Cable surface defect detection method

Abstract

The invention relates to the technical field of cable surface defect detection, and provides a cable surface defect detection method, in terms of detection efficiency, a hybrid quantification technology and a hardware acceleration means are adopted, the model reasoning speed is greatly improved, a system can rapidly process a large amount of detection data, a unique dynamic tracking technology reduces unnecessary calculation amount, and the detection efficiency is improved. The detection process is more efficient and can adapt to the detection rhythm of high-speed production of cables; in terms of detection precision and robustness, a multi-modal feature fusion strategy fully integrates advantage information of different sensors, and fusion from pixels and feature patterns to decision levels effectively enhances the capture and analysis ability of defect features and the structural optimization and innovative training mode of a deep learning model, significantly improves the detection ability of micro defects, and improves the detection accuracy and robustness of the micro defects. And a relatively low false detection rate can still be kept in a complex environment.

Description

Technical Field

[0001] The present invention relates to the technical field of cable surface defect detection, in particular to a cable surface defect detection method. Background Art

[0002] In today's power industry, cables are the key carrier of power transmission, and their surface defect detection is of vital importance. Traditional cable surface defect detection methods have many difficult-to-overcome technical difficulties.

[0003] Manual visual inspection relies on the subjective judgment and experience of the inspector, resulting in extremely low inspection efficiency and difficulty meeting the needs of modern large-scale, high-speed cable production. Furthermore, long inspection periods can easily lead to visual fatigue, making missed inspections and false detections highly likely, seriously impacting inspection accuracy.

[0004] While existing automated inspection technologies, such as single-use infrared or ultrasonic testing, have improved detection efficiency to a certain extent, they still have significant shortcomings. For one thing, these methods are less adaptable in complex environments. For example, in environments with fluctuating lighting, oil coverage, or electromagnetic interference, detection accuracy can drop significantly, making it impossible to accurately identify and locate defects. Furthermore, their ability to detect minor defects is limited, making it difficult to detect tiny flaws such as fine cracks and pinholes that could pose serious safety risks.

[0005] Therefore, a cable surface defect detection method is proposed, which aims to achieve efficient, accurate and real-time cable surface defect detection, improve detection efficiency and accuracy, enhance the robustness of the detection system in complex environments, and effectively detect tiny defects to ensure the quality of the cable and the safety and reliability of power transmission. Summary of the Invention

[0006] Technical problems solved

[0007] In view of the deficiencies in the prior art, the present invention provides a method for detecting cable surface defects.

[0008] Technical Solution

[0009] To achieve the above-mentioned solution, the present invention provides the following technical solution: a method for detecting cable surface defects, comprising the following steps:

[0010] S1. Build a professional inspection system, select linear array CCD cameras, infrared thermal imagers, laser vibrometers, and microwave sensors, determine their layout based on the cable production line conditions, and use computing equipment to build an edge computing processing unit and configure the software architecture.

[0011] S2. Implement multi-sensor data acquisition and spatio-temporal alignment, achieve μs-level time synchronization using the IEEE 1588 Precision Clock Protocol, and perform spatial calibration using an improved Zhang Zhengyou calibration method combined with checkerboards and circular targets.

[0012] S3. Promote multi-modal feature fusion, perform pixel-level weighted fusion on RGB images and thermal imaging images, perform feature map-level fusion through global average pooling, fully connected layers, and activation functions, and perform decision-level fusion using the D-S evidence theory.

[0013] S4. Construct and train a lightweight deep learning model, improve EfficientNet-Lite3 by introducing the CSP structure, add a dedicated detection branch for micro-defects to the multi-scale YOLOv8-Nano detection head, and use a defect simulation data augmentation method based on physical rendering and an improved FocalLoss+DiceLoss as the loss function.

[0014] S5. Achieve real-time optimization of the detection process, use FP16→INT8 mixed quantization technology and FPGA hardware acceleration to improve the inference speed, and use Kalman filtering to dynamically track the ROI to reduce the computational load.

[0015] S6. Conduct online detection, collect data in real time by multiple sensors, input the processed data into the model for detection, output the results through a three-level early warning mechanism, and use them for quality traceability and model update iteration.

[0016] Preferably, when using the IEEE 1588 Precision Clock Protocol for time synchronization, the slave clock calculates the clock offset according to the timestamps of the master clock sending the synchronization message, the slave clock receiving the synchronization message, the slave clock sending the response message, and the master clock receiving the response message, according to the formula where t1 is the moment when the master clock sends the synchronization message, t2 is the moment when the slave clock receives the synchronization message, t3 is the moment when the slave clock sends the response message, and t4 is the moment when the master clock receives the response message, and then calibrates the time of data collected by each sensor.

[0017] Preferably, when performing spatial calibration using an improved Zhang Zhengyou calibration method combined with checkerboards and circular targets, the projection process of the camera involves pixel coordinates in the image plane, world coordinate system coordinates, scale factor, camera internal parameter matrix, rotation matrix, and translation vector. By performing specific matrix operations on these parameters, the spatial data collected by different sensors is unified into the same coordinate system.

[0018] Preferably, when performing pixel-level weighted fusion on RGB images and thermal imaging images, the fused pixel value is calculated by the formula I f =αI R +(1-α)I T where IR is the pixel value of the RGB image, I T is the pixel value of the thermal imaging image, and α is the weighting coefficient determined through a large number of experiments.

[0019] Preferably, when performing feature map-level fusion, first perform global average pooling on the feature map, that is, add the feature values at each position in the feature map channel and then divide by the product of the height and width of the feature map to obtain the channel descriptor. Then, calculate the channel attention weight through a fully connected layer and an activation function. The activation function includes the Sigmoid function and the ReLU function, and the fully connected layer has a corresponding weight matrix. Multiply the calculated channel attention weight by the original feature map to obtain the fused feature map.

[0020] Preferably, when using the D-S evidence theory for decision-level fusion, for the detection results of multiple sensors, the degree of trust of each sensor in the proposition is represented by the basic probability assignment function. The fused basic probability assignment function is obtained by first multiplying the basic probability assignment functions of all sensors for the relevant proposition, then summing these products that meet specific conditions, and finally dividing the summation result by a normalization constant.

[0021] Preferably, when introducing the CSP structure into the improved EfficientNet-Lite3, in the CSP module, the input feature map is divided into two parts. One part directly passes through the convolutional layer, and the other part undergoes a series of convolutional operations, and then these two parts are concatenated.

[0022] Preferably, when adding a dedicated detection branch for micro defects to the multi-scale YOLOv8-Nano detection head, this branch adopts a mixed receptive field structure. The formula for calculating the output feature map elements through convolutional operations is y = W1·x + b1 + W2·x + b2, where W1 and W2 are the weight matrices of different convolutional kernels, b1 and b2 are the corresponding bias vectors, and x is the input feature map element.

[0023] Preferably, when using physical-rendering-based defect simulation for data augmentation, through the formulas I s = I o ×(1 + ò1), S s = S o ×(1 + ò2), K s = K o ×(1 + ò3) to generate simulated defect images, where I o , S o , K o are the original light intensity, the proportion of the oil stain area, and the size of the motion blur kernel respectively, and ò1, ò2, ò3 are random perturbation coefficients.

[0024] Preferably, when using Kalman filtering to dynamically track the ROI, the state transition equation is x k = Fx k-1 + Q, and the observation equation is z k = Hx k + R, where x k is the state vector at the k-th moment, including the position and velocity information of the cable, F is the state transition matrix, Q is the process noise, z k is the observation vector at the k-th moment, H is the observation matrix, R is the observation noise, and the predicted value of the cable movement trajectory is obtained through the prediction and update steps to achieve dynamic cropping of the region of interest.

[0025] Advantages

[0026] Compared with the prior art, the present invention provides a method for detecting cable surface defects, having the following

[0027] advantages:

[0028] 1. For this cable surface defect detection method, in terms of detection efficiency, by adopting the hybrid quantization technology and hardware acceleration means, the model inference speed is greatly improved, enabling the system to quickly process a large amount of detection data. The unique dynamic tracking technology reduces unnecessary computational volume, making the detection process more efficient and capable of adapting to the detection rhythm of high-speed cable production.

[0029] 2. For this cable surface defect detection method, in terms of detection accuracy and robustness, the multi-modal feature fusion strategy fully integrates the advantageous information of different sensors. From pixel, feature map to decision-level fusion, it effectively enhances the ability to capture and analyze defect features. The structural optimization and innovative training method of the deep learning model significantly improve the detection ability for micro defects and still maintain a low false detection rate in complex environments. Description of the Drawings

[0030] Figure 1 is the schematic diagram of the step flow of the present invention;

[0031] Figure 2 is the schematic diagram of the method framework of the present invention. Detailed Embodiments

[0032] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0033] Please refer to Figures 1 to 2, the present invention provides a method for detecting cable surface defects, including the following:

[0034] S1. Build a professional detection system

[0035] 1.1 Sensor selection and layout

[0036] For the constructed cable surface defect detection system, a line array CCD camera, an infrared thermal imager, a laser vibrometer, and a microwave sensor are selected. Each sensor, relying on its unique performance advantages, realizes the all-round monitoring of the cable state.

[0037] The line array CCD camera has an ultra-high resolution of 12K and a frame rate of 2000FPS, and supports the automatic switching between visible light and near-infrared light. On the high-speed cable production line, its high resolution can clearly capture the fine texture and geometric features on the cable surface, and the high frame rate ensures that during the rapid movement of the cable, images can be collected completely and continuously, avoiding image blurring and data omission. The multi-spectral automatic switching function can obtain cable surface information from different spectral dimensions, greatly improving the recognition of various production-related defects, such as surface abnormalities caused by material inhomogeneity or process problems.

[0038] The infrared thermal imager is equipped with an imaging unit of 640×512 pixels, with a temperature resolution of 0.05°C and a response time of less than 10ms. During the cable production process, this device can sensitively capture the subtle temperature changes in the cable caused by internal structural abnormalities or production process defects. For example, local overheating may imply hidden defects such as internal bubbles and insufficient fusion, providing a key basis for timely adjustment of the production process.

[0039] The accuracy of the laser vibrometer is ±0.1μm, and the sampling frequency is 10kHz. In the production links such as cable stretching and forming, it can monitor the dynamic deformation of the cable in real time and accurately, and promptly detect abnormal cable deformation caused by mechanical stress, avoiding quality problems caused by excessive stretching or extrusion.

[0040] The microwave sensor operates in the 24GHz frequency band, with a penetration depth of 5 - 10mm and a resolution of 0.5mm. During the manufacturing process of the cable insulation layer, this sensor can deeply detect the inside of the insulation layer and effectively discover defects such as uneven insulation layer thickness, bubbles, and impurities, ensuring that the cable insulation performance meets the production standards.

[0041] When arranging the sensors, fully consider the spatial structure of the cable production line, the running speed of the cable, and the detection focus. Through precise calculation and repeated debugging, determine the optimal installation positions and angles of each sensor. Ensure that the detection ranges of each sensor can cover the entire surface of the cable, collect high-quality data without blind spots, and provide a solid data foundation for subsequent detection and analysis.

[0042] 1.2 Construction and Configuration of Edge Computing Processing Unit

[0043] NVIDIA Jetson AGX Orin is adopted as the core device for edge computing. It has a powerful computing power of 275 TOPS, which can meet the stringent requirements of complex data processing for computing power. With the FPGA acceleration module, the parallel computing advantages of the FPGA are fully utilized, significantly improving the speed and efficiency of data processing, and realizing real-time and efficient processing of the massive data collected by multiple sensors.

[0044] At the software level, a complete data processing architecture is built, including a spatio-temporal alignment module, a feature pyramid fusion module, a lightweight neural network module, a defect real-time classification and localization module, and an early warning output / parameter adjustment module. The spatio-temporal alignment module uses precise algorithms to ensure that the data collected by multiple sensors is highly consistent in the time and space dimensions, eliminating data errors caused by time and space differences; the feature pyramid fusion module deeply fuses features of different scales and types to fully mine useful information in the data; the lightweight neural network module is based on advanced deep learning algorithms to efficiently analyze the fused features and extract key features related to cable surface defects; the defect real-time classification and localization module can accurately identify the defect type and precisely determine its location; the early warning output / parameter adjustment module issues corresponding early warnings in a timely manner according to the detection results and dynamically adjusts the system parameters to adapt to different detection scenarios and working conditions changes.

[0045] S2. Implement Multi-Sensor Data Acquisition and Spatio-Temporal Alignment

[0046] 2.1 Achieve High-Precision Time Synchronization

[0047] The system uses the IEEE 1588 Precision Clock Protocol to achieve sub-microsecond high-precision time synchronization between sensors. In each synchronization cycle, the master clock sends a synchronization message to the slave clock. When the slave clock receives and sends the message, it records the corresponding timestamps respectively. The slave clock calculates the clock offset Δt from the master clock according to the following formula:

[0048]

[0049] In this formula, t1 represents the moment when the master clock sends the synchronization message, t2 is the moment when the slave clock receives the synchronization message, t3 is the moment when the slave clock sends the response message, and t4 is the moment when the master clock receives the response message. Each sensor accurately calibrates the time of the collected data according to the calculated clock offset, ensuring the consistency of multi-sensor data in the time dimension, avoiding data errors caused by time asynchronization, and providing an accurate and reliable time basis for subsequent data fusion and analysis.

[0050] 2.2 Conduct improved spatial calibration

[0051] Adopt the improved Zhang-Zhengyou calibration method and combine a checkerboard and a circular target to perform spatial calibration on the camera. The projection process of the camera is described by the following equations:

[0052]

[0053] In the formula, (u, v) represents the coordinates of pixels on the image plane, reflecting the position information of the object in the image; (X, Y, Z) are the coordinates in the world coordinate system, used to determine the position of the object in the real space; s is the scale factor, used to unify the coordinate scale and facilitate the conversion between different coordinate systems; K is the camera internal parameter matrix, containing key parameters such as the focal length and the position of the principal point of the camera, and these parameters determine the geometric characteristics of camera imaging; R is the rotation matrix, describing the rotation relationship of the camera coordinate system relative to the world coordinate system; t is the translation vector, determining the translation position of the camera coordinate system relative to the world coordinate system. Through this calibration method, the spatial data collected by different sensors are unified into the same coordinate system, and the calibration error is strictly controlled to be less than 0.02 mm, providing a reliable spatial basis for the subsequent fusion and accurate analysis of multi-modal data, and ensuring the accuracy and reliability of the detection results.

[0054] S3. Promote multi-modal feature fusion

[0055] 3.1 Implement pixel-level weighted fusion

[0056] For RGB images and thermal imaging images, the weighted average method is used for pixel-level fusion to fully integrate the advantageous information of the two images and highlight the defect features on the cable surface. The pixel value I of the fused image f is calculated by the following formula:

[0057] I f = αI R + (1 - α)I T

[0058] where, I R is the pixel value of the RGB image, which can intuitively reflect rich detail information such as the color and texture of the object and plays an important role in discovering defects caused by abnormal surface texture; I T is the pixel value of the thermal imaging image, which can display the temperature distribution of the object and helps to detect defects caused by abnormal temperature; α is the weighting coefficient, determined through a large number of experiments, aiming to reasonably balance the contributions of the two images in the fusion result during the fusion process, extract defect features to the greatest extent, and provide more comprehensive and accurate data support for subsequent feature analysis and defect judgment.

[0059] 3.2 Conduct feature map-level fusion

[0060] In the feature map level fusion stage, first, a global average pooling operation is performed on the feature map to obtain the channel descriptor z:

[0061]

[0062] Here, x(i,j) represents the feature value of the feature map channel at the position (i,j), and H and W are the height and width of the feature map respectively. Through global average pooling, the channel descriptor z can synthesize the feature information of the entire channel and reflect the global features of the channel in the image.

[0063] Subsequently, through two fully connected layers and activation functions, the channel attention weight a is calculated:

[0064] a = σ(W2δ(W1z))

[0065] Among them, σ is the Sigmoid function, which maps the output value to between 0 and 1, intuitively representing the importance degree of the channel. The closer the value is to 1, the more important the features contained in this channel are for defect detection; δ is the ReLU function, as a non-linear activation function, effectively increases the expression ability of the model, enabling the model to better learn and capture complex feature relationships; W1 and W2 are the weight matrices of the fully connected layers, and the parameters of these two matrices are obtained through training and learning, which determine the attention degree and weight allocation of the model to different features.

[0066] Finally, multiply the channel attention weight a by the original feature map to obtain the fused feature map:

[0067]

[0068] Through this feature map level fusion method based on the attention mechanism, the model can automatically focus on and highlight the features of important channels, suppress the interference of irrelevant information, effectively improve the pertinence and accuracy of feature extraction, and enhance the model's ability to identify and extract defect features in complex scenarios.

[0069] 3.3 Perform decision level fusion

[0070] The D-S evidence theory is adopted for decision level fusion to fully integrate the detection information of multiple sensors and improve the accuracy and reliability of defect detection. For the detection results of n sensors, the degree of trust of each sensor in the proposition A is represented by the basic probability assignment function m i (A). Through the Dempster combination rule, the fused basic probability assignment function m(A) is calculated:

[0071]

[0072] Among them, is a normalization constant, whose function is to avoid a zero denominator and ensure the stability and effectiveness of the calculation process. This synthesis rule can fully integrate the information of multiple sensors, effectively solve the uncertainty and limitations existing in single-sensor detection, significantly improve the accuracy and reliability of defect detection, and make the detection results more scientific and persuasive.

[0073] S4. Construct and train a lightweight deep learning model

[0074] 4.1 Optimize the network structure

[0075] Improve EfficientNet-Lite3 by introducing a Cross Stage Partial (CSP) structure, aiming to reduce the computational amount of the model while enhancing the efficiency of feature propagation. In the CSP module, the input feature map is divided into two parts. One part directly passes through the convolutional layer, and the other part undergoes a series of convolutional operations and then is concatenated with the former. Specifically, it is expressed as:

[0076] o = Concat(Conv1(x1), Conv2(x2))

[0077] where x1 and x2 are the two parts of the input feature map, and Conv1 and Conv2 represent different convolutional operations respectively. Through this structural design, not only the number of model parameters is reduced, the computational complexity is decreased, and the computational burden is alleviated, but also the features can be fully extracted and fused in different paths, effectively improving the performance of the model and enabling it to better adapt to the actual needs of cable surface defect detection.

[0078] In the multi-scale YOLOv8-Nano detection head, add a dedicated detection branch for tiny defects, and this branch adopts a 1×1 + 5×5 hybrid receptive field structure. Taking the convolutional operation to calculate the output feature map element y as an example, its calculation formula is as follows:

[0079] y = W1·x + b1 + W2·x + b2

[0080] where W1 and W2 are the weight matrices of the 1×1 and 5×5 convolutional kernels respectively, b1 and b2 are the corresponding bias vectors, and x is the input feature map element. This hybrid receptive field structure can take into account feature information of different scales, enabling the model to capture both the local detailed features of tiny defects and the context information within a larger range, thereby better detecting tiny defects and significantly improving the detection ability and accuracy for tiny defects.

[0081] 4.2 Implement innovative training strategies

[0082] In the data enhancement phase, defect simulation is performed based on physical rendering. By randomly changing parameters such as light intensity I, oil stain area ratio S, and motion blur kernel size K, simulated defect images are generated. The specific formula is as follows:

[0083] I s =I o ×(1+ò1)

[0084] S s =S o ×(1+ò2)

[0085] K s =K o ×(1+ò3)

[0086] Among them, I o 、S o , K o where ò1, ò2, and ò3 are the original light intensity, oil stain area ratio, and motion blur kernel size, respectively. This greatly expands the diversity of the training dataset, allowing the model to be exposed to a variety of complex scenarios and defect morphologies during training, thereby improving the model's generalization ability and enabling it to better adapt to the complex and variable nature of actual inspection environments.

[0087] The improved FocalLoss+DiceLoss is used as the loss function, and the calculation formula of the total loss L is:

[0088] L=αL F +(1-α)L D

[0089] Among them, α is the weighting coefficient, which is used to balance the contribution of the two loss functions in the total loss; L F FocalLoss, L D for DiceLoss.

[0090] The definition of FocalLoss is as follows:

[0091]

[0092] Where N is the number of samples, α t is a category balance factor, which is used to deal with the problem of sample category imbalance and ensure that the model pays enough attention to minority class samples during training, thereby improving the detection ability of various defects; t is the model's predicted probability for the sample, and γ is the focusing parameter. By adjusting the value of γ, the model can pay more attention to samples that are difficult to classify and enhance the model's classification ability for complex samples.

[0093] DiceLoss is defined as:

[0094]

[0095] Among them, y i is the true label, and [label] is the model prediction result. DiceLoss is used to measure the similarity between the model prediction result and the true label. By minimizing this loss function, the model's prediction result can be made closer to the real situation and the detection accuracy of the model can be improved.

[0096] S5. Realize the real-time optimization of the detection process

[0097] 5.1 Accelerate model inference

[0098] To improve the inference speed of the model and meet the requirements of real-time detection of cable surface defects, the FP16→INT8 hybrid quantization technology is adopted. The FP16 data is converted into INT8 data through linear mapping, and the conversion formula is as follows:

[0099] Q = round(S·F + Z)

[0100] Among them, Q is the quantized value, S is the scale factor used to scale the data range to ensure that the accuracy of the quantized data can meet the actual requirements; F is the original FP16 value, and Z is the zero-point offset, whose function is to ensure the accuracy of the quantization process. Through this quantization method, the inference speed can be increased by 2.3 times, effectively reducing the demand of model inference for computing resources and enabling the model to achieve fast inference with limited hardware resources.

[0101] In terms of FPGA hardware acceleration, taking two-dimensional convolution as an example, the calculation process of the output feature map element y is as follows:

[0102]

[0103] Among them, w mn is the convolution kernel weight, x i+m,j+n is the input feature map element, M and N are the convolution kernel sizes, and b is the bias. The FPGA realizes a 40% improvement in the calculation efficiency of the convolution layer through parallel computing, making full use of its hardware parallelism advantage, greatly shortening the model inference time, and meeting the strict time requirements of real-time detection.

[0104] 5.2 Dynamically track ROI

[0105] To reduce the calculation amount and improve the detection efficiency, the Kalman filter is used to predict the cable movement trajectory. The state transition equation is as follows:

[0106] x k = Fx k-1 + Q

[0107] where, x k is the state vector at the k-th moment, which contains key information such as the position and velocity of the cable; F is the state transition matrix, which describes the relationship between the state and time; Q is the process noise, reflecting the uncertainty existing in the actual motion process.

[0108] The observation equation is:

[0109] z k = Hx k + R

[0110] where, z k is the observation vector at the k-th moment, obtained by sensor measurement; H is the observation matrix, which maps the state vector to the observation space; R is the observation noise, considering the errors existing in the sensor measurement process.

[0111] Through the prediction and update steps of the Kalman filter, the predicted value of the cable motion trajectory can be obtained. Based on this, dynamic cropping of the region of interest (ROI) is realized, reducing the computational amount by 30% - 50%. This technology not only improves the detection efficiency, but also enables the model to focus more on the regions that may have defects, improving the accuracy and pertinence of detection. While ensuring the detection accuracy, it effectively reduces the consumption of computing resources.

[0112] S6. Conduct online detection

[0113] During the production process of the cable, multi-sensors collect cable surface data in real time. The collected data is first subjected to spatio-temporal alignment and preprocessing to remove noise and outliers to ensure the data quality. Then, multi-modal feature fusion is carried out to fuse the data from different sensors into a unified feature representation, fully mining the useful information in the data. Then, the fused features are input into a lightweight deep learning model for defect detection.

[0114] The detection results are output through a three-level early warning mechanism: Level I early warning marks the suspicious region, indicating the possible position of the defect; Level II identifies the defect type, clarifying the nature of the defect; Level III outputs the defect location coordinates, providing accurate position information for subsequent repair and treatment. At the same time, the detection data is used for quality traceability. New defect samples are uploaded to the cloud knowledge base through federated learning, and the model is updated and iterated monthly to continuously improve the accuracy and adaptability of the detection method to better meet the ever-developing requirements of the cable industry for surface defect detection, providing all-round and multi-level technical support for the stable operation and safety guarantee of the cable.

[0115] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting surface defects of a cable, characterized in that: It includes the following steps: S1. Build a professional detection system, select a line array CCD camera, an infrared thermal imager, a laser vibrometer, and a microwave sensor, determine its layout according to the cable production line conditions, and build an edge computing processing unit using a computing device and configure the software architecture; S2. Implement multi-sensor data acquisition and spatio-temporal alignment, use the IEEE1588 precision clock protocol to achieve μs-level time synchronization, and use an improved Zhang Zhengyou calibration method combined with a checkerboard and a circular target for spatial calibration; S3. Promote multi-modal feature fusion, perform pixel-level weighted fusion on RGB images and thermal imaging images, perform feature map-level fusion through global average pooling, fully connected layers, and activation functions, and use the D-S evidence theory for decision-level fusion; S4. Build and train a lightweight deep learning model, improve EfficientNet-Lite3 by introducing the CSP structure, add a dedicated detection branch for micro-defects to the multi-scale YOLOv8-Nano detection head, and use a defect simulation data augmentation method based on physical rendering and an improved FocalLoss+DiceLoss as the loss function; S5. Achieve real-time optimization of the detection process, use the FP16→INT8 mixed quantization technology and FPGA hardware acceleration to improve the inference speed, and use Kalman filtering to dynamically track the ROI to reduce the calculation amount; S6. Conduct online detection, multi-sensors collect data in real time, input the processed data into the model for detection, output the results through a three-level early warning mechanism, and use them for quality traceability and model update and iteration.

2. The method for detecting surface defects of a cable according to claim 1, characterized in that: When performing time synchronization using the IEEE 1588 Precision Time Protocol, the slave clock calculates the clock offset according to the timestamps of the master clock sending the synchronization message, the slave clock receiving the synchronization message, the slave clock sending the response message, and the master clock receiving the response message, according to the formula where t1 is the time when the master clock sends the synchronization message, t2 is the time when the slave clock receives the synchronization message, t3 is the time when the slave clock sends the response message, and t4 is the time when the master clock receives the response message, and then calibrates the time of the data collected by each sensor.

3. A method for detecting surface defects of a cable according to claim 1, characterized in that: When using the improved Zhang Zhengyou calibration method combined with a checkerboard and a circular target for spatial calibration, the projection process of the camera involves image plane pixel coordinates, world coordinate system coordinates, scale factor, camera internal parameter matrix, rotation matrix, and translation vector. By performing specific matrix operations on these parameters, the spatial data collected by different sensors is unified into the same coordinate system.

4. A method for detecting surface defects of a cable according to claim 1, characterized in that: When performing pixel-level weighted fusion on the RGB image and the thermal imaging image, the pixel value after fusion is calculated by the formula I f = αI R + (1 - α)I T where I R is the pixel value of the RGB image, I T is the pixel value of the thermal imaging image, and α is the weighting coefficient determined through a large number of experiments.

5. A method for detecting surface defects of a cable according to claim 1, characterized in that: When performing feature map-level fusion, first perform a global average pooling operation on the feature map, that is, add the feature values at each position in the feature map channel and then divide by the product of the height and width of the feature map to obtain a channel descriptor. Calculate the channel attention weight through a fully connected layer and an activation function. The activation function includes the Sigmoid function and the ReLU function, and the fully connected layer has a corresponding weight matrix. Multiply the calculated channel attention weight by the original feature map to obtain the fused feature map.

6. The method for detecting surface defects of a cable according to claim 1, characterized in that: When using the D-S evidence theory for decision-level fusion, for the detection results of multiple sensors, the degree of trust of each sensor in the proposition is represented by the basic probability assignment function. The fused basic probability assignment function is obtained by first multiplying the basic probability assignment functions of all sensors for the relevant proposition, then summing these products that meet specific conditions, and finally dividing the summation result by a normalization constant.

7. A method for detecting surface defects of a cable according to claim 1, characterized in that: When the improved EfficientNet-Lite3 introduces the CSP structure, in the CSP module, the input feature map is divided into two parts. One part directly passes through the convolutional layer, and the other part undergoes a series of convolutional operations, and then these two parts are concatenated.

8. A method for detecting surface defects of a cable according to claim 1, characterized in that: In the multi-scale YOLOv8-Nano detection head, a dedicated detection branch for tiny defects is added. This branch adopts a hybrid receptive field structure. The formula for calculating the elements of the output feature map through convolutional operations is y = W1·x + b1 + W2·x + b2, where W1 and W2 are the weight matrices of different convolutional kernels, b1 and b2 are the corresponding bias vectors, and x is the element of the input feature map.

9. A method for detecting surface defects of a cable according to claim 1, characterized in that: When performing data augmentation using defect simulation based on physically based rendering, through the formula I s = I o ×(1 + ò1), S s = S o ×(1 + ò2), K s = K o ×(1 + ò3) to generate a simulated defect image, where I o , S o , K o are the original light intensity, the proportion of the oil stain area, and the size of the motion blur kernel respectively, and ò1, ò2, ò3 are random perturbation coefficients.

10. A method for detecting surface defects of a cable according to claim 1, characterized in that: When using Kalman filtering to dynamically track the ROI, the state transition equation is x k = Fx k-1 + Q, and the observation equation is z k = Hx k + R, where x k is the state vector at the k-th moment, including the position and velocity information of the cable, F is the state transition matrix, Q is the process noise, z k is the observation vector at the k-th moment, H is the observation matrix, R is the observation noise, and the predicted value of the cable motion trajectory is obtained through the prediction and update steps to achieve dynamic cropping of the region of interest.

Images