Steel wire rope defect intelligent detection and early warning system based on convolutional neural network

Abstract

The application discloses a steel wire rope defect intelligent detection and early warning system based on a convolutional neural network, relates to the technical field of steel wire rope defect detection, and by introducing a hollow convolution with different expansion rates to construct a multi-scale parallel channel group, the network obtains synchronous perception ability for local details and overall structure under the premise of keeping the calculation overhead unchanged, and fundamentally improves the capture ability for defect characteristics of different scales, and is especially suitable for the identification of non-intact broken wires and staggered wear areas. In the system design, a boundary perception attention fusion mechanism is introduced, multi-scale feature responses and edge gradient intensity are dynamically combined, the model automatically allocates boundary weights in the fusion stage, and the response effect to fuzzy and non-closed edges is significantly enhanced. This mechanism effectively suppresses the false boundary activation phenomenon caused by surface contamination, oil stains and weak reflection, and improves the boundary precision and stability of the detection result.

Description

Technical Field

[0001] This invention relates to the field of wire rope defect detection technology, specifically to an intelligent detection and early warning system for wire rope defects based on convolutional neural networks. Background Technology

[0002] In scenarios such as steel and mineral transportation, hoisting, and deep well drilling, wire ropes serve as critical load-bearing components, and their condition directly impacts equipment safety and personnel protection. In this inspection process, subtle defects such as broken wires, slight flattening, and discontinuous wear are often difficult to accurately identify using traditional image processing or manual inspection methods. Therefore, in practical systems, there is an urgent need to introduce higher-resolution, more expressive feature extraction mechanisms to achieve automatic identification and early warning of such abnormal details.

[0003] Currently, convolutional neural networks are widely used for defect recognition in the field of industrial steel wire rope image inspection. However, most existing methods rely on convolutional structures with fixed receptive fields, which easily lead to decreased recognition accuracy or boundary misjudgment when faced with complex shapes such as blurred wire break boundaries and discontinuous expansion of flattened areas. Especially when image resolution is limited or the shooting angle is offset, traditional convolutional layers cannot simultaneously take into account local texture details and global contour information, resulting in problems such as blurred recognition and inaccurate boundary localization when the system encounters complex defects such as "irregular fractures," "multi-directional diagonal scratches," and "anisotropic wear fracture fusion areas." This lack of scale adaptability in convolutional structures has become a key bottleneck for further improvement of current intelligent steel wire rope image inspection systems.

[0004] This situation arises from two main factors. First, traditional network structures lack the ability to fuse multi-scale information. In particular, when faced with images containing both strong and weak boundaries, minor texture discontinuities, or blurred defect areas, convolutional layers with a single receptive field cannot simultaneously respond to abrupt boundary changes and gradual transitions. Second, most existing systems do not introduce boundary-sensitive mechanisms, resulting in insensitivity to edge information. This can easily lead to problems such as response compression and boundary loss, especially in images with complex defect morphologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent detection and early warning system for steel wire rope defects based on convolutional neural networks, which solves the problems mentioned in the background technology.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: a wire rope defect intelligent detection and early warning system based on convolutional neural network, including an image input and standardization preprocessing module, a multi-scale receptive field feature extraction module, a boundary-aware attention fusion module, a defect boundary response curvature estimation module, a boundary response enhancement index calculation module, and a defect category estimation and early warning output module. The image input and normalization preprocessing module acquires raw steel wire rope images through a camera and performs preprocessing to obtain the steel wire image tensor Istd; The multi-scale receptive field feature extraction module is based on the wire image tensor Istd, constructs parallel convolutional paths, and introduces dilated convolutional kernels with different dilation rates to construct receptive field channel groups of different scales to obtain the fused response tensor Fr(x,y). The boundary-aware attention fusion module extracts features from the acquired fusion response tensor Fr(x,y) and fuses them to obtain the boundary enhancement feature map Fus(x,y). The defect boundary response curvature estimation module extracts the high response region in the boundary enhancement feature map Fus(x,y) and calculates the boundary curvature response kb(x,y). The boundary response enhancement index calculation module calculates the boundary enhancement index Ψedge based on the boundary curvature response kb(x,y). The defect category estimation and early warning output module combines the boundary enhancement feature map Fus(x,y) with the boundary enhancement index Ψedge to judge the defects of the wire rope and provide feedback early warning.

[0007] Preferably, the image input and normalization preprocessing module includes a color structure unification and interference suppression processing unit and a contrast adaptive stretching and tensor construction processing unit; The color structure unification and interference suppression processing unit acquires the original wire rope image through an industrial camera, performs color space conversion and high-frequency interference removal, and obtains the processed wire rope image Yfi(x,y). The color space conversion process is as follows: the original steel wire rope image captured by the industrial camera is converted from the RGB color space to the luminance-color difference color space, and the structural luminance information and color perturbation information are separated. The conversion formula is as follows: Y(x,y)=0.299R(x,y)+0.587G(x,y)+0.114B(x,y); Cb(x,y)=128-0.169R(x,y)-0.331G(x,y)+0.5B(x,y); Cr(x,y)=128+0.5R(x,y)-0.419G(x,y)-0.081B(x,y); In the formula, Y(x,y) represents the luminance channel, Cb(x,y) represents the blue color difference component, Cr(x,y) represents the red color difference component, (x,y) is the image space coordinate position, and R, G and B represent the pixels of the red, green and blue channels respectively; High-frequency interference removal is based on Fourier spectrum analysis. The image is filtered for high-frequency energy to remove periodic texture artifacts such as uneven lighting and surface scratch reflections. A band-limited filter is used to remove noise from the original wire rope image. The contrast-adaptive stretching and tensor construction processing unit standardizes the processed wire rope image Yfi(x,y) by using local histogram enhancement and tensor normalization to obtain the wire rope image tensor Istd; The local histogram enhancement method involves applying a contrast-limited adaptive histogram equalization method to perform local contrast stretching on the processed wire rope image Yfi(x,y) to enhance local texture differences and suppress the masking of texture responses by overly bright / dark areas; the formula is as follows: Ycl(x,y)=CLAHE(Yfi(x,y);clip,tile); In the formula, CLAHE represents the histogram equalization function, Ycl(x,y) represents the enhanced wire image, clip represents limiting the maximum histogram height to avoid overstretching, and tile represents dividing the image into small grids for local equalization. The tensor normalization construction process is as follows: the enhanced wire image Ycl(x,y) is converted into the standard tensor format of the convolutional network, and an additional color difference signal is added to obtain the wire image tensor Istd; The wire image tensor Istd is obtained using the following formula: ; In the formula, Istd(x,y) represents the wire image tensor at the image space coordinate position (x,y), QYcl(x,y) represents the brightness gradient map, which is used to preserve micro-boundary texture, and Cb(x,y) / 255 and Cr(x,y) / 255 represent the blue color difference signal and the red color difference signal, respectively.

[0008] The brightness gradient map QYcl(x, y) is obtained using the following formula: ; In the formula, Y represents the luminance channel, and PD represents the sign of the partial derivative; Preferably, the multi-scale receptive field feature extraction module includes a multi-path convolutional feature extraction unit and a receptive field enhancement and feature rearrangement mapping unit; The multi-path convolution feature extraction unit is based on the wire image tensor Istd and constructs different dilation rate dilated convolution kernels Kr in three different parallel paths, where r∈{1,2,4}. For each path, extract the initial response graph cFr(x, y); The initial response plot cFr(x, y) is obtained using the following formula: ; In the formula, k represents the convolution kernel radius, (i, j) represents the convolution kernel position, Kr(i, j) represents the convolution kernel weight at (i, j), Istd(xr×i, yr×j) represents the pixel position participating in convolution in the input wire image tensor, and cFr(x, y) represents the initial response map of the image spatial coordinate position (x, y), specifically including three scales: F1(x, y) represents focusing on the boundary of broken wire texture and burr microstructure, F2(x, y) represents capturing shallow scratches and crack segments, and F4(x, y) represents identifying large-area flattening and fatigue deformation contours; The receptive field enhancement and feature rearrangement mapping unit stacks the three scales in the initial response map cFr(x,y) along the channel dimension to form a scale fusion tensor. The stacked tensor is then adjusted for consistency in spatial coordinate mapping to obtain the fused response tensor Fr(x,y). The fusion response tensor Fr(x, y) is obtained using the following formula: Fr(x,y)=Concat(F1(x,y),F,2(x,y),F4(x,y)); In the formula, Concat represents the tensor dimension concatenation function.

[0009] Preferably, the boundary-aware attention fusion module includes a scale boundary gradient extraction and attention factor generation unit and a boundary attention-guided fusion unit; The scale boundary gradient extraction and attention factor generation unit calculates the spatial gradient for each fused response tensor Fr(x,y) to obtain the boundary intensity map Gr(x,y). The boundary strength map Gr(x, y) is obtained using the following formula: ; In the formula, Gr(x, y) represents the boundary intensity map, specifically the boundary intensity map at the image spatial coordinate position (x, y), and PD represents the partial derivative sign; The boundary intensity maps at all scales are normalized to obtain the attention factor ηr(x,y), which is used to measure the boundary contribution intensity at the current spatial location at that scale. The attention factor ηr(x,y) is obtained by summing up the boundary intensity maps Gr(x,y) at the same spatial coordinate position (x,y) for all three scales to obtain the total boundary intensity; then, the boundary intensity map Gr(x,y) of a single scale is divided by the total boundary intensity, and then added to a non-zero constant to obtain the attention factor ηr(x,y) for that single scale.

[0010] Preferably, the boundary attention-guided fusion unit obtains the boundary enhancement feature map Fus(x,y) by weighting the fusion response tensor Fr(x,y) using the attention factor ηr(x,y) at each scale. The boundary enhancement feature map Fus(x,y) is obtained as follows: at the spatial coordinate position (x,y), the fusion response tensor Fr(x,y) at three different scales is first obtained; at the same spatial coordinate position (x,y), the fusion response tensor Fr(x,y) at each scale is multiplied by the attention factor ηr(x,y); finally, the weighted results at all scales are summed at this position to obtain the boundary enhancement feature map Fus(x,y). The acquired boundary enhancement feature map Fus(x,y) is compared with the preset boundary enhancement threshold Tedge to determine the defect status at the spatial coordinate position (x,y). The judgment method is as follows: When the boundary enhancement feature map Fus(x,y) < the boundary enhancement threshold Tedge, it indicates that the spatial coordinate position is in the normal region. When the boundary enhancement feature map Fus(x,y) ≥ the boundary enhancement threshold Tedge, it indicates that the spatial coordinate position is in the defect region and is marked as a high response region.

[0011] Preferably, the defect boundary response curvature estimation module includes a high-response region extraction unit and a boundary curvature response calculation unit; The high-response region extraction unit filters out the significant boundary regions, i.e. high-response regions, from the boundary enhancement feature map Fus(x,y). The high-response regions are then processed by using a Laplacian edge enhancement filter to enhance the second-order change response of the boundary pixels and obtain the second-order Laplacian map L(x,y). The second-order Laplace plot L(x, y) is obtained using the following formula: ; By using an edge detection algorithm framework, information is extracted from the second-order Laplacian graph L(x,y) to extract boundary linear information and obtain the boundary binary graph B(x,y). The steps to obtain the boundary binary graph B(x, y) are as follows: Z-score normalization is applied to the grayscale values ​​of the second-order Laplacian plot L(x,y) to enhance the dynamic range contrast between the boundary and the background; extremely low response regions are filtered out to remove noise-dominated regions. Set non-linear dual thresholds Tlow and Thigh; the high threshold Thigh is used to identify strong boundaries, and the low threshold Tlow is used for weak boundaries, and is only retained when connected to a strong boundary; Tlow = 0.4 × Thigh ~ 0.5 × Thigh; The neighboring pixels are checked along the edge gradient direction, and only the current pixel that is a local maximum is retained; the purpose is to refine the boundary lines and remove the response of blurry areas. Mark all points greater than the high threshold Thigh as strong edge points; discard all points less than the low threshold Tlow; retain points between the low threshold Tlow and the high threshold Thigh as edges if they are connected to strong edge points. ; The boundary curvature response calculation unit calculates the boundary curvature response kb(x,y) for the pixels on the extracted boundary binary map B(x,y) by using a geometric curvature estimation method based on the second derivative of the image. The boundary curvature response kb(x, y) is obtained using the following formula: ; In the formula, FU=Fus(x,y) specifically represents the boundary enhancement feature map at the spatial coordinate position (x,y).

[0012] Preferably, the boundary response enhancement index calculation module includes a curvature density joint modeling unit and a boundary enhancement index integral calculation unit; The curvature density co-modeling unit extracts all locations that satisfy the boundary existence condition from the boundary curvature response kb(x,y) to obtain the boundary point set Ω={(x,y)|B(x,y)=1}; In the boundary point set Ω, the distribution density function p(x,y) in the local region is statistically analyzed by analyzing the boundary enhancement feature map Fus(x,y). The distribution density function p(x, y) is obtained as follows: First, at each spatial coordinate position (x, y), calculate the gradient magnitude of the boundary enhancement feature map Fus(x, y) to obtain the intensity of the boundary change at the current point; then, find the maximum value among all gradient magnitudes in the entire set of boundary points Ω; finally, divide the gradient magnitude at the current spatial coordinate position (x, y) by the maximum value among the gradient magnitudes to obtain the distribution density function p(x, y).

[0013] Preferably, the boundary enhancement index integral calculation unit calculates the integral of the product of all boundary curvature responses kb(x,y) and distribution density function p(x,y) in the boundary point set Ω to obtain the boundary enhancement index Ψedge; The boundary enhancement index Ψedge is obtained using the following formula: ; In the formula, AΩ represents the total area of ​​the boundary point set Ω, i.e. the number of effective boundary points, and d represents the integral sign.

[0014] Preferably, the defect category estimation and early warning output module includes a defect feature fusion and classification decision unit and an early warning level determination and response triggering unit; The defect feature fusion and classification decision unit compresses the boundary enhancement feature map Fus(x,y) into a vector form by using a spatial compression method of global average pooling: boundary enhancement vector Fvec∈RF. Do ; The boundary enhancement vector Fvec∈RF Do The defect feature vector zX is obtained by concatenating the boundary enhancement index Ψedge with the boundary enhancement index. The formula for obtaining the defect feature vector zX is: ; In the formula, RF represents the set of real tensors, and Do represents the number of feature channels contained in the image; The defect feature vector zX is input into a fully connected neural network to classify the defect type, and the probability output Ss of each defect category is obtained. The classification probability Pdefect is then calculated. The formula for obtaining the probability output Ss is as follows: Ss = Ws × z + bs, Ss ∈ RF 4 ; In the formula, Ws represents the weight matrix, and Ws∈RF 4×（Do+1） bs represents the bias vector, and bs∈RF 4 Ss represents the probability output of the defect category. Ss=[Ss1, Ss2, Ss3, Ss4] correspond to the scores of broken wire, flattened, worn and normal, respectively. The classification probability Pdefect is obtained as follows: First, for each defect category, the raw score is obtained through the output of the fully connected neural network; then, the scores of all defect categories are exponentially calculated to obtain a set of positive values; finally, the exponential score of each defect category is divided by the sum of the exponential scores of all categories to obtain the classification probability Pdefect of each defect category.

[0015] Preferably, the early warning level determination and response triggering unit combines the defect category classification probability Pdefect with the boundary enhancement index Ψedge to perform graded processing on the detection results and determine whether to trigger an early warning. The determination method is as follows: When the classification probability Pdefect > 0.8 and the boundary enhancement index Ψedge > the preset index threshold Tnga, it indicates high risk and triggers a high-level alarm command: mark it as a broken wire and record the alarm log. When 0.8 ≥ classification probability Pdefect > 0.4 and boundary enhancement index Ψedge ≤ preset index threshold Tnga, it indicates low risk and triggers the manual review process: it is marked as flattened and worn, sent to the review cache pool, and awaits manual confirmation; When 0.4 ≥ classification probability Pdefect, it indicates no risk and the image passes normally: it is marked as a normal image, and the next frame is collected.

[0016] This invention provides an intelligent detection and early warning system for steel wire rope defects based on convolutional neural networks, which has the following beneficial effects: (1) During system operation, multi-scale parallel channel groups are constructed by introducing dilated convolutions with different dilation rates. This enables the network to achieve synchronous perception of local details and overall structure while maintaining constant computational overhead, fundamentally improving the ability to capture defect features at different scales. This is particularly suitable for identifying incomplete broken wires and interlaced wear areas. The system design introduces a boundary-aware attention fusion mechanism, which dynamically combines multi-scale feature responses with edge gradient intensity. This allows the model to automatically allocate boundary weights during the fusion stage, significantly enhancing the response to blurred and non-closed edges. This mechanism effectively suppresses false boundary activation caused by surface contamination, oil stains, and weak reflectivity, improving the boundary accuracy and stability of the detection results.

[0017] (2) By extracting the spatial gradient of the fusion response tensor at each scale, a scale boundary intensity map is constructed, and then an attention factor is formed using normalization. This mechanism overcomes the limitation of traditional fixed convolution channels having weak response capability to discontinuous textures, enabling the model to maintain high boundary resolution when facing blurred wear, burrs, broken wires, or locally deformed contours in wire rope images. Especially in low-contrast edge regions such as tangled wires and micro-bending perturbations, the attention modulation driven by boundary intensity effectively enhances the boundary response, allowing incomplete defect regions to be fully expressed.

[0018] By utilizing multi-scale response channel groups formed by dilated convolutions with different expansion rates, the full range of boundary structure responses, from microscopic cracks to large-scale flattening, is preserved. Based on this, an attention factor, formed after boundary intensity normalization, is used to weight and fuse the responses at each scale. This results in a final boundary enhancement feature map that not only possesses fine-grained resolution of microstructures but also maintains a complete morphological representation of the structural contour.

[0019] (3) When constructing the boundary enhancement index, the system not only considers the curvature response of each boundary point, but also integrates the gradient density distribution of the region where the point is located, establishing a product mapping relationship of boundary change intensity × local density. Finally, by integrating the response of the entire boundary region, the transition from local pixel-level changes to the overall regional comprehensive performance is realized. This joint modeling method makes the boundary enhancement index Ψedge not only usable for the quantitative description of defect boundaries, but also as a strong response indicator for subsequent classification and early warning modules.

[0020] (4) By introducing a joint judgment mechanism of classification probability and boundary response index, the risk level is divided into multiple levels based on the judgment result, and mapped to three decision paths: high risk triggering alarm, low risk submission for review, and no risk normal passage. This realizes dynamic response control of the credibility of the detection result, and is especially suitable for abnormal handling of ambiguous states such as equipment vibration, image blur or boundary breakage and discontinuity in actual field deployment.

[0021] By constructing independent scoring paths for various defects through vector decoupling based on fully connected neural networks, deconstructive recognition of broken wire textures, flattened contours, and worn edges is achieved. This enables accurate classification of image samples with complex shapes, avoiding fuzzy judgments caused by feature aliasing, and is especially suitable for complex working conditions with multiple types of mixed defects. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the block diagram of the intelligent detection and early warning system for steel wire rope defects based on convolutional neural networks of the present invention. Figure 2 This is a schematic diagram illustrating the process of obtaining the classification probability in this invention; Figure 3 This is a trend chart of the boundary enhancement index of the present invention. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Example 1

[0024] This invention provides an intelligent detection and early warning system for steel wire rope defects based on convolutional neural networks. Please refer to [link / reference]. Figure 1 It includes an image input and normalization preprocessing module, a multi-scale receptive field feature extraction module, a boundary-aware attention fusion module, a defect boundary response curvature estimation module, a boundary response enhancement index calculation module, and a defect category estimation and early warning output module. The image input and normalization preprocessing module acquires the original steel wire rope image through the camera and performs preprocessing to obtain the steel wire image tensor Istd; The multi-scale receptive field feature extraction module is based on the wire image tensor Istd, constructs parallel convolutional paths, and introduces dilated convolutional kernels with different dilation rates to construct receptive field channel groups of different scales to obtain the fused response tensor Fr(x,y). The boundary-aware attention fusion module extracts features from the acquired fusion response tensor Fr(x,y) and fuses them to obtain the boundary-enhanced feature map Fus(x,y). The defect boundary response curvature estimation module extracts the high response region in the boundary enhancement feature map Fus(x,y) and calculates the boundary curvature response kb(x,y). The boundary response enhancement index calculation module calculates the boundary enhancement index Ψedge based on the boundary curvature response kb(x,y); The defect category estimation and early warning output module combines the boundary enhancement feature map Fus(x,y) with the boundary enhancement index Ψedge to judge the defects of the wire rope and provide feedback early warning.

[0025] In this embodiment, traditional wire rope image recognition networks typically employ a fixed receptive field. When faced with complex defect structures, such as broken wire edge details and flattened contour stretching, they exhibit problems such as blurred response and unclear structural discrimination. This system constructs multi-scale parallel channel groups by introducing dilated convolutions with different expansion rates. This allows the network to achieve simultaneous perception of local details and overall structure while maintaining constant computational overhead, fundamentally improving its ability to capture defect features at different scales. It is particularly suitable for recognizing incomplete broken wires and interlaced wear areas.

[0026] The system design incorporates a boundary-aware attention fusion mechanism, dynamically combining multi-scale feature responses with edge gradient intensities. This allows the model to automatically assign boundary weights during the fusion phase, significantly enhancing the response to blurred and non-closed edges. This mechanism effectively suppresses false boundary activation caused by surface contamination, oil stains, and weak reflectivity, improving the boundary accuracy and stability of the detection results.

[0027] By introducing a curvature response calculation module based on the second derivative of the image into the high-response boundary region and further constructing a boundary response enhancement index, this system achieves a geometric-level quantitative representation of defect morphology, avoiding the single-method judgment of "whether it is a defect" in traditional methods. This index not only provides a sensitivity criterion for defects such as broken wires and flattening, but also serves as a quantitative indicator for defect level classification and alarm grading, enhancing the interpretability and risk correlation of the system output. Using the boundary enhancement feature map and the boundary response index together as input features, a structure-geometry joint vector input classification network is formed, which can effectively distinguish defect types with similar textures but different structural essences. Combining the classification probability and structural strength dual-threshold grading judgment logic, an intelligent response strategy of "high-risk immediate warning, medium-risk review and caching, and low-risk automatic release" is realized, improving the overall practicality, reliability, and decision-making closed-loop capability of the system. Example 2

[0028] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically: the image input and normalization preprocessing module includes a color structure unification and interference suppression processing unit and a contrast adaptive stretching and tensor construction processing unit; The color structure unification and interference suppression processing unit acquires the original wire rope image through an industrial camera, performs color space conversion and high-frequency interference removal, and obtains the processed wire rope image Yfi(x,y). The color space conversion process is as follows: the original steel wire rope image captured by the industrial camera is converted from the RGB color space to the luminance-color difference color space, and the structural luminance information and color perturbation information are separated. The conversion formula is as follows: Y(x,y)=0.299R(x,y)+0.587G(x,y)+0.114B(x,y); Cb(x,y)=128-0.169R(x,y)-0.331G(x,y)+0.5B(x,y); Cr(x,y)=128+0.5R(x,y)-0.419G(x,y)-0.081B(x,y); In the formula, Y(x,y) represents the luminance channel, Cb(x,y) represents the blue color difference component, Cr(x,y) represents the red color difference component, (x,y) is the image space coordinate position, and R, G and B represent the pixels of the red, green and blue channels respectively; High-frequency interference removal is based on Fourier spectrum analysis. High-frequency energy is screened in the image to remove periodic texture artifacts, and a band-limited filter is used to remove noise from the original wire rope image. The contrast-adaptive stretching and tensor construction processing unit standardizes the processed wire rope image Yfi(x,y) by using local histogram enhancement and tensor normalization to obtain the wire rope image tensor Istd; The local histogram enhancement method involves applying a contrast-limited adaptive histogram equalization method to perform local contrast stretching on the processed wire rope image Yfi(x,y) to enhance local texture differences and suppress the masking of texture responses by overly bright / dark areas; the formula is as follows: Ycl(x,y)=CLAHE(Yfi(x,y);clip,tile); In the formula, CLAHE represents the histogram equalization function, Ycl(x,y) represents the enhanced wire image, clip represents limiting the maximum histogram height, and tile represents dividing the image into small grids for local equalization; The tensor normalization construction process is as follows: the enhanced wire image Ycl(x,y) is converted into the standard tensor format of the convolutional network, and an additional color difference signal is added to obtain the wire image tensor Istd; The wire image tensor Istd is obtained using the following formula: ; In the formula, Istd(x,y) represents the wire image tensor at the image space coordinate position (x,y), QYcl(x,y) represents the brightness gradient map, and Cb(x,y) / 255 and Cr(x,y) / 255 represent the blue color difference signal and the red color difference signal, respectively.

[0029] The multi-scale receptive field feature extraction module includes a multi-path convolutional feature extraction unit and a receptive field enhancement and feature rearrangement mapping unit; The multi-path convolution feature extraction unit is based on the wire image tensor Istd and constructs different dilation rate dilated convolution kernels Kr in three different parallel paths, where r∈{1,2,4}. For each path, extract the initial response graph cFr(x, y); The initial response plot cFr(x, y) is obtained using the following formula: ; In the formula, k represents the convolution kernel radius, (i, j) represents the convolution kernel position, Kr(i, j) represents the convolution kernel weight at (i, j), Istd(xr×i, yr×j) represents the pixel position participating in convolution in the input wire image tensor, and cFr(x, y) represents the initial response map of the image spatial coordinate position (x, y), specifically including three scales: F1(x, y) represents focusing on the boundary of broken wire texture and burr microstructure, F2(x, y) represents capturing shallow scratches and crack segments, and F4(x, y) represents identifying large-area flattening and fatigue deformation contours; The receptive field enhancement and feature rearrangement mapping unit stacks the three scales in the initial response map cFr(x,y) along the channel dimension to form a scale fusion tensor. The stacked tensor is then adjusted for consistency in spatial coordinate mapping to obtain the fused response tensor Fr(x,y). The fusion response tensor Fr(x, y) is obtained using the following formula: Fr(x,y)=Concat(F1(x,y),F,2(x,y),F4(x,y)); In the formula, Concat represents the tensor dimension concatenation function.

[0030] In this embodiment, a conversion from RGB to the luminance-color difference (Y-Cb-Cr) color space is employed to effectively separate the structural and color information of the image. This allows subsequent processing to focus on texture and boundaries, rather than being affected by lighting or color cast. The introduction of Fourier frequency domain analysis and band-limited filtering in the high-frequency interference suppression stage effectively removes periodic pseudo-textures and mechanical friction stripes, significantly reducing false edge responses in industrial images and enhancing the accuracy of true boundary extraction. This structure avoids edge misjudgment caused by uneven lighting and jitter without altering the original image, providing a stable and unified input foundation for subsequent convolutional modules.

[0031] Contrast-limited local histogram enhancement adaptively strengthens the contrast in areas of texture variation, making it particularly suitable for enhancing low-contrast areas such as tiny cracks and broken wires, without easily causing overexposure or loss of detail. After tensor normalization, the enhanced image retains not only luminance information but also normalized chromatic aberration signals, forming a three-dimensional tensor format that includes structural gradients and weak chromatic aberration features. This format better meets the input requirements of neural networks and facilitates subsequent deep feature extraction.

[0032] This module employs three parallel paths to construct the receptive field, each using dilated convolutional kernels with different dilation rates to focus on the texture range of different defect morphologies: the small receptive field path focuses on sharp broken wire edges and fine cracks; the medium receptive field path identifies surface scratches and early cracks; and the large receptive field path captures the boundary contours of flattened and fatigue-deformed areas. Compared to traditional convolution, this structure has stronger structural discrimination capabilities, can accommodate multiple defect scale features, and is particularly adaptable in the identification of irregular defect regions.

[0033] The response maps obtained from different convolutional paths are stacked along the channel dimension to form a unified scale fusion tensor, and a coordinate mapping consistency operation is further performed to ensure that the registration relationship between features of different scales is maintained. This fusion method avoids spatial displacement errors caused by multiple paths, which helps to construct a more coherent representation of defect features, and is particularly suitable for the identification of discontinuous and non-continuous boundary structures. Example 3

[0034] This embodiment is an explanation based on Embodiment 2. Please refer to it. Figure 1 Specifically: the boundary-aware attention fusion module includes a scale boundary gradient extraction and attention factor generation unit and a boundary attention guided fusion unit; The scale boundary gradient extraction and attention factor generation unit calculates the spatial gradient for each fused response tensor Fr(x,y) to obtain the boundary intensity map Gr(x,y). The boundary strength map Gr(x, y) is obtained using the following formula: ; In the formula, Gr(x, y) represents the boundary intensity map, specifically the boundary intensity map at the image spatial coordinate position (x, y), and PD represents the partial derivative sign; The boundary intensity maps at all scales are normalized to obtain the attention factor ηr(x,y). The attention factor ηr(x,y) is obtained by summing up the boundary intensity maps Gr(x,y) at the same spatial coordinate position (x,y) for all three scales to obtain the total boundary intensity; then, the boundary intensity map Gr(x,y) of a single scale is divided by the total boundary intensity, and then added to a non-zero constant to obtain the attention factor ηr(x,y) for that single scale.

[0035] The boundary attention-guided fusion unit obtains the boundary enhancement feature map Fus(x,y) by weighting the fusion response tensor Fr(x,y) using the attention factor ηr(x,y) at each scale. The boundary enhancement feature map Fus(x,y) is obtained as follows: at the spatial coordinate position (x,y), the fusion response tensor Fr(x,y) at three different scales is first obtained; at the same spatial coordinate position (x,y), the fusion response tensor Fr(x,y) at each scale is multiplied by the attention factor ηr(x,y); finally, the weighted results at all scales are summed at the position to obtain the boundary enhancement feature map Fus(x,y). The acquired boundary enhancement feature map Fus(x,y) is compared with the preset boundary enhancement threshold Tedge to determine the defect status at the spatial coordinate position (x,y). The judgment method is as follows: When the boundary enhancement feature map Fus(x,y) < the boundary enhancement threshold Tedge, it indicates that the spatial coordinate position is in the normal region. When the boundary enhancement feature map Fus(x,y) ≥ the boundary enhancement threshold Tedge, it indicates that the spatial coordinate position is in the defect region and is marked as a high response region.

[0036] In this embodiment, spatial gradient extraction is performed on the fused response tensor at each scale to construct a scale boundary intensity map. Then, a normalization method is used to form an attention factor. This mechanism overcomes the limitation of traditional fixed convolution channels in their weak response to discontinuous textures, enabling the model to maintain high boundary resolution even when faced with blurred wear, burrs, broken wires, or locally deformed contours in wire rope images. Particularly in low-contrast edge regions such as tangled wires and micro-bending perturbations, the boundary intensity-driven attention modulation effectively enhances the boundary response, allowing even incomplete defect regions to be fully represented.

[0037] By utilizing multi-scale response channel groups formed through convolution with varying dilation rates, the system preserves boundary structure responses across a full range, from microscopic cracks to large-scale flattening. Furthermore, an attention factor, generated after boundary intensity normalization, is used to weight and fuse the responses at each scale. This results in an enhanced boundary feature map that not only possesses fine-grained resolution of microstructures but also maintains a complete morphological representation of the structural outline. This cross-scale response fusion approach improves the system's ability to identify and distinguish complex defect types such as anisotropic wear, nonlinear indentations, and fine cracks.

[0038] After generating the boundary enhancement feature map, the system introduces a preset boundary enhancement threshold as a control condition for response level determination. By comparing each spatial coordinate point in the feature map with this threshold point by point, spatial discrimination between normal and abnormal regions can be effectively achieved, and high-response region marking results can be generated to support accurate input for subsequent defect curvature estimation and category classification. This mechanism avoids misjudgment caused by fluctuations in global response intensity, improves the stability and reliability of the model, and maintains judgment accuracy even under backgrounds of multi-illumination interference and uneven image intensity. Example 4

[0039] This embodiment is an explanation based on Embodiment 3. Please refer to it. Figure 1 Specifically: the defect boundary response curvature estimation module includes a high response region extraction unit and a boundary curvature response calculation unit; The high-response region extraction unit filters out the significant boundary regions, i.e. high-response regions, from the boundary enhancement feature map Fus(x,y). The high-response regions are then processed by using a Laplacian edge enhancement filter to enhance the second-order change response of the boundary pixels and obtain the second-order Laplacian map L(x,y). The second-order Laplace plot L(x, y) is obtained using the following formula: ; By using an edge detection algorithm framework, information is extracted from the second-order Laplacian graph L(x,y) to extract boundary linear information and obtain the boundary binary graph B(x,y). The boundary curvature response calculation unit calculates the boundary curvature response kb(x,y) for the pixels on the extracted boundary binary map B(x,y) by using a geometric curvature estimation method based on the second derivative of the image. The boundary curvature response kb(x, y) is obtained using the following formula: ; In the formula, FU=Fus(x,y) specifically represents the boundary enhancement feature map at the spatial coordinate position (x,y).

[0040] The boundary response enhancement index calculation module includes a curvature density joint modeling unit and a boundary enhancement index integral calculation unit; The curvature density co-modeling unit extracts all locations that satisfy the boundary existence condition from the boundary curvature response kb(x,y) to obtain the boundary point set Ω={(x,y)|B(x,y)=1}; In the boundary point set Ω, the distribution density function p(x,y) in the local region is statistically analyzed by analyzing the boundary enhancement feature map Fus(x,y). The distribution density function p(x, y) is obtained as follows: First, at each spatial coordinate position (x, y), calculate the gradient magnitude of the boundary enhancement feature map Fus(x, y) to obtain the intensity of the boundary change at the current point; then, find the maximum value among all gradient magnitudes in the entire set of boundary points Ω; finally, divide the gradient magnitude at the current spatial coordinate position (x, y) by the maximum value among the gradient magnitudes to obtain the distribution density function p(x, y).

[0041] The boundary enhancement index integral calculation unit calculates the integral of the product of all boundary curvature responses kb(x,y) and distribution density function p(x,y) in the boundary point set Ω to obtain the boundary enhancement index Ψedge; The boundary enhancement index Ψedge is obtained using the following formula: ; In the formula, AΩ represents the total area of ​​the boundary point set Ω (i.e., the number of effective boundary points), and d represents the integral sign.

[0042] In this embodiment, a second-order Laplacian operator is used to enhance the high-response region, amplifying the variation trend of boundary pixels in the image at the second derivative level. This enhances the ability to recognize slight curvature changes, burr edges, and secondary wear textures. This processing method is particularly suitable for irregular microstructures such as nonlinear scratches and curled cracks in wire rope images, enabling the model to extract defect contours in a finer-grained space.

[0043] By performing image curvature calculation on the binary image after boundary line extraction, the edge is no longer merely determined by its position, but rather by its specific curvature response value. This value can then be used to describe defect structural attributes such as "degree of curling," "knotting and bending," and "convex and concave deformation." This geometric feature-based representation method is highly adaptable to detecting defect types with prominent boundary structural changes, such as incomplete broken wires, torn openings, and anisotropic deformation, breaking through the limitations of coarse judgment methods that rely solely on grayscale gradients.

[0044] When constructing the boundary enhancement index, the system not only considers the curvature response of each boundary point but also integrates the gradient density distribution of the region where that point is located, establishing a product mapping relationship of "boundary change intensity × local density". Finally, by integrating the response of the entire boundary region, a transition from local pixel-level changes to a comprehensive representation of the overall region is achieved. This joint modeling approach allows the boundary enhancement index Ψedge to not only be used for the quantitative description of defect boundaries but also as a strong response indicator for subsequent classification and early warning modules.

[0045] The boundary enhancement index Ψedge has a clear physical meaning and computational closure. It does not rely on subjective model settings and can be directly used as an important quantitative indicator to determine whether there are boundary anomalies or whether the boundaries are continuously broken in the wire rope image. It provides a stable and reliable structural input for the downstream early warning decision and category recognition modules of the system. Example 5

[0046] This embodiment is an explanation based on Embodiment 4. Please refer to it. Figure 1 Specifically: the defect category estimation and early warning output module includes a defect feature fusion and classification decision unit and an early warning level determination and response triggering unit; The defect feature fusion and classification decision unit compresses the boundary enhancement feature map Fus(x,y) into a vector form by using a spatial compression method of global average pooling: boundary enhancement vector Fvec∈RF. Do ; The boundary enhancement vector Fvec∈RF Do The defect feature vector zX is obtained by concatenating the boundary enhancement index Ψedge with the boundary enhancement index. The formula for obtaining the defect feature vector zX is: ; In the formula, RF represents the set of real tensors, and Do represents the number of feature channels contained in the image; The defect feature vector zX is input into a fully connected neural network to classify the defect type, and the probability output Ss of each defect category is obtained. The classification probability Pdefect is then calculated. The formula for obtaining the probability output Ss is as follows: Ss = Ws × z + bs, Ss ∈ RF 4 ; In the formula, Ws represents the weight matrix, and Ws∈RF 4×（Do+1） bs represents the bias vector, and bs∈RF 4 Ss represents the probability output of the defect category. Ss=[Ss1, Ss2, Ss3, Ss4] correspond to the scores of broken wire, flattened, worn and normal, respectively. The classification probability Pdefect is obtained as follows: First, for each defect category, the raw score is obtained through the output of the fully connected neural network; then, the scores of all defect categories are exponentially calculated to obtain a set of positive values; finally, the exponential score of each defect category is divided by the sum of the exponential scores of all categories to obtain the classification probability Pdefect of each defect category.

[0047] The warning level determination and response triggering unit combines the defect category classification probability Pdefect with the boundary enhancement index Ψedge to perform graded processing on the detection results and determine whether to trigger a warning. The determination method is as follows: When the classification probability Pdefect > 0.8 and the boundary enhancement index Ψedge > the preset index threshold Tnga, it indicates high risk and triggers a high-level alarm command: mark it as a broken wire and record the alarm log. When 0.8 ≥ classification probability Pdefect > 0.4 and boundary enhancement index Ψedge ≤ preset index threshold Tnga, it indicates low risk and triggers the manual review process: it is marked as flattened and worn, sent to the review cache pool, and awaits manual confirmation; When 0.4 ≥ classification probability Pdefect, it indicates no risk and passes normally: marked as a normal image.

[0048] In this embodiment, classification no longer relies solely on the hidden layer feature vectors of the image network. Instead, the boundary enhancement index obtained by the preceding module is used as a structural sensitivity factor and concatenated with the image feature vector to form a composite defect expression vector. This takes into account both the data representation of texture depth information and structural boundary response, enabling the classification model to have a stronger ability to express differences when identifying irregular boundary deformations and avoiding misjudgments caused by a single channel.

[0049] This embodiment introduces a joint judgment mechanism of classification probability and boundary response index, and classifies risk levels into multiple levels based on the judgment results. These levels are mapped to three decision paths: high risk triggers alarm, low risk submits for review, and no risk allows normal passage. This enables dynamic response control of the credibility of the detection results, and is especially suitable for handling abnormalities such as equipment vibration, image blurring, or boundary breakage and discontinuity in actual field deployments.

[0050] This module constructs independent scoring paths for various defects through vector decoupling based on fully connected neural networks, enabling deconstructive recognition of broken wire textures, flattened contours, and worn edges. This allows for accurate classification of image samples with complex shapes, avoiding fuzzy judgments caused by feature aliasing, and is particularly suitable for complex working conditions with multiple types of mixed defects.

[0051] In this embodiment, an alarm command is automatically triggered and the identification result is recorded under high-risk conditions; under medium-risk conditions, a manual review process is guided and image data is retained; and under no-risk conditions, the test is passed directly. This mechanism realizes graded response and dynamic storage control of detection results. Combined with the system log management interface, a complete defect identification and traceability link can be formed, meeting the requirements for accurate management and response to high-risk defect events in industrial scenarios.

[0052] Although embodiments of the invention have been shown and described, it will be understood by those skilled 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 invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart detection and early warning system for steel wire rope defects based on convolutional neural networks, characterized in that: It includes an image input and normalization preprocessing module, a multi-scale receptive field feature extraction module, a boundary-aware attention fusion module, a defect boundary response curvature estimation module, a boundary response enhancement index calculation module, and a defect category estimation and early warning output module; The image input and normalization preprocessing module acquires the original steel wire rope image through the camera and performs preprocessing to obtain the steel wire image tensor; The multi-scale receptive field feature extraction module is based on the wire image tensor and constructs parallel convolutional paths. By introducing dilated convolutional kernels with different dilation rates, it constructs receptive field channel groups of different scales to obtain the fused response tensor. The boundary-aware attention fusion module extracts features from the acquired fusion response tensor and fuses them to obtain a boundary-enhanced feature map. The defect boundary response curvature estimation module extracts the boundary of high-response regions in the boundary enhancement feature map and calculates the boundary curvature response. The boundary response enhancement index calculation module calculates and obtains the boundary enhancement index based on the boundary curvature response. The defect category estimation and early warning output module combines the boundary enhancement feature map with the boundary enhancement index to judge the defects of the wire rope and provide feedback early warning.

2. The intelligent detection and early warning system for steel wire rope defects based on convolutional neural networks according to claim 1, characterized in that: The image input and normalization preprocessing module includes a color structure unification and interference suppression processing unit and a contrast adaptive stretching and tensor construction processing unit; The color structure unification and interference suppression processing unit acquires the original wire rope image through an industrial camera, performs color space conversion and high-frequency interference removal, and obtains the processed wire rope image Yfi(x,y). The color space conversion process is as follows: the original steel wire rope image captured by the industrial camera is converted from the RGB color space to the luminance-color difference color space, and the structural luminance information and color perturbation information are separated. The conversion formula is as follows: Y(x,y)=0.299R(x,y)+0.587G(x,y)+0.114B(x,y); Cb(x,y)=128-0.169R(x,y)-0.331G(x,y)+0.5B(x,y); Cr(x,y)=128+0.5R(x,y)-0.419G(x,y)-0.081B(x,y); In the formula, Y(x,y) represents the luminance channel, Cb(x,y) represents the blue color difference component, Cr(x,y) represents the red color difference component, (x,y) is the image space coordinate position, and R, G and B represent the pixels of the red, green and blue channels respectively; High-frequency interference removal is based on Fourier spectrum analysis. High-frequency energy is screened in the image to remove periodic texture artifacts, and a band-limited filter is used to remove noise from the original wire rope image. The contrast-adaptive stretching and tensor construction processing unit standardizes the processed wire rope image Yfi(x,y) by using local histogram enhancement and tensor normalization to obtain the wire rope image tensor Istd; The local histogram enhancement method involves applying a contrast-limited adaptive histogram equalization method to perform local contrast stretching on the processed wire rope image Yfi(x,y) to enhance local texture differences and suppress the masking of texture responses by overly bright / dark areas; the formula is as follows: Ycl(x,y)=CLAHE(Yfi(x,y);clip,tile); In the formula, CLAHE represents the histogram equalization function, Ycl(x,y) represents the enhanced wire image, clip represents limiting the maximum histogram height, and tile represents dividing the image into small grids for local equalization; The tensor normalization construction process is as follows: the enhanced wire image Ycl(x,y) is converted into the standard tensor format of the convolutional network, and an additional color difference signal is added to obtain the wire image tensor Istd; The wire image tensor Istd is obtained using the following formula: ； In the formula, Istd(x,y) represents the wire image tensor at the image space coordinate position (x,y), QYcl(x,y) represents the brightness gradient map, and Cb(x,y) / 255 and Cr(x,y) / 255 represent the blue color difference signal and the red color difference signal, respectively.

3. The intelligent detection and early warning system for steel wire rope defects based on convolutional neural networks according to claim 2, characterized in that: The multi-scale receptive field feature extraction module includes a multi-path convolutional feature extraction unit and a receptive field enhancement and feature rearrangement mapping unit; The multi-path convolution feature extraction unit is based on the wire image tensor Istd and constructs different dilation rate dilated convolution kernels Kr in three different parallel paths, where r∈{1,2,4}. For each path, extract the initial response graph cFr(x, y); The initial response plot cFr(x, y) is obtained using the following formula: ； In the formula, k represents the convolution kernel radius, (i, j) represents the convolution kernel position, Kr(i, j) represents the convolution kernel weight at (i, j), Istd(xr×i, yr×j) represents the pixel position participating in convolution in the input wire image tensor, and cFr(x, y) represents the initial response map of the image spatial coordinate position (x, y), specifically including three scales: F1(x, y) represents focusing on the boundary of broken wire texture and burr microstructure, F2(x, y) represents capturing shallow scratches and crack segments, and F4(x, y) represents identifying large-area flattening and fatigue deformation contours; The receptive field enhancement and feature rearrangement mapping unit stacks the three scales in the initial response map cFr(x,y) along the channel dimension to form a scale fusion tensor. The stacked tensor is then adjusted for consistency in spatial coordinate mapping to obtain the fused response tensor Fr(x,y). The fusion response tensor Fr(x, y) is obtained using the following formula: Fr(x,y)=Concat(F1(x,y),F,2(x,y),F4(x,y)); In the formula, Concat represents the tensor dimension concatenation function.

4. The intelligent detection and early warning system for steel wire rope defects based on convolutional neural networks according to claim 3, characterized in that: The boundary-aware attention fusion module includes a scale boundary gradient extraction and attention factor generation unit and a boundary attention-guided fusion unit; The scale boundary gradient extraction and attention factor generation unit calculates the spatial gradient for each fused response tensor Fr(x,y) to obtain the boundary intensity map Gr(x,y). The boundary strength map Gr(x, y) is obtained using the following formula: ； In the formula, Gr(x, y) represents the boundary intensity map, specifically the boundary intensity map at the image spatial coordinate position (x, y), and PD represents the partial derivative sign; The boundary intensity maps at all scales are normalized to obtain the attention factor ηr(x,y). The attention factor ηr(x,y) is obtained by summing up the boundary intensity maps Gr(x,y) at the same spatial coordinate position (x,y) for all three scales to obtain the total boundary intensity; then, the boundary intensity map Gr(x,y) of a single scale is divided by the total boundary intensity, and then added to a non-zero constant to obtain the attention factor ηr(x,y) for that single scale.

5. The intelligent detection and early warning system for steel wire rope defects based on convolutional neural networks according to claim 4, characterized in that: The boundary attention-guided fusion unit obtains the boundary enhancement feature map Fus(x,y) by weighting the fusion response tensor Fr(x,y) using the attention factor ηr(x,y) at each scale. The boundary enhancement feature map Fus(x,y) is obtained as follows: at the spatial coordinate position (x,y), the fusion response tensor Fr(x,y) at three different scales is first obtained; at the same spatial coordinate position (x,y), the fusion response tensor Fr(x,y) at each scale is multiplied by the attention factor ηr(x,y); finally, the weighted results at all scales are summed at the position to obtain the boundary enhancement feature map Fus(x,y). The acquired boundary enhancement feature map Fus(x,y) is compared with the preset boundary enhancement threshold Tedge to determine the defect status at the spatial coordinate position (x,y). The judgment method is as follows: When the boundary enhancement feature map Fus(x,y) < the boundary enhancement threshold Tedge, it indicates that the spatial coordinate position is in the normal region. When the boundary enhancement feature map Fus(x,y) ≥ the boundary enhancement threshold Tedge, it indicates that the spatial coordinate position is in the defect region and is marked as a high response region.

6. The intelligent detection and early warning system for steel wire rope defects based on convolutional neural networks according to claim 5, characterized in that: The defect boundary response curvature estimation module includes a high-response region extraction unit and a boundary curvature response calculation unit; The high-response region extraction unit filters out the significant boundary regions, i.e. high-response regions, from the boundary enhancement feature map Fus(x,y). The high-response regions are then processed by using a Laplacian edge enhancement filter to enhance the second-order change response of the boundary pixels and obtain the second-order Laplacian map L(x,y). The second-order Laplace plot L(x, y) is obtained using the following formula: ； By using an edge detection algorithm framework, information is extracted from the second-order Laplacian graph L(x,y) to extract boundary linear information and obtain the boundary binary graph B(x,y). The boundary curvature response calculation unit calculates the boundary curvature response kb(x,y) for the pixels on the extracted boundary binary map B(x,y) by using a geometric curvature estimation method based on the second derivative of the image. The boundary curvature response kb(x, y) is obtained using the following formula: ； In the formula, FU=Fus(x,y) specifically represents the boundary enhancement feature map at the spatial coordinate position (x,y).

7. The intelligent detection and early warning system for steel wire rope defects based on convolutional neural networks according to claim 6, characterized in that: The boundary response enhancement index calculation module includes a curvature density joint modeling unit and a boundary enhancement index integral calculation unit; The curvature density co-modeling unit extracts all locations that satisfy the boundary existence condition from the boundary curvature response kb(x,y) to obtain the boundary point set Ω={(x,y)|B(x,y)=1}; In the boundary point set Ω, the distribution density function p(x,y) in the local region is statistically analyzed by analyzing the boundary enhancement feature map Fus(x,y). The distribution density function p(x, y) is obtained as follows: First, at each spatial coordinate position (x, y), calculate the gradient magnitude of the boundary enhancement feature map Fus(x, y) to obtain the intensity of the boundary change at the current point; then, find the maximum value among all gradient magnitudes in the entire set of boundary points Ω; finally, divide the gradient magnitude at the current spatial coordinate position (x, y) by the maximum value among the gradient magnitudes to obtain the distribution density function p(x, y).

8. The intelligent detection and early warning system for steel wire rope defects based on convolutional neural networks according to claim 7, characterized in that: The boundary enhancement index integral calculation unit calculates the integral of the product of all boundary curvature responses kb(x,y) and distribution density function p(x,y) in the boundary point set Ω to obtain the boundary enhancement index Ψedge; The boundary enhancement index Ψedge is obtained using the following formula: ； In the formula, AΩ represents the total area of ​​the boundary point set Ω, and d represents the integral sign.

9. The intelligent detection and early warning system for steel wire rope defects based on convolutional neural networks according to claim 8, characterized in that: The defect category estimation and early warning output module includes a defect feature fusion and classification decision unit and an early warning level determination and response trigger unit; The defect feature fusion and classification decision unit compresses the boundary enhancement feature map Fus(x,y) into a vector form by using a spatial compression method of global average pooling: boundary enhancement vector Fvec∈RF. Do ; The boundary enhancement vector Fvec∈RF Do The defect feature vector zX is obtained by concatenating the boundary enhancement index Ψedge with the boundary enhancement index. The formula for obtaining the defect feature vector zX is: ； In the formula, RF represents the set of real tensors, and Do represents the number of feature channels contained in the image; The defect feature vector zX is input into a fully connected neural network to classify the defect type, and the probability output Ss of each defect category is obtained. The classification probability Pdefect is then calculated. The formula for obtaining the probability output Ss is as follows: Ss = Ws × z + bs, Ss ∈ RF 4 ; In the formula, Ws represents the weight matrix, and Ws∈RF 4×（Do+1） bs represents the bias vector, and bs∈RF 4 Ss represents the probability output of the defect category. Ss=[Ss1, Ss2, Ss3, Ss4] correspond to the scores of broken wire, flattened, worn and normal, respectively. The classification probability Pdefect is obtained as follows: First, for each defect category, the raw score is obtained through the output of the fully connected neural network; then, the scores of all defect categories are exponentially calculated to obtain a set of positive values; finally, the exponential score of each defect category is divided by the sum of the exponential scores of all categories to obtain the classification probability Pdefect of each defect category.

10. The intelligent detection and early warning system for steel wire rope defects based on convolutional neural networks according to claim 9, characterized in that: The warning level determination and response triggering unit combines the defect category classification probability Pdefect with the boundary enhancement index Ψedge to perform graded processing on the detection results and determine whether to trigger a warning. The determination method is as follows: When the classification probability Pdefect > 0.8 and the boundary enhancement index Ψedge > the preset index threshold Tnga, it indicates high risk and triggers a high-level alarm command: mark it as a broken wire and record the alarm log. When 0.8 ≥ classification probability Pdefect > 0.4 and boundary enhancement index Ψedge ≤ preset index threshold Tnga, it indicates low risk and triggers the manual review process: it is marked as flattened and worn, sent to the review cache pool, and awaits manual confirmation; When 0.4 ≥ classification probability Pdefect, it indicates no risk and passes normally: marked as a normal image.

Images