Steel wire rope broken wire detection method and system based on double-branch cross-level Mamba network

Abstract

The application discloses a steel wire rope broken wire detection method and system based on a double-branch cross-level Mamba network. The method comprises the following steps: collecting a magnetic flux leakage signal; performing adaptive sliding window and standardization on the magnetic flux leakage signal based on the lay length of the steel wire rope; constructing and training a steel wire rope broken wire magnetic flux leakage quantitative detection network, including a double-branch feature extraction backbone network, a cross-level one-dimensional Mamba enhancement module and a discrete broken wire quantitative mapping module connected in sequence; and performing steel wire rope broken wire detection based on the detection network. The double-branch feature extraction backbone network extracts local mutation features and global trend features respectively; the cross-level one-dimensional Mamba enhancement module adaptively truncates the transmission of baseline distortion; and the discrete broken wire quantitative mapping module aggregates and maps the global extreme value into the number of broken wires. The application breaks through the limitation of traditional networks that are prone to distortion in the face of strong lifting interference, and uses mathematical parameter mapping and physical state updating mechanism to strip noise, thereby adapting to complex industrial online detection environments.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of electromagnetic nondestructive testing data processing, sequence modeling and industrial artificial intelligence applications, and specifically relates to a method and system for detecting broken wires in steel wire ropes based on a dual-branch cross-level Mamba network. Background Technology

[0002] Mining wire ropes are the "lifeline" of coal mine hoisting systems. Under harsh working conditions and long-term service, these ropes are highly susceptible to fatigue breakage. Magnetic flux leakage (MFL) testing, due to its non-contact and highly sensitive characteristics, has become the most widely used non-destructive testing method for wire ropes.

[0003] However, in the practical implementation and engineering application of advanced online non-destructive testing systems for single-rope hoists' wire ropes, quantitative magnetic flux leakage detection faces extreme lift-off interference. The system requires stable detection within a high-speed operating range of 0-10 m / s, targeting an effective extraction distance of 0-10 mm on one side. However, under harsh mine conditions, the severe lateral vibration of the wire rope causes a random dynamic lift-off abrupt change of 0-10 mm between the probe and the rope surface. This physical disturbance causes a significant change in spatial magnetic reluctance, inducing a low-frequency background baseline drift with extremely high amplitude and a wide frequency band in the sensor. Simultaneously, on-site quantitative magnetic flux leakage detection requires the ability to accurately distinguish the evolution of damage levels across a wide range, from single broken wires to severely broken wires. Under these extreme conditions, the high-frequency distortion spikes in magnetic flux leakage caused by minor wire breaks are extremely weak and easily swallowed up by the massive lift-off baseline drift. Traditional methods relying on multi-channel acquisition and superposition, cascaded denoising, and conventional deep learning methods struggle to achieve accurate physical-layer decoupling of weak signals, leading to a sharp drop in the accuracy of quantitative detection of minor wire breaks.

[0004] With the development of deep learning, one-dimensional convolutional neural networks (1D-CNN) and Transformer-based attention mechanisms have been introduced into signal processing. However, existing deep networks have significant limitations when applied to magnetic flux leakage signals: First, traditional CNN models have poor global perception. CNNs rely on sliding windows of fixed size to extract features, making it difficult to capture the global patterns of slowly changing, long-period baseline drift, leading to feature entanglement. Second, traditional self-attention mechanisms are prone to distortion. Softmax-based attention mechanisms are extremely sensitive to large-amplitude signals. When encountering severe lift-off noise, the attention weights are forcibly dominated by the noise, suppressing the true small defect peaks. Third, the computational complexity is high. The Transformer architecture has a quadratic computational complexity with respect to the sequence length, greatly hindering its real-time deployment in portable detectors and lower-level hardware. Summary of the Invention

[0005] To address the shortcomings of existing technologies, one of the objectives of this invention is to provide a wire rope broken wire detection method based on a dual-branch, multi-level Mamba network. By designing a dual-branch, multi-scale backbone and a multi-level Mamba enhancement module, high-precision, real-time quantitative detection of minor wire rope broken wires can be achieved.

[0006] The second objective of this invention is to provide a system for implementing the wire rope breakage detection method based on a dual-branch cross-level Mamba network.

[0007] This invention provides a method for detecting broken wires in steel wire ropes based on a dual-branch, multi-level Mamba network, comprising the following steps:

[0008] Original magnetic flux leakage sequence signals of steel wire rope under different damage states and dynamic lifting disturbance conditions were collected, and a multi-channel one-dimensional magnetic flux leakage sequence dataset was constructed and labeled.

[0009] Based on the prior physical parameters of the steel wire rope, the collected one-dimensional magnetic flux leakage sequence dataset is preprocessed by adaptive sliding window truncation and physical dimension alignment to construct a one-dimensional magnetic flux leakage tensor containing the complete spatial magnetic field jump cycle.

[0010] An initial wire rope broken wire leakage magnetic flux quantitative detection network is constructed. This network comprises a bi-branch feature extraction backbone network, a multi-level one-dimensional Mamba enhancement module, and a discrete broken wire quantitative mapping module, connected sequentially. The network input is a one-dimensional leakage magnetic flux tensor. The input feature data is first processed by the bi-branch feature extraction backbone network to extract high-dimensional local features and global background features. Then, the multi-level one-dimensional Mamba enhancement module adaptively filters out baseline drift to obtain a broken wire enhanced feature sequence. Finally, the discrete broken wire quantitative mapping module performs spatial information compression and nonlinear mapping on the broken wire enhanced feature sequence to obtain the probability distribution of the number of broken wires as the network output.

[0011] A joint ordered regression loss function is introduced to perform end-to-end iterative training on the constructed initial wire rope broken wire leakage magnetic flux quantitative detection network to obtain the wire rope broken wire leakage magnetic flux quantitative detection network.

[0012] A magnetic flux leakage detection network for broken wires in steel wire ropes is loaded, and the magnetic flux leakage signals collected in real time are inferred to obtain the precise number of broken wires at the current position of the steel wire rope.

[0013] The original magnetic flux leakage sequence signal was collected using a ring magnetizer and a multi-channel Hall sensor array under different rope speeds and different shaking amplitudes of mining steel wire ropes; then, the actual number of broken wires was used for labeling to obtain a multi-channel one-dimensional magnetic flux leakage sequence dataset.

[0014] The one-dimensional leakage magnetic tensor is obtained as follows: a fixed-length physical sliding window is used based on the nominal lay length parameter of the wire rope to be tested, so that each set slice window contains a complete spatial magnetization cycle; then, based on the median, the amplitude of the channel is scaled independently using Z-score to obtain the one-dimensional leakage magnetic tensor.

[0015] The dual-branch feature extraction backbone network includes parallel local mutation branches and global trend branches;

[0016] The local mutation branch uses cascaded one-dimensional convolutional modules with a preset receptive field size, and the processing is represented by the following formula: The global trend branch uses a pre-defined receptive field or a dilated one-dimensional convolutional module, and the receptive field of the global trend branch is larger than that of the local mutation branch; the processing of the local mutation branch is represented by the following formula: ;in, The input is a one-dimensional leakage flux tensor; combined with the physical characteristics of wire rope leakage flux, This indicates a high-resolution convolution with a small kernel, whose receptive field physical span corresponds to the microscale spatial leakage magnetic distortion field generated by a single broken wire of a steel wire rope, aiming to accurately extract high-frequency extreme peaks at the microsecond / millimeter level. This indicates a low-resolution convolution with a large kernel or dilation rate, whose receptive field physical span corresponds to macroscopic lift-off changes of tens of millimeters. It aims to filter out local abrupt changes and simply fit the wide low-frequency envelope caused by the low-frequency mechanical vibration and slow change of spatial magnetoresistance of the probe during high-speed sports car operation. This represents the SiLU activation function; BN is the batch normalization operation.

[0017] The operational logic of the cross-level one-dimensional Mamba enhancement module includes:

[0018] Features output by the global trend branch After linear projection, a discretized step size is generated that depends on the current low-frequency lift-off noise fluctuation. The state input matrix B and the observation matrix C are represented by the following formula: ; ; ;in, The activation function for smooth rectification is defined as follows: ,default ; For linear projection, and They have the same projection structure but learn their parameters independently, each playing a different role as state input and observation mapping;

[0019] Calculate the discretized state transition parameters using the zero-order preserve method and input parameters It can be expressed using the following formula: ; Where I is the identity matrix, Given the dimension and state transition matrix;

[0020] Using the features output by the local mutation branch as the input sequence, the state transition parameters after discretization... and input parameters Selective scanning and updating of features driven by the process is expressed by the following formula: ; ;in, In hidden state, Output the clean features for the current time step;

[0021] The clean feature outputs from different time steps are integrated in temporal order to obtain a clean feature sequence.

[0022] The multi-level one-dimensional Mamba enhancement module utilizes low-frequency features that characterize large-scale flux envelope drift. To cross-level dominant step size The generation of; when encountering a violent lift-off fluctuation that causes a sudden change in the magnetic field baseline, due to The low-frequency peaks containing large amplitudes are forced by linear mapping to adjust the step size. The exponential term tends towards a maximum or minimum adjustment, thereby causing the exponential term to... A sharp decline occurs, approaching zero; at this point, the mathematical model is equivalent to triggering a gated forgetting mechanism. The global low-frequency magnetic flux accumulation memory caused by the lift-off is adaptively truncated and canceled from the integral space, ensuring that the hidden state is updated only with the local high-frequency leakage magnetic distortion caused by the actual wire breakage. Its input term is .

[0023] The discrete broken wire quantitative mapping module includes an adaptive global extreme value pooling layer and a multilayer perceptron classification head. The discrete broken wire quantitative mapping module takes the pure feature sequence output by the cross-level one-dimensional Mamba enhancement module as input, extracts the global maximum extreme value of the sequence in the adaptive global extreme value pooling layer and flattens it, and then inputs the result into the multilayer perceptron classification head to obtain the raw score of the probability of broken wire number.

[0024] Furthermore, the initial wire rope broken wire leakage magnetic flux quantitative detection network also includes a dynamic fusion gating module; the dynamic fusion gating module will output the results of local abrupt branches. After linear mapping, it is compared with the global trend branch output. The linear mapping results are fused, and the fusion weights are generated using the Sigmoid activation function, expressed by the following formula: ; For fusion weights; For local linear projection layers, The global linear projection layer is used, while the local linear projection layer does not share parameters with the global linear projection layer. Local features and global features are mapped to the same fused feature space, respectively. For the Sigmoid function; For the Sigmoid function;

[0025] Then, based on the fusion weights, the output of the local mutation branch is... Global trend branch output Fusion yields fusion characteristics , will integrate features The output of the replacement local mutation branch , and the output of the global trend branch As input to the cross-level one-dimensional Mamba enhancement module; the fusion feature Express it using the following formula: .

[0026] Furthermore, based on the idea of ​​cross-layer connections in U-Net, the output features of different convolutional blocks in the dual-branch feature extraction backbone network are directly passed to the output of the cross-level one-dimensional Mamba enhancement module through skip connections, performing multi-scale cross-level interaction to obtain the final output. As input to the discrete broken wire quantitative mapping module, it is represented by the following formula: ;in, The clean feature sequence output by the cross-level one-dimensional Mamba enhancement module; This represents the output features of the first convolutional block in the local mutation branch; The output features of the i-th convolutional block in the local mutation branch; This refers to the output features of the first convolutional block in the global trend branch; This represents the output feature of the j-th convolutional block in the global trend branch.

[0027] The joint ordered regression loss function comprises two parts: cross-entropy classification loss and mathematical expectation regression loss, specifically:

[0028] Calculate the classification loss of the network output using the cross-entropy loss function. For samples with a batch size of N, it is defined as: ;in For the first The true label of each sample For the network prediction of the first The probability of a class;

[0029] Perform a Softmax operation on the raw scores output by the network to generate a probability distribution sequence. The expected number of broken wires predicted by the network is calculated by summing the dot product of this number and the increasing tensor k of the broken wire category. It can be expressed using the following formula: ;

[0030] Calculate the expected number of broken wires Compared with the actual broken wire label value The mean squared error is used as the expected regression loss. It can be expressed using the following formula: ;

[0031] The classification loss and expected regression loss are weighted and summed according to their respective weight factors to obtain the joint ordered regression loss function, expressed by the following formula: ;in The weighting coefficients for balancing the two types of losses (within the range of 0 to 1, and in this example, a value of 1 is used) ).

[0032] The present invention also provides a system for implementing the wire rope broken wire detection method based on a dual-branch cross-level Mamba network, including a data acquisition module, a data preprocessing module, a network construction module, a network training module, and a broken wire count prediction module;

[0033] The data acquisition module collects the original magnetic flux leakage sequence signals of the wire rope under different damage states and dynamic lifting disturbance conditions, constructs and labels a multi-channel one-dimensional magnetic flux leakage sequence dataset, and uploads the data to the data preprocessing module.

[0034] The data preprocessing module, based on the prior physical parameters of the steel wire rope, performs adaptive sliding window truncation and physical dimension alignment preprocessing on the collected one-dimensional leakage magnetic field sequence dataset, constructs a one-dimensional leakage magnetic field tensor containing the complete spatial magnetic field jump cycle, and uploads the data to the network training module.

[0035] The network construction module builds the initial wire rope broken wire leakage magnetic flux quantitative detection network and uploads the data to the network training module;

[0036] The network training module introduces a joint ordered regression loss function to perform end-to-end iterative training on the constructed initial wire rope broken wire leakage magnetic flux quantitative detection network, thereby obtaining the wire rope broken wire leakage magnetic flux quantitative detection network, and uploading the data to the broken wire number prediction module;

[0037] The broken wire count prediction module is loaded with a wire rope broken wire leakage magnetic flux quantitative detection network. It infers the real-time collected leakage magnetic flux signal to obtain the accurate number of broken wires at the current position of the wire rope.

[0038] This invention discloses a method and system for detecting broken wires in steel wire ropes based on a dual-branch, multi-level Mamba network, which has the following beneficial effects:

[0039] (1) The network constructed in this invention is verified by an adaptive electromagnetic baseline cancellation mechanism. The dual-branch framework completes the physical isolation of the spatial frequency of electromagnetic signals in the initial stage of feature extraction through convolution operations with receptive fields of different sizes.

[0040] (2) The network constructed in this invention utilizes the pure low-frequency noise numerical form at the computational level to actively eliminate the integral accumulation of low-frequency lift-off baseline in the state machine; at the same time, it retains the broken wire sharp corner feature of the high-frequency branch input; the network structure realizes the reverse control of the baseline integral by using the form of lift-off magnetoresistive distortion from the bottom physical meaning of electromagnetics, thereby ensuring high sensitivity and accurate quantitative ability to the extreme point of broken wire under extreme 0-10mm dynamic lift-off interference.

[0041] (3) This invention not only retains extremely high high-frequency resolution at the feature extraction end, but also introduces a difficult sample weighting mechanism for error compensation at the loss function end. Compared with traditional regression analysis and conventional deep learning methods, it significantly improves the global recall and precision of quantitative judgment of minute defects. Attached Figure Description

[0042] Figure 1 This is a schematic flowchart of the method of the present invention;

[0043] Figure 2 This is a schematic diagram of the system structure of the present invention;

[0044] Figure 3 This is a schematic diagram of the dual-branch feature extraction backbone network structure in an embodiment of the present invention;

[0045] Figure 4 This is a comparison diagram of the cross-level one-dimensional Mamba enhancement module and the traditional self-dependent Mamba network in an embodiment of the present invention; where (a) is the traditional self-dependent Mamba module and (b) is the two-branch cross-level Mamba module;

[0046] Figure 5 This is a diagram illustrating the effect of decoupling the original signal through a dual-branch design in an embodiment of the present invention; where the sub- Figure 5 (a) is the raw leakage magnetic signal directly picked up by the sensor, sub Figure 5 (b) represents the low-frequency background field output by the global trend branch, sub- Figure 5 (c) is the high-frequency wire breakage signal output by the local mutation branch. Detailed Implementation

[0047] This invention provides a method for detecting broken wires in steel wire ropes based on a dual-branch, multi-level Mamba network, the flowchart of which is shown below. Figure 1 As shown, it includes the following steps:

[0048] Original magnetic flux leakage sequence signals of steel wire rope under different damage states and dynamic lifting disturbance conditions were collected, and a multi-channel one-dimensional magnetic flux leakage sequence dataset was constructed and labeled.

[0049] Based on the prior physical parameters of the steel wire rope, the collected one-dimensional magnetic flux leakage sequence dataset is preprocessed by adaptive sliding window truncation and physical dimension alignment to construct a one-dimensional magnetic flux leakage tensor containing the complete spatial magnetic field jump cycle.

[0050] An initial wire rope broken wire leakage magnetic flux quantitative detection network is constructed. This network comprises a bi-branch feature extraction backbone network, a multi-level one-dimensional Mamba enhancement module, and a discrete broken wire quantitative mapping module, connected sequentially. The network input is a one-dimensional leakage magnetic flux tensor. The input feature data is first processed by the bi-branch feature extraction backbone network to extract high-dimensional local features and global background features. Then, the multi-level one-dimensional Mamba enhancement module adaptively filters out baseline drift to obtain a broken wire enhanced feature sequence. Finally, the discrete broken wire quantitative mapping module performs spatial information compression and nonlinear mapping on the broken wire enhanced feature sequence to obtain the probability distribution of the number of broken wires as the network output.

[0051] A joint ordered regression loss function is introduced to perform end-to-end iterative training on the constructed initial wire rope broken wire leakage magnetic flux quantitative detection network to obtain the wire rope broken wire leakage magnetic flux quantitative detection network.

[0052] A magnetic flux leakage detection network for broken wires in steel wire ropes is loaded, and the magnetic flux leakage signals collected in real time are inferred to obtain the precise number of broken wires at the current position of the steel wire rope.

[0053] The original magnetic flux leakage sequence signal was collected using a ring magnetizer and a multi-channel Hall sensor array under different rope speeds and different shaking amplitudes of mining steel wire ropes; then, the actual number of broken wires was used for labeling to obtain a multi-channel one-dimensional magnetic flux leakage sequence dataset.

[0054] The one-dimensional leakage magnetic tensor is obtained as follows: a fixed-length physical sliding window is used based on the nominal lay length parameter of the wire rope to be tested, so that each set slice window contains a complete spatial magnetization cycle; then, based on the median, the amplitude of the channel is scaled independently using Z-score to obtain the one-dimensional leakage magnetic tensor.

[0055] The dual-branch feature extraction backbone network includes parallel local mutation branches and global trend branches;

[0056] The local mutation branch uses cascaded one-dimensional convolutional modules with a preset receptive field size, and the processing is represented by the following formula: The global trend branch uses a pre-defined receptive field or a dilated one-dimensional convolutional module, and the receptive field of the global trend branch is larger than that of the local mutation branch; the processing of the local mutation branch is represented by the following formula: ;in, The input is a one-dimensional leakage flux tensor; combined with the physical characteristics of wire rope leakage flux, This indicates a high-resolution convolution with a small kernel, whose receptive field physical span corresponds to the microscale spatial leakage magnetic distortion field generated by a single broken wire of a steel wire rope, aiming to accurately extract high-frequency extreme peaks at the microsecond / millimeter level. This indicates a low-resolution convolution with a large kernel or dilation rate, whose receptive field physical span corresponds to macroscopic lift-off changes of tens of millimeters. It aims to filter out local abrupt changes and simply fit the wide low-frequency envelope caused by the low-frequency mechanical vibration and slow change of spatial magnetoresistance of the probe during high-speed sports car operation. This represents the SiLU activation function; BN is the batch normalization operation.

[0057] The operational logic of the cross-level one-dimensional Mamba enhancement module includes:

[0058] Features output by the global trend branch After linear projection, a discretized step size is generated that depends on the current low-frequency lift-off noise fluctuation. The state input matrix B and the observation matrix C are represented by the following formula: ; ; ;in, The activation function for smooth rectification is defined as follows: ,default ; For linear projection, and They have the same projection structure but learn their parameters independently, each playing a different role as state input and observation mapping;

[0059] Calculate the discretized state transition parameters using the zero-order preserve method and input parameters It can be expressed using the following formula: ; Where I is the identity matrix, Given the dimension and state transition matrix;

[0060] Using the features output by the local mutation branch as the input sequence, the state transition parameters after discretization... and input parameters Selective scanning and updating of features driven by the process is expressed by the following formula: ; ;in, In hidden state, Output the clean features for the current time step;

[0061] The clean feature outputs from different time steps are integrated in temporal order to obtain a clean feature sequence.

[0062] The multi-level one-dimensional Mamba enhancement module utilizes low-frequency features that characterize large-scale flux envelope drift. To cross-level dominant step size The generation of; when encountering a violent lift-off fluctuation that causes a sudden change in the magnetic field baseline, due to The low-frequency peaks containing large amplitudes are forced by linear mapping to adjust the step size. The exponential term tends towards a maximum or minimum adjustment, thereby causing the exponential term to... A sharp decline occurs, approaching zero; at this point, the mathematical model is equivalent to triggering a gated forgetting mechanism. The global low-frequency magnetic flux accumulation memory caused by the lift-off is adaptively truncated and canceled from the integral space, ensuring that the hidden state is updated only with the local high-frequency leakage magnetic distortion caused by the actual wire breakage. Its input term is .

[0063] The discrete broken wire quantitative mapping module includes an adaptive global extreme value pooling layer and a multilayer perceptron classification head. The discrete broken wire quantitative mapping module takes the pure feature sequence output by the cross-level one-dimensional Mamba enhancement module as input, extracts the global maximum extreme value of the sequence in the adaptive global extreme value pooling layer and flattens it, and then inputs the result into the multilayer perceptron classification head to obtain the raw score of the probability of broken wire number.

[0064] Furthermore, the initial wire rope broken wire leakage magnetic flux quantitative detection network also includes a dynamic fusion gating module; the dynamic fusion gating module will output the results of local abrupt branches. After linear mapping, it is compared with the global trend branch output. The linear mapping results are fused, and the fusion weights are generated using the Sigmoid activation function, expressed by the following formula: ; For fusion weights; For local linear projection layers, The global linear projection layer is used, while the local linear projection layer does not share parameters with the global linear projection layer. Local features and global features are mapped to the same fused feature space, respectively. For the Sigmoid function;

[0065] Then, based on the fusion weights, the output of the local mutation branch is... Global trend branch output Fusion yields fusion characteristics , will integrate features The output of the replacement local mutation branch , and the output of the global trend branch As input to the cross-level one-dimensional Mamba enhancement module; the fusion feature Express it using the following formula: .

[0066] Furthermore, based on the idea of ​​cross-layer connections in U-Net, the output features of different convolutional blocks in the dual-branch feature extraction backbone network are directly passed to the output of the cross-level one-dimensional Mamba enhancement module through skip connections, performing multi-scale cross-level interaction to obtain the final output. As input to the discrete broken wire quantitative mapping module, it is represented by the following formula: ;in, The clean feature sequence output by the cross-level one-dimensional Mamba enhancement module; This represents the output features of the first convolutional block in the local mutation branch; The output features of the i-th convolutional block in the local mutation branch; This refers to the output features of the first convolutional block in the global trend branch; This represents the output feature of the j-th convolutional block in the global trend branch.

[0067] The joint ordered regression loss function comprises two parts: cross-entropy classification loss and mathematical expectation regression loss, specifically:

[0068] Calculate the classification loss of the network output using the cross-entropy loss function. For samples with a batch size of N, it is defined as: ;in For the first The true label of each sample For the network prediction of the first The probability of a class;

[0069] Perform a Softmax operation on the raw scores output by the network to generate a probability distribution sequence. The expected number of broken wires predicted by the network is calculated by summing the dot product of this number and the increasing tensor k of the broken wire category. It can be expressed using the following formula: ;

[0070] Calculate the expected number of broken wires Compared with the actual broken wire label value The mean squared error is used as the expected regression loss. It can be expressed using the following formula: ;

[0071] The classification loss and expected regression loss are weighted and summed according to their respective weight factors to obtain the joint ordered regression loss function, expressed by the following formula: ;in The weighting coefficients for balancing the two types of losses (within the range of 0 to 1, and in this example, a value of 1 is used) ).

[0072] This invention also provides a system for implementing the wire rope broken wire detection method based on a dual-branch cross-level Mamba network, the structural schematic of which is shown below. Figure 2 As shown, it includes a data acquisition module, a data preprocessing module, a network construction module, a network training module, and a broken wire count prediction module;

[0073] The data acquisition module collects the original magnetic flux leakage sequence signals of the wire rope under different damage states and dynamic lifting disturbance conditions, constructs and labels a multi-channel one-dimensional magnetic flux leakage sequence dataset, and uploads the data to the data preprocessing module.

[0074] The data preprocessing module, based on the prior physical parameters of the steel wire rope, performs adaptive sliding window truncation and physical dimension alignment preprocessing on the collected one-dimensional leakage magnetic field sequence dataset, constructs a one-dimensional leakage magnetic field tensor containing the complete spatial magnetic field jump cycle, and uploads the data to the network training module.

[0075] The network construction module builds the initial wire rope broken wire leakage magnetic flux quantitative detection network and uploads the data to the network training module;

[0076] The network training module introduces a joint ordered regression loss function to perform end-to-end iterative training on the constructed initial wire rope broken wire leakage magnetic flux quantitative detection network, thereby obtaining the wire rope broken wire leakage magnetic flux quantitative detection network, and uploading the data to the broken wire number prediction module;

[0077] The broken wire count prediction module is loaded with a wire rope broken wire leakage magnetic flux quantitative detection network. It infers the real-time collected leakage magnetic flux signal to obtain the accurate number of broken wires at the current position of the wire rope.

[0078] The method of the present invention will be further described below with reference to an embodiment:

[0079] A portable wire rope magnetic flux leakage detector based on an ARM or DSP architecture was constructed, equipped with a ring magnetizer and a multi-channel Hall sensor array. Raw magnetic flux leakage sequences of mining wire ropes were collected under different rope speeds and different shaking amplitudes (lifting variations: 5mm to 15mm). The actual number of broken wires was recorded (0 to 10, with a focus on 1-3), forming a supervised dataset.

[0080] A fixed-length physical sliding window is used based on the nominal lay length parameters of the wire rope to be tested. This ensures that each slice window... It contains a complete spatial magnetization cycle. To eliminate the physical polarization differences of multi-channel probes, Z-score (based on median) is used for channel-independent amplitude scaling to generate the input one-dimensional tensor. (C is the number of channels, L is the sequence length).

[0081] A quantitative detection network for magnetic flux leakage from broken wires in steel wire ropes is constructed based on the method of this invention. The structural diagram of the dual-branch feature extraction backbone network is shown below. Figure 3 As shown, the input one-dimensional leakage magnetic tensor is processed by a local mutation branch to extract high-frequency local mutation features and a global trend branch to extract low-frequency background features, thereby achieving physical isolation of the spatial frequency of the leakage magnetic signal. The local mutation branch includes a small-size receptive field convolution, batch normalization, and SiLU activation function connected in sequence. The global trend branch includes a large-size receptive field convolution kernel or a one-dimensional dilated convolution, batch normalization, and SiLU activation function.

[0082] For local mutation branches, cascaded one-dimensional convolutional modules with small receptive fields are used to capture the spike mutation of a single broken filament in response to high-frequency magnetic pole jump regions.

[0083] In the basic implementation, the convolution kernel size is set to... or Output features .

[0084] In a preferred embodiment, to further enhance the ability to fit the asymmetric leakage magnetic spikes of 1-2 weak broken wires, this branch employs Dynamic Snake Convolution (DSConv), whose convolution kernel path can adaptively bend according to local gradient changes, mathematically expressed as: .

[0085] For the global trend branch, a large receptive field or a dilated one-dimensional convolution module is used to ignore local small jumps and specifically fit the wide-range low-frequency lift-off envelope caused by the lateral jitter of the sensor.

[0086] In the basic implementation, the dilation rate of the dilated convolution is set. or large convolution kernel Output features .

[0087] In another preferred embodiment, to achieve more efficient global modeling, this branch employs a lightweight Mamba module: only the first 64 dimensions of the feature channels are scanned in the state space, while the remaining channels are directly connected. This design reduces the number of parameters by approximately 40% and inference latency by 25% while maintaining low-frequency envelope fitting capabilities.

[0088] For a cross-level one-dimensional Mamba enhancement module, let the high-frequency characteristics of the output of the local mutation branch be: The low-frequency lift-off feature of the global trend branch output is This will contain prior information about lifting fluctuations. The data is fed into a linear projection layer to calculate the discretization step size for generating a continuous-time Mamba state-space model. Input state matrix and observation matrix .

[0089] The core mathematical and physical mechanism of this internal structural design lies in the introduction of an adaptive time-scale mechanism. The step size Δ in the continuous state space discretization process essentially determines the sampling frequency of the network's historical memory. This is achieved by allowing features containing large-scale noise to be extracted. The cross-level dominant calculation Δ enables the network to sense the physical lift-off drift velocity. When the probe experiences severe lateral jitter causing a sudden change in the low-frequency baseline, The maxima are forced through nonlinear mappings such as Softplus. This causes a dramatic change, which is equivalent to an instantaneous shortening of the state transition time step in the mathematical integral domain. This results in the old low-frequency baseline memory being rapidly cleared due to exponential decay, thus achieving noise suppression inversely using noise morphology from the pure algorithm level to the physical signal level.

[0090] A schematic diagram comparing the mechanisms of the cross-level one-dimensional Mamba enhancement module and traditional self-dependent Mamba is shown below. Figure 4 As shown, (a) is a traditional self-dependent Mamba module, and (b) is a two-branch cross-level Mamba module.

[0091] Fixed reference state matrix based on continuous system Using the zero-order preserve method (ZOH), at a discretization step size dominated by low-frequency features, Under the action, the continuous state parameters are discretized into and : ; .

[0092] High-frequency local features As the actual input sequence to the state machine, the hidden state is executed under the drive of the aforementioned cross-level discretization parameters. Updated and clean features Output: ; .

[0093] To further facilitate information exchange between high- and low-frequency branches, dynamic fusion gating is introduced at the input of the Mamba module. This will enable local high-frequency branch outputs to... After linear mapping, it is compared with the global low-frequency branch features. Perform cross-attention fusion to generate fusion weights. : ; ;

[0094] in The Sigmoid function is used. The fused features... Replace the original As input to the Mamba module, it guides high-frequency details and low-frequency trends to each other, enhancing the ability to preserve weak wire breakage features under lift-off interference.

[0095] Drawing inspiration from the cross-layer connectivity concept of U-Net, the intermediate layer features from different receptive fields in step 3 are combined. and (in Indicates the first Each convolutional block is directly passed to the corresponding output layer of the Mamba enhancement module via skip connections, forming an encoder-decoder type multi-scale cross-level interaction. The specific operation is as follows: ;

[0096] This structure effectively prevents the loss of spatial information in deep networks, and in particular improves the ability to accurately locate the spatial position of broken wires.

[0097] A dual-branch network is used to decouple the leakage magnetic signal at high and low frequencies under strong lift-off interference. The effect is shown in the figure below. Figure 5 As shown.

[0098] Figure 5 This invention demonstrates the effectiveness of physically decoupling the original leakage magnetic field signal under strong lift-off interference using a dual-branch, multi-stage Mamba network, where the sub- Figure 5 (a) is the raw leakage magnetic signal directly picked up by the sensor, sub Figure 5 (b) represents the low-frequency background field output by the global trend branch, sub- Figure 5 (c) is the high-frequency wire breakage signal output by the local mutation branch.

[0099] like Figure 5 As shown in (a), in the original magnetic flux leakage signal, the periodic strand oscillations caused by the spiral strand structure of the steel wire rope and the low-frequency baseline drift caused by the violent lifting jitter are superimposed on each other. The transient magnetic flux leakage pulse generated by the broken wire is completely submerged in the strong background noise because of its extremely small amplitude and short duration (only 5 to 20 sampling points). The location and number of broken wires cannot be directly identified from the figure.

[0100] like Figure 5 As shown in (b), the global trend branch uses a large receptive field convolution module to filter out local abrupt changes and output a smooth low-frequency background field signal. This signal is highly consistent with the periodic wave pattern and baseline drift in the original signal in terms of morphology, fully preserving the evolution trend of the low-frequency background field, but eliminating the high-frequency transient components caused by broken fibers, thus achieving pure extraction of the background field.

[0101] like Figure 5As shown in (c), the local mutation branch uses a small receptive field convolution module to remove the low-frequency background field from the original signal. The output high-frequency broken wire signal maintains a zero baseline in the non-defect region, and only bursts with sharp high-frequency narrow pulse peaks at the broken wire location. More importantly, the pulse amplitudes of 1, 2, and 3 broken wires show a clear gradient increasing relationship, and the pulse duration corresponds to the fracture width. This physical consistency proves that the features extracted by the model are not random noise responses, but truly reflect the physical superposition law of the number of broken wires and the fracture size.

[0102] pass Figure 5 A longitudinal comparison of the three sub-figures (a), (b), and (c) reveals that the broken wire pulse, which was completely unrecognizable in the original signal (a), is isolated as a background field in the low-frequency branch output (b) after decoupling via the dual-branch architecture, and is clearly presented in the high-frequency branch output (c), achieving a physical division of labor of "one path extracting the background and one path extracting the defect." This result intuitively verifies at the signal level that the core advantage of the dual-branch cross-level Mamba architecture proposed in this invention lies in completing the physical isolation of different spatial frequency components in the leakage magnetic signal at the initial stage of feature extraction, fundamentally avoiding the mutual entanglement of the background field and the defect field in the feature space, thereby ensuring high-sensitivity capture and accurate quantitative identification of extremely weak broken wire pulses under strong dynamic interference removal.

[0103] The clean, enhanced feature sequence is then fed into an adaptive global extremum pooling layer. After extracting the global maximum extremum, the sequence is flattened and then fed into a multilayer perceptron as a classification head. Using extremum pooling instead of average pooling avoids the dilution of sharp, small broken wire leakage magnetic flux peaks by a long-range flat baseline, thus preserving the true physical amplitude features. The problem is then transformed into a problem with a number of classes... Discrete classification problems (such as 0 to 10 broken wires).

[0104] A joint ordered regression loss function is constructed for end-to-end training, employing a weighted sum of cross-entropy classification loss and expected regression loss. First, the cross-entropy loss is calculated. Then, a Softmax algorithm is applied to the logits output by the network to generate a probability distribution. Calculate the expected number of broken wires. Then calculate the actual number of broken wires. mean square error The total loss is: ;

[0105] in Set to 0.5. This loss function forces the network to learn the physical law of increasing number of broken wires, especially enhancing its ability to distinguish between 1-3 small broken wires. During training, higher backpropagation weights are assigned to difficult samples (1-3 broken wires), and the optimal network model parameters that meet the preset accuracy index are saved.

[0106] The trained network model is converted to ONNX format and embedded into the ARM / DSP hardware of the portable magnetic flux leakage detector. Sliding window inference is performed on the real-time acquired magnetic flux leakage signals to output the precise number of broken wires at the current position of the steel wire rope, and warnings and displays are provided through a human-machine interface.

Claims

1. A method for detecting broken wires in steel wire ropes based on a dual-branch, multi-level Mamba network, characterized in that, Includes the following steps: Original magnetic flux leakage sequence signals of steel wire rope under different damage states and dynamic lifting disturbance conditions were collected, and a multi-channel one-dimensional magnetic flux leakage sequence dataset was constructed and labeled. Based on the prior physical parameters of the steel wire rope, the collected one-dimensional magnetic flux leakage sequence dataset is preprocessed by adaptive sliding window truncation and physical dimension alignment to construct a one-dimensional magnetic flux leakage tensor containing the complete spatial magnetic field jump cycle. An initial wire rope broken wire leakage magnetic flux quantitative detection network is constructed. The initial wire rope broken wire leakage magnetic flux quantitative detection network includes a bi-branch feature extraction backbone network, a cross-level one-dimensional Mamba enhancement module, and a discrete broken wire quantitative mapping module connected in sequence. The input of the network is a one-dimensional leakage magnetic flux tensor. The input feature data is first processed by the bi-branch feature extraction backbone network to extract high-dimensional local features and global background features. Then, the baseline drift is adaptively filtered out by the cross-level one-dimensional Mamba enhancement module to obtain the broken wire enhancement feature sequence; finally, the spatial information compression and nonlinear mapping of the broken wire enhancement feature sequence are performed by the discrete broken wire quantitative mapping module to obtain the broken wire number probability distribution as the output of the network. A joint ordered regression loss function is introduced to perform end-to-end iterative training on the constructed initial wire rope broken wire leakage magnetic flux quantitative detection network to obtain the wire rope broken wire leakage magnetic flux quantitative detection network. A magnetic flux leakage detection network for broken wires in steel wire ropes is loaded, and the magnetic flux leakage signals collected in real time are inferred to obtain the precise number of broken wires at the current position of the steel wire rope. The operational logic of the cross-level one-dimensional Mamba enhancement module includes: global background features After linear projection, a discretized step size is generated that depends on the current low-frequency lift-off noise fluctuation. The state input matrix B and the observation matrix C are represented by the following formula: ; ; ;in, The activation function for smooth rectification is defined as follows: ,default ; For linear projection; Calculate the discretized state transition parameters using the zero-order preserve method and input parameters It can be expressed using the following formula: ; Where I is the identity matrix, Given the dimension and state transition matrix; Using high-dimensional local features as input sequences, the state transition parameters after discretization... and input parameters Selective scanning and updating of features driven by the process is expressed by the following formula: ; ;in, In hidden state, Output the clean features for the current time step; The clean feature outputs from different time steps are integrated in temporal order to obtain a clean feature sequence.

2. The method for detecting broken wires in steel wire ropes based on a dual-branch, multi-level Mamba network according to claim 1, characterized in that, The original magnetic flux leakage sequence signal was collected using a ring magnetizer and a multi-channel Hall sensor array under different rope speeds and different shaking amplitudes of mining steel wire ropes; then, the actual number of broken wires was used for labeling to obtain a multi-channel one-dimensional magnetic flux leakage sequence dataset.

3. The method for detecting broken wires in steel wire ropes based on a dual-branch, multi-level Mamba network according to claim 1, characterized in that, The one-dimensional leakage magnetic tensor is obtained as follows: a fixed-length physical sliding window is used based on the nominal lay length parameter of the wire rope to be tested, so that each set slice window contains a complete spatial magnetization cycle; then, based on the median, the amplitude of the channel is scaled independently using Z-score to obtain the one-dimensional leakage magnetic tensor.

4. The method for detecting broken wires in steel wire ropes based on a dual-branch, multi-level Mamba network according to claim 1, characterized in that, The dual-branch feature extraction backbone network includes parallel local mutation branches and global trend branches; The local mutation branch uses cascaded one-dimensional convolutional modules with a preset receptive field size, and the processing is represented by the following formula: The global trend branch uses a pre-defined receptive field or a dilated one-dimensional convolutional module, and the receptive field of the global trend branch is larger than that of the local mutation branch. The global trend branching process is represented by the following formula: ;in, The input is a one-dimensional leakage flux tensor; combined with the physical characteristics of wire rope leakage flux, This indicates that it has a preset kernel size. The convolution of the receptive field has a physical span corresponding to the microscale spatial leakage magnetic distortion field generated by a single broken wire of the steel wire rope. This indicates that the kernel has a preset size. Or the convolution of the dilation rate, whose receptive field physical span corresponds to macroscopic lift-off changes in the range of 10~100 mm; This represents the SiLU activation function; BN is the batch normalization operation.

5. The method for detecting broken wires in steel wire ropes based on a dual-branch, multi-level Mamba network according to claim 1, characterized in that, The discrete broken wire quantitative mapping module includes an adaptive global extreme value pooling layer and a multilayer perceptron classification head; The discrete broken wire quantitative mapping module takes the pure feature sequence output by the cross-level one-dimensional Mamba enhancement module as input, extracts the global maximum extremum of the sequence in the adaptive global extremum pooling layer and flattens it, and then inputs the result into the multilayer perceptron classification head to obtain the raw score of the probability of broken wire number.

6. The method for detecting broken wires in steel wire ropes based on a dual-branch, multi-level Mamba network according to claim 1, characterized in that, The initial wire rope broken wire leakage magnetic flux quantitative detection network also includes a dynamic fusion gating module; the dynamic fusion gating module will output the local mutation branches. After linear mapping, it is compared with the global trend branch output. The linear mapping results are fused, and the fusion weights are generated using the Sigmoid activation function, expressed by the following formula: ; For weighting; For local linear projection layers, The global linear projection layer is used, while the local linear projection layer does not share parameters with the global linear projection layer. Local features and global features are mapped to the same fused feature space, respectively. For the Sigmoid function; Then, based on the fusion weights, the output of the local mutation branch is... Global trend branch output Fusion yields fusion characteristics , will integrate features The output of the replacement local mutation branch , and the output of the global trend branch As input to the cross-level one-dimensional Mamba enhancement module; the fusion feature Express it using the following formula: .

7. The method for detecting broken wires in steel wire ropes based on a dual-branch, multi-level Mamba network according to claim 1, characterized in that, The initial wire rope broken wire leakage magnetic flux quantitative detection network also includes skip connections. Based on the U-Net cross-layer connection concept, the output features of different layers of convolutional blocks in the dual-branch feature extraction backbone network are directly passed to the output of the cross-level one-dimensional Mamba enhancement module through skip connections, performing multi-scale cross-level interaction to obtain the final output. As input to the discrete broken wire quantitative mapping module, it is represented by the following formula: ;in, The clean feature sequence output by the cross-level one-dimensional Mamba enhancement module; This represents the output features of the first convolutional block in the local mutation branch; The output features of the i-th convolutional block in the local mutation branch; This refers to the output features of the first convolutional block in the global trend branch; This represents the output feature of the j-th convolutional block in the global trend branch.

8. The method for detecting broken wires in steel wire ropes based on a dual-branch, multi-level Mamba network according to claim 1, characterized in that, The joint ordered regression loss function comprises two parts: cross-entropy classification loss and mathematical expectation regression loss, specifically: Calculate the classification loss of the network output using the cross-entropy loss function. For samples with a batch size of N, it is defined as: ;in For the first The true label of each sample For the network prediction of the first The probability of a class; Perform a Softmax operation on the raw scores output by the network to generate a probability distribution sequence. The expected number of broken wires predicted by the network is calculated by summing the dot product of this number and the increasing tensor k of the broken wire category. It can be expressed using the following formula: ; Calculate the expected number of broken wires Compared with the actual broken wire label value The mean squared error is used as the expected regression loss. It can be expressed using the following formula: ; The classification loss and expected regression loss are weighted and summed according to their respective weight factors to obtain the joint ordered regression loss function, expressed by the following formula: ;in Weighting coefficients to balance the two types of losses.

9. A system for implementing the wire rope broken wire detection method based on a dual-branch cross-level Mamba network as described in any one of claims 1 to 8, characterized in that, It includes a data acquisition module, a data preprocessing module, a network construction module, a network training module, and a broken wire count prediction module; The data acquisition module collects the original magnetic flux leakage sequence signals of the wire rope under different damage states and dynamic lifting disturbance conditions, constructs and labels a multi-channel one-dimensional magnetic flux leakage sequence dataset, and uploads the data to the data preprocessing module. The data preprocessing module, based on the prior physical parameters of the steel wire rope, performs adaptive sliding window truncation and physical dimension alignment preprocessing on the collected one-dimensional leakage magnetic field sequence dataset, constructs a one-dimensional leakage magnetic field tensor containing the complete spatial magnetic field jump cycle, and uploads the data to the network training module. The network construction module builds the initial wire rope broken wire leakage magnetic flux quantitative detection network and uploads the data to the network training module; The network training module introduces a joint ordered regression loss function to perform end-to-end iterative training on the constructed initial wire rope broken wire leakage magnetic flux quantitative detection network, thereby obtaining the wire rope broken wire leakage magnetic flux quantitative detection network, and uploading the data to the broken wire number prediction module; The broken wire count prediction module is loaded with a wire rope broken wire leakage magnetic flux quantitative detection network. It infers the real-time collected leakage magnetic flux signal to obtain the accurate number of broken wires at the current position of the wire rope.

Images