-pixel registration algorithm to obtain standard image data; The second production module is configured to perform joint diagnosis on the cable gas injection structure, insulation layer quality and surface defects based on the standard image data through a multi-task analysis engine to obtain a defect diagnosis result; The third production module is configured to dynamically adjust the traction speed, extrusion temperature and gas injection pressure based on the defect diagnosis result by using a fuzzy PID controller optimized by reinforcement learning to form an online process parameter closed-loop control; The fourth production module is configured to transfer historical process knowledge through cross-line federated learning and model distillation technology to perform rapid specification switching under a small sample condition when the line is changed; The multi-spectral imaging unit is configured to acquire three-modal image data containing visible light, infrared and polarization at the gas injection section, extrusion section and forming section of the data cable, and perform spatial alignment on the three-modal image data through a sub-pixel registration algorithm to obtain standard image data, specifically comprising: A rotary encoder is configured to acquire a cable movement position signal in real time, and a FPGA is configured to generate a synchronous trigger pulse signal to control the three-modal camera to perform adaptive adjustment of the exposure time and synchronously acquire visible light images, infrared images and polarization images when the cable enters the detection area; First, second and third SURF feature point sets are extracted from the visible light images, infrared images and polarization images, respectively; Based on the Hessian matrix response value, key feature regions are selected from the first SURF feature point set; The second and third SURF feature point sets are subjected to Taylor expansion correction to obtain sub-pixel level position correction amounts; Based on the key feature regions and the sub-pixel level position correction amounts, a coordinate mapping relationship is calculated through an affine transformation model, and the registration result is verified by normalized mutual information to output the spatially aligned standard image data.

9. A computer device, comprising: The system specifically comprises: A memory and a processor, and a computer program stored in the memory, when the computer program is executed on the processor, implements the image analysis-based data cable adaptive production method according to any one of claims 1 to 7.

Images