Steel wire rope internal defect on-line detection system based on AI multi-sensor fusion

Abstract

The invention discloses a steel wire rope internal defect on-line detection system based on AI multi-sensor fusion, and relates to the technical field of defect detection, the steel wire rope internal defect on-line detection system comprises a data preprocessing module used for synchronously collecting and preprocessing signal data streams of a plurality of heterogeneous sensors in the steel wire rope on-line operation process; the feature decoupling module is used for constructing a causal characterization learning network, inputting the preprocessed signal data flow into the causal characterization learning network, performing fusion and decoupling decomposition, generating a low-dimensional causal feature tensor and a high-dimensional hybrid feature tensor, and calculating a decoupling loss value according to the separation degree of the low-dimensional causal feature tensor and the high-dimensional hybrid feature tensor; and the on-line detection module continuously collects new signal data streams, inputs the new signal data streams into the updated discriminator network, and outputs an on-line detection report of the internal defects of the steel wire rope. According to the method, the causal characterization learning network is constructed, the features after multi-sensor fusion are decomposed, and accurate extraction of defect essential features and effective isolation of interference signals are achieved.

Description

Technical Field

[0001] This invention relates to the field of defect detection technology, and in particular to an online detection system for internal defects in steel wire ropes based on AI multi-sensor fusion. Background Technology

[0002] In recent years, non-destructive testing technology for wire ropes based on multi-sensor fusion has gradually become a research hotspot in the field of industrial safety monitoring. By integrating heterogeneous sensors such as magnetic, acoustic, and optical sensors, complementary information about the internal state of the wire rope can be obtained, and combined with deep learning algorithms, automatic extraction and classification of defect features can be achieved. Existing technologies mostly use convolutional neural networks or recurrent neural networks to perform feature-level or decision-level fusion of multi-source data, aiming to improve the accuracy and robustness of defect identification. In addition, generative adversarial networks have been introduced to enhance the model's adaptability to complex working conditions.

[0003] However, existing methods generally lack explicit modeling of causal relationships during feature fusion, making the extracted features susceptible to interference from non-defect-related factors such as environmental noise and load fluctuations. Confounding factors are highly coupled with real defect features at the data level, making it difficult for the model to distinguish between essential causal features and superficial correlation features. Especially in online detection scenarios, dynamically changing working conditions further amplify the feature confusion problem, restricting the generalization ability and reliability of the detection model. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an online detection system for internal defects in wire ropes based on AI multi-sensor fusion to solve the problems of insufficient generalization ability and high false alarm rate caused by feature confusion in existing wire rope defect detection methods.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides an online detection system for internal defects in steel wire ropes based on AI multi-sensor fusion. The system includes: a data preprocessing module, which simultaneously collects signal data streams from multiple heterogeneous sensors during the online operation of the steel wire rope and performs preprocessing; a feature decoupling module, which constructs a causal representation learning network, inputs the preprocessed signal data streams into the causal representation learning network, performs fusion and decoupling decomposition to generate a low-dimensional causal feature tensor and a high-dimensional hybrid feature tensor, and calculates a decoupling loss value based on the degree of separation between the low-dimensional causal feature tensor and the high-dimensional hybrid feature tensor; an adversarial generation module, which combines the high-dimensional hybrid feature tensor with a preset random noise vector and inputs it into a generator network to output adversarial feature samples; a parameter update module, which inputs the adversarial feature samples and the low-dimensional causal feature tensor into a discriminator network and outputs the difference in discrimination results; calculates an adversarial loss value based on the difference in discrimination results, and updates the parameters of the generator network and the discriminator network together with the adversarial loss value and the decoupling loss value; and an online detection module, which continuously collects new signal data streams and inputs them into the updated discriminator network to output an online detection report of internal defects in the steel wire rope.

[0008] As a preferred embodiment of the online detection system for internal defects of steel wire rope based on AI multi-sensor fusion described in this invention, the multiple heterogeneous sensors include magnetic sensors, acoustic sensors and optical sensors;

[0009] The signal data stream includes magnetic flux leakage signal data, ultrasonic echo signal data, and surface visual image data;

[0010] The preprocessing includes spatiotemporal registration, noise reduction filtering, and amplitude normalization.

[0011] As a preferred embodiment of the online detection system for internal defects of steel wire rope based on AI multi-sensor fusion described in this invention, the steps for constructing the causal representation learning network are as follows:

[0012] A one-dimensional convolutional neural network is used to build a temporal feature encoding layer, a two-dimensional convolutional neural network is used to build a spatial feature encoding layer, and a feature fusion and decoupling layer is built through a cross-modal attention mechanism;

[0013] The temporal feature encoding layer, spatial feature encoding layer, and feature fusion and decoupling layer are processed and fused in parallel to generate a causal representation learning network.

[0014] As a preferred embodiment of the online detection system for internal defects of steel wire rope based on AI multi-sensor fusion described in this invention, the steps for generating the low-dimensional causal feature tensor and the high-dimensional hybrid feature tensor are as follows:

[0015] The preprocessed signal data stream is input into the temporal feature coding layer and the spatial feature coding layer to extract temporal and spatial features.

[0016] Temporal and spatial features are input features that are fused and decoupled to perform cross-modal attention fusion, generating preliminary fused features;

[0017] Based on the preliminary fusion features, the importance scores of each feature dimension in the preliminary fusion features are calculated, and the preliminary fusion features are weighted according to the importance scores to obtain the causal enhancement features;

[0018] The causal enhancement features are input into the decoupled encoder, and dimension mapping and regularization constraints are performed through two parallel fully connected layers to generate a low-dimensional causal feature tensor and a high-dimensional hybrid feature tensor.

[0019] As a preferred embodiment of the online detection system for internal defects of steel wire rope based on AI multi-sensor fusion described in this invention, the steps for calculating the decoupling loss value are as follows:

[0020] A global correlation measure between low-dimensional causal feature tensors and high-dimensional mixed feature tensors is calculated using a dynamic weighted multi-scale mutual information estimator.

[0021] By calling historical online detection results and evaluating the causal driving strength of the low-dimensional causal feature tensor based on the mapping relationship between historical online detection results and the low-dimensional causal feature tensor;

[0022] We weight and fuse the global relevance measure and the causal driving strength to generate a decoupling loss value.

[0023] As a preferred embodiment of the online detection system for internal defects of steel wire rope based on AI multi-sensor fusion described in this invention, the steps for outputting adversarial feature samples are as follows:

[0024] By performing multi-scale spectral analysis on the high-dimensional hybrid feature tensor, the energy distribution feature vector is extracted, and the preset random noise vector is mapped into a conditional vector through a conditional coding network.

[0025] A dynamic weight matrix is ​​generated based on the energy distribution feature vector and condition vector;

[0026] The high-dimensional hybrid feature tensor and the preset random noise vector are input into the backbone encoder of the generator network to obtain the initial feature map.

[0027] The initial feature map is conditionally spectrally modulated using a dynamic weight matrix to generate a modulated feature map. The modulated feature map is then input into the decoder of the generator network to output adversarial feature samples.

[0028] As a preferred embodiment of the online detection system for internal defects of steel wire rope based on AI multi-sensor fusion described in this invention, the steps for determining the difference in output discrimination results are as follows:

[0029] The discriminator network uses an interactive attention mechanism built into it to fuse adversarial feature samples with low-dimensional causal feature tensors to generate interactive enhanced features.

[0030] The processing path is dynamically selected based on the local information entropy and global variance of the interaction enhancement features. The detailed features and context features in the interaction enhancement features are differentiated to generate path modulation features.

[0031] Multi-scale gradient statistical features of path modulation are extracted, and attention-weighted multivariate nonlinear mapping is performed on the multi-scale gradient statistical features to generate discrimination result differences.

[0032] As a preferred embodiment of the online detection system for internal defects of steel wire rope based on AI multi-sensor fusion described in this invention, the steps for calculating the adversarial loss value based on the difference in discrimination results are as follows:

[0033] Multi-scale temporal decomposition of the discrimination results is performed to extract the difference components at different time scales;

[0034] The weighting coefficients are calculated based on the variance of each difference component and the covariance of the difference between each difference component and the discrimination result.

[0035] By integrating and fusing the differential components using weighted coefficients, adversarial loss values ​​are generated.

[0036] As a preferred embodiment of the online detection system for internal defects of steel wire rope based on AI multi-sensor fusion described in this invention, the steps of updating the parameters of the generator network and the discriminator network by jointly using adversarial loss values ​​and decoupling loss values ​​are as follows:

[0037] The adversarial loss value and the decoupling loss value are input into a dual-channel weight allocation network to extract the dynamic features of each loss value, and a dynamic fusion weight is generated using a cross-attention mechanism.

[0038] The adversarial loss value and the decoupling loss value are weighted and fused into a total loss function based on dynamic fusion weights, and the parameters of the generator network and the discriminator network are updated synchronously through the gradient descent algorithm.

[0039] As a preferred embodiment of the online detection system for internal defects of steel wire rope based on AI multi-sensor fusion described in this invention, the steps for outputting an online detection report of internal defects of the steel wire rope are as follows:

[0040] New signal data streams will be continuously collected and input into the updated discriminator network, outputting defect detection results. The defect detection results will be verified in a multi-scale time series and real-time confidence assessment results will be generated.

[0041] Based on the theory of multi-source evidence fusion, the real-time confidence assessment results are synthesized to generate a comprehensive confidence distribution.

[0042] A structured detection report is generated by fusing multimodal information from the comprehensive confidence distribution and the low-dimensional causal feature tensor through a joint decision-making mechanism.

[0043] The beneficial effects of this invention are as follows: By constructing a causal representation learning network, the features after multi-sensor fusion are decomposed, achieving accurate extraction of the essential features of defects and effective isolation of interference signals; by using a decoupling loss function to constrain the feature separation process, it is ensured that causal features are only related to the essential factors of defects inside the wire rope, while hybrid features carry non-critical information such as environmental interference, improving the purity and interpretability of feature representation, enabling the discriminator network to focus on the essential representation of defects, effectively suppressing misjudgments caused by fluctuations in operating conditions. At the same time, by quantifying the degree of separation between causal features and hybrid features, a clear guide is provided for model optimization, enhancing stability and adaptability in complex industrial scenarios. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic diagram of an online detection system for internal defects in steel wire ropes based on AI multi-sensor fusion.

[0046] Figure 2 This is a flowchart for outputting the preprocessed signal data stream.

[0047] Figure 3 A flowchart for generating low-dimensional causal feature tensors and high-dimensional hybrid feature tensors.

[0048] Figure 4 A flowchart for generating a structured inspection report. Detailed Implementation

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0050] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0051] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0052] Reference Figures 1-4 This is one embodiment of the present invention, which provides an online detection system for internal defects in steel wire ropes based on AI multi-sensor fusion, comprising the following steps:

[0053] The data preprocessing module simultaneously collects signal data streams from multiple heterogeneous sensors during the online operation of the wire rope and performs preprocessing.

[0054] Multiple heterogeneous sensors include magnetic sensors, acoustic sensors, and optical sensors;

[0055] It should be noted that magnetic sensors are sensors used to detect changes in magnetic fields, thereby reflecting the internal properties of materials such as defects; acoustic sensors are devices that assess the integrity of the internal structure of materials by capturing the propagation behavior of ultrasonic waves; and optical sensors are devices that use phenomena such as light reflection and scattering to detect the surface condition and minute defects of objects.

[0056] The signal data stream includes magnetic flux leakage signal data, ultrasonic echo signal data, and surface visual image data;

[0057] It should be noted that magnetic flux leakage signal data refers to the leakage magnetic field signal generated by the magnetic sensor detecting magnetic field distortion caused by internal defects (such as broken wires or corrosion) in a magnetized steel wire rope.

[0058] Ultrasonic echo signal data refers to the echo signal reflected back from internal defects in a steel wire rope by sound waves emitted from an ultrasonic probe, which is used to characterize abnormalities in the internal structure of the material.

[0059] Surface visual image data refers to high-definition images of the outer surface of a steel wire rope captured by an optical camera, used to observe information such as surface wear, deformation, or visible damage.

[0060] Preprocessing includes spatiotemporal registration, noise reduction filtering, and amplitude normalization.

[0061] Specifically, based on the linear velocity of the wire rope during online operation and the installation positions of each sensor, the magnetic flux leakage signal data, ultrasonic echo signal data, and surface visual image data are aligned on the time axis to the sampling time corresponding to the same physical location. Spatially, the coordinate reference of the sensing areas of each sensor is unified, ensuring that data from different sources correspond to the same segment of the wire rope in both time and space. Noise reduction and filtering are then applied to the spatiotemporally aligned magnetic flux leakage signal data, ultrasonic echo signal data, and surface visual image data. The magnetic flux leakage signal data uses a low-pass filter to suppress high-frequency electromagnetic interference, the ultrasonic echo signal data uses a time-domain moving average to filter out random acoustic noise, and the surface visual image data uses median filtering to eliminate isolated noise points in the image. Amplitude standardization is then applied to each of the noise-reduced magnetic flux leakage signal data, ultrasonic echo signal data, and surface visual image data. The amplitude range of the magnetic flux leakage signal data is scaled to a unified interval, the energy intensity of the ultrasonic echo signal data is normalized to a reference level, and the pixel brightness of the surface visual image data is adjusted to a consistent dynamic range, generating a pre-processed signal data stream.

[0062] The feature decoupling module constructs a causal representation learning network. The preprocessed signal data stream is input into the causal representation learning network for fusion and decoupling decomposition, generating a low-dimensional causal feature tensor and a high-dimensional mixed feature tensor. The decoupling loss value is calculated based on the degree of separation between the low-dimensional causal feature tensor and the high-dimensional mixed feature tensor.

[0063] A one-dimensional convolutional neural network is used to build a temporal feature encoding layer, a two-dimensional convolutional neural network is used to build a spatial feature encoding layer, and a feature fusion and decoupling layer is built through a cross-modal attention mechanism;

[0064] Specifically, the magnetic flux leakage signal data and ultrasonic echo signal data in the preprocessed signal data stream are arranged in a time series. A one-dimensional convolutional neural network is used to extract local temporal variation patterns layer by layer in the time dimension, forming a representation of the dynamic defect response during wire rope operation, thus completing the construction of the temporal feature encoding layer. The surface visual image data in the preprocessed signal data stream are organized in the form of a two-dimensional pixel matrix. A two-dimensional convolutional neural network is used to capture local texture and structural anomaly features layer by layer in the spatial dimension, forming a spatial distribution representation of the wire rope surface state, thus completing the construction of the spatial feature encoding layer. The temporal variation patterns and... Spatial distribution representations are fed into a cross-modal attention mechanism, which uses attention weights to dynamically measure the correlation strength between features of different modalities and generate a joint representation that integrates multi-source information. The joint representation is explicitly decomposed through dual-path decoupling. One path extracts causal feature tensors directly related to defects inside the wire rope through a low-dimensional bottleneck structure (such as a fully connected layer compressed to 32 dimensions) and enhances its invariance by minimizing its distribution differences under different operating conditions (such as MMD loss). The other path retains the high-dimensional representation (such as 256 dimensions) to capture the hybrid feature tensors affected by environmental interference and sensor noise, thereby completing the construction of the feature fusion and decoupling layer.

[0065] The temporal feature encoding layer, spatial feature encoding layer, and feature fusion and decoupling layer are processed and fused in parallel to generate a causal representation learning network.

[0066] Specifically, the preprocessed signal data stream is fed into the temporal feature coding layer and the spatial feature coding layer respectively to obtain the corresponding temporal variation pattern and spatial distribution representation. The temporal variation pattern and spatial distribution representation are then fed into the feature fusion and decoupling layer, which integrates the two representations across modes to form a causal representation learning network.

[0067] The preprocessed signal data stream is input into the temporal feature coding layer and the spatial feature coding layer to extract temporal and spatial features.

[0068] Specifically, the preprocessed signal data stream is fed into the temporal feature encoding layer, where the magnetic flux leakage signal data and ultrasonic echo signal data are arranged in chronological order. A one-dimensional convolutional neural network is used to capture local change patterns in the time dimension layer by layer to form temporal features. At the same time, the surface visual image data in the preprocessed signal data stream is fed into the spatial feature encoding layer, which is organized in the form of a two-dimensional pixel matrix. A two-dimensional convolutional neural network is used to extract local structure and texture information in the spatial dimension layer by layer to form spatial features.

[0069] Temporal and spatial features are input features that are fused and decoupled to perform cross-modal attention fusion, generating preliminary fused features;

[0070] Specifically, temporal and spatial features are fed into a feature fusion and decoupling layer. A cross-modal attention mechanism is used to jointly model the temporal and spatial features. Attention weights are used to dynamically measure the correlation strength between each time step in the temporal features and each spatial location in the spatial features. This aligns the temporal variation patterns from magnetic flux leakage signal data and ultrasonic echo signal data with the spatial structure information from surface visual image data in a unified representation space, and they mutually enhance each other to form preliminary fused features. These preliminary fused features are then used to explicitly decompose them into low-dimensional causal feature tensors and high-dimensional hybrid feature tensors to decouple the essential information of defects from environmental interference factors.

[0071] Based on the preliminary fusion features, the importance scores of each feature dimension in the preliminary fusion features are calculated, and the preliminary fusion features are weighted according to the importance scores to obtain the causal enhancement features;

[0072] Specifically, based on the preliminary fusion features, the importance score of each feature dimension in the preliminary fusion features is calculated; using the importance score as the weight, the corresponding feature dimensions in the preliminary fusion features are weighted and adjusted so that the feature dimensions that are more closely related to the essence of the defect occupy a larger proportion in the representation, thereby obtaining the causal enhancement features.

[0073] Based on the preliminary fusion features, the importance score of each feature dimension in the preliminary fusion features is calculated, as expressed by:

[0074] ;

[0075] In the formula, Indicates the first Importance scores for each feature dimension Indicates preliminary fusion characteristics The first in Each feature dimension Indicates preliminary fusion characteristics The first in The variance of each feature dimension Indicates the total number of feature dimensions. Indicates the first Each feature dimension index is used to iterate through all feature dimensions during the summation. This indicates that during the summation traversal, features are initially merged. The first in The variance of each feature dimension Indicates preliminary fusion characteristics The first in Each feature dimension and target variable The absolute value of the correlation strength between them.

[0076] It should be noted that the correlation strength The absolute value is used to measure the initial feature fusion. The first in Each feature dimension and target variable The formula focuses on the strength of the correlation between features, regardless of whether they are positive or negative. Therefore, absolute values ​​are used in the formula to ensure that the "importance score" of each feature is always non-negative.

[0077] It should be noted that the target variable In machine learning or data analysis, the target variable typically refers to the final result or output variable that is expected to be predicted or explained. It originates from experimental data; for example, in a defect detection task, the target variable might be the state of a wire rope, such as whether it has a defect (e.g., "defective" or "no defect"). In the formula, the target variable... Its function is to communicate with Correlation between This measures the degree of influence of each feature dimension on the target outcome, thereby assessing the causal importance of each feature dimension.

[0078] It should be noted that the ratios in the formula It is the first The ratio of the variance of one feature dimension to the sum of the variances of all feature dimensions is dimensionless, thus maintaining dimensional uniformity. It is the absolute value of the correlation strength, and it is a dimensionless value. As an importance score, it is dimensionless because it is obtained by multiplying the dimensionless variance ratio and the dimensionless correlation value. Therefore, all parameters in the entire formula maintain the same dimension, ensuring the rationality and consistency of the calculation.

[0079] The causal enhancement features are input into the decoupled encoder, and dimension mapping and regularization constraints are performed through two parallel fully connected layers to generate a low-dimensional causal feature tensor and a high-dimensional hybrid feature tensor.

[0080] Specifically, the causal enhancement features are fed into the decoupled encoder, which contains two parallel fully connected layers. One fully connected layer performs dimensional compression and structural constraints on the causal enhancement features to form a low-dimensional causal feature tensor. The other fully connected layer performs high-dimensional expansion and distribution constraints on the causal enhancement features to form a high-dimensional hybrid feature tensor. The two fully connected layers operate independently in structure, and perform representation separation for causal related factors and non-causal interference factors, respectively.

[0081] It should be noted that the low-dimensional causal feature tensor refers to the feature subspace that is directly related to the internal structural defects of the wire rope (such as broken wires, corrosion, and diameter reduction) and remains unchanged under different working conditions; the high-dimensional hybrid feature tensor refers to the non-essential features that change due to external disturbances (such as vibration, temperature drift, sensor noise, and surface oil).

[0082] The global correlation measure between the low-dimensional causal feature tensor and the high-dimensional confounding feature tensor is calculated using a dynamically weighted multi-scale mutual information estimator, expressed as follows:

[0083] ;

[0084] In the formula, Represents a low-dimensional causal feature tensor With high-dimensional hybrid feature tensors Global correlation measure between This represents a low-dimensional causal feature tensor. This represents a high-dimensional hybrid feature tensor. Represents a low-dimensional causal feature tensor The total number of dimensions, Represents a low-dimensional causal feature tensor The Middle Each component index is used to traverse the low-dimensional feature tensor. Each component in, Represents a high-dimensional hybrid feature tensor The total number of dimensions, Represents a high-dimensional hybrid feature tensor The Middle Each component index is used to traverse high-dimensional features. Each component in, Indicates the first The component and the first Weighting coefficients between components Represents a low-dimensional causal feature tensor The Middle Each component and the high-dimensional hybrid feature tensor The Middle A measure of local correlation between components.

[0085] It should be noted that the first The component and the first The weighting coefficients between the components are based on the calculation of the low-dimensional causal feature tensor. and high-dimensional hybrid feature tensors The correlation value between them is used to determine the weighting coefficient. Through normalization, the weights are made to meet actual needs; the example value is 0.8, which is based on the correlation strength between features and their contribution to the target variable. Stronger correlations will be given larger weights in order to highlight the influence of important features.

[0086] It should be noted that, As a measure of local correlation, it is calculated using standardized covariance or correlation, and therefore is dimensionless. As a weighting coefficient, its function is to weight local correlations, and therefore it is also dimensionless. It is also dimensionless. It is a global correlation measure among all feature tensors, and therefore is dimensionless. The entire formula maintains dimensional consistency, ensuring that all calculation results are dimensionless.

[0087] By calling historical online detection results and evaluating the causal driving strength of the low-dimensional causal feature tensor based on the mapping relationship between historical online detection results and the low-dimensional causal feature tensor;

[0088] Specifically, historical online detection results are retrieved and matched with low-dimensional causal feature tensors. Based on the consistency between the two in multiple detection records, the stability of the low-dimensional causal feature tensor's response to the internal defect state of the wire rope is determined, thus forming an assessment of the causal driving strength of the low-dimensional causal feature tensor.

[0089] It should be noted that the historical online detection results are derived from labeled datasets that have been manually reviewed or verified by offline high-precision detection equipment, ensuring that they are consistent with the actual defect status and used to supervise the effectiveness evaluation of causal features.

[0090] We weight and fuse the global relevance measure and the causal driving strength to generate a decoupling loss value.

[0091] Specifically, the global correlation measure and the causal driving strength are assigned different weights according to their respective impact on the feature decoupling target. The information of the two is integrated by weighted combination to form a decoupling loss value used to measure the separation effect of low-dimensional causal feature tensors and high-dimensional mixed feature tensors.

[0092] The adversarial generation module combines a high-dimensional mixed feature tensor with a preset random noise vector, inputs it into the generator network, and outputs adversarial feature samples.

[0093] By performing multi-scale spectral analysis on the high-dimensional hybrid feature tensor, the energy distribution feature vector is extracted, and the preset random noise vector is mapped into a conditional vector through a conditional coding network.

[0094] Specifically, by performing spectral decomposition on the high-dimensional hybrid feature tensor at different scales, the energy concentration of each frequency band is observed, forming an energy distribution feature vector that reflects its frequency domain energy distribution characteristics; at the same time, a preset random noise vector is fed into the conditional coding network, which transforms the preset random noise vector according to the current training state to obtain a conditional vector related to the generation task.

[0095] It should be noted that the preset random noise vector is an initial vector used to introduce randomness to enhance sample diversity during the adversarial generation process. The setting of the preset random noise vector is based on the standard random sampling method. The specific setting steps are as follows: determine that the noise dimension is consistent with the input requirements of the generator network; select a random number generation method that follows a standard normal distribution or a uniform distribution; generate a set of independent and identically distributed random values ​​according to the selected distribution and dimension to form the preset random noise vector.

[0096] A dynamic weight matrix is ​​generated based on the energy distribution feature vector and condition vector;

[0097] Specifically, by concatenating the energy distribution feature vector and the condition vector along the feature dimension or interacting element-wise, a joint representation is formed. Then, the joint representation is mapped through a multilayer perceptual structure to obtain a set of weight parameters that matches the feature map size and number of channels of the generator network backbone encoder. The set of weight parameters is the dynamic weight matrix.

[0098] The high-dimensional hybrid feature tensor and the preset random noise vector are input into the backbone encoder of the generator network to obtain the initial feature map.

[0099] Specifically, the high-dimensional mixed feature tensor and the preset random noise vector are fed into the backbone encoder of the generator network. The backbone encoder abstracts the joint representation of the two through multi-layer convolution or fully connected structures, and incorporates the random perturbation of the preset random noise vector while retaining the structural information of the high-dimensional mixed feature tensor, thus forming the initial feature map.

[0100] It should be noted that the high-dimensional hybrid feature tensor structure information refers to the high-dimensional representation form and its internal dimensional organization relationship that are retained in the high-dimensional hybrid feature tensor after the magnetic flux leakage signal data, ultrasonic echo signal data and surface visual image data are decoupled by the causal representation learning network and are related to non-defect factors such as environmental interference and load fluctuation.

[0101] The initial feature map is conditionally spectrally modulated using a dynamic weight matrix to generate a modulated feature map. The modulated feature map is then input into the decoder of the generator network to output adversarial feature samples.

[0102] Specifically, the initial feature map is adjusted in the frequency domain by channel-by-channel or position-by-position using a dynamic weight matrix, so that the spectral components of the initial feature map change in a targeted manner according to the weights determined by the energy distribution feature vector and the condition vector, forming a modulated feature map. The modulated feature map is then fed into the decoder of the generator network, and the decoder gradually restores the spatial resolution through upsampling or deconvolution structures to generate adversarial feature samples.

[0103] The parameter update module inputs adversarial feature samples and low-dimensional causal feature tensors into the discriminator network and outputs the difference in discrimination results. Based on the difference in discrimination results, it calculates the adversarial loss value and updates the parameters of the generator network and the discriminator network together with the adversarial loss value and the decoupling loss value.

[0104] The discriminator network uses an interactive attention mechanism built into it to fuse adversarial feature samples with low-dimensional causal feature tensors to generate interactive enhanced features.

[0105] Specifically, through the interactive attention mechanism built into the discriminator network, adversarial feature samples and low-dimensional causal feature tensors are weighted in terms of importance in the channel dimension or spatial location, so that each feature dimension or location receives different weights according to its relevance to the final detection result, and then the two types of features are combined to form interactive enhancement features.

[0106] It should be noted that the final detection result refers to the judgment output generated by the discriminator network based on adversarial feature samples and low-dimensional causal feature tensors, which is used to characterize whether there are defects and their state inside the wire rope.

[0107] The processing path is dynamically selected based on the local information entropy and global variance of the interaction enhancement features. The detailed features and context features in the interaction enhancement features are differentiated to generate path modulation features.

[0108] Specifically, the local information entropy and global variance of the interaction enhancement features are standardized to obtain standardized local information entropy and standardized global variance. The standardized local information entropy is compared with a first threshold, and the standardized global variance is compared with a second threshold. If the standardized local information entropy is greater than the first threshold and the standardized global variance is less than the second threshold, the detail enhancement path is used to operate on the interaction enhancement features. In other cases, the context aggregation path is used to operate on the interaction enhancement features. The differential operation of detail features and context features in the interaction enhancement features is completed through the detail enhancement path or the context aggregation path to form path modulation features.

[0109] It should be noted that the first and second thresholds are set based on the statistical distribution of the local information entropy and global variance corresponding to the interactive enhancement features in historical detection data. The specific setting steps are as follows: collect a batch of interactive enhancement feature samples of wire ropes in defect-free and defective states, verified by manual review; calculate the local information entropy and global variance of each sample; independently sort the local information entropy sequence and global variance sequence and determine their empirical distribution; set the first threshold to the median of the local information entropy sequence or select the local information entropy value that achieves the highest detection accuracy using a grid search method on the validation set; set the second threshold to the median of the global variance sequence or the optimal global variance value determined using the same grid search method; the exemplary range of the first threshold is 0.45 to 0.65, and the exemplary range of the second threshold is... The example value ranges from 0.30 to 0.50. The basis for this value is that the selection of detail enhancement path and context aggregation path can effectively match the feature distribution of the actual defect mode and interference mode of the wire rope, so that the path modulation feature retains key details and maintains structural consistency. If the first threshold is below 0.45, the detail enhancement path will be triggered by a slight fluctuation in local information entropy, resulting in sensitivity to noise and an increase in misjudgments. If the first threshold is above 0.65, it is difficult to activate the detail enhancement path, resulting in missed detection of minor defects. If the second threshold is below 0.30, the detail path will be suppressed by a slight change in global variance, resulting in an excessive bias towards context aggregation and a weakening of the ability to respond to local anomalies. If the second threshold is above 0.50, the context aggregation path is difficult to activate, the overall structural information integration is insufficient, and the stability of long-range defect mode recognition is reduced.

[0110] Multi-scale gradient statistical features of path modulation are extracted, and attention-weighted multivariate nonlinear mapping is performed on the multi-scale gradient statistical features to generate discrimination result differences.

[0111] Specifically, gradients are extracted from path modulation features at multiple scales to obtain gradient distribution patterns at different resolutions, forming multi-scale gradient statistical features. These multi-scale gradient statistical features are then fed into an attention mechanism, which assigns different weights based on the importance of each scale gradient to the discrimination task. Through a multivariate nonlinear mapping structure, the weighted multi-scale gradient statistical features are transformed into discrimination result differences that reflect the degree of inconsistency between adversarial feature samples and low-dimensional causal feature tensors.

[0112] Multi-scale temporal decomposition of the discrimination results is performed to extract the difference components at different time scales;

[0113] Specifically, a multi-scale decomposition method is applied to the differences in the discrimination results along the time dimension, dividing the differences in the discrimination results into multiple components that reflect short-term fluctuations (the most recent 10 sampling points (about 0.1 seconds)), medium-term changes (the most recent 100 sampling points (about 1 second)), and long-term trends (the most recent 1000 sampling points (about 10 seconds)). Each component corresponds to a difference component at a time scale, thereby obtaining the difference components at different time scales.

[0114] Based on the variance of each difference component and the covariance of the difference between each difference component and the discrimination result, the weighting coefficient is calculated, and the expression is as follows:

[0115] ;

[0116] In the formula, Indicates the first The weighting coefficients of the difference components. Indicates the first The variance of each difference component, This represents the adjustment coefficient. Indicates the first The difference between the individual differential components and the discrimination result Covariance between This represents the total number of differential components. This represents the index of the difference component during the summation process, used to iterate through all difference components. This indicates that during the summation process, the traversal of the first... The variance of each difference component, This indicates that during the summation process, the traversal of the first... The difference between the individual differential components and the discrimination result Covariance between This represents extremely small positive numbers to prevent the denominator from being 0.

[0117] It should be noted that, , , and For quantities of the same dimensions, the adjustment coefficient It is dimensionless. Since all variances and covariances are on the same dimension, the addition and summation operations in the formula are valid, ensuring the uniformity of dimensions. The weighting coefficients also maintain a consistent dimension.

[0118] It should be noted that the adjustment coefficient The design is based on a trade-off of the contributions of different factors. The specific steps are as follows: First, identify the two factors that need to be balanced, such as the impact of variance and covariance on the final result; second, evaluate the importance of these two factors through data analysis, for example, a larger variance may require more attention, or the covariance may have a stronger relationship with the target, thus increasing its weight; third, conduct experiments on different... Based on the test results, select the optimal weighting ratio to improve overall performance; adjust according to the experimental results. This allows the final result to find the optimal balance among different features or factors; the example value is... The values ​​are determined through experiments, with the aim of finding the optimal weight allocation to achieve the most ideal overall effect.

[0119] By integrating and fusing the differential components using weighted coefficients, adversarial loss values ​​are generated.

[0120] Specifically, weighting coefficients are used to weight and accumulate each difference component over time. The difference components at different time scales are integrated according to their corresponding importance ratios to form a scalar value that comprehensively reflects the degree of discrimination bias of the discriminator. The scalar value is the adversarial loss value.

[0121] The adversarial loss value and the decoupling loss value are input into a dual-channel weight allocation network to extract the dynamic features of each loss value, and a dynamic fusion weight is generated using a cross-attention mechanism.

[0122] Specifically, the adversarial loss value and the decoupling loss value are fed into two independent channels of the dual-channel weight allocation network. The dual-channel weight allocation network extracts dynamic features reflecting the stability of generative adversarial processes from the adversarial loss value and dynamic features reflecting the degree of separation between causal and confounding features from the decoupling loss value. These two sets of dynamic features are fed into the cross-attention mechanism, which measures the importance of each other to the overall optimization goal through mutual query and response, forming dynamic fusion weights.

[0123] The adversarial loss value and the decoupling loss value are weighted and fused into a total loss function based on dynamic fusion weights, and the parameters of the generator network and the discriminator network are updated synchronously through the gradient descent algorithm.

[0124] Specifically, based on dynamic fusion weights, adversarial loss values ​​and decoupling loss values ​​are linearly combined according to their respective weight ratios to form a total loss function. The gradient descent algorithm is adopted, with the total loss function as the optimization objective, and the parameters of each layer in the generator network and the discriminator network are adjusted at the same time, so that the two networks gradually improve their ability to characterize and distinguish internal defects of wire ropes under the joint constraints of adversarial and decoupling.

[0125] The online detection module continuously collects new signal data streams and inputs them into the updated discriminator network, outputting an online detection report of internal defects in the wire rope.

[0126] New signal data streams will be continuously collected and input into the updated discriminator network, outputting defect detection results. The defect detection results will be verified in a multi-scale time series and real-time confidence assessment results will be generated.

[0127] Specifically, the continuously collected new signal data stream is fed into the updated discriminator network. The discriminator network generates defect detection results based on the learned causal features and discrimination capabilities. The consistency of the defect detection results is tested at multiple time scales, and its stability performance in short-term, medium-term and long-term windows is observed. The reliability of the current judgment is measured based on the consistency of the results at each scale, and the measurement result is generated. The measurement result is the real-time confidence assessment result.

[0128] Based on the theory of multi-source evidence fusion, the real-time confidence assessment results are synthesized to generate a comprehensive confidence distribution.

[0129] Specifically, based on multi-source evidence fusion theory (such as Dempster-Shafer evidence theory), real-time confidence assessment results from different time scales and different sensor modalities are regarded as independent evidence sources. Trust weights are assigned according to the reliability of each independent evidence source, and the real-time confidence assessment results are integrated through evidence synthesis rules (such as Dempster synthesis rules) to form a comprehensive confidence distribution covering multiple possibilities.

[0130] It should be noted that multi-source evidence fusion theory is a theoretical method used to integrate uncertain evidence from multiple information sources to form a unified and credible judgment. Its function is to collaboratively synthesize real-time confidence assessment results from different sources or scales to generate a more robust and comprehensive integrated confidence distribution.

[0131] Evidence synthesis rules are mathematical mechanisms in multi-source evidence fusion theory used to coordinate and merge the confidence levels of multiple evidence bodies. Their function is to fuse real-time confidence assessment results corresponding to different time scales or sensor modalities according to their reliability, and generate a consistent and comprehensive integrated confidence distribution.

[0132] A structured detection report is generated by fusing multimodal information from the comprehensive confidence distribution and the low-dimensional causal feature tensor through a joint decision-making mechanism.

[0133] Specifically, the joint decision-making mechanism compares the comprehensive confidence distribution with the low-dimensional causal feature tensor. When the confidence of the defect category in the comprehensive confidence distribution is higher than the preset high confidence threshold, and the discrimination response intensity of the low-dimensional causal feature tensor on the corresponding defect category is higher than the preset response threshold, the current state of the wire rope is determined to be a high confidence defect state. A structured inspection report containing defect type, spatial location, severity level and confidence level is generated by combining the temporal location information.

[0134] It should be noted that the high confidence threshold is set based on the consistency statistics between the comprehensive confidence distribution and the verification results in historical inspections. The setting steps are as follows: collect a batch of verified wire rope defect samples and their corresponding comprehensive confidence distributions; statistically analyze the range of comprehensive confidence values ​​under different defect categories; and select the confidence value that achieves the optimal balance between the true positive rate and the false positive rate as the high confidence threshold. An exemplary value range is 0.85 to 0.95. The basis for this value is that the comprehensive confidence distribution can effectively distinguish between real defects and interference signals, ensuring the reliability of the structured inspection report. If the high confidence threshold is lower than 0.85, too many low-confidence results will be included in the high confidence judgment, leading to an increase in false alarms and a decrease in report confidence. If the high confidence threshold is higher than 0.95, only a very small number of strong responses will be accepted, resulting in a large number of real defects being missed and reducing detection sensitivity.

[0135] The response threshold is set based on the distribution of the discriminative response intensity of a low-dimensional causal feature tensor under known defect states. The setting steps are as follows: using historical data labeled with real defect types, obtain the discriminative response values ​​of the low-dimensional causal feature tensor on various defects, plot the empirical cumulative distribution curve of the response intensity, and select the response intensity value that can cover the main defect samples and exclude most normal samples as the response threshold. The exemplary value range is 0.80 to 0.90. The basis for the value is that the low-dimensional causal feature tensor has sufficient discriminative power to represent the essential features of defects and can effectively support high-confidence judgment. If the response threshold is lower than 0.80, normal states or weak interference may trigger defect judgment, increasing false alarms. If the response threshold is higher than 0.90, it only responds to extremely significant defects, weakening the detection ability of early or minor defects and affecting practicality.

[0136] It should be noted that the joint decision-making mechanism is a decision-making method that integrates multimodal and multidimensional information to form a unified judgment result. Its function is to collaboratively integrate the comprehensive confidence distribution and the low-dimensional causal feature tensor to generate a structured detection report that is complete in content, logically consistent, and interpretable.

[0137] In summary, this invention achieves accurate extraction of essential defect features and effective isolation of interference signals by constructing a causal representation learning network and decomposing the features after multi-sensor fusion; by using a decoupling loss function to constrain the feature separation process, it ensures that causal features are only related to the essential factors of internal defects in the wire rope, while hybrid features carry non-critical information such as environmental interference, thus improving the purity and interpretability of feature representation. This allows the discriminator network to focus on the essential representation of defects, effectively suppressing misjudgments caused by fluctuations in operating conditions. At the same time, by quantifying the degree of separation between causal features and hybrid features, it provides clear guidance for model optimization, enhancing stability and adaptability in complex industrial scenarios.

[0138] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An online detection system for internal defects in steel wire ropes based on AI multi-sensor fusion, characterized in that: include, The data preprocessing module simultaneously collects signal data streams from multiple heterogeneous sensors during the online operation of the wire rope and performs preprocessing. The feature decoupling module constructs a causal representation learning network, inputs the preprocessed signal data stream into the causal representation learning network, performs fusion and decoupling decomposition, generates a low-dimensional causal feature tensor and a high-dimensional mixed feature tensor, and calculates the decoupling loss value based on the degree of separation between the low-dimensional causal feature tensor and the high-dimensional mixed feature tensor. The adversarial generation module combines a high-dimensional hybrid feature tensor with a preset random noise vector, inputs it into the generator network, and outputs adversarial feature samples. The parameter update module inputs adversarial feature samples and low-dimensional causal feature tensors into the discriminator network and outputs the difference in discrimination results. It calculates the adversarial loss value based on the difference in discrimination results and updates the parameters of the generator network and the discriminator network together with the adversarial loss value and the decoupling loss value. The online detection module continuously collects new signal data streams and inputs them into the updated discriminator network, outputting an online detection report of internal defects in the wire rope.

2. The online detection system for internal defects of steel wire rope based on AI multi-sensor fusion as described in claim 1, characterized in that: The plurality of heterogeneous sensors include magnetic sensors, acoustic sensors and optical sensors; The signal data stream includes magnetic flux leakage signal data, ultrasonic echo signal data, and surface visual image data; The preprocessing includes spatiotemporal registration, noise reduction filtering, and amplitude normalization.

3. The online detection system for internal defects of steel wire rope based on AI multi-sensor fusion as described in claim 2, characterized in that: The steps for constructing the causal representation learning network are as follows: A one-dimensional convolutional neural network is used to build a temporal feature encoding layer, a two-dimensional convolutional neural network is used to build a spatial feature encoding layer, and a feature fusion and decoupling layer is built through a cross-modal attention mechanism; The temporal feature encoding layer, spatial feature encoding layer, and feature fusion and decoupling layer are processed and fused in parallel to generate a causal representation learning network.

4. The online detection system for internal defects of steel wire rope based on AI multi-sensor fusion as described in claim 3, characterized in that: The steps for generating the low-dimensional causal feature tensor and the high-dimensional mixed feature tensor are as follows. The preprocessed signal data stream is input into the temporal feature coding layer and the spatial feature coding layer to extract temporal and spatial features. Temporal and spatial features are input features that are fused and decoupled to perform cross-modal attention fusion, generating preliminary fused features; Based on the preliminary fusion features, the importance scores of each feature dimension in the preliminary fusion features are calculated, and the preliminary fusion features are weighted according to the importance scores to obtain the causal enhancement features; The causal enhancement features are input into the decoupled encoder, and dimension mapping and regularization constraints are performed through two parallel fully connected layers to generate a low-dimensional causal feature tensor and a high-dimensional hybrid feature tensor.

5. The online detection system for internal defects of steel wire rope based on AI multi-sensor fusion as described in claim 4, characterized in that: The steps for calculating the decoupling loss are as follows: A global correlation measure between low-dimensional causal feature tensors and high-dimensional mixed feature tensors is calculated using a dynamic weighted multi-scale mutual information estimator. By calling historical online detection results and evaluating the causal driving strength of the low-dimensional causal feature tensor based on the mapping relationship between historical online detection results and the low-dimensional causal feature tensor; We weight and fuse the global relevance measure and the causal driving strength to generate a decoupling loss value.

6. The online detection system for internal defects of steel wire rope based on AI multi-sensor fusion as described in claim 5, characterized in that: The steps for outputting adversarial feature samples are as follows: By performing multi-scale spectral analysis on the high-dimensional hybrid feature tensor, the energy distribution feature vector is extracted, and the preset random noise vector is mapped into a conditional vector through a conditional coding network. A dynamic weight matrix is ​​generated based on the energy distribution feature vector and condition vector; The high-dimensional hybrid feature tensor and the preset random noise vector are input into the backbone encoder of the generator network to obtain the initial feature map. The initial feature map is conditionally spectrally modulated using a dynamic weight matrix to generate a modulated feature map. The modulated feature map is then input into the decoder of the generator network to output adversarial feature samples.

7. The online detection system for internal defects of steel wire rope based on AI multi-sensor fusion as described in claim 6, characterized in that: The steps for determining the difference in output discrimination results are as follows: The discriminator network uses an interactive attention mechanism built into it to fuse adversarial feature samples with low-dimensional causal feature tensors to generate interactive enhanced features. The processing path is dynamically selected based on the local information entropy and global variance of the interaction enhancement features. The detailed features and context features in the interaction enhancement features are differentiated to generate path modulation features. Multi-scale gradient statistical features of path modulation are extracted, and attention-weighted multivariate nonlinear mapping is performed on the multi-scale gradient statistical features to generate discrimination result differences.

8. The online detection system for internal defects of steel wire rope based on AI multi-sensor fusion as described in claim 7, characterized in that: The steps for calculating the adversarial loss value based on the difference in discrimination results are as follows: Multi-scale temporal decomposition of the discrimination results is performed to extract the difference components at different time scales; The weighting coefficients are calculated based on the variance of each difference component and the covariance of the difference between each difference component and the discrimination result. By integrating and fusing the differential components using weighted coefficients, adversarial loss values ​​are generated.

9. The online detection system for internal defects of steel wire rope based on AI multi-sensor fusion as described in claim 8, characterized in that: The steps for updating the parameters of the generator network and the discriminator network by jointly using adversarial loss and decoupling loss are as follows: The adversarial loss value and the decoupling loss value are input into a dual-channel weight allocation network to extract the dynamic features of each loss value, and a dynamic fusion weight is generated using a cross-attention mechanism. The adversarial loss value and the decoupling loss value are weighted and fused into a total loss function based on dynamic fusion weights, and the parameters of the generator network and the discriminator network are updated synchronously through the gradient descent algorithm.

10. The online detection system for internal defects of steel wire rope based on AI multi-sensor fusion as described in claim 9, characterized in that: The online detection report for internal defects in the output wire rope is generated through the following steps. New signal data streams will be continuously collected and input into the updated discriminator network, outputting defect detection results. The defect detection results will be verified in a multi-scale time series and real-time confidence assessment results will be generated. Based on the theory of multi-source evidence fusion, the real-time confidence assessment results are synthesized to generate a comprehensive confidence distribution. A structured detection report is generated by fusing multimodal information from the comprehensive confidence distribution and the low-dimensional causal feature tensor through a joint decision-making mechanism.

Images