Multi-scene steel wire rope defect intelligent detection method and system based on transfer learning

Abstract

The invention discloses a multi-scene steel wire rope defect intelligent detection method and system based on transfer learning, and relates to the technical field of intelligent detection, and the method comprises the steps: collecting a surface image set of a steel wire rope in real time through an industrial camera, and carrying out image processing to generate a standard image set; the method comprises the following steps: constructing a transfer learning backbone network model, extracting a structural feature map, outputting a convolution feature data set through a dimension reduction coding function, performing spatial feature analysis on the structural feature map according to the convolution feature data set, performing semantic feature distribution analysis when the structural feature map does not meet a structural matching requirement, and evaluating a semantic offset condition of an image. And by fitting with a spatial feature analysis result, comprehensive adaptability analysis is further performed on the structural feature pattern, and cross-scene intelligent detection of the defects of the steel wire rope is realized. The method can effectively solve the problem of detecting the defects of the steel wire rope in different working scenes, improves the accuracy and adaptability of detection, and has a wide practical application prospect.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology, specifically to a method and system for intelligent detection of defects in steel wire ropes in multiple scenarios based on transfer learning. Background Technology

[0002] In practical industrial applications, steel wire ropes, as commonly used load-bearing and traction components, are widely used in critical scenarios such as hoisting and transportation, mining, elevators, and cable trays. Their operational safety has a significant impact on the stability of the overall system. Traditional steel wire rope defect detection relies on manual visual inspection, magnetic leakage detection, or image recognition methods in single scenarios, which have limitations. In recent years, with the development of computer vision and artificial intelligence, deep learning-based defect recognition has been gradually applied to steel wire rope inspection tasks. However, these methods often rely on large amounts of training data under single-scenario, high-quality, and uniform conditions, lacking generalization ability. Transfer learning, as a machine learning technique to improve the adaptability of models in small samples and heterogeneous scenarios, enables models to quickly adapt to new environments through shared knowledge transfer and structural parameter optimization, gradually becoming an important solution for intelligent steel wire rope defect detection tasks in multiple scenarios.

[0003] The actual application environment of steel wire ropes is complex and diverse. Factors such as image acquisition angle, lighting conditions, and background clutter vary significantly under different working conditions. Traditional single-model approaches suffer from decreased recognition rate and increased false positive rate when facing scene changes, especially in heterogeneous scenes where the model's generalization ability degrades significantly. Furthermore, the lack of unified image standards across multiple scenes leads to large differences in image quality from different sources, further exacerbating the difficulty of model discrimination. While some existing methods attempt to introduce standardized image processing, they only perform normalization at the low-level feature level, such as size and brightness, making it difficult to perform feature transfer and discrimination at the deep semantic level. This hinders the identification of potential minute structural defects across scenes, affecting the accuracy and reliability of the final recognition. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a multi-scenario intelligent detection method and system for steel wire rope defects based on transfer learning, which solves the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a multi-scenario intelligent detection method for steel wire rope defects based on transfer learning, comprising the following steps:

[0006] S1. Real-time acquisition of surface images of the steel wire rope by industrial cameras deployed in the target steel wire rope detection scene, and image processing to obtain a standard image set;

[0007] S2. Construct a transfer learning backbone network model. Input the standard image set into the transfer learning backbone network model to extract structural feature maps and map them to the shared feature library for reference feature indexing. Then extract the reference structural feature maps and output the convolutional feature data set through the dimensionality reduction encoding function.

[0008] S3. Perform spatial feature analysis on the structural feature map based on the convolutional feature data set, and generate feature matching evaluation based on the analysis results;

[0009] S4. When the feature matching evaluation indicates that the structural feature map does not meet the structural matching requirements, construct a semantic response data set and perform semantic feature distribution analysis on the structural feature map. Generate a semantic offset evaluation based on the analysis results.

[0010] S5. When the semantic offset assessment indicates that the semantics have cross-scene semantic expression, the spatial feature analysis results are fitted with the semantic feature distribution analysis results. Further comprehensive adaptive analysis is performed on the structural feature map, and a comprehensive adaptive assessment is generated based on the analysis results.

[0011] Preferably, S1 includes S11 and S12;

[0012] S11. Real-time acquisition of surface image set of steel wire rope by industrial cameras deployed in the target steel wire rope detection scene, and simultaneous binding of image metadata while acquiring surface image set;

[0013] The image metadata includes image acquisition timestamp information, acquisition angle label, and scene number label;

[0014] S12. Perform image processing on the surface image set to obtain a standard image set;

[0015] The image processing includes size normalization, brightness normalization, edge structure enhancement, and background noise removal;

[0016] The resolution and brightness space of the surface image set are standardized through size normalization and brightness normalization operations. The size normalization operation uses an interpolation resampling method to reset the image to a fixed size, and the brightness normalization uses a linear stretching method to normalize the brightness distribution to a set range. After the surface image set is normalized, edge structure enhancement and background noise removal are performed to extract the edge contours of defects on the wire rope surface and suppress background interference. The edge structure enhancement uses the Laplacian operator for second-order differential enhancement processing, and the background noise removal uses a median filter to filter out isolated pixels and salt-and-pepper noise in the surface image set.

[0017] Preferably, S2 includes S21 and S22;

[0018] S21. Based on the size characteristics and spatial distribution patterns of defects in the standard image set, define the basic hierarchical structure of the convolutional neural network, including an input layer, multi-level convolutional layers, downsampling layers, and feature aggregation layers. Select a convolutional neural network architecture with general perceptual capabilities as the basic model to construct a transfer learning backbone network model. Input all historical standard image sets with defects into the transfer learning backbone network model for model training. Then, input the acquired standard image sets into the transfer learning backbone network model for layer-by-layer feature encoding. The layer-by-layer feature encoding involves inputting the standard image sets into the input layer of the transfer learning backbone network. The input layer performs size alignment and channel alignment processing on the images. The aligned standard image sets are then sequentially fed into multi-level convolutional layers. Each convolutional layer performs convolution and activation operations on the input standard image sets to extract image structural features. The image structural features are then spatially downsampled to transform the image structural features from pixel level to structural level, obtaining the feature response tensor of the corresponding level. The structural layer feature response tensor of the current standard image set is extracted to form a structural feature map.

[0019] For each feature response tensor in the structural feature map, an aggregation operation is performed on its corresponding spatial dimension to obtain a single response value corresponding to each feature channel. The single response values ​​of each feature channel are arranged in channel order to form a convolutional feature response value F used to represent the structural feature state of the current structural feature map.

[0020] S22. Identify the current working environment type of the wire rope by reading the scene identification information attached to the image acquisition task, and map it to the shared feature library to retrieve the reference feature index that matches the current scene. Extract the reference structure feature map in the current state, and generate the standard convolutional feature response value F through a dimensionality reduction coding function. ref ;

[0021] Compare the convolutional feature response value F with the standard convolutional feature response value F of the same working environment type. ref Together they form a convolutional feature data set.

[0022] Preferably, S3 includes S31;

[0023] S31. Based on the convolutional feature data set, spatial feature analysis is performed on the structural feature map. The feature residual matching index Map is obtained by pixel-by-pixel squared difference operation. It represents the sum of the mean squared residuals between the local structure of the structural feature map in the two-dimensional pixel space and the standard image. It reflects the difference strength between the input structural feature map and the reference structural feature map in the shared feature library at the spatial feature layer, as follows.

[0024] ;

[0025] Where log represents the logarithmic function, H represents the number of structural samples in the vertical direction of the structural feature map, W represents the number of structural samples in the horizontal direction of the structural feature map, and F... h，w This represents the convolutional feature response value in the h-th row and w-th column of the current structural feature map. This represents the standard convolutional feature response value of the reference structural feature map in row h and column w.

[0026] Preferably, S3 further includes S32;

[0027] S32. Calculate the mean and standard deviation of the feature residual matching index Map when the historical structural feature map meets the structural matching requirements according to the statistical method, and set the sum of the mean and standard deviation as the structural matching residual threshold Zm. Compare it with the feature residual matching index Map obtained in real time, and generate a feature matching evaluation based on the comparison results.

[0028] When the feature residual matching index Map ≤ the structure matching residual threshold Zm, it means that the current structure feature map meets the structure matching requirements. At this time, the structure feature map is directly used for wire rope defect identification, the classification result label is extracted and a defect judgment report is generated for recording and reporting.

[0029] When the feature residual matching index Map > the structure matching residual threshold Zm, it means that the current structure feature map does not meet the structure matching requirements, and semantic feature distribution analysis is triggered.

[0030] Preferably, S4 includes S41;

[0031] S41. When the feature matching evaluation indicates that the structure matching requirement is not met, extract the structural feature map that does not meet the structural matching requirement, calculate the mean response u of all semantic channels, and simultaneously extract the reference structural feature map of the current match to calculate the mean response u of all semantic channels. ref and response standard deviation σ ref Together, they form a semantic response data set.

[0032] Preferably, S4 further includes S42 and S43;

[0033] S42. Based on the semantic response data set, perform semantic feature distribution analysis on the structural feature map and construct the semantic feature compression index Sem, which represents the offset strength between the structural feature map in the feature channel dimension and the channel distribution of the reference structural feature map, reflecting the abnormal offset of the structural feature map at the channel-level semantic expression level, as follows;

[0034] ;

[0035] Where d represents the number of feature channels, max(0,·) represents the ReLU activation function, which suppresses negative values ​​to 0, and u iThis represents the mean response of the i-th semantic channel in the structural feature map. This represents the mean response of the i-th semantic channel of the reference structural feature map. The standard deviation of the activation of the reference structural feature map in the i-th semantic channel is represented by ln, and ln represents the logarithmic function.

[0036] S43. Calculate the semantic feature compression index Sem of all historical structural feature maps and corresponding reference structural feature maps using statistical methods, and set the median value as the semantic compression deviation threshold Zs. Compare it with the semantic feature compression index Sem obtained in real time, and generate a semantic offset evaluation based on the comparison results.

[0037] When the semantic feature compression index Sem ≤ semantic compression deviation threshold Zs, it indicates that the structural feature map has cross-scene semantic expression, and at this time, comprehensive adaptive analysis is triggered;

[0038] When the semantic feature compression index Sem > the semantic compression deviation threshold Zs, it indicates that the structural feature map does not have cross-scene semantic expression. At this time, the current image is stored in the transfer supplement training pool.

[0039] Preferably, S5 includes S51;

[0040] S51. When the semantic offset evaluation indicates that the structure feature map has cross-scene semantic expression, the feature residual matching index Map and the semantic feature compression index Sem are fitted together. Further comprehensive adaptability analysis is performed on the structure feature map with cross-scene semantic expression, and a comprehensive adaptability index Ada is constructed to represent the overall adaptability of the structure feature map in the multi-scene shared feature library, reflecting the comprehensive performance of structure features and semantic features in a specific environment, as follows.

[0041] ;

[0042] Where exp represents the exponential decay function, Ten is the scene perturbation factor, representing the stability of the structural feature map in the current feature space, and its value is 0 < Ten < 1.

[0043] Preferably, S5 further includes S52;

[0044] S52. Calculate the mean and standard deviation of the comprehensive adaptability index Ada when the historical structural feature map has feature adaptability according to the statistical method, and set the sum of the mean and standard deviation as the adaptability judgment threshold Za. Compare it with the comprehensive adaptability index Ada obtained in real time, and generate a comprehensive adaptability assessment based on the comparison results.

[0045] When the comprehensive fitness index Ada < fitness threshold Za, it means that the structural feature map does not have feature fitness. At this time, the current image is judged to be a cross-scene mapping failure image and is not recognized. It is marked as a difference image and stored in the transfer supplement training pool.

[0046] When the comprehensive fitness index Ada is greater than or equal to the fitness threshold Za, it indicates that the structural feature map has feature fitness, and the identification label and location coordinates are directly output.

[0047] A multi-scenario intelligent detection system for steel wire rope defects based on transfer learning includes an image acquisition module, a feature extraction module, a spatial feature analysis module, a semantic channel evaluation module, and an adaptive decision-making module.

[0048] The image acquisition module acquires a set of surface images of the steel wire rope in real time through industrial cameras deployed in the target steel wire rope detection scene, and performs image processing to obtain a standard image set.

[0049] The feature extraction module is used to construct a transfer learning backbone network model. It inputs a standard image set into the transfer learning backbone network model to extract structural feature maps and maps them to a shared feature library for reference feature indexing. Then, it extracts reference structural feature maps and outputs convolutional feature data sets through a dimensionality reduction encoding function.

[0050] The spatial feature analysis module is used to perform spatial feature analysis on the structural feature map based on the convolutional feature data set, and generate feature matching evaluation based on the analysis results;

[0051] The semantic channel evaluation module is used to construct a semantic response data group and perform semantic feature distribution analysis on the structural feature map when the feature matching evaluation shows that the structural feature map does not meet the structural matching requirements, and generate a semantic offset evaluation based on the analysis results.

[0052] The adaptive decision module is used to fit the spatial feature analysis results with the semantic feature distribution analysis results when the semantic offset evaluation indicates that it has cross-scene semantic expression, further perform comprehensive adaptive analysis on the structural feature map, and generate a comprehensive adaptive evaluation based on the analysis results.

[0053] This invention provides a method and system for intelligent detection of defects in steel wire ropes across multiple scenarios based on transfer learning. It offers the following advantages:

[0054] (1) This method acquires a set of surface images of the wire rope in real time using an industrial camera, and then performs size normalization, brightness normalization, edge structure enhancement, and background noise removal to obtain a standard image set. Size and brightness normalization help eliminate differences in illumination and resolution under different shooting conditions, ensuring image standardization and enabling subsequent feature extraction to be performed on a more stable image quality basis. Edge structure enhancement is performed using the Laplacian operator, while background noise removal uses a median filter to clean up stray noise and isolated pixels in the image. The preprocessing steps enhance the visibility of surface defects on the wire rope, laying a solid foundation for subsequent deep feature learning and detection.

[0055] (2) Based on the size characteristics of the standard image set and the spatial distribution pattern of defects, this method defines the basic hierarchical structure of the convolutional neural network, selects a convolutional neural network architecture with general perception capabilities as the basic model to construct a transfer learning backbone network model, and trains the model using historical defect images. The standard image set is input into the transfer learning backbone network model for layer-by-layer feature encoding. These features are mapped to a shared feature library for reference matching, and the reference structural feature map of the current state is extracted. A convolutional feature data set is generated through a dimensionality reduction encoding function. Based on the convolutional feature data set, the feature residual matching index Map is calculated to generate a feature matching evaluation. When the structural matching requirements are not met, the semantic features of the image are analyzed through the semantic response data set, and a semantic feature compression index Sem is constructed to generate a semantic offset evaluation to determine whether the current image has cross-scene adaptability. Through this dual analysis of spatial and semantic features, the model can accurately determine whether the wire rope defect meets the detection standard and maintain high detection accuracy when the scene changes.

[0056] (3) This method fits the feature residual matching index Map with the semantic feature compression index Sem, further performs a comprehensive adaptive analysis on the structural feature map with cross-scene semantic expression, and constructs a comprehensive adaptive index Ada. Combining the results of feature residual matching and semantic shift, it further improves the accuracy of cross-scene detection. When an image has high adaptability, it will automatically output defect identification labels and their location coordinates, and generate a detection report; when an image fails to meet the adaptability requirements, the image will be marked as a difference image and stored in the transfer supplement training pool for further training. Through this intelligent training and adaptation mechanism, this method can improve the performance of the model in different working environments, ensure the reliability and accuracy of the detection results, and promote the application of intelligent detection of wire rope defects in actual industrial environments. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the steps of the intelligent detection method for multi-scenario steel wire rope defects based on transfer learning in this invention;

[0058] Figure 2 This is a schematic diagram of the process of the intelligent detection system for multi-scenario steel wire rope defects based on transfer learning of the present invention;

[0059] Figure 3 The present invention presents a logic block diagram of a multi-scenario intelligent detection method for steel wire rope defects based on transfer learning. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Example 1

[0062] Please see Figure 1 This invention provides an intelligent detection method for steel wire rope defects in multiple scenarios based on transfer learning. To achieve the above objectives, this invention is implemented through the following technical solution, including the following steps:

[0063] S1. Real-time acquisition of surface images of the steel wire rope by industrial cameras deployed in the target steel wire rope detection scene, and image processing to obtain a standard image set;

[0064] S2. Construct a transfer learning backbone network model. Input the standard image set into the transfer learning backbone network model to extract structural feature maps and map them to the shared feature library for reference feature indexing. Then extract the reference structural feature maps and output the convolutional feature data set through the dimensionality reduction encoding function.

[0065] S3. Perform spatial feature analysis on the structural feature map based on the convolutional feature data set, and generate feature matching evaluation based on the analysis results;

[0066] S4. When the feature matching evaluation indicates that the structural feature map does not meet the structural matching requirements, construct a semantic response data set and perform semantic feature distribution analysis on the structural feature map. Generate a semantic offset evaluation based on the analysis results.

[0067] S5. When the semantic offset assessment indicates that the semantics have cross-scene semantic expression, the spatial feature analysis results are fitted with the semantic feature distribution analysis results. Further comprehensive adaptive analysis is performed on the structural feature map, and a comprehensive adaptive assessment is generated based on the analysis results.

[0068] In this embodiment, in S1, industrial cameras deployed in the target wire rope inspection scene acquire a set of surface images of the wire rope in real time, and perform image processing to obtain a standard image set. The image processing steps include size normalization, brightness normalization, edge structure enhancement, and background noise removal, eliminating the influence of different scenes and lighting conditions, ensuring the consistency and high quality of image data, and providing a stable foundation for subsequent deep learning feature extraction and analysis. Compared with traditional manual screening or simple image processing methods, this step improves the standardization level of image data and reduces interference caused by environmental changes or differences in image quality. In S2, a transfer learning backbone network model is constructed, and the standard image set is input into the transfer learning backbone network for feature extraction. The transfer learning backbone network model extracts the structural feature maps of the images, maps them to a shared feature library for reference feature indexing, extracts the reference structural feature maps, and finally outputs convolutional feature data sets through a dimensionality reduction coding function. Through this step, not only is deep feature learning performed on the wire rope images, but universal features are also extracted from historical defect images, realizing feature transfer and adaptation for multiple scenes. This process enhances cross-scene adaptability, maintaining high detection accuracy in diverse working environments. Step S3 performs spatial feature analysis on the structural feature map using convolutional feature data sets and generates a feature matching evaluation based on the analysis results, indicating the degree of matching between the current image's structural features and historical images. If the feature matching evaluation results show that the current structural feature map does not meet the structural matching requirements, Step S4 is initiated to construct a semantic response data set and perform semantic feature distribution analysis on the structural feature map. By generating a semantic shift evaluation, this step improves the image's cross-scene detection capability, identifying semantic shifts caused by environmental changes in different scenes, maintaining efficient detection capabilities even in diverse working environments. In Step S5, the spatial feature analysis results are fitted with the semantic feature distribution analysis results to further perform a comprehensive adaptability analysis on the structural feature map. Based on the analysis results, a comprehensive adaptability evaluation is generated to assess the adaptability of the structural feature map in a multi-scene shared feature library, reflecting the combined performance of structural and semantic features in a specific environment. Based on the evaluation results, it is determined whether the current image is suitable for the current detection scene. If the structural feature map possesses feature adaptability, it indicates that the current image has strong cross-scene adaptability, and the recognition label and its location coordinates are directly output. If the structural feature map does not possess feature adaptability, it indicates that the current image cannot meet the requirements of cross-scene processing, and the image will be labeled as a difference image and stored in the transfer learning supplement training pool. This mechanism ensures continuous optimization and updates, enhancing the model's intelligent learning and adaptive capabilities.

[0069] Example 2

[0070] Please refer to Figure 3 Specifically: S1 includes S11 and S12;

[0071] S11. Real-time acquisition of surface image set of steel wire rope by industrial cameras deployed in the target steel wire rope detection scene, and simultaneous binding of image metadata while acquiring surface image set;

[0072] The image metadata includes image acquisition timestamp information, acquisition angle label, and scene number label;

[0073] S12. Perform image processing on the surface image set to obtain a standard image set;

[0074] The image processing includes size normalization, brightness normalization, edge structure enhancement, and background noise removal;

[0075] The resolution and brightness space of the surface image set are standardized through size normalization and brightness normalization operations. The size normalization operation uses an interpolation resampling method to reset the image to a fixed size, and the brightness normalization uses linear stretching to normalize the brightness distribution to a set range, reducing the interference caused by different scene lighting conditions on the image feature expression. After the surface image set is normalized, edge structure enhancement and background noise removal are performed to extract the edge contour of the steel wire rope surface defects and suppress background interference. The edge structure enhancement uses the Laplacian operator for second-order differential enhancement processing, and the background noise removal uses a median filter to filter out isolated pixels and salt-and-pepper noise in the surface image set.

[0076] In this embodiment, step S11 involves deploying industrial cameras to acquire real-time images of the wire rope surface and simultaneously binding image metadata, providing background data for subsequent feature extraction and analysis. This not only ensures the timeliness and consistency of the data but also facilitates subsequent multi-scene transfer learning and data tracking. Subsequently, step S12 obtains a standard image set by performing size normalization, brightness normalization, edge structure enhancement, and background noise removal on the surface image set, ensuring the consistency and high quality of the surface image set. Size and brightness normalization effectively eliminate the influence of different lighting and resolutions on image feature expression, while edge structure enhancement and background noise removal optimize the extraction of defect features, making the identification of wire rope surface defects more accurate. These image preprocessing operations improve the ability of the subsequent transfer learning model to identify wire rope defects and enhance its adaptability and stability in complex and variable environments.

[0077] Example 3

[0078] Please refer to Figure 3 Specifically: S2 includes S21 and S22;

[0079] S21. Based on the size characteristics and spatial distribution patterns of defects in the standard image set, define the basic hierarchical structure of the convolutional neural network, including an input layer, multi-level convolutional layers, downsampling layers, and feature aggregation layers. Select a convolutional neural network architecture with general perceptual capabilities as the basic model to construct a transfer learning backbone network model. Input all historical standard image sets with defects into the transfer learning backbone network model for model training. Then, input the acquired standard image sets into the transfer learning backbone network model for layer-by-layer feature encoding. The layer-by-layer feature encoding involves inputting the standard image sets into the input layer of the transfer learning backbone network. The input layer performs size alignment and channel alignment on the images. The aligned standard image sets are then sequentially fed into multi-level convolutional layers. Each convolutional layer performs convolution and activation operations on the input standard image sets to extract image structural features. The image structural features are then spatially downsampled to transform the image structural features from pixel level to structural level, obtaining the feature response tensor of the corresponding layer. The structural layer feature response tensor of the current standard image set is extracted to form a structural feature map with a size of H×W×d.

[0080] For each feature response tensor in the structural feature map, an aggregation operation is performed on its corresponding spatial dimension to obtain a single response value corresponding to each feature channel. The single response values ​​of each feature channel are arranged in channel order to form a convolutional feature response value F used to represent the structural feature state of the current structural feature map.

[0081] The configuration input layer receives a standard image set and sets the input size to a fixed H0×W0×C, where H0 represents the original height of the input image, which refers to the number of pixel rows of the image before it is sent to the network, W0 represents the original width of the input image, which is the number of pixel columns of the image, and C is the number of channels.

[0082] In the multi-level convolutional layer, each level of the convolutional layer consists of a convolutional kernel, an activation function, and a normalization unit, which extracts texture, edge, and structural information at different scales of the standard image set;

[0083] The downsampling layer achieves downsampling by setting stride convolution, forming a feature map sequence that gradually reduces spatial dimension and increases semantic density;

[0084] The feature aggregation layer is used to fuse features extracted from different levels by multi-level convolutional layers and downsampling layers;

[0085] S22. Identify the current working environment type of the wire rope by reading the scene identification information attached to the image acquisition task, and map it to the shared feature library to retrieve the reference feature index that matches the current scene. Extract the reference structure feature map in the current state, and generate the standard convolutional feature response value F through a dimensionality reduction coding function. ref ;

[0086] Compare the convolutional feature response value F with the standard convolutional feature response value F of the same working environment type. ref Together they form a convolutional feature data set.

[0087] In this embodiment, in S21, a transfer learning backbone network model is constructed by defining the basic hierarchical structure of a convolutional neural network and combining the size features of the standard image set and the spatial distribution patterns of defects. The basic hierarchical structure includes a configuration input layer, multi-level convolutional layers, downsampling layers, and feature aggregation layers to extract the texture, edge, and structural features of the wire rope image at different scales. Through layer-by-layer feature encoding, the structural feature map of the wire rope surface is extracted and quantized into a convolutional feature response value F, providing efficient and accurate feature data for subsequent detection and matching. In S22, by reading the scene identification information in the image acquisition task, the working environment of the current wire rope is identified, and a reference structural feature map matching the current scene is retrieved from the shared feature library. The standard convolutional feature response value F is then generated through a dimensionality reduction encoding function. ref This further improves the model's cross-scene adaptability. Finally, the convolutional feature response value F of the current image is compared with the standard convolutional feature response value F of the reference image. ref This approach constructs convolutional feature data sets. It improves the adaptability and accuracy of wire rope defect detection across multiple scenarios, and enhances the model's intelligent learning and cross-scenario transfer capabilities.

[0088] Example 4

[0089] Please refer to Figure 3 Specifically: S3 includes S31;

[0090] S31. Based on the convolutional feature data set, spatial feature analysis is performed on the structural feature map. The feature residual matching index Map is obtained by pixel-by-pixel squared difference operation. It represents the sum of the mean squared residuals between the local structure of the structural feature map in the two-dimensional pixel space and the standard image. It reflects the difference strength between the input structural feature map and the reference structural feature map in the shared feature library at the spatial feature layer, as follows.

[0091] ;

[0092] Where log represents the logarithmic function, H represents the number of structural samples in the vertical direction of the structural feature map, W represents the number of structural samples in the horizontal direction of the structural feature map, and F... h，w This represents the convolutional feature response value in the h-th row and w-th column of the current structural feature map. This represents the standard convolutional feature response value of the reference structural feature map in row h and column w.

[0093] S3 further includes S32;

[0094] S32. Calculate the mean and standard deviation of the feature residual matching index Map when the historical structural feature map meets the structural matching requirements according to the statistical method, and set the sum of the mean and standard deviation as the structural matching residual threshold Zm. Compare it with the feature residual matching index Map obtained in real time, and generate a feature matching evaluation based on the comparison results.

[0095] When the feature residual matching index Map ≤ the structure matching residual threshold Zm, it means that the current structure feature map meets the structure matching requirements. At this time, the structure feature map is directly used for wire rope defect identification, the classification result label is extracted and a defect judgment report is generated for recording and reporting.

[0096] When the feature residual matching index Map > the structure matching residual threshold Zm, it means that the current structure feature map does not meet the structure matching requirements, and semantic feature distribution analysis is triggered.

[0097] In this embodiment, in S31, spatial feature analysis is performed on the structural feature map based on the convolutional feature data set. A feature residual matching index Map is obtained through pixel-by-pixel squared difference calculation. This map measures the intensity of the difference between the input structural feature map and the reference structural feature map at the spatial feature layer, evaluating the matching degree between the local structure of the structural feature map in the two-dimensional pixel space and the standard image. The feature residual matching index Map reflects the local feature differences between the input image and the reference image, quantifying the degree of feature difference between each pixel in the two-dimensional space and between the input image and the standard reference image. This represents the feature difference value between the input structural feature map and the reference structural feature map at a specific location. This formula is used to sum the differences across the entire image (i.e., each pixel), obtaining the total difference across all locations and capturing the matching quality of the entire structural feature map. The derivation of this formula is based on a classic formula combining mean squared error and logarithmic transform. Mean squared error is a commonly used method to measure the difference between two images or two sets of data, widely applied in image processing, computer vision, and pattern recognition. Its formula is: In image processing, logarithmic transformation is often used to reduce the impact of extreme values ​​on calculation results, especially when dealing with large differences. The logarithmic function smooths data, reduces the impact of large errors on the final calculation result, and enhances the detail representation of differences. Therefore, applying logarithmic transformation is to compress the range of differences, making feature matching more smooth and stable. Logarithmic transformation is commonly used in image enhancement and image difference quantization to help avoid computational instability caused by large errors when processing image data. Its form is log(1+AbsDifference), a common logarithmic compression operation, avoiding the inapplicability of the logarithmic function when there is zero difference (AbsDifference=0), while the "1+" term ensures that the logarithmic transformation is always effective; finally, a feature residual matching index map is constructed. S32 calculates the mean and standard deviation of the feature residual matching index Map of historical structural feature maps using statistical methods, sets a structural matching residual threshold Zm, and compares it with the real-time acquired feature residual matching index Map to generate a feature matching evaluation. If the current structural feature map meets the structural matching requirements, it will be directly used for wire rope defect identification, generating classification labels and a defect judgment report. If the current structural feature map does not meet the structural matching requirements, subsequent semantic feature distribution analysis is triggered. This method, through precise feature matching and threshold determination, not only improves the accuracy of defect detection but also automatically adapts to different scene changes in diverse working environments.

[0098] Example 5

[0099] Please refer to Figure 3 Specifically: S4 includes S41;

[0100] S41. When the feature matching evaluation indicates that the structure matching requirement is not met, extract the structural feature map that does not meet the structural matching requirement, calculate the mean response u of all semantic channels, and simultaneously extract the reference structural feature map of the current match to calculate the mean response u of all semantic channels. ref and response standard deviation σ ref Together, they form a semantic response data set.

[0101] S4 further includes S42 and S43;

[0102] S42. Based on the semantic response data set, perform semantic feature distribution analysis on the structural feature map and construct the semantic feature compression index Sem, which represents the offset strength between the structural feature map in the feature channel dimension and the channel distribution of the reference structural feature map, reflecting the abnormal offset of the structural feature map at the channel-level semantic expression level, as follows;

[0103] ;

[0104] Where d represents the number of feature channels, max(0,·) represents the ReLU activation function, which suppresses negative values ​​to 0 and only considers the difference in activation of the current image that is greater than the reference value, u i This represents the mean response of the i-th semantic channel in the structural feature map. This represents the mean response of the i-th semantic channel of the reference structural feature map. The standard deviation of the activation of the reference structural feature map in the i-th semantic channel is represented by ln, and ln represents the logarithmic function.

[0105] S43. Calculate the semantic feature compression index Sem of all historical structural feature maps and corresponding reference structural feature maps using statistical methods, and set the median value as the semantic compression deviation threshold Zs. Compare it with the semantic feature compression index Sem obtained in real time, and generate a semantic offset evaluation based on the comparison results.

[0106] When the semantic feature compression index Sem ≤ semantic compression deviation threshold Zs, it indicates that the structural feature map has cross-scene semantic expression, and at this time, comprehensive adaptive analysis is triggered;

[0107] When the semantic feature compression index Sem > the semantic compression deviation threshold Zs, it indicates that the structural feature map does not have cross-scene semantic expression. At this time, the current image is stored in the transfer supplement training pool.

[0108] In this embodiment, in S41, when the feature matching evaluation determines that the structure matching requirement is not met, the structural feature map that does not meet the matching requirement is extracted and the mean response u of each semantic channel is calculated. Simultaneously, the reference structural feature map of the current match is extracted and the mean response u of all semantic channels is calculated. ref and response standard deviation σ ref S42 constructs a semantic response data set. Based on the semantic response data set, it performs semantic feature distribution analysis on the structural feature map and constructs a semantic feature compression index Sem to quantify the offset intensity between the structural feature map and the reference structural feature map in the channel dimension, reflecting the abnormal offset of the image at the semantic expression level.

[0109] The derivation of this formula is based on the ReLU activation function and the concept of "information gain" in information theory. The mathematical expression for the ReLU activation function is f(x) = max(0,x), which is widely used in deep learning to suppress negative values ​​to 0, thus avoiding the influence of negative values ​​on subsequent calculations. When used to calculate feature channel offsets, it ensures that the offset is non-negative. The concept of "information gain" in information theory typically uses a logarithmic function to smooth and non-linearly compress the data, smoothing the information content according to certain rules to avoid the influence of noise in the calculation. For image feature offset processing, a logarithmic transformation is applied to avoid excessive bias interfering with model evaluation. A summation operation is used to combine the offsets of all feature channels to obtain an overall semantic offset index. The offsets of all feature channels are averaged to ensure that each channel's contribution has equal weight in the calculation. S43 combines historical data statistical analysis and sets the median value as the semantic compression bias threshold Zs. By comparing this with the real-time acquired semantic feature compression index Sem, a semantic offset evaluation is generated based on the comparison results to determine whether the image possesses cross-scene semantic expression. If it meets the cross-scene expression requirements, it enters the comprehensive adaptability analysis stage; if not, the image is stored in the transfer learning supplementary training pool to enhance the model's self-learning ability. This process improves the cross-scene adaptability of the detection and, through continuous model optimization, ensures the identification of wire rope defects in changing working environments.

[0110] Example 6

[0111] Please refer to Figure 3 Specifically: S5 includes S51;

[0112] S51. When the semantic offset evaluation indicates that the structure feature map has cross-scene semantic expression, the feature residual matching index Map and the semantic feature compression index Sem are fitted together. Further comprehensive adaptability analysis is performed on the structure feature map with cross-scene semantic expression, and a comprehensive adaptability index Ada is constructed to represent the overall adaptability of the structure feature map in the multi-scene shared feature library, reflecting the comprehensive performance of structure features and semantic features in a specific environment, as follows.

[0113] ;

[0114] Where exp represents the exponential decay function, Ten is the scene perturbation factor, representing the stability of the structural feature map in the current feature space, and its value is 0 < Ten < 1.

[0115] S5 also includes S52;

[0116] S52. Calculate the mean and standard deviation of the comprehensive adaptability index Ada when the historical structural feature map has feature adaptability according to the statistical method, and set the sum of the mean and standard deviation as the adaptability judgment threshold Za. Compare it with the comprehensive adaptability index Ada obtained in real time, and generate a comprehensive adaptability assessment based on the comparison results.

[0117] When the comprehensive fitness index Ada < fitness threshold Za, it means that the structural feature map does not have feature fitness. At this time, the current image is judged to be a cross-scene mapping failure image and is not recognized. It is marked as a difference image and stored in the transfer supplement training pool.

[0118] When the comprehensive fitness index Ada is greater than or equal to the fitness threshold Za, it indicates that the structural feature map has feature fitness, and the identification label and location coordinates are directly output.

[0119] In this embodiment, in S51, the feature residual matching index Map is fitted with the semantic feature compression index Sem to further perform a comprehensive adaptability analysis on the structural feature map with cross-scene semantic expression, and a comprehensive adaptability index Ada is constructed to evaluate its adaptability in a multi-scene shared feature library. This process measures the stability of the structural feature map in a specific environment by introducing an exponential decay function and a scene perturbation factor Ten, maintaining consistency and efficiency in a changing working environment. The derivation of this formula is based on an exponential decay-based measurement method. In classical physics and mathematics, exponential functions are often used to describe the decay or adaptability process of a system. In this formula, the exponential function is used to compare the adaptability of the structural feature map in different scenarios. In machine learning and transfer learning, weighted summation and exponential decay are used to quantify the adaptability of the model. Especially in the process of transfer learning, the model needs to adapt quickly to a new environment, which is usually achieved by measuring the difference between the current environment and the target environment. The exponential decay function is one of the common tools for measuring the degree of difference. The feature residual matching index Map measures the degree of matching between the current image and the reference image at the spatial feature level. Feature residual refers to the error or gap between the current image and the reference image in the feature space. The semantic feature compression index Sem represents the degree of compression of the structural feature map at the semantic level, reflecting the difference between the semantic information of the current image and the reference image. Its function is to quantify the semantic shift across scenes and describe the abnormal shift of the current image at the feature channel level. Map×Sem represents the joint effect of feature matching and semantic shift, reflecting the total difference between the current image and the reference image. The introduction of the scene perturbation factor Ten considers the impact of scene changes on adaptability. The addition of 1 is to avoid the division by zero when the scene perturbation factor Ten is 0. S52 uses statistical methods to calculate the mean and standard deviation of the comprehensive adaptability index Ada of the historical structural feature maps, and sets the sum of the mean and standard deviation as the adaptability judgment threshold Za. It is compared with the real-time acquired comprehensive adaptability index Ada to determine whether the image has feature adaptability. When the comprehensive fitness index Ada is greater than or equal to the fitness decision threshold Za, the recognition result is directly output; when the comprehensive fitness index Ada is lower than the fitness decision threshold Za, the image is marked as a difference image and stored in the supplementary training pool. This mechanism improves the model's cross-scene adaptability and intelligent learning ability, ensuring that wire rope defect detection can efficiently and accurately identify defects in different working environments.

[0120] Example 7

[0121] Please refer to Figure 2A multi-scenario intelligent detection system for steel wire rope defects based on transfer learning includes an image acquisition module, a feature extraction module, a spatial feature analysis module, a semantic channel evaluation module, and an adaptive decision-making module.

[0122] The image acquisition module acquires a set of surface images of the steel wire rope in real time through industrial cameras deployed in the target steel wire rope detection scene, and performs image processing to obtain a standard image set.

[0123] The feature extraction module is used to construct a transfer learning backbone network model. It inputs a standard image set into the transfer learning backbone network model to extract structural feature maps and maps them to a shared feature library for reference feature indexing. Then, it extracts reference structural feature maps and outputs convolutional feature data sets through a dimensionality reduction encoding function.

[0124] The spatial feature analysis module is used to perform spatial feature analysis on the structural feature map based on the convolutional feature data set, and generate feature matching evaluation based on the analysis results;

[0125] The semantic channel evaluation module is used to construct a semantic response data group and perform semantic feature distribution analysis on the structural feature map when the feature matching evaluation shows that the structural feature map does not meet the structural matching requirements, and generate a semantic offset evaluation based on the analysis results.

[0126] The adaptive decision module is used to fit the spatial feature analysis results with the semantic feature distribution analysis results when the semantic offset evaluation indicates that it has cross-scene semantic expression, further perform comprehensive adaptive analysis on the structural feature map, and generate a comprehensive adaptive evaluation based on the analysis results.

[0127] 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 alterations 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 multi-scenario intelligent detection method for steel wire rope defects based on transfer learning, characterized in that: Includes the following steps: S1. Real-time acquisition of surface images of the steel wire rope by industrial cameras deployed in the target steel wire rope detection scene, and image processing to obtain a standard image set; S2. Construct a transfer learning backbone network model. Input the standard image set into the transfer learning backbone network model to extract structural feature maps and map them to the shared feature library for reference feature indexing. Then extract the reference structural feature maps and output the convolutional feature data set through the dimensionality reduction encoding function. S3. Perform spatial feature analysis on the structural feature map based on the convolutional feature data set, and generate feature matching evaluation based on the analysis results; S4. When the feature matching evaluation indicates that the structural feature map does not meet the structural matching requirements, construct a semantic response data set and perform semantic feature distribution analysis on the structural feature map. Generate a semantic offset evaluation based on the analysis results. S5. When the semantic offset assessment indicates that the semantics have cross-scene semantic expression, the spatial feature analysis results are fitted with the semantic feature distribution analysis results. Further comprehensive adaptive analysis is performed on the structural feature map, and a comprehensive adaptive assessment is generated based on the analysis results.

2. The intelligent detection method for multi-scenario steel wire rope defects based on transfer learning according to claim 1, characterized in that: S1 includes S11 and S12; S11. Real-time acquisition of surface image set of steel wire rope by industrial cameras deployed in the target steel wire rope detection scene, and simultaneous binding of image metadata while acquiring surface image set; The image metadata includes image acquisition timestamp information, acquisition angle label, and scene number label; S12. Perform image processing on the surface image set to obtain a standard image set; The image processing includes size normalization, brightness normalization, edge structure enhancement, and background noise removal; The resolution and brightness space of the surface image set are standardized through size normalization and brightness normalization operations. The size normalization operation uses an interpolation resampling method to reset the image to a fixed size, and the brightness normalization uses a linear stretching method to normalize the brightness distribution to a set range. After the surface image set is normalized, edge structure enhancement and background noise removal are performed to extract the edge contours of defects on the wire rope surface and suppress background interference. The edge structure enhancement uses the Laplacian operator for second-order differential enhancement processing, and the background noise removal uses a median filter to filter out isolated pixels and salt-and-pepper noise in the surface image set.

3. The intelligent detection method for multi-scenario steel wire rope defects based on transfer learning according to claim 2, characterized in that: S2 includes S21 and S22; S21. Based on the size characteristics and spatial distribution patterns of defects in the standard image set, define the basic hierarchical structure of the convolutional neural network, including an input layer, multi-level convolutional layers, downsampling layers, and feature aggregation layers. Select a convolutional neural network architecture with general perceptual capabilities as the basic model to construct a transfer learning backbone network model. Input all historical standard image sets with defects into the transfer learning backbone network model for model training. Then, input the acquired standard image sets into the transfer learning backbone network model for layer-by-layer feature encoding. The layer-by-layer feature encoding involves inputting the standard image sets into the input layer of the transfer learning backbone network. The input layer performs size alignment and channel alignment processing on the images. The aligned standard image sets are then sequentially fed into multi-level convolutional layers. Each convolutional layer performs convolution and activation operations on the input standard image sets to extract image structural features. The image structural features are then spatially downsampled to transform the image structural features from pixel level to structural level, obtaining the feature response tensor of the corresponding level. The structural layer feature response tensor of the current standard image set is extracted to form a structural feature map. For each feature response tensor in the structural feature map, an aggregation operation is performed on its corresponding spatial dimension to obtain a single response value corresponding to each feature channel. The single response values ​​of each feature channel are arranged in channel order to form a convolutional feature response value F used to represent the structural feature state of the current structural feature map. S22. Identify the current working environment type of the wire rope by reading the scene identification information attached to the image acquisition task, and map it to the shared feature library to retrieve the reference feature index that matches the current scene. Extract the reference structure feature map in the current state, and generate the standard convolutional feature response value F through a dimensionality reduction coding function. ref ; Compare the convolutional feature response value F with the standard convolutional feature response value F of the same working environment type. ref Together they form a convolutional feature data set.

4. The intelligent detection method for multi-scenario steel wire rope defects based on transfer learning according to claim 3, characterized in that: S3 includes S31; S31. Based on the convolutional feature data set, spatial feature analysis is performed on the structural feature map. The feature residual matching index Map is obtained by pixel-by-pixel squared difference operation. It represents the sum of the mean squared residuals between the local structure of the structural feature map in the two-dimensional pixel space and the standard image. It reflects the difference strength between the input structural feature map and the reference structural feature map in the shared feature library at the spatial feature layer, as follows. ； Where log represents the logarithmic function, H represents the number of structural samples in the vertical direction of the structural feature map, W represents the number of structural samples in the horizontal direction of the structural feature map, and F... h，w This represents the convolutional feature response value in the h-th row and w-th column of the current structural feature map. This represents the standard convolutional feature response value of the reference structural feature map in row h and column w.

5. The intelligent detection method for multi-scenario steel wire rope defects based on transfer learning according to claim 4, characterized in that: S3 further includes S32; S32. Calculate the mean and standard deviation of the feature residual matching index Map when the historical structural feature map meets the structural matching requirements according to the statistical method, and set the sum of the mean and standard deviation as the structural matching residual threshold Zm. Compare it with the feature residual matching index Map obtained in real time, and generate a feature matching evaluation based on the comparison results. When the feature residual matching index Map ≤ the structure matching residual threshold Zm, it means that the current structure feature map meets the structure matching requirements. At this time, the structure feature map is directly used for wire rope defect identification, the classification result label is extracted and a defect judgment report is generated for recording and reporting. When the feature residual matching index Map > the structure matching residual threshold Zm, it means that the current structure feature map does not meet the structure matching requirements, and semantic feature distribution analysis is triggered.

6. The intelligent detection method for multi-scenario steel wire rope defects based on transfer learning according to claim 5, characterized in that: S4 includes S41; S41. When the feature matching evaluation indicates that the structure matching requirement is not met, extract the structural feature map that does not meet the structural matching requirement, calculate the mean response u of all semantic channels, and simultaneously extract the reference structural feature map of the current match to calculate the mean response u of all semantic channels. ref and response standard deviation σ ref Together, they form a semantic response data set.

7. The intelligent detection method for multi-scenario steel wire rope defects based on transfer learning according to claim 6, characterized in that: S4 further includes S42 and S43; S42. Based on the semantic response data set, perform semantic feature distribution analysis on the structural feature map and construct the semantic feature compression index Sem, which represents the offset strength between the structural feature map in the feature channel dimension and the channel distribution of the reference structural feature map, reflecting the abnormal offset of the structural feature map at the channel-level semantic expression level, as follows; ； Where d represents the number of feature channels, max(0,·) represents the ReLU activation function, which suppresses negative values ​​to 0, and u i This represents the mean response of the i-th semantic channel in the structural feature map. This represents the mean response of the i-th semantic channel of the reference structural feature map. The standard deviation of the activation of the reference structural feature map in the i-th semantic channel is represented by ln, and ln represents the logarithmic function. S43. Calculate the semantic feature compression index Sem of all historical structural feature maps and corresponding reference structural feature maps using statistical methods, and set the median value as the semantic compression deviation threshold Zs. Compare it with the semantic feature compression index Sem obtained in real time, and generate a semantic offset evaluation based on the comparison results. When the semantic feature compression index Sem ≤ semantic compression deviation threshold Zs, it indicates that the structural feature map has cross-scene semantic expression, and at this time, comprehensive adaptive analysis is triggered; When the semantic feature compression index Sem > the semantic compression deviation threshold Zs, it indicates that the structural feature map does not have cross-scene semantic expression. At this time, the current image is stored in the transfer supplement training pool.

8. The intelligent detection method for multi-scenario steel wire rope defects based on transfer learning according to claim 7, characterized in that: S5 includes S51; S51. When the semantic offset evaluation indicates that the structure feature map has cross-scene semantic expression, the feature residual matching index Map and the semantic feature compression index Sem are fitted together. Further comprehensive adaptability analysis is performed on the structure feature map with cross-scene semantic expression, and a comprehensive adaptability index Ada is constructed to represent the overall adaptability of the structure feature map in the multi-scene shared feature library, reflecting the comprehensive performance of structure features and semantic features in a specific environment, as follows. ； Where exp represents the exponential decay function, Ten is the scene perturbation factor, representing the stability of the structural feature map in the current feature space, and its value is 0 < Ten < 1.

9. The intelligent detection method for multi-scenario steel wire rope defects based on transfer learning according to claim 8, characterized in that: S5 also includes S52; S52. Calculate the mean and standard deviation of the comprehensive adaptability index Ada when the historical structural feature map has feature adaptability according to the statistical method, and set the sum of the mean and standard deviation as the adaptability judgment threshold Za. Compare it with the comprehensive adaptability index Ada obtained in real time, and generate a comprehensive adaptability assessment based on the comparison results. When the comprehensive fitness index Ada < fitness threshold Za, it means that the structural feature map does not have feature fitness. At this time, the current image is judged to be a cross-scene mapping failure image and is not recognized. It is marked as a difference image and stored in the transfer supplement training pool. When the comprehensive fitness index Ada is greater than or equal to the fitness threshold Za, it indicates that the structural feature map has feature fitness, and the identification label and location coordinates are directly output.

10. A multi-scenario intelligent detection system for steel wire rope defects based on transfer learning, comprising the multi-scenario intelligent detection method for steel wire rope defects based on transfer learning as described in any one of claims 1-9, characterized in that: It includes an image acquisition module, a feature extraction module, a spatial feature analysis module, a semantic channel evaluation module, and an adaptive decision-making module; The image acquisition module acquires a set of surface images of the steel wire rope in real time through industrial cameras deployed in the target steel wire rope detection scene, and performs image processing to obtain a standard image set. The feature extraction module is used to construct a transfer learning backbone network model. It inputs a standard image set into the transfer learning backbone network model to extract structural feature maps and maps them to a shared feature library for reference feature indexing. Then, it extracts reference structural feature maps and outputs convolutional feature data sets through a dimensionality reduction encoding function. The spatial feature analysis module is used to perform spatial feature analysis on the structural feature map based on the convolutional feature data set, and generate feature matching evaluation based on the analysis results; The semantic channel evaluation module is used to construct a semantic response data group and perform semantic feature distribution analysis on the structural feature map when the feature matching evaluation shows that the structural feature map does not meet the structural matching requirements, and generate a semantic offset evaluation based on the analysis results. The adaptive decision module is used to fit the spatial feature analysis results with the semantic feature distribution analysis results when the semantic offset evaluation indicates that it has cross-scene semantic expression, further perform comprehensive adaptive analysis on the structural feature map, and generate a comprehensive adaptive evaluation based on the analysis results.

Images