Steel wire rope broken wire detection method and system based on electromagnetic guidance cross-modal attention

Abstract

The present application belongs to the cross technical field of steel wire rope nondestructive testing and computer vision, and provides a steel wire rope broken wire detection method and system based on electromagnetic guidance cross-modal attention. Multi-scale time-frequency features are extracted from the magnetic flux leakage signal to obtain an electromagnetic feature vector. The electromagnetic feature vector is mapped into a dynamic weight matrix corresponding to the multi-level of the preset target detection model feature pyramid. After normalizing each dynamic weight matrix, a constraint is introduced to make the weight distribution concentrate in the broken wire sensitive area. The multi-scale feature map is determined according to the image signal. The dynamic weight matrix is applied to the multi-level of the target detection model feature pyramid, and the different scale feature maps are adaptively modulated through element-by-element weighting to obtain the modulated multi-scale feature map. The modulated multi-scale feature map is input into the detection head to obtain the broken wire detection result. The active guidance of the electromagnetic signal to the visual feature and the cross-modal collaborative detection of the electromagnetic-visual information are realized, and the detection accuracy is improved.

Description

Technical Field

[0001] This invention belongs to the field of nondestructive testing of steel wire ropes and computer vision, and particularly relates to a method and system for detecting broken wires in steel wire ropes based on electromagnetically guided cross-modal attention. Background Technology

[0002] Among the current methods for detecting broken wires in steel wire ropes, magnetic flux leakage detection can effectively identify magnetic field distortion caused by internal broken wires, but it is difficult to accurately describe the spatial morphology of the broken wires; machine vision detection can clearly capture the surface morphology of broken wires, but it cannot see through the inside of the steel wire rope and is powerless to detect internal broken wires.

[0003] The information dimensions of magnetic flux leakage detection and machine vision detection techniques are fundamentally different, making direct fusion difficult. In existing technologies, there are methods that convert electromagnetic signals into images and then use target detection models to detect broken wires. However, these methods do not achieve active guidance of visual features by electromagnetic signals, nor can they achieve adaptive modulation of hierarchical features of the target detection model. As a result, the fusion effect is poor, and the detection accuracy needs to be improved. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a method and system for detecting broken wires in steel wire ropes based on electromagnetically guided cross-modal attention. This invention utilizes the physical sensitivity of leakage magnetic signals to broken wire defects to generate multi-scale dynamic weights, which are then adaptively modulated to the feature pyramid of the target detection model. This achieves active guidance of visual features by electromagnetic signals and cross-modal collaborative detection of electromagnetic-visual information, thereby improving detection accuracy.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: In a first aspect, the present invention provides a method for detecting broken wires in a steel wire rope based on electromagnetically guided cross-modal attention, comprising: Acquire the image signal and leakage magnetic signal of the wire rope; After denoising the leakage magnetic signal, multi-scale time-frequency features are extracted; the multi-scale time-frequency features are encoded to obtain electromagnetic feature vectors; the electromagnetic feature vectors are mapped to dynamic weight matrices corresponding to multiple levels of the feature pyramid of the preset target detection model through multiple parallel fully connected branches; after normalizing each dynamic weight matrix, constraints are introduced to concentrate the weight distribution in the wire breakage sensitive area. Multi-scale feature maps are determined based on image signals; dynamic weight matrices are applied to multiple levels of the feature pyramid of the target detection model, and adaptive modulation of feature maps at different scales is performed by weighting each element to obtain modulated multi-scale feature maps; the modulated multi-scale feature maps are input into the detection head to obtain the broken wire detection results.

[0006] Furthermore, the extraction of the multi-scale time-frequency features includes: performing successive variational mode decomposition on the leakage magnetic signal for adaptive denoising to eliminate wave noise and baseline drift; extracting multi-scale time-frequency features from the denoised leakage magnetic signal; the multi-scale time-frequency features include peak amplitude, peak-to-peak spacing, energy spectrum, rise steepness, and peak symmetry.

[0007] Furthermore, the target detection model is YOLOv8, and the mapping of the dynamic weight matrix includes: using a one-dimensional convolutional neural network and a bidirectional long short-term memory network to encode multi-scale time-frequency features to obtain electromagnetic feature vectors; mapping the electromagnetic feature vectors through three parallel fully connected branches to dynamic weight matrices corresponding to the P3, P4 and P5 layers of the YOLOv8 feature pyramid, respectively.

[0008] Furthermore, the dimensions of each weight matrix are matched with the number of channels in the corresponding feature pyramid layer.

[0009] Furthermore, the adaptive modulation process is represented as follows: ; in, For the first Feature map after layer modulation; For the first Original feature map of the layer; This is the corresponding dynamic weight matrix. denoted as the learnable scaling coefficient; ⊙ represents the element-wise product; when the modulation of the layer contributes to reducing the loss, the corresponding learnable scaling system obtains a positive gradient and gradually increases; if the modulation of the layer does not help reduce the loss, the corresponding learnable scaling system obtains a negative gradient and gradually decreases.

[0010] Furthermore, the broken wire detection includes: inputting the modulated multi-scale feature map into the detection head of the target detection model, performing category prediction and bounding box regression of the broken wire target, and obtaining the broken wire location, confidence level and number of broken wires.

[0011] Secondly, the present invention also provides a wire rope breakage detection system based on electromagnetically guided cross-modal attention, comprising: The data acquisition module is configured to acquire image signals and magnetic leakage signals of the wire rope; The dynamic weight matrix determination module is configured to: extract multi-scale time-frequency features after denoising the leakage magnetic signal; encode the multi-scale time-frequency features to obtain electromagnetic feature vectors; map the electromagnetic feature vectors through multiple parallel fully connected branches to dynamic weight matrices corresponding to multiple levels of the feature pyramid of the preset target detection model; and introduce constraints after normalizing each dynamic weight matrix to concentrate the weight distribution in the wire breakage sensitive area. The detection module is configured to: determine multi-scale feature maps based on image signals; apply dynamic weight matrices to multiple levels of the feature pyramid of the target detection model, and adaptively modulate the feature maps of different scales by weighting each element to obtain modulated multi-scale feature maps; input the modulated multi-scale feature maps into the detection head to obtain the broken wire detection results.

[0012] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the wire rope breakage detection method based on electromagnetically guided cross-modal attention as described in the first aspect.

[0013] Fourthly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the steps of the wire rope breakage detection method based on electromagnetically guided cross-modal attention as described in the first aspect.

[0014] Fifthly, the present invention also provides a computer program product, the computer program product comprising a computer program, which, when executed by a processor, implements the steps of the wire rope breakage detection method based on electromagnetically guided cross-modal attention as described in the first aspect.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention first denoises the magnetic flux leakage signal and extracts multi-scale time-frequency features; then encodes these multi-scale time-frequency features to obtain electromagnetic feature vectors; these electromagnetic feature vectors are mapped through multiple parallel fully connected branches to dynamic weight matrices corresponding to multiple levels of the feature pyramid of a preset target detection model. After normalizing each dynamic weight matrix, constraints are introduced to concentrate the weight distribution in the wire-breaking sensitive area; then, multi-scale feature maps are determined based on the image signal; the dynamic weight matrices are applied to multiple levels of the feature pyramid of the target detection model, and adaptive modulation of the feature maps at different scales is performed through element-wise weighting to obtain modulated multi-scale feature maps; the modulated multi-scale feature maps are input into the detection head to obtain the wire-breaking detection result; by utilizing the physical sensitivity of the magnetic flux leakage signal to wire-breaking defects, multi-scale dynamic weights are generated to adaptively modulate the feature pyramid of the target detection model, realizing the active guidance of electromagnetic signals for visual features and cross-modal collaborative detection of electromagnetic-visual information, thus improving detection accuracy. Attached Figure Description

[0016] The accompanying drawings, which form part of this embodiment, are used to provide a further understanding of this embodiment. The illustrative embodiments and their descriptions are used to explain this embodiment and do not constitute an improper limitation of this embodiment.

[0017] Figure 1This is a flowchart of the method in Embodiment 1 of the present invention; Figure 2 This refers to the electromagnetically guided cross-modal attention module of Embodiment 1 of the present invention; Figure 3 This is the adaptive modulation of the feature pyramid in Embodiment 1 of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0020] Example 1: As a core load-bearing component in critical fields such as mine hoisting, port loading and unloading, and passenger ropeways, wire rope breakage directly affects equipment operational safety and the safety of people's lives and property. Currently, wire rope breakage detection mainly faces the following technical challenges: Single detection methods have inherent limitations. Magnetic flux leakage detection can effectively identify magnetic field distortions caused by internal wire breaks, but it is difficult to accurately describe the spatial morphology (location, orientation, and opening width) of the broken wire. Machine vision detection can clearly capture the surface morphology of the broken wire, but it cannot see into the inside of the wire rope and is powerless to detect internal broken wires. The two technologies have fundamentally different information dimensions and are difficult to directly integrate.

[0021] Multimodal information fusion lacks an effective guidance mechanism. Existing electromagnetic-visual fusion methods mostly adopt simple feature splicing or weighted fusion strategies, failing to fully utilize the physical sensitivity of electromagnetic signals to wire breakage defects to guide visual feature extraction, resulting in visual models having difficulty focusing on wire breakage areas in complex backgrounds.

[0022] Visual detection models are not good at detecting small targets such as broken wires. Broken wires in steel ropes are typical small targets in images (only 1-3mm wide). Although the feature pyramids of general target detection models such as YOLOv8 have multi-scale designs, they still have the problem of missing detection for such slender, low-contrast small targets.

[0023] In existing technologies, electromagnetic signals are converted into images and then the target detection model (YOLOv8) is used to detect broken wires. However, this method only processes images and does not realize the active guidance of electromagnetic signals to visual features. In addition, some studies have proposed damage detection methods based on YOLOv8, but none of them involve cross-modal attention fusion mechanisms.

[0024] To address at least one of the aforementioned problems, this embodiment provides a wire rope breakage detection method based on electromagnetically guided cross-modal attention. Utilizing the physical sensitivity of leakage magnetic signals to wire breakage defects, multi-scale dynamic weights are generated and adaptively modulated onto the YOLOv8 feature pyramid. This achieves cross-modal collaborative detection of electromagnetic and visual information, resolving issues such as the lack of effective guidance in electromagnetic-visual information fusion and insufficient perception of small targets like broken wires in existing wire rope breakage detection methods. Figure 1 As shown, the method includes the following specific steps: S1. Synchronous acquisition of multi-source data: The encoder and FPGA hard synchronization trigger circuit control three line scanning cameras and a magnetic flux leakage detector to synchronously acquire surface image signal sequences and magnetic flux leakage signal sequences at the same spatial position of the wire rope, with the synchronization error controlled within 1mm.

[0025] In one specific embodiment, at the wire rope inspection site, the inspection probe is installed on the wire rope's running path. The wire rope displacement is acquired in real time via a rotary encoder. The FPGA divides the encoder pulses and synchronously triggers three line-scan cameras (2048×2048 pixels resolution) and a magnetic flux leakage detector (10kHz sampling rate). The acquired images are denoised and illumination-corrected before being input into the YOLOv8 Backbone network to extract multi-scale features. The acquired magnetic flux leakage signal is decomposed using SVMD (decomposition layer K=6, penalty factor α=2000), and the three intrinsic mode components with the highest correlation to the defect signal are selected to reconstruct the signal, extracting 8-dimensional time-frequency features such as peak amplitude, peak-to-peak spacing, and energy spectrum.

[0026] S2. Preprocessing and feature extraction of leakage magnetic signal: The acquired magnetic flux leakage signal was subjected to successive variational mode decomposition (SVMD) adaptive denoising to eliminate wave noise and baseline drift. Multi-scale time-frequency features, including peak amplitude, peak-to-peak spacing, energy spectrum, rise steepness, and peak symmetry, were extracted from the denoised magnetic flux leakage signal. Multi-scale feature maps were also extracted from the surface image signal.

[0027] In one specific embodiment, the 8-dimensional time-frequency features extracted in step S1 are input into a cascaded network of 1D-CNN (convolution kernel size 3, number of channels 64→128→256) and Bi-LSTM (128-dimensional hidden layer) to output a 256-dimensional electromagnetic feature vector Fmag.

[0028] S3. Construct an electromagnetically guided cross-modal attention module (EMCA): like Figure 2 As shown, the EMCA module includes an electromagnetic feature encoder submodule, a multi-scale dynamic weight generator submodule, and a weight normalization and sparsification submodule.

[0029] S3.1 Electromagnetic Feature Encoder: A one-dimensional convolutional neural network (1D-CNN) and a bidirectional long short-term memory network (Bi-LSTM) are cascaded to encode the multi-scale time-frequency features extracted in step S2 and output the electromagnetic feature vector Fmag.

[0030] S3.2, Multi-scale dynamic weight generator: The electromagnetic eigenvector Fmag is mapped to dynamic weight matrices corresponding to layers P3, P4, and P5 of the YOLOv8 feature pyramid via three parallel fully connected branches. , and The dimensions of each weight matrix are matched with the number of channels in the corresponding feature pyramid layer.

[0031] S3.3 Weight Normalization and Sparsity: After normalizing each dynamic weight matrix with Sigmoid, a sparsity constraint is introduced to concentrate the weight distribution in the wire breakage sensitive area and suppress the weight response to background noise.

[0032] The wire breakage sensitive area is a localized region on the steel wire rope where there is a wire breakage or a high risk of wire breakage. Within this region, the physical characteristics of the magnetic flux leakage signal (such as magnetic field distortion and sudden changes in magnetic flux) are strongly correlated with the wire breakage damage. For example, a wire breakage can cause a sudden change in local permeability, which manifests as waveform distortion in the magnetic flux leakage signal. During training, the model automatically learns from a large amount of data labeled with wire breakage locations. The model continuously adjusts the weights through backpropagation, gradually enhancing the response to these regions. Each feature component in the electromagnetic feature vector Fmag can be traced back to a specific spatial location on the steel wire rope; therefore, when the model identifies a certain electromagnetic feature that is highly correlated with a wire breakage, the spatial location corresponding to that feature is marked as a sensitive area.

[0033] Optionally, sparsity constraints can be applied to the output of the dynamic weight generator via L1 regularization, for example, during training, to the multi-scale dynamic weights. , and Applying L1 regularization, the mathematical properties of which push a large number of unimportant weights toward zero, retains only the few weights that contribute most to the loss function. The model optimizes both the detection loss and the sparsity loss through backpropagation. For non-broken wire regions (background, wave noise, normal wire regions), the corresponding weights gradually approach zero under the influence of L1 regularization; while for broken wire sensitive regions, the corresponding weights retain larger values ​​because they significantly contribute to reducing the detection loss.

[0034] In one specific embodiment, the 256-dimensional electromagnetic feature vector Fmag is input into three parallel fully connected branches (each containing two fully connected layers with ReLU activation function), which output weight vectors matching the dimensions of the YOLOv8 feature pyramids P3 (256 channels), P4 (512 channels), and P5 (1024 channels), respectively. After Sigmoid normalization, the weight vectors are obtained. , and L1 sparse regularization is introduced during training, with a sparse coefficient λ=0.001.

[0035] Adaptive modulation of S4 and YOLOv8 feature pyramids: like Figure 3 As shown, the dynamic weight matrix generated in step S3 , and The modulation is applied to layers P3, P4, and P5 of the YOLOv8 feature pyramid, respectively, to achieve adaptive modulation of feature maps at different scales through element-wise weighting. The modulation process is represented as follows: ; in, For the first Feature map after layer modulation; For the first Original feature map of the layer; This is the corresponding dynamic weight matrix. is a learnable scaling factor; ⊙ represents element-wise multiplication.

[0036] Among them, the dynamic weight matrix The purpose of this matrix is ​​to assign attention to different regions of the feature map, with initial values ​​that are random. During training, the model compares image features with labels indicating whether a broken wire is present. A weight in a region is strengthened if it helps correctly identify a broken wire; conversely, a weight is weakened if it targets irrelevant background data (such as a normal wire). After training on a large amount of labeled data (marking the precise location of the broken wire), this matrix spontaneously evolves into a spatial mask. Values ​​become large (close to 1) in locations where broken wires are likely to occur, while values ​​become small (close to 0) in background regions.

[0037] Learnable scale regulation system The function of the number is to dynamically adjust the multi-scale dynamic weights. , and The modulation intensity of each level in the YOLOv8 feature pyramid can be dynamically adjusted to change the importance of features at different levels based on the size of the broken wire target. This intensity is also learned through the network and is a variable related to the input data content. A learnable scaling system is also included. The model "customizes" the optimal detection scale for different broken wire targets. Based on the current input features, the model automatically determines which scale of feature map is most effective for detecting broken wires.

[0038] During model initialization, , and All values ​​were set to 0.5, indicating that the initial modulation intensity was consistent across all layers. During training, end-to-end joint optimization was performed using backpropagation and gradient descent. The optimization is not arbitrary; the direction and magnitude of its changes are entirely determined by the detection loss function. The gradient determines: ; in, The classification loss is used to determine whether the region contains a broken wire. The bounding box regression loss is used to accurately locate the broken wire position. This is the weight sparsity loss (L1 regularization).

[0039] When modulation at a certain level contributes to reducing the aforementioned loss, the corresponding learnable scaling system... A positive gradient will be obtained and gradually increased; conversely, if modulation at a certain level does not help reduce the loss or even introduces noise, the corresponding learnable scale adjustment system will be affected. It will acquire a negative gradient, which will be gradually reduced and even approach zero. For example, in the early stages of training, the modulation of layer P3 (high resolution, responsible for small targets) makes a significant contribution to the detection of small broken filaments, therefore... It will gradually increase from 0.5 (e.g., approaching 0.8); while the P5 layer (low resolution, responsible for large targets) contributes limitedly to small broken filaments. It may gradually decrease (e.g., approaching 0.2). The final value is the result of natural convergence driven by the core task of "making the broken wire target most accurately classified and located", thus effectively improving the detection accuracy.

[0040] This embodiment uses a dynamic weight matrix. By focusing attention on the broken wire area, interference from background noise such as oil stains and rust is suppressed, resulting in higher "purity" of the features extracted by the model and fundamentally reducing the risk of false detections. Learnable scale adjustment system. This allows the model to flexibly adapt to targets of different sizes. Whether it's a tiny single broken filament or extensive wear, it can be captured at the most suitable scale, significantly improving recall and reducing false negatives. L1 sparsity allows the dynamic matrix to focus on key regions, eliminating the need for the model to process all information in the entire image. This enables the model to concentrate limited computational resources on the most critical broken filament features, achieving more accurate judgments.

[0041] In one specific embodiment, dynamic weights , and Acting on the corresponding feature maps respectively , and The modulation formula is ,in, , and The initial values ​​were all set to 0.5 and were optimized during the training process.

[0042] S5. Cross-modal collaborative detection and output: S5.1 Model Training: Modulated multi-scale feature maps , and Input the YOLOv8 detector head to perform category prediction and bounding box regression for the broken wire target, and output the broken wire location, confidence level, and number of broken wires.

[0043] In one specific embodiment, the modulated multi-scale feature map , and The model takes a YOLOv8 detector as input and outputs the bounding box of the broken wire and its class confidence score. It is trained on a dataset containing 2000 labeled samples using the SGD optimizer with an initial learning rate of 0.01, momentum of 0.937, and weight decay of 0.0005 for 200 epochs. The loss function includes bounding box regression loss (CIoU), classification loss (BCE), and weight sparsity regularization loss.

[0044] S5.2 Broken wire detection and result output: The trained model is deployed on an edge computing device (NVIDIA Jetson AGX Orin) to perform inference on real-time collected data and output the location (spatial coordinates) of broken wires, the number of broken wires, and the detection confidence.

[0045] This embodiment utilizes leakage magnetic signals to guide a visual detection model to achieve cross-modal collaborative detection of broken wires in steel wire ropes. It can be applied to intelligent detection of broken wire damage in ferromagnetic steel wire ropes such as mine hoisting steel wire ropes, port crane steel wire ropes, and passenger ropeway steel wire ropes.

[0046] Example 2: This embodiment provides a wire rope breakage detection system based on electromagnetically guided cross-modal attention, including: The data acquisition module is configured to acquire image signals and magnetic leakage signals of the wire rope; The dynamic weight matrix determination module is configured to: extract multi-scale time-frequency features after denoising the leakage magnetic signal; encode the multi-scale time-frequency features to obtain electromagnetic feature vectors; map the electromagnetic feature vectors through multiple parallel fully connected branches to dynamic weight matrices corresponding to multiple levels of the feature pyramid of the preset target detection model; and introduce constraints after normalizing each dynamic weight matrix to concentrate the weight distribution in the wire breakage sensitive area. The detection module is configured to: determine multi-scale feature maps based on image signals; apply dynamic weight matrices to multiple levels of the feature pyramid of the target detection model, and adaptively modulate the feature maps of different scales by weighting each element to obtain modulated multi-scale feature maps; input the modulated multi-scale feature maps into the detection head to obtain the broken wire detection results.

[0047] The working method of the system is the same as that of the wire rope broken wire detection method based on electromagnetically guided cross-modal attention in Example 1, and will not be repeated here.

[0048] Example 3: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the wire rope breakage detection method based on electromagnetically guided cross-modal attention described in Embodiment 1.

[0049] Example 4: This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the program, it implements the steps of the wire rope breakage detection method based on electromagnetically guided cross-modal attention described in Embodiment 1.

[0050] Example 5: This embodiment provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the steps of the wire rope broken wire detection method based on electromagnetically guided cross-modal attention described in Embodiment 1.

[0051] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A method for detecting broken wires in steel wire ropes based on electromagnetically guided cross-modal attention, characterized in that, include: Acquire the image signal and leakage magnetic signal of the wire rope; Multi-scale time-frequency features are extracted after denoising the magnetic flux leakage signal; Encoding multi-scale time-frequency features yields electromagnetic feature vectors; Will The electromagnetic feature vector is mapped to dynamic weight matrices corresponding to multiple levels of the feature pyramid of the preset target detection model through multiple parallel fully connected branches. After normalizing each dynamic weight matrix, constraints are introduced to concentrate the weight distribution in the wire breakage sensitive area. Multi-scale feature maps are determined based on image signals; dynamic weight matrices are applied to multiple levels of the feature pyramid of the target detection model, and adaptive modulation of feature maps at different scales is performed by weighting each element to obtain modulated multi-scale feature maps; the modulated multi-scale feature maps are input into the detection head to obtain the broken wire detection results.

2. The method for detecting broken wires in a steel wire rope based on electromagnetically guided cross-modal attention as described in claim 1, characterized in that, The extraction of the multi-scale time-frequency features includes: performing successive variational mode decomposition on the leakage magnetic signal for adaptive denoising to eliminate wave noise and baseline drift; extracting multi-scale time-frequency features from the denoised leakage magnetic signal; the multi-scale time-frequency features include peak amplitude, peak-to-peak spacing, energy spectrum, rise steepness, and peak symmetry.

3. The method for detecting broken wires in a steel wire rope based on electromagnetically guided cross-modal attention as described in claim 1, characterized in that, The target detection model is YOLOv8, and the mapping of the dynamic weight matrix includes: using a one-dimensional convolutional neural network and a bidirectional long short-term memory network to encode multi-scale time-frequency features to obtain electromagnetic feature vectors; mapping the electromagnetic feature vectors through three parallel fully connected branches to dynamic weight matrices corresponding to the P3, P4 and P5 layers of the YOLOv8 feature pyramid, respectively.

4. The method for detecting broken wires in a steel wire rope based on electromagnetically guided cross-modal attention as described in claim 3, characterized in that, The dimensions of each weight matrix are matched with the number of channels in the corresponding feature pyramid layer.

5. The method for detecting broken wires in a steel wire rope based on electromagnetically guided cross-modal attention as described in claim 1, characterized in that, The adaptive modulation process is represented as follows: ； in, For the first Feature map after layer modulation; For the first Original feature map of the layer; This is the corresponding dynamic weight matrix. denoted as the learnable scaling coefficient; ⊙ represents the element-wise product; when the modulation of the layer contributes to reducing the loss, the corresponding learnable scaling system obtains a positive gradient and gradually increases; if the modulation of the layer does not help reduce the loss, the corresponding learnable scaling system obtains a negative gradient and gradually decreases.

6. The method for detecting broken wires in a steel wire rope based on electromagnetically guided cross-modal attention as described in claim 3, characterized in that, The broken wire detection includes: inputting the modulated multi-scale feature map into the detection head of the target detection model, performing category prediction and bounding box regression of the broken wire target, and obtaining the broken wire location, confidence level and number of broken wires.

7. A wire rope broken wire detection system based on electromagnetically guided cross-modal attention, characterized in that, include: The data acquisition module is configured to acquire image signals and magnetic leakage signals of the wire rope; The dynamic weight matrix determination module is configured to: extract multi-scale time-frequency features after denoising the leakage magnetic signal; encode the multi-scale time-frequency features to obtain electromagnetic feature vectors; map the electromagnetic feature vectors through multiple parallel fully connected branches to dynamic weight matrices corresponding to multiple levels of the feature pyramid of the preset target detection model; and introduce constraints after normalizing each dynamic weight matrix to concentrate the weight distribution in the wire breakage sensitive area. The detection module is configured to: determine multi-scale feature maps based on image signals; apply dynamic weight matrices to multiple levels of the feature pyramid of the target detection model, and adaptively modulate the feature maps of different scales by weighting each element to obtain modulated multi-scale feature maps; input the modulated multi-scale feature maps into the detection head to obtain the broken wire detection results.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the steps of the wire rope breakage detection method based on electromagnetically guided cross-modal attention as described in any one of claims 1-6.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the program, it implements the steps of the wire rope breakage detection method based on electromagnetically guided cross-modal attention as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the wire rope breakage detection method based on electromagnetically guided cross-modal attention as described in any one of claims 1-6.

Images