Method for detecting surface cracks of main shaft of mine-used motor based on transient response of pulse eddy current

Abstract

The present application relates to the surface crack detection method of mine motor main shaft based on pulse eddy current transient response, including acquiring the multi-channel transient time sequence signal and synchronous pose in the rotation process of main shaft by using array type pulse eddy current probe;Extracting the lift-off intersection feature to solve the absolute lift-off value and generating the lift-off offset;Using spatial difference algorithm to calculate the transient change rate of spatial dimension, combined with spatial projection algorithm to construct two-dimensional quasi-plane mapping, convert the cylindrical surface electromagnetic diffusion response into curvature neutralized data set;Based on the lift-off offset, construct the dynamic adjustment parameter to carry out amplitude positive, extract the multi-dimensional feature matrix and input the inversion engine to calculate the crack depth, length and strike angle.The present application overcomes the lift-off fluctuation interference under the bad working condition by lift-off compensation, decouples the residual stress and real crack signal by using spatial difference, and eliminates the surface divergence effect by means of spatial projection, realizes the three-dimensional quantitative analysis of main shaft crack.

Description

Technical Field

[0001] This invention relates to a method for detecting surface cracks in the spindle of a mining motor based on pulsed eddy current transient response. Background Technology

[0002] Mining motors are core power equipment in harsh underground working environments such as coal mines. During long-term service, the main shaft of a mining motor must withstand high-frequency variable loads, high torque, and severe mechanical vibrations. Under this complex alternating stress, fatigue cracks are highly likely to initiate on the surface of the main shaft. If not detected in time, these cracks will propagate inwards, potentially leading to shaft breakage or even serious mine production safety accidents. Therefore, efficient and accurate non-destructive testing of surface and near-surface cracks on mining motor main shafts is of great engineering significance.

[0003] The actual working environment of the main shaft of a mining motor is extremely harsh. Its surface is often covered with oil, dust, and rust, and is also subject to roughness changes due to mechanical wear. In the traditional single-frequency eddy current testing process, this complex working condition makes it difficult to maintain an absolutely constant lift-off distance between the probe and the main shaft surface. Small lift-off fluctuations can cause strong background noise from the lift-off effect, which can easily mask weak real crack signals. In addition, traditional single-frequency eddy current testing can usually only obtain the amplitude and phase information of the impedance plane, resulting in a single characteristic dimension. Furthermore, since the main shaft is mostly made of high-strength ferromagnetic alloy steel, the local permeability non-uniformity caused by its internal residual stress can easily be confused with the conductivity changes caused by cracks. Due to the lack of rich characteristic frequency band signals for effective decoupling, this method is difficult to establish an accurate mapping relationship with the three-dimensional geometric dimensions such as crack depth and width, which often limits the detection to a qualitative judgment of the presence or absence of defects, making it difficult to complete the three-dimensional quantitative analysis of defects. In view of the above-mentioned shortcomings, the present invention aims to create a surface crack detection method for mining motor spindles based on pulsed eddy current transient response, so as to make it more valuable for industrial application. Summary of the Invention

[0004] To address the aforementioned technical problems, the purpose of this invention is to provide a method for detecting surface cracks in the spindle of a mining motor based on pulsed eddy current transient response.

[0005] The present invention provides a method for detecting surface cracks in the spindle of a mining motor based on pulsed eddy current transient response, comprising: The multi-channel induced voltage transient timing signal during the spindle rotation process is acquired using an array-type pulsed eddy current probe, and the real-time rotation angle coordinates and axial displacement fed back by the encoder are acquired simultaneously; the lift-off intersection feature is extracted from the multi-channel induced voltage transient timing signal, and the lift-off offset under the corresponding spatial pose is calculated; The multi-channel induced voltage transient timing signal is input into the spatial difference algorithm to calculate the transient change rate of the spatial dimension and generate a spatial gradient transient feature vector. Combined with the real-time rotation coordinates and axial displacement input into the spatial projection algorithm, a quasi-plane coordinate system mapping logic is constructed based on the known principal axis radius to convert the electromagnetic diffusion response on the cylindrical surface under the corresponding spatial coordinates into a curvature-neutralized transient response dataset. Based on the lift-off offset, a dynamic adjustment parameter is constructed and introduced into the amplitude correction calculation. The curvature-neutralized transient response dataset is processed to extract a crack multidimensional feature matrix containing the scattering vector deflection angle. The crack multidimensional feature matrix is ​​then input into the parameter inversion engine to calculate the crack depth, length, and orientation angle.

[0006] Furthermore, the multi-channel induced voltage transient timing signal is a two-dimensional spatiotemporal data matrix composed of discrete voltage values ​​collected by each independent receiving channel in the array-type pulse eddy current probe at a preset sampling frequency within the pulse excitation period. One dimension represents the spatial channel position of the array-type pulse eddy current probe, and the other dimension represents the transient decay time point. The spatial channel position is the relative physical coordinate sequence of each independent receiving channel inside the array-type pulse eddy current probe arranged linearly along the main axis. The coordinate difference between adjacent spatial channel positions is equal to the center geometric distance between adjacent independent receiving channels in the array.

[0007] Furthermore, the liftoff offset calculation process involves calculating the differential timing sequence between the multi-channel induced voltage transient timing signal and the preset defect-free reference signal, taking the transient decay time point corresponding to the zero amplitude in the differential timing sequence as the liftoff intersection feature; extracting the time coordinates corresponding to the liftoff intersection feature, inputting them into the pre-trained liftoff evaluation model, solving for the absolute liftoff value under the corresponding spatial pose, and generating the liftoff offset by calculating the deviation between the absolute liftoff value and the rated working liftoff value.

[0008] Furthermore, the lift-off evaluation model is a prediction model pre-trained based on a backpropagation neural network; the prediction model is trained and converged from a sample set containing multiple known absolute lift-off values ​​and their corresponding lift-off intersection time coordinates, with the time coordinates corresponding to the lift-off intersection features as the model input node and the absolute lift-off value under the corresponding spatial pose as the model output node.

[0009] Furthermore, the spatial difference algorithm takes the voltage amplitude of the multi-channel induced voltage transient timing signal at adjacent spatial channel positions as input, extracts the spatial difference of voltage amplitude at the same transient decay time point through differential operation, and divides the spatial difference of voltage amplitude by the center geometric distance between adjacent independent receiving channels in the array to calculate the transient rate of change in the spatial dimension; the transient rates of change corresponding to each spatial channel position are arranged and combined sequentially according to the spatial arrangement order of the probe array to generate the spatial gradient transient feature vector.

[0010] Furthermore, the spatial projection algorithm involves multiplying the known principal axis radius by the real-time rotation coordinates to calculate the equivalent circumferential unfolded arc length. A two-dimensional quasi-plane coordinate system is established using the equivalent circumferential unfolded arc length as the first dimension coordinate and the axial displacement as the second dimension coordinate. The spatial gradient transient feature vector is then mapped to the two-dimensional quasi-plane coordinate system according to the corresponding coordinates, eliminating the electromagnetic field geometric divergence effect caused by the principal axis cylindrical surface. The electromagnetic diffusion response on the cylindrical surface is then converted into a curvature-neutralized transient response dataset with a spatial distribution equivalent to a plane.

[0011] Furthermore, the amplitude correction operation involves determining dynamic adjustment parameters based on the lift-off offset; extracting the first-direction transient gradient component and the second-direction transient gradient component along the two dimensions of the two-dimensional quasi-plane coordinate system from the coordinate-expanded quasi-plane transient response dataset; and using the dynamic adjustment parameters to perform amplitude correction on the first-direction transient gradient component and the second-direction transient gradient component to obtain the corrected first-direction component and the second-direction component, which constitute an orthogonal gradient feature set. The arctangent function is used to calculate the deflection angle of the scattering vector by combining the corrected second directional component with the first directional component. The square root of the sum of the squares of the corrected first directional component and the second directional component is used as the vector magnitude of the orthogonal gradient feature set. If both the corrected first directional component and the second directional component are lower than a preset noise threshold, the corresponding sampling point is marked as a low confidence point. The vector magnitude of the orthogonal gradient feature set, the deflection angle of the scattering vector, and the corresponding first-dimensional coordinates and second-dimensional coordinates are merged to generate the crack multidimensional feature matrix.

[0012] Furthermore, the dynamic adjustment parameter determination process involves capturing the initial response time window of the multi-channel induced voltage transient timing signal after the pulse excitation is turned off, calculating the sum of squares of the discrete voltage amplitudes within the initial response time window as the transient peak energy; obtaining the standard peak energy calculated from a pre-stored defect-free reference signal within the same time window; and dividing the standard peak energy by the transient peak energy to calculate the amplitude attenuation ratio caused by the lift-off effect. The amplitude attenuation ratio and the liftoff offset are normalized to obtain a normalized attenuation ratio and a normalized liftoff offset; a spatial diffusion compensation factor is calculated by exponentiation with the natural constant as the base and the normalized liftoff offset as the exponent; the dynamic adjustment parameter is calculated by multiplying the normalized attenuation ratio and the spatial diffusion compensation factor.

[0013] By means of the above-described solution, the present invention has at least the following advantages: 1. This application extracts the liftoff intersection features from the transient time-series signal of pulsed eddy current and calculates the absolute liftoff value by combining it with a pre-trained evaluation model to generate the liftoff offset. Then, it constructs dynamic adjustment parameters by combining the energy decay law of the initial response to correct the amplitude of the feature data. This mechanism effectively removes and compensates for the liftoff fluctuation interference caused by oil, dust, rust and mechanical wear on the spindle surface from the bottom layer of signal processing. It overcomes the problem of background noise from strong liftoff effect masking weak real crack signals under harsh working conditions, enabling the flaw detection method to adapt to complex downhole operating environments.

[0014] 2. This application introduces a spatial difference algorithm to calculate the transient rate of change of the magnetic field component in the spatial dimension using the multi-channel signal of the array probe, thereby generating a spatial gradient transient feature vector. This spatial difference processing is equivalent to performing spatial high-pass filtering on the signal, which can effectively filter out the slowly varying background signal caused by the local permeability non-uniformity due to the residual stress inside the high-strength ferromagnetic alloy steel of the main shaft, and realize the physical-level decoupling of residual stress interference and the conductivity mutation signal caused by the real crack.

[0015] 3. This application introduces a spatial projection algorithm and combines real-time rotation coordinates and axial displacement to construct a two-dimensional quasi-planar coordinate system mapping logic, converting the electromagnetic diffusion response on the cylindrical surface into a curvature-neutralized transient response dataset. This coordinate dimension transformation eliminates the electromagnetic field geometric divergence and distortion effects caused by the principal axis cylindrical surface. Based on this, by extracting the transient gradient components and scattering vector deflection angles in the first / second directions, a multi-dimensional feature matrix of the crack is constructed, successfully establishing a mapping relationship between the electromagnetic feature signal and the three-dimensional spatial morphology of the crack. This enables the flaw detection method to overcome the limitation of traditional single-frequency eddy current testing, which can only perform qualitative judgments on the presence or absence of defects, and to possess the three-dimensional quantitative analysis capability to calculate crack depth, length, and orientation angle.

[0016] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show a certain embodiment of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic diagram of the process for surface crack detection of mining motor spindle based on pulsed eddy current transient response; Figure 2 This is a schematic diagram of the liftoff offset calculation process according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the spatial projection algorithm mapping process in an embodiment of the present invention. Detailed Implementation

[0019] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0020] Example 1: Please see Figure 1 This invention provides a method for detecting surface cracks in the spindle of a mining motor based on pulsed eddy current transient response. The technical solution is as follows: A method for detecting surface cracks in the spindle of a mining motor based on pulsed eddy current transient response includes: The multi-channel induced voltage transient timing signal during the spindle rotation process is acquired using an array-type pulsed eddy current probe, and the real-time rotation angle coordinates and axial displacement fed back by the encoder are acquired simultaneously; the lift-off intersection feature is extracted from the multi-channel induced voltage transient timing signal, and the lift-off offset under the corresponding spatial pose is calculated; The multi-channel induced voltage transient timing signal is input into the spatial difference algorithm to calculate the transient change rate of the spatial dimension and generate a spatial gradient transient feature vector. Combined with the real-time rotation coordinates and axial displacement input into the spatial projection algorithm, a quasi-plane coordinate system mapping logic is constructed based on the known principal axis radius to convert the electromagnetic diffusion response on the cylindrical surface under the corresponding spatial coordinates into a curvature-neutralized transient response dataset. Based on the lift-off offset, a dynamic adjustment parameter is constructed and introduced into the amplitude correction calculation. The curvature-neutralized transient response dataset is processed to extract a crack multidimensional feature matrix containing the scattering vector deflection angle. The crack multidimensional feature matrix is ​​then input into the parameter inversion engine to calculate the crack depth, length, and orientation angle.

[0021] The multi-channel induced voltage transient timing signal is a two-dimensional spatiotemporal data matrix composed of discrete voltage values ​​collected by each independent receiving channel in the array-type pulse eddy current probe at a preset sampling frequency within the pulse excitation period. One dimension represents the spatial channel position of the probe array, and the other dimension represents the transient decay time point. The spatial channel position is the relative physical coordinate sequence of each independent receiving channel inside the array-type pulse eddy current probe arranged linearly along the main axis. The coordinate difference between adjacent spatial channel positions is equal to the center geometric distance between adjacent independent receiving channels in the array.

[0022] Specifically, the array-type pulsed eddy current probe has 16 independent receiving channels and a pulse excitation period of 2ms. When entering the pulse excitation period, each independent receiving channel is driven to start signal reception synchronously, and the preset sampling frequency is specifically set to 500KHz, so that each independent receiving channel continuously and synchronously collects the discrete voltage values ​​generated by eddy current attenuation at the frequency of 500KHz. The discrete voltage values ​​acquired by 16 independent receiving channels within the same pulse excitation cycle are arranged sequentially to construct a two-dimensional spatiotemporal data matrix of 16 rows and 1000 columns. In the structural definition of this matrix, one dimension (row dimension) is defined as representing the spatial channel positions of the probe array, such that rows 1 to 16 correspond to the physical indices of the 16 independent receiving channels; the other dimension (column dimension) is defined as representing transient decay time points, such that columns 1 to 1000 correspond sequentially to each sampling instant from the beginning to the end of the pulse excitation cycle; the transient decay time points are the time node sequence of each discrete voltage sampling operation within the pulse excitation cycle, and in this embodiment, the time interval between adjacent time nodes is 2. The two-dimensional spatiotemporal data matrix is ​​the multi-channel induced voltage transient timing signal. The quantification of the spatial channel position is achieved by constructing a relative physical coordinate sequence. In the implementation process, the physical state of the linear arrangement of each independent receiving channel inside the array-type pulse eddy current probe along the main shaft of the mining motor is first confirmed. The geometric center of the first independent receiving channel located at the front end of the probe array is taken as the position reference origin, and its relative physical coordinate is marked as 0mm. The axial distance of the geometric center of the subsequent 15 independent receiving channels relative to the origin is measured and recorded in sequence, thereby forming a relative physical coordinate sequence containing 16 position values. Each value in the relative physical coordinate sequence corresponds one-to-one with each row in the two-dimensional spatiotemporal data matrix. When generating the relative physical coordinate sequence, the coordinate difference between adjacent spatial channel positions is controlled to be always equal to the center geometric distance between adjacent independent receiving channels in the array. In this embodiment, the center geometric distance between adjacent independent receiving channels in the array is set to 2mm. During initialization, the coordinate sequence is automatically generated according to the arithmetic step size of 2mm, that is, the generated sequence is specifically 0mm, 2mm, 4mm up to 30mm.

[0023] The system synchronously acquires real-time angular coordinates and axial displacement fed back by the encoder. The encoder consists of a rotary encoder and a linear displacement encoder configured on the inspection and scanning platform of the mining motor spindle. The rotary encoder is coaxially mechanically connected to the end of the tested mining motor spindle or to a rotary drive chuck to monitor the circumferential rotation of the spindle. The resolution of the rotary encoder is specifically set to 3600 pulses per revolution. Simultaneously, the linear displacement encoder is installed parallel to the linear scanning guide rail on which the array-type pulse eddy current probe moves axially along the spindle. This is used to monitor the linear movement of the probe array, and the displacement resolution of the linear displacement encoder is specifically set to 0.1 mm per pulse. During the uniform rotation of the mining motor spindle driven by the motor, the electrical pulse signals emitted by the rotary encoder are received in real time, and pulse counting begins with a physical reference line marked on the spindle surface as the circumferential zero point. Based on the resolution setting of 3600 pulses per revolution, the accumulated pulse count value is mapped to an absolute angle value, that is, each pulse corresponds to a fixed 0.1 degree rotation angle increment, and the pulse count value is converted into a circumferential angle value to generate the real-time rotation angle coordinates; During the process of the array-type pulse eddy current probe being driven by the transmission mechanism to translate along the scanning guide, it receives the electrical pulse signal emitted by the linear displacement encoder in real time, and counts the pulses using the mechanical limit start end of the scanning guide as the axial zero position; based on the resolution setting of 0.1mm per pulse, the accumulated pulse count value is multiplied by the physical displacement corresponding to a single pulse (i.e., 0.1mm) to convert it into an absolute linear distance, thereby generating the axial displacement. At the initial trigger moment of each preset 2ms pulse excitation cycle of the probe, while sending an excitation command to the probe to obtain the transient timing signal of the multi-channel induced voltage, a hardware latch pulse is sent to the encoder. Upon receiving the latch pulse, the encoder immediately freezes and reads the count values ​​of the current rotary encoder and linear displacement encoder, and then calculates the real-time angular coordinates and axial displacement at this moment; A two-dimensional spatiotemporal data matrix based on high-frequency sampling and precise geometric arrangement was constructed to achieve high-fidelity digital reconstruction of the transient decay process of pulsed eddy currents in terms of temporal evolution and spatial topology. At the same time, a low-level hardware clock latching mechanism was introduced to accurately and synchronously freeze high-resolution angle and displacement encoder data at the moment of each pulse excitation trigger, eliminating mechanical motion deviation and communication delay errors in the dynamic scanning process.

[0024] See Figure 2 Extract the lift-off intersection features from the multi-channel induced voltage transient timing signal and calculate the lift-off offset under the corresponding spatial pose; The liftoff offset calculation process involves calculating the differential timing sequence between the multi-channel induced voltage transient timing signal and the preset defect-free reference signal, and taking the transient decay time point corresponding to the amplitude of zero in the differential timing sequence as the liftoff intersection feature. Extract the time coordinates corresponding to the lift-off intersection features, input them into the pre-trained lift-off evaluation model, calculate the absolute lift-off value under the corresponding spatial pose, and generate the lift-off offset by calculating the deviation between the absolute lift-off value and the rated working lift-off value.

[0025] Specifically, the defect-free reference signal refers to a reference discrete voltage sequence pre-collected on a defect-free standard test block made of the same material as the main shaft of the mining motor under test, maintaining the rated working lift-off value, and using a 2ms pulse excitation period and a 500KHz sampling frequency that are exactly the same as those used in actual detection. The rated working lift-off value is specifically set to 2mm. For each independent receiving channel data in the multi-channel induced voltage transient timing signal, the discrete voltage values ​​corresponding to 1000 transient decay time points collected in real time are subtracted point by point from the reference voltage amplitude at the same time node in the preset defect-free reference signal; by calculating the amplitude difference between the two at each sampling instant, a differential timing sequence containing 1000 data points is generated. The system automatically retrieves feature points with zero amplitude in the differential time series. Because the pulsed eddy current induced voltage exhibits a lift-off crossover phenomenon during decay—meaning that at a specific time point, the amplitude of the induced voltage is unaffected by changes in the lift-off distance—this embodiment identifies the transient decay time point corresponding to a zero amplitude in the differential time series as the lift-off crossover feature. In the specific data processing logic, if there is no absolute zero value in the differential sequence, linear interpolation is used to find the zero point between two adjacent sampling points where the sign reverses, thus determining the time coordinate of the zero point.

[0026] The time coordinates are input into a pre-trained lift-off evaluation model, which is a prediction model generated by pre-training a backpropagation neural network. The prediction model is trained and converged from a sample set containing multiple known absolute lift-off values ​​and their corresponding lift-off intersection time coordinates. The time coordinates corresponding to the lift-off intersection features are used as the model input nodes, and the absolute lift-off values ​​under the corresponding spatial poses are used as the model output nodes.

[0027] Specifically, the known absolute lift-off value is obtained by setting multiple known physical gaps above the surface of the standard part of the mining motor spindle without defects, and using the distance value of the physical gaps as the known absolute lift-off value. In this embodiment, the range of the known absolute lift-off value is preset to be 0.5mm to 5.0mm, the step size is set to 0.1mm, and a total of 46 known absolute lift-off value status points are obtained. At each known absolute lift-off value state point, the control probe collects the induced voltage signal at a preset sampling frequency, and calculates the differential time sequence between it and the preset defect-free reference signal under the reference lift-off state. The transient decay time point corresponding to the amplitude of zero in the differential time sequence is extracted as the lift-off intersection time coordinate corresponding to the known absolute lift-off value. All the known absolute lift-off values ​​and their corresponding lift-off intersection time coordinates are combined into data pairs to construct an initial sample set. The initial sample set is then randomly divided into a training sample set and a test sample set according to a ratio of 80% and 20%. The construction process of the backpropagation neural network is as follows: the backpropagation neural network is configured to consist of an input layer, a hidden layer, and an output layer; in this embodiment, the number of nodes in the input layer is set to 1, which is used to receive the time coordinates corresponding to the lift-off intersection features; the hidden layer is set to have two layers, in this embodiment, the first hidden layer is preset to contain 10 neurons and the second hidden layer contains 5 neurons, and the neurons in the hidden layer all adopt a logarithmic nonlinear smooth activation function; the number of nodes in the output layer is set to 1, which is used to output the absolute lift-off value under the corresponding spatial pose, and the output layer adopts a linear activation function to ensure the continuity and unboundedness of the output value range. The corresponding spatial pose is the relative physical position state between the probe and the spindle surface determined by the real-time rotational coordinates of the motor spindle and the axial displacement of the probe when the array pulsed eddy current probe acquires the current multi-channel induced voltage transient timing signal, which is synchronously acquired and fed back by the encoder. The pre-training and convergence process of the prediction model specifically involves initializing the connection weight matrix and bias vector between each neuron node in the backpropagation neural network. In this embodiment, the initialization is performed using random values ​​within the range of -1 to +1. The time coordinates of each lift-off intersection point in the training sample set are input to the model input node one by one. The numerical signal is forward-propagated to the model output node through the hidden layer to calculate the predicted lift-off value of the forward output. The predicted lift-off value is compared with the known absolute lift-off value corresponding to the training sample set, and the mean square error between the two is calculated. When the calculated mean square error is greater than the error threshold of 0.001 preset in this embodiment, the error signal is backpropagated layer by layer from the output layer to the input layer according to the error gradient descent principle. The learning rate is set to 0.01 preset in this embodiment, and the connection weights and bias parameters between each layer of neurons are dynamically updated and adjusted. The forward propagation of data and the backward propagation of error are repeated continuously until the maximum number of iterations of training preset in this embodiment, 1000, is reached. At this point, the update of the network parameters is stopped, and the model training is determined to be converged.

[0028] The solidification and actual solution process of the lift-off evaluation model specifically involves saving all network connection weights and bias parameters under the above training convergence state to generate a fixed lift-off evaluation model. In the actual execution stage of surface crack detection of the mining motor spindle, the time coordinates corresponding to the lift-off intersection features extracted in real time during spindle rotation and probe movement are input to the model input node of the solidified prediction model. After the time coordinate data undergoes linear weighted summation and nonlinear activation mapping operations on the parameters of each layer of the network, the model output node outputs the predicted value, which is used as the absolute lift-off value under the corresponding spatial pose obtained by the solution. The deviation between the calculated absolute lift-off value and the preset rated working lift-off value is used to generate the lift-off offset. The rated working lift-off value is the optimal detection distance preset based on the probe sensitivity in this embodiment. In this embodiment, it is set to 2mm. When the absolute lift-off value is 2.2mm, the calculated lift-off offset is 0.2mm.

[0029] By constructing a lift-off evaluation model using a pre-trained backpropagation neural network and leveraging data-driven nonlinear mapping capabilities, the complex mechanism modeling challenges in traditional analytical electromagnetic field methods are effectively avoided, improving the efficiency and accuracy of lift-off value calculation. Furthermore, by selecting the time coordinates corresponding to the lift-off intersection features as the model input, and considering the inherent immunity of these features to fluctuations in the electromagnetic parameters of the spindle material and surface roughness, the model can still calculate the absolute lift-off value under the corresponding spatial pose even in the complex operating conditions of mining motors.

[0030] The multi-channel induced voltage transient timing signal is input into a spatial difference algorithm to calculate the transient rate of change of the magnetic field component in the spatial dimension and generate a spatial gradient transient feature vector. The spatial difference algorithm takes the voltage amplitude of the multi-channel induced voltage transient timing signal at adjacent spatial channel positions as input, extracts the spatial difference of voltage amplitude at the same transient decay time point through differential operation, and divides the spatial difference of voltage amplitude by the center geometric distance between adjacent independent receiving channels in the array to calculate the transient rate of change in the spatial dimension; the transient rates of change corresponding to each spatial channel position are arranged and combined sequentially according to the spatial arrangement order of the probe array to generate the spatial gradient transient feature vector; Specifically, for any given transient decay time point in the two-dimensional spatiotemporal data matrix, the spatial difference algorithm extracts the discrete voltage amplitudes of two adjacent rows of corresponding independent receiving channels along the preset axial physical arrangement direction of the probe array and performs difference calculations. The specific difference logic is set to subtract the voltage amplitude of the previous spatial channel from the voltage amplitude of the next spatial channel position, that is, calculate the subtraction of the corresponding channel in the 2nd row from the corresponding channel in the 1st row, the subtraction of the corresponding channel in the 3rd row from the corresponding channel in the 2nd row, and so on, until the calculation of the subtraction of the corresponding channel in the 16th row from the corresponding channel in the 15th row is completed. Through the above layer-by-layer traversal difference operation, 15 consecutive spatial differences of the voltage amplitude are extracted at each determined transient decay time point. The pre-stored array probe physical parameters are called, and each specific voltage amplitude spatial difference is divided by the center geometric spacing of adjacent independent receiving channels in the array, which is preset in this embodiment. In this embodiment, the center geometric spacing is explicitly set to 2mm. This division operation is to divide the relative voltage difference by a constant 2, thereby mapping the voltage difference to voltage attenuation gradient data per unit physical length. The quotient obtained from each division operation is directly used as the transient rate of change in the spatial dimension of the corresponding adjacent channel interval at the same transient attenuation time point. For the transient change rates corresponding to each adjacent channel interval calculated above, they are sequentially arranged and spliced ​​according to the spatial arrangement order of each independent receiving channel within the probe array from channel 1 to channel 16. At any single transient decay time point, the 15 calculated transient change rates are arranged into a one-dimensional vector component containing 15 elements according to spatial order. In the time dimension, all 1000 transient decay time points of the two-dimensional spatiotemporal data matrix are traversed, and the above steps of subtraction, division, and sequential arrangement are repeated cyclically to calculate and generate 1000 one-dimensional vector components. Finally, these 1000 one-dimensional vector components containing spatial gradient information are spliced ​​and fused according to the temporal order of the transient decay time points to generate the spatial gradient transient feature vector that completely covers the entire pulse excitation cycle. By introducing a spatial differential algorithm and combining it with the physical spacing of the probes for differential division, the discrete absolute voltage signal is accurately mapped into a spatial attenuation gradient vector with real physical meaning. This spatial differential processing is equivalent to constructing a spatial high-pass filter at the signal level, which can effectively filter out the slowly varying background interference signal caused by residual stress or local permeability inhomogeneity inside the high-strength ferromagnetic alloy steel of the main shaft. This achieves physical-level decoupling between residual stress interference and the conductivity mutation signal caused by real micro-cracks, highlighting the characteristic saliency of the crack.

[0031] Combining the real-time rotation coordinates and axial displacement input spatial projection algorithm, a quasi-plane coordinate system mapping logic is constructed based on the known principal axis radius to convert the electromagnetic diffusion response on the cylindrical surface under the corresponding spatial coordinates into a curvature-neutralized transient response dataset. See Figure 3 The spatial projection algorithm is to multiply the known principal axis radius by the real-time rotation coordinate to calculate the equivalent circumferential unfolded arc length, and establish a two-dimensional quasi-plane coordinate system with the equivalent circumferential unfolded arc length as the first dimension coordinate and the axial displacement as the second dimension coordinate. The spatial gradient transient feature vector is mapped to the two-dimensional quasi-plane coordinate system according to the corresponding coordinates, eliminating the electromagnetic field geometric divergence effect caused by the principal axis cylindrical surface, and converting the electromagnetic diffusion response on the cylindrical surface into the curvature-neutralized transient response dataset whose spatial distribution is equivalent to that of a plane.

[0032] Specifically, the physical outer diameter of the mining motor spindle in a stopped state is measured multiple times using a laser diameter gauge, and the average value is divided by 2 to obtain the spindle radius. The known spindle radius is then set as... The known principal axis radius is multiplied by the real-time rotation coordinate to calculate the equivalent circumferential arc length on the physical circumferential surface. If the known principal axis radius is... Real-time rotation coordinates are The equivalent circumferential arc length is The specific calculation formula is as follows:

[0033] The two-dimensional quasi-planar coordinate system uses the direction of the equivalent circumferential arc length as the first dimension coordinate (i.e., the horizontal column index of the corresponding two-dimensional data matrix) and the direction of the synchronously read axial displacement as the second dimension coordinate (i.e., the vertical row index of the corresponding two-dimensional data matrix). To map continuous physical spatial coordinates to discrete computer storage space, this embodiment presets the spatial grid division resolution of the two-dimensional quasi-planar coordinate system to 1mm × 1mm, that is, a physical spacing of 1mm is used as the digital discrete step length of the coordinate axes in both the first and second dimensions, establishing a discrete orthogonal data storage grid in memory. The spatial gradient transient feature vector is read. Since the underlying hardware uses a unified global timestamp when acquiring raw data, the algorithm uses a time matching mechanism to bind each spatial gradient transient feature vector to the spatial coordinate information corresponding to its generation time, forming a data unit carrying specific spatial pose information. During the mapping operation, based on the values ​​of the first and second dimension coordinates bound to each spatial gradient transient feature vector, memory addressing instructions are used to write the values ​​in the spatial gradient transient feature vector to the corresponding grid nodes in the two-dimensional quasi-plane coordinate system. When the calculated physical coordinate point cannot perfectly coincide with a preset 1mm integer grid node, the microprocessor calls a bilinear interpolation algorithm to calculate the weight allocation based on the inverse Euclidean distance ratio from the non-integer physical coordinate point to its four nearest standard integer grid nodes. The values ​​of the spatial gradient transient feature vector are then split according to this weight ratio and accumulated and written to the memory addresses corresponding to the four adjacent integer grid nodes. By performing the aforementioned point-by-point coordinate transformation, addressing and tracking, and interpolation mapping on all feature vectors generated during the full scan, the transient feature data originally distributed in a three-dimensional cylindrical coordinate system is spread out to a two-dimensional coordinate system grid on a plane. Numerical resampling and matrix filling are performed to expand the sampling points on the circumferential surface towards the straight line direction. After the data traversal mapping and filling of all coordinate points are completed, the two-dimensional matrix region is encapsulated and backed up, thereby eliminating the electromagnetic field geometric divergence effect caused by the principal axis cylindrical surface. Thus, the electromagnetic diffusion response on the cylindrical surface is converted into a transient response dataset with a spatial distribution equivalent to a plane with neutral curvature. By establishing a two-dimensional quasi-plane coordinate system through a spatial projection algorithm, the problem of signal distortion caused by changes in principal axis curvature in traditional flaw detection methods is solved. The transient response data, which was originally attached to the three-dimensional cylindrical coordinate system, is cleverly flattened and mapped into a two-dimensional curvature neutralization matrix. This dimension reduction and resampling operation not only completely eliminates the electromagnetic field geometric divergence effect caused by the cylindrical surface of the principal axis from a physical mechanism perspective, but also eliminates curvature interference caused by principal axis radii of different specifications. At the same time, it transforms the complex three-dimensional spatial calculation into standard two-dimensional matrix processing, reducing computational complexity.

[0034] Based on the lift-off offset, a dynamic adjustment parameter is constructed and introduced into the directional weighted calculation to process the curvature-neutralized transient response dataset and extract the crack multidimensional feature matrix containing the scattering vector deflection angle. The amplitude correction operation involves determining dynamic adjustment parameters based on the lift-off offset; extracting transient gradient components in the first direction and the second direction along the two dimensions of the two-dimensional quasi-plane coordinate system from the coordinate-expanded quasi-plane transient response dataset; and using the dynamic adjustment parameters to perform amplitude correction on the transient gradient components in the first direction and the transient gradient components in the second direction to obtain the corrected first direction components and the second direction components, which constitute an orthogonal gradient feature set. The arctangent function is used to calculate the deflection angle of the scattering vector by combining the corrected second directional component with the first directional component. The square root of the sum of the squares of the corrected first directional component and the second directional component is used as the vector magnitude of the orthogonal gradient feature set. If both the corrected first directional component and the second directional component are lower than a preset noise threshold, the corresponding sampling point is marked as a low confidence point. The vector magnitude of the orthogonal gradient feature set, the deflection angle of the scattering vector, and the corresponding first-dimensional coordinates and second-dimensional coordinates are merged to generate the crack multidimensional feature matrix.

[0035] The process of determining the dynamic adjustment parameters involves extracting the initial response time window of the multi-channel induced voltage transient timing signal after the pulse excitation is turned off, calculating the sum of squares of the discrete voltage amplitudes within the initial response time window as the transient peak energy, and obtaining the standard peak energy calculated from the pre-stored defect-free reference signal under the same time window; dividing the standard peak energy by the transient peak energy to calculate the amplitude attenuation ratio caused by the lift-off effect. The amplitude attenuation ratio and the liftoff offset are normalized to obtain a normalized attenuation ratio and a normalized liftoff offset; a spatial diffusion compensation factor is calculated by exponentiation with the natural constant as the base and the normalized liftoff offset as the exponent; the dynamic adjustment parameter is calculated by multiplying the normalized attenuation ratio and the spatial diffusion compensation factor. Specifically, the acquired multi-channel induced voltage transient timing signal is truncated along the time axis. In this embodiment, it is preset to truncate the signal 10 after the pulse excitation is turned off. Up to 50 The time period is set as the initial response time window; Extract all discrete voltage amplitude data within the initial response time window, square each of these discrete voltage amplitude data, then sum all the squared values, and store the final sum of squared voltage amplitudes as the transient peak energy. Recall the pre-stored defect-free reference signal from internal memory, following the same 10... Up to 50 The data is captured within a time window and the standard peak energy is calculated using the same sum of squares method. In this embodiment, the baseline value of the standard peak energy is preset to be 500. Then, the obtained standard peak energy is divided by the transient peak energy calculated in real time, and the amplitude attenuation ratio caused by the probe lift-off effect is obtained through division. The maximum and minimum values ​​of the amplitude attenuation ratio and the liftoff offset in the measurement batch data are obtained respectively. The maximum and minimum values ​​of the amplitude attenuation ratio and the liftoff offset are linearly mapped to the numerical range of zero to one using the maximum-minimum linear normalization method, thereby obtaining the dimensionless normalized attenuation ratio and the normalized liftoff offset respectively.

[0036] Using the natural constant as the base, the value of the normalized lift-off offset obtained above is used as the exponent term, and an exponentiation operation is performed. The result of the operation is used as the spatial diffusion compensation factor. The normalized attenuation ratio and the calculated spatial diffusion compensation factor are directly multiplied, and the product result is generated and output as the dynamic adjustment parameter. Based on the quasi-plane transient response dataset after coordinate expansion, spatial difference calculations are performed along the two dimensions of the two-dimensional quasi-plane coordinate system. Specifically, the transient response data difference at adjacent first-dimensional coordinates, i.e., equivalent circumferential expanded arc length coordinates, is calculated as the transient gradient component in the first direction. At the same time, the transient response data difference at adjacent second-dimensional coordinates, i.e., axial displacement coordinates, is calculated as the transient gradient component in the second direction. The dynamically adjusted parameters are multiplied by the extracted transient gradient components in the first and second directions, respectively. The product results are used to replace the original component values ​​to correct the magnitudes of the gradient components in both directions. The corrected gradient component data in the two directions are then combined to form the orthogonal gradient feature set. Extract the corrected first direction component and second direction component from the orthogonal gradient feature set, and use the arctangent function to calculate the angle between the corrected second direction component and the first direction component to obtain the angle data of the synthesized spatial vector. Define and output the angle data as the deflection angle of the scattering vector. The weighted and corrected transient gradient components in the first and second directions are squared and summed. The square root of the sum is then taken to calculate the vector magnitude of the orthogonal gradient feature set. Before merging to generate the feature matrix, data verification is performed. If both the corrected first and second direction components are below a preset noise threshold, the corresponding sampling point is marked as a low-confidence point and discarded or temporarily deferred. The noise threshold is a specific multiple of the background noise standard deviation, usually three to five times, to ensure that only when the signal strength significantly exceeds the background random fluctuations (i.e., has a sufficient signal-to-noise ratio) is it judged as a valid crack feature. The vector magnitude of the orthogonal gradient feature set corresponding to the verified analysis data point, the deflection angle of the scattering vector, and the first and second dimension coordinate values ​​of the current analysis data point in the two-dimensional quasi-plane coordinate system are concatenated and combined in a fixed row and column order to generate the crack multidimensional feature matrix used to invert the crack depth and width direction.

[0037] By extracting the transient peak energy within the initial response time window and combining it with the standard peak energy of the defect-free reference signal, the amplitude attenuation caused by the probe lift-off effect was quantified. An exponential spatial diffusion compensation factor based on the normalized lift-off offset was introduced to generate dynamic adjustment parameters, effectively overcoming the dynamic lift-off interference caused by mechanical vibration or radial runout of the mining motor spindle during rotary flaw detection. At the same time, the dynamic adjustment parameters were used to perform targeted amplitude weighting correction on the orthogonal transient gradient components in the two-dimensional quasi-plane, and based on this, the deflection angle and vector magnitude of the scattering vector with clear physical direction were synthesized. This successfully transformed the scalar amplitude signal, which is susceptible to environmental interference, into a highly robust spatial vector feature matrix, fundamentally eliminating the nonlinear error caused by the spatial electromagnetic divergence effect.

[0038] The multidimensional feature matrix of the crack is input into the parameter inversion engine to calculate the depth, length, and orientation angle of the crack.

[0039] In this embodiment, the parameter inversion engine is pre-trained using known crack sample data to form a convergent deep convolutional neural network model containing convolutional extraction layers and fully connected regression layers. The known crack sample data contains a large number of four-channel input feature matrices corresponding to standard cracks and their corresponding true depth, length, and orientation angle labels. Before performing the inversion calculation, the parameter inversion engine is loaded into the memory of the main control unit, and the pre-fixed network topology and trained weights and bias parameters are called to complete the model initialization.

[0040] The vector magnitude, scattering vector deflection angle, and distribution patterns of the first and second dimension coordinates of each coordinate point within the entire detection area contained in the crack multidimensional feature matrix are used as a two-dimensional input feature layer for four channels. In order to eliminate the difference in absolute numerical magnitude between the above-mentioned different physical quantities due to the different dimensions, batch normalization is first performed on the input feature layer, and then the normalized data is synchronously sent into the initial convolutional layer of the engine. The parameter inversion engine performs forward propagation calculations on the imported crack multidimensional feature matrix. In this embodiment, the convolutional extraction layer of the parameter inversion engine is preset to include a structure consisting of three cascaded two-dimensional convolutional layers and two max-pooling layers; firstly, multiple internally set two-dimensional convolutional kernels are used to perform sliding multiplication and addition operations on the matrix, and each convolutional layer selects... By using convolutional kernels of a specific size and combining them with the ReLU activation function, the implicit gradient abrupt boundary and vector deflection local topological features in the feature matrix are extracted, and then... The max pooling operation of the sliding window is used to filter redundant information and perform data dimensionality reduction. The high-dimensional abstract feature data after convolution and pooling is flattened and reorganized into a one-dimensional feature sequence, which is then input into the fully connected layer at the end of the network. Nonlinear regression fitting calculation is performed through the activation function of the neurons in the fully connected layer to completely transform the spatial distribution pattern and electromagnetic vector features into physical geometric dimensions.

[0041] The three independent linear output nodes at the end of the fully connected layer synchronously calculate continuous predicted values ​​for three target dimensions. The output value of the first output node is scaled and assigned to millimeters as the calculated crack depth. The output value of the second output node is scaled and assigned to millimeters as the calculated crack length. The output value of the third output node is scaled and assigned to degrees as the calculated crack orientation angle. The scale transformation is determined based on the numerical range of the label data used during training. The mapping weights within the first, second, and third output nodes are formed according to the following logic, corresponding to physical dimensions: During the pre-training phase, a multi-task loss function is constructed, composed of depth error components, length error components, and orientation angle error components; the multi-task loss function... The calculation formula is defined as follows:

[0042] in, , , Weight coefficients for depth, length, and orientation angle are used to compensate for backpropagation gradient imbalances caused by different physical dimensions and numerical ranges. These coefficients are determined by pre-calculating the global variances of the true depth, true length, and true orientation angle labels in the training set, with each weight coefficient set as the reciprocal of its corresponding global variance. After each forward computation, the difference between the predicted value of the first output node and the true depth label value is independently calculated as the depth error component. The difference between the predicted value of the second output node and the value of the actual length label is calculated as the length error component. The difference between the predicted value of the third output node and the value of the actual azimuth angle label is calculated as the azimuth angle error component. ; Subsequently, through the backpropagation mechanism, the depth error component is used to specifically drive and adjust the weights of the hidden layer connected to the first output node, forcing the node to focus on the amplitude attenuation feature in the input matrix that is highly correlated with the crack depth; similarly, the length error component and the orientation angle error component are used to independently drive and adjust the weights connected to the second and third output nodes, so that they focus on the spatial broadening feature and vector deflection feature that are related to the crack length and orientation, respectively. The known crack sample data are collected from standard crack specimens with the same material as the main shaft of the mining motor under test or with electromagnetic parameters within the same preset range. Each sample includes a four-channel two-dimensional input feature map consisting of vector magnitude, scattering vector deflection angle, first-dimensional coordinates, and second-dimensional coordinates, as well as corresponding crack depth labels, crack length labels, and crack orientation angle labels. The crack depth and crack length labels are recorded in millimeters, and the crack orientation angle labels are recorded in degrees. During training, the four-channel two-dimensional input feature map is normalized and input into the convolutional extraction layer. After convolution, pooling, flattening, and fully connected regression, three continuous predicted values ​​are output. The network parameters are updated by a multi-task loss function composed of weighted depth error, length error, and orientation angle error until the validation set error meets the preset threshold or reaches the preset number of iterations.

[0043] Since the input true depth and true length labels are measured in millimeters and the true orientation angle label is measured in degrees during the construction of the training set, the network not only learns the feature mapping relationship during the convergence of the multi-task loss function, but also implicitly completes the calibration of numerical units, forcing the output distribution of these three nodes to be in the same linear space as the real physical values; therefore, in the actual inversion calculation stage, no additional scaling is required, and the three independent linear output nodes at the end of the fully connected layer can directly and synchronously output continuous predicted values ​​of the three target dimensions that strictly correspond to the preset physical units.

[0044] By extracting the features of the lift-off intersection and constructing dynamic adjustment parameters for directional amplitude weighting correction, the interference of lift-off fluctuations caused by harsh working conditions on the spindle surface (such as oil stains, rust, and mechanical wear) is effectively removed and compensated, overcoming the problem of strong lift-off effect background noise masking weak real crack signals. A spatial difference algorithm is introduced to generate spatial gradient transient feature vectors, which can perform spatial high-pass filtering at the signal processing level, realizing physical-level decoupling between the slowly varying background interference caused by the residual stress of the spindle and the abrupt change signal of the real crack. By constructing a two-dimensional quasi-plane coordinate system with the help of a spatial projection algorithm, not only is the electromagnetic field geometric divergence and distortion effect caused by the cylindrical surface of the spindle completely eliminated, but the mapping relationship between the electromagnetic feature signal and the three-dimensional spatial morphology of the crack is also successfully established, enabling the flaw detection method to break through the limitation of traditional qualitative judgment of the presence or absence of defects.

[0045] Example 2: This embodiment applies a surface crack detection method for mining motor spindles based on pulsed eddy current transient response to the spindle of a high-power coal mining machine's cutting section in a coal mine. The spindle material is high-strength ferromagnetic alloy steel, and the measured spindle radius is set as follows: mm, in the actual working environment, the spindle surface is covered with a thick layer of coal dust and is accompanied by roughness changes caused by mechanical wear; An array-type pulsed eddy current probe was used to acquire multi-channel induced voltage transient timing signals during the spindle rotation process, and real-time coordinates were simultaneously acquired. To meet the detection requirements of this spindle model, the number of independent receiving channels was set to 24, the pulse excitation period to 1ms, and a preset sampling frequency of 1MHz was used. Each independent receiving channel continuously and synchronously acquired discrete voltage values ​​within the pulse period, constructing a 24-row by 1000-column two-dimensional spatiotemporal data matrix. The center-to-center geometric distance between adjacent independent receiving channels in the array was set to 1.5mm, generating a relative physical coordinate sequence increasing from 0mm in 1.5mm increments to 34.5mm. Encoders were configured: the rotary encoder resolution was set to 7200 pulses per revolution (corresponding to a 0.05-degree rotation increment), and the linear displacement encoder resolution was set to 0.05mm per pulse. At the moment of pulse triggering, the pulse was latched by hardware, synchronously freezing and calculating the real-time rotational coordinates and axial displacement. The lift-off intersection feature is extracted from the transient time-series signal of multi-channel induced voltage. The absolute lift-off value under the corresponding spatial pose is calculated and the lift-off offset is generated. The rated working lift-off value for this scenario is set to 1.5mm. The differential time-series sequence between the real-time signal and the preset defect-free reference signal is calculated. The sign reversal point is found using linear interpolation. The transient decay time coordinate corresponding to the amplitude of zero is extracted as the lift-off intersection feature. This time coordinate is input into a pre-trained backpropagation neural network model. In this embodiment, the first hidden layer of the network is set to 12 neurons and the second hidden layer is set to 8 neurons. Assuming that the model outputs an absolute lift-off value of 1.8mm under the current pose, the lift-off offset caused by coal dust and vibration at this spatial position is 0.3mm, calculated by the deviation from the rated lift-off value of 1.5mm. The multi-channel induced voltage transient timing signal is input into a spatial difference algorithm to extract the spatial difference of voltage amplitude and generate a spatial gradient transient feature vector. For a 24-row × 1000-column two-dimensional matrix, at the same transient decay time point, the voltage amplitude of the previous channel is subtracted from the voltage amplitude of the next channel, and the difference is divided by the center geometric spacing of 1.5mm, accurately mapping the discrete voltage to the decay gradient over a unit physical length. After traversing all 24 channels, a vector component containing 23 elements is generated at a single time point. Data from 1000 time points are concatenated on the time axis to generate a spatial gradient transient feature vector that fully covers a 1ms period. By combining real-time rotation coordinates and axial displacement input spatial projection algorithms, a two-dimensional quasi-plane coordinate system is established and transformed to generate a curvature-neutralized transient response dataset; the principal axis radius is calculated. The equivalent circumferential unfolded arc length is obtained by multiplying mm with the real-time rotation coordinate. Using this arc length as the first dimension coordinate and the axial displacement as the second dimension coordinate, a discrete orthogonal data grid with a spatial resolution of 0.5mm×0.5mm is constructed in memory. The bilinear interpolation algorithm is called to map and distribute the feature vectors carrying spatial pose information to the two-dimensional grid according to the weights, thus flattening the electromagnetic diffusion response that was originally attached to the three-dimensional cylindrical surface. Dynamic adjustment parameters are constructed based on the lift-off offset, and a directional weighted operation is introduced to extract the multidimensional feature matrix of the crack. The crack feature matrix is ​​extracted after the pulse is turned off. Up to 30 The time period is defined as the initial response time window. The sum of squares of the voltage amplitudes within the window is calculated to obtain the transient peak energy. The standard peak energy within the same window is obtained. The standard peak energy is divided by the transient peak energy to obtain an amplitude attenuation ratio of 1.25. This ratio is normalized with a liftoff offset of 0.3 mm and then subjected to exponential operation to generate a spatial diffusion compensation factor. The two are multiplied to obtain a dynamic adjustment parameter. This parameter is used to perform amplitude weighted correction on the transient gradient components in the first and second directions within the alignment plane. The arctangent function is solved by calculating the ratio of the two to obtain the scattering vector deflection angle. This angle is then concatenated with the vector magnitude and the corresponding two-dimensional coordinates to generate a crack multidimensional feature matrix. The multidimensional feature matrix of the crack is input into the parameter inversion engine, and forward propagation is performed to calculate the three-dimensional quantitative size of the microcrack. The feature matrix, as a two-dimensional input layer with four dimensions, is sent to a three-layer input engine after batch normalization. The convolutional neural network with convolutional kernels extracts topological features, reorganizes them into a one-dimensional sequence, and enters the fully connected layer for nonlinear regression fitting. Since the model has completed the implicit calibration of physical dimensions through the multi-task loss function, the three independent linear output nodes at the end of the fully connected layer directly and synchronously output absolute predicted values. Through inversion calculation, the depth of the latent fatigue crack on the surface of the coal mining machine main shaft at this point is accurately identified as 1.2 mm, the crack length is 15.5 mm, and the spatial orientation angle of the crack is 42.5 degrees.

[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting surface cracks in the spindle of a mining motor based on pulsed eddy current transient response, characterized in that, include: The multi-channel induced voltage transient timing signal during the spindle rotation process is acquired using an array-type pulsed eddy current probe, and the real-time rotation angle coordinates and axial displacement fed back by the encoder are acquired simultaneously; the lift-off intersection feature is extracted from the multi-channel induced voltage transient timing signal, and the lift-off offset under the corresponding spatial pose is calculated; The multi-channel induced voltage transient timing signal is input into the spatial difference algorithm to calculate the transient change rate of the spatial dimension and generate a spatial gradient transient feature vector. Combined with the real-time rotation coordinates and axial displacement input into the spatial projection algorithm, a quasi-plane coordinate system mapping logic is constructed based on the known principal axis radius to convert the electromagnetic diffusion response on the cylindrical surface under the corresponding spatial coordinates into a curvature-neutralized transient response dataset. Based on the lift-off offset, a dynamic adjustment parameter is constructed and introduced into the amplitude correction calculation. The curvature-neutralized transient response dataset is processed to extract a multidimensional feature matrix of cracks containing the deflection angle of the scattering vector. The multidimensional feature matrix of the crack is input into the parameter inversion engine to calculate the depth, length, and orientation angle of the crack.

2. The method for surface crack detection of mining motor spindle based on pulsed eddy current transient response according to claim 1, characterized in that, The multi-channel induced voltage transient timing signal is a two-dimensional spatiotemporal data matrix composed of discrete voltage values ​​collected by each independent receiving channel in the array-type pulse eddy current probe at a preset sampling frequency within the pulse excitation period. One dimension represents the spatial channel position of the array-type pulse eddy current probe, and the other dimension represents the transient decay time point. The spatial channel position is the relative physical coordinate sequence of each independent receiving channel inside the array-type pulse eddy current probe arranged linearly along the main axis. The coordinate difference between adjacent spatial channel positions is equal to the center geometric distance between adjacent independent receiving channels in the array.

3. The method for detecting surface cracks in the spindle of a mining motor based on pulsed eddy current transient response according to claim 1, characterized in that, The liftoff offset calculation process involves calculating the differential timing sequence between the multi-channel induced voltage transient timing signal and the preset defect-free reference signal, and taking the transient decay time point corresponding to the amplitude of zero in the differential timing sequence as the liftoff intersection feature. Extract the time coordinates corresponding to the lift-off intersection features, input them into the pre-trained lift-off evaluation model, calculate the absolute lift-off value under the corresponding spatial pose, and generate the lift-off offset by calculating the deviation between the absolute lift-off value and the rated working lift-off value.

4. The method for detecting surface cracks in the spindle of a mining motor based on pulsed eddy current transient response according to claim 3, characterized in that, The lift-off evaluation model is a prediction model pre-trained based on a backpropagation neural network. The prediction model is trained and converged from a sample set containing multiple known absolute lift-off values ​​and their corresponding lift-off intersection time coordinates. The time coordinates corresponding to the lift-off intersection features are used as the model input nodes, and the absolute lift-off values ​​under the corresponding spatial poses are used as the model output nodes.

5. The method for detecting surface cracks in the spindle of a mining motor based on pulsed eddy current transient response according to claim 1, characterized in that, The spatial difference algorithm takes the voltage amplitude of the multi-channel induced voltage transient timing signal at adjacent spatial channel positions as input, extracts the spatial difference of voltage amplitude at the same transient decay time point through differential operation, and divides the spatial difference of voltage amplitude by the center geometric distance between adjacent independent receiving channels in the array to calculate the transient rate of change in the spatial dimension; the transient rates of change corresponding to each spatial channel position are arranged and combined in sequence according to the spatial arrangement order of the probe array to generate the spatial gradient transient feature vector.

6. The method for surface crack detection of mining motor spindle based on pulsed eddy current transient response according to claim 1, characterized in that, The spatial projection algorithm calculates the equivalent circumferential unfolded arc length by multiplying the known principal axis radius by the real-time rotation coordinates. A two-dimensional quasi-plane coordinate system is established using the equivalent circumferential unfolded arc length as the first dimension coordinate and the axial displacement as the second dimension coordinate. The spatial gradient transient feature vector is mapped to the two-dimensional quasi-plane coordinate system according to the corresponding coordinates, thereby eliminating the electromagnetic field geometric divergence effect caused by the principal axis cylindrical surface and converting the electromagnetic diffusion response on the cylindrical surface into a curvature-neutralized transient response dataset with a spatial distribution equivalent to a plane.

7. The method for surface crack detection of mining motor spindle based on pulsed eddy current transient response according to claim 1, characterized in that, The amplitude correction operation is to determine the dynamic adjustment parameters based on the lift-off offset; and to extract the transient gradient components in the first direction and the transient gradient components in the second direction along the two dimensions of the two-dimensional quasi-plane coordinate system from the coordinate unfolded quasi-plane transient response dataset. The amplitude of the transient gradient component in the first direction and the transient gradient component in the second direction are corrected using the dynamic adjustment parameters to obtain the corrected first direction component and the second direction component, which constitute an orthogonal gradient feature set. The scattering vector deflection angle is calculated by combining the arctangent function with the corrected second directional component and the first directional component, and the square root of the sum of the squares of the corrected first directional component and the second directional component is calculated as the vector magnitude of the orthogonal gradient feature set. If the corrected first directional component and the second directional component are both below the preset noise threshold, the corresponding sampling point is marked as a low confidence point; the vector magnitude of the orthogonal gradient feature set, the deflection angle of the scattering vector, and the corresponding first-dimensional coordinates and second-dimensional coordinates are merged to generate the crack multidimensional feature matrix.

8. The method for surface crack detection of mining motor spindle based on pulsed eddy current transient response according to claim 7, characterized in that, The process of determining the dynamic adjustment parameters involves extracting the initial response time window of the multi-channel induced voltage transient timing signal after the pulse excitation is turned off, and calculating the sum of squares of the discrete voltage amplitudes within the initial response time window as the transient peak energy. The standard peak energy is calculated using a pre-stored defect-free reference signal within the same time window; the standard peak energy is divided by the transient peak energy to calculate the amplitude attenuation ratio caused by the lift-off effect. The amplitude attenuation ratio and the liftoff offset are normalized to obtain the normalized attenuation ratio and the normalized liftoff offset; Using the natural constant as the base and the normalized lift-off offset as the exponent, a power operation is performed to calculate and generate the spatial diffusion compensation factor; The normalized attenuation ratio is multiplied by the spatial diffusion compensation factor to generate the dynamic adjustment parameter.

Images