Track crack in-transit nondestructive testing and quantitative evaluation method based on electromagnetic eddy current

Abstract

The invention relates to the technical field of track nondestructive testing and crack quantitative evaluation, in particular to a track crack in-transit nondestructive testing and quantitative evaluation method based on electromagnetic eddy current, which comprises the following steps of: synchronously acquiring eddy current response voltage signals of a dual-band sensor array through a high-speed data acquisition module, associating track space position coordinates, and determining the track crack in-transit nondestructive testing and quantitative evaluation. Dynamically adjusting the sampling frequency according to the advancing speed; vibration noise is eliminated through wavelet noise reduction, in-phase and orthogonal components of all frequency bands are decoupled and separated through a digital lock-in amplifier, and a three-dimensional feature vector set is constructed; a double-cascade evaluation model is input, a first-stage convolutional neural network identifies crack existence and a spatial distribution mode, and a second-stage physical constraint inversion network calculates crack depth, length and inclination angle; risk grades are divided according to a multi-parameter decision rule, a structured report containing standard railway mileage information is generated and transmitted to a ground system, the detection accuracy and stability are improved, and a reliable basis is provided for track maintenance.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing and quantitative assessment of cracks in rail tracks, specifically, to a method for in-transit nondestructive testing and quantitative assessment of rail cracks based on electromagnetic eddy currents. Background Technology

[0002] Nondestructive testing (NDT) and quantitative crack assessment of railway tracks is an important technology, specifically applied to the in-transit detection and safety assessment of track cracks. Its core principle is to improve detection accuracy and assessment reliability by precisely processing detection signals and quantifying crack parameters, thus meeting the core needs of railway operation and maintenance for early detection and handling of track safety hazards. In in-transit track crack detection, the speed of the detection equipment changes dynamically, track vibration generates interference noise, and adjacent sensors in the sensor array are prone to electromagnetic coupling. These factors collectively cause distortion of the acquired eddy current response voltage signal, making it impossible to effectively separate crack-sensitive characteristic quantities. This affects the accurate calculation of key parameters such as crack depth, length, and inclination angle, making it difficult to scientifically classify risk levels and posing safety hazards to track operation and maintenance. To solve this technical problem, we provide an in-transit NDT and quantitative assessment method for track cracks based on electromagnetic eddy currents. Summary of the Invention

[0003] The purpose of this invention is to provide a method for nondestructive testing and quantitative evaluation of track cracks in transit based on electromagnetic eddy currents, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, one of the objectives of this invention is to provide a method for in-transit nondestructive testing and quantitative evaluation of track cracks based on electromagnetic eddy currents, comprising the following steps:

[0005] S1. A high-speed data acquisition module is used to synchronously acquire the eddy current response voltage signals of each sensor unit, and the encoder is used to associate the spatial position coordinates of the acquisition point relative to the track in real time. At the same time, the signal sampling frequency is dynamically adjusted according to the travel speed fed back by the speed sensor.

[0006] S2. The wavelet denoising algorithm is used to preprocess the eddy current response voltage signal to eliminate track vibration noise. The preprocessed eddy current response voltage signal is decoupled by a digital lock-in amplifier to separate the in-phase and quadrature components corresponding to each frequency band. Crack sensitive feature quantities are calculated based on the in-phase and quadrature components. The crack sensitive feature quantities are then used to construct a three-dimensional feature vector set according to the spatial position coordinates.

[0007] S3. Input the three-dimensional feature vector set into the dual-cascade evaluation model. The first stage uses a convolutional neural network to identify the existence and spatial distribution pattern of cracks in the three-dimensional feature vectors. The second stage calculates the crack parameters through a physical constraint inversion network. The crack depth is output by the phase offset and the rate of change of impedance amplitude through a pre-calibrated nonlinear mapping function. The crack length is calculated by the continuous distribution length of the three-dimensional feature vectors in space. The crack inclination angle is determined by the direction angle of the impedance change gradient. The risk level is automatically classified according to the crack depth.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0009] This invention utilizes a dual-band, layered electromagnetic eddy current sensor array, combined with phase synchronization control to eliminate electromagnetic coupling interference between adjacent sensors. It simultaneously covers the detection needs of surface and subsurface cracks on the track. A speed adaptive compensation mechanism dynamically adjusts the sampling frequency to ensure a constant sampling density per unit track length. Combined with an adaptive threshold wavelet denoising algorithm and digital phase-locked loop decoupling processing, it effectively eliminates track vibration noise, accurately separates in-phase and orthogonal components in each frequency band, extracts crack-sensitive features, and constructs a three-dimensional feature vector set. A dual-cascaded evaluation model accurately identifies the existence and direction of cracks through a convolutional neural network. A physical constraint inversion network, combined with an electromagnetic field analytical model, quantifies crack depth, length, and inclination angle. Multi-parameter decision rules scientifically classify risk levels, and an online self-calibration module dynamically optimizes detection accuracy. A structured report accurately correlates crack parameters with railway mileage and rapidly transmits them to the operation and maintenance system. This improves the accuracy of on-the-road track crack detection and parameter quantification, reduces the risk of misjudgment and missed detection, and provides timely and reliable technical support for safe railway operation and maintenance. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the overall workflow of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] Please see Figure 1 As shown, this embodiment provides a method for in-transit nondestructive testing and quantitative evaluation of track cracks based on electromagnetic eddy currents, including the following steps:

[0013] S1. A high-speed data acquisition module is used to synchronously acquire the eddy current response voltage signals of each sensor unit, and the encoder is used to associate the spatial position coordinates of the acquisition point relative to the track in real time. At the same time, the signal sampling frequency is dynamically adjusted according to the travel speed fed back by the speed sensor.

[0014] S2. Wavelet denoising algorithm is used to preprocess the eddy current response voltage signal to eliminate track vibration noise. The preprocessed eddy current response voltage signal is decoupled by digital lock-in amplifier to separate the in-phase and quadrature components corresponding to each frequency band. Crack sensitive feature quantities are calculated based on the in-phase and quadrature components. A three-dimensional feature vector set is constructed according to the spatial position coordinates of the crack sensitive feature quantities.

[0015] S3. Input the three-dimensional feature vector set into the dual-cascade evaluation model. The first stage uses a convolutional neural network to identify the existence and spatial distribution pattern of cracks in the three-dimensional feature vectors. The second stage calculates the crack parameters through a physical constraint inversion network. The crack depth is output by the phase offset and the rate of change of impedance amplitude through a pre-calibrated nonlinear mapping function. The crack length is calculated by the continuous distribution length of the three-dimensional feature vectors in space. The crack inclination angle is determined by the direction angle of the impedance change gradient. The risk level is automatically classified according to the crack depth.

[0016] The electromagnetic eddy current sensor array connected to the high-speed data acquisition module adopts a dual-band layered layout structure. The operating frequency range of the low-frequency sensor unit covers the subsurface crack detection requirements, and its U-shaped magnetic core is installed at a preset tilt angle in the rail waist area to enhance the magnetic field penetration capability. The operating frequency range of the high-frequency sensor unit covers the surface crack detection requirements, and its rectangular magnetic core symmetrically covers the working surface of the rail head. The detection areas of adjacent sensor units overlap in the track extension direction. The phase synchronization control module of the multi-channel signal generator eliminates the electromagnetic coupling interference between adjacent sensor units.

[0017] The operation of dynamically adjusting the signal sampling frequency includes a speed adaptive compensation mechanism. Based on the real-time feedback of the travel speed value from the speed sensor, the sampling rate of the high-speed data acquisition module is dynamically adjusted within a preset sampling frequency range through an interpolation algorithm, so that the density of signal sampling points within a unit track length remains constant. At the same time, the acquisition time of the encoder's mileage pulse trigger signal and the eddy current response voltage signal is synchronized.

[0018] The wavelet denoising algorithm employs an adaptive threshold optimization strategy. By analyzing the spectral distribution characteristics of track vibration noise, it performs dynamic threshold calculations on the coefficients of each frequency band after wavelet decomposition. The high-frequency noise coefficients are nonlinearly compressed according to energy weights, while the low-frequency signal coefficients retain their full amplitude. The final reconstructed denoised signal retains the original characteristics of the crack-sensitive frequency band.

[0019] The decoupling operation of the digital lock-in amplifier includes two-stage processing of quadrature demodulation and frequency band separation. The first-stage mixer multiplies the noise-reduced eddy current response voltage signal with the local reference signal to generate an in-phase component. The second-stage mixer multiplies the phase-shifted reference signal with the voltage signal to generate a quadrature component. For different frequency band signals in the composite excitation, a configurable bandpass filter bank is used to separate the in-phase and quadrature components corresponding to each frequency band.

[0020] The decoupling operation of the digital lock-in amplifier includes two-stage processing of quadrature demodulation and frequency band separation. The first-stage mixer multiplies the noise-reduced eddy current response voltage signal with the local reference signal to generate an in-phase component, and the second-stage mixer multiplies the phase-shifted reference signal with the voltage signal to generate a quadrature component. For different frequency band signals in the composite excitation, a configurable bandpass filter bank is used to separate the in-phase and quadrature components corresponding to each frequency band.

[0021] The first-level convolutional neural network of the dual-cascade evaluation model adopts a multi-scale feature fusion structure. It extracts the spatial distribution pattern of the three-dimensional feature vector set through parallel convolutional kernel groups and outputs the crack existence probability and orientation classification results. The second-level physical constraint inversion network embeds the crack depth mapping relationship derived from the electromagnetic field analytical model and forces the crack depth output value to meet the preset function constraints of phase offset and impedance amplitude change rate.

[0022] The crack length is calculated using a spatial continuity discrimination algorithm. After identifying the crack region, the impedance amplitude change rate is differentially calculated along the track extension direction. The lengths of spatial intervals that continuously exceed the sensitivity threshold are accumulated as the crack length. The crack inclination angle is calculated using a gradient direction detection algorithm to calculate the spatial angle of the impedance change vector, and cross-validation is performed by combining the orientation classification results of the first stage output.

[0023] The risk level classification introduces a multi-parameter decision rule. When the crack depth reaches the intermediate risk threshold and the crack inclination angle exceeds the critical angle, the risk level is automatically upgraded. When the structured report is generated, the spatial location coordinates are converted into standard railway mileage information. Crack parameters and risk levels are encapsulated into a tree data structure and transmitted to the ground engineering management system through a dedicated wireless communication protocol.

[0024] It also includes an online self-calibration module. When the detection system enters the defect-free track section, it automatically collects the eddy current response signal under the current environmental parameters to update the defect-free reference impedance database. When the output confidence of the physical constraint inversion network is lower than the set threshold, it triggers the parameter optimization process of the crack depth mapping relationship and adjusts the function constraint coefficient through the error backpropagation mechanism.

[0025] It needs further explanation that, before starting the high-speed data acquisition module to acquire signals, in order to simultaneously cover the crack detection needs of both the track surface and subsurface, and to ensure that cracks at different locations and depths can be accurately captured, avoiding detection blind spots caused by a single frequency band or layout, the electromagnetic eddy current sensor array connected to the high-speed data acquisition module adopts a dual-frequency layered layout structure. The dual-frequency layered layout structure refers to dividing the sensor array into two groups according to the operating frequency: low frequency and high frequency, which are deployed for track subsurface and surface cracks respectively, forming a spatially layered and functionally complementary detection system.The low-frequency sensor unit operates within a frequency range that covers the needs of subsurface crack detection. This low-frequency sensor unit refers to an electromagnetic eddy current sensor with an operating frequency concentrated between 1 kHz and 10 kHz. This frequency band has strong electromagnetic field penetration capabilities, reaching depths of 1 to 5 millimeters below the track surface, perfectly matching the needs of subsurface crack detection. Subsurface crack detection requires accurately identifying cracks below the track surface that do not penetrate into the structure. These cracks are easily overlooked but can affect track strength over time. The U-shaped magnetic core is installed at a preset tilt angle in the rail web area to enhance magnetic field penetration. The U-shaped magnetic core is the core magnetic circuit component of the low-frequency sensor; the U-shaped structure can concentrate magnetic field energy, improving... To enhance the magnetic field strength, the preset tilt angle, calibrated through extensive experiments, is set to 30-45 degrees. This angle ensures the magnetic field direction is obliquely intersecting the rail surface, reducing reflection losses from the rail surface material and allowing the magnetic field to penetrate more easily to the subsurface. The rail waist region, the central connecting part of the track, experiences complex stresses and is prone to subsurface cracks. Installing the sensor here allows for precise targeting. Enhanced magnetic field penetration is achieved through optimization of the core structure and installation angle, slowing down the attenuation rate of the electromagnetic field within the track. This ensures effective sensing of eddy current changes caused by subsurface cracks. The high-frequency sensor unit's operating frequency range covers the surface crack detection requirements. Set at 100 kHz to 1 MHz, this frequency band concentrates the electromagnetic field within the 0 to 1 mm range of the track surface, exhibiting extremely high sensitivity to micro-cracks. The surface crack detection requirement refers to identifying cracks on the exposed surface of the track. These cracks are directly exposed to the external environment and are easily affected by rain, wear, etc., leading to rapid expansion. Priority detection and early warning are necessary. The rectangular magnetic core symmetrically covers the rail head working surface. The rectangular magnetic core features uniform magnetic field distribution and wide coverage, allowing the electromagnetic field to act evenly on the rail head working surface. The rail head working surface is a critical area of ​​contact between the track and the train wheels, subjected to long-term compression and friction, making it a high-risk area for surface cracks. Symmetrical coverage refers to the installation of two high-frequency sensors respectively... On both sides of the rail head, the detection ranges are symmetrically superimposed to ensure that there are no blind spots on the working surface of the rail head and to avoid missed detection of edge cracks due to unilateral installation. In addition, the detection areas of adjacent sensor units have an overlapping layout in the direction of track extension. Adjacent sensor units include sensors arranged one after the other in the same frequency band and adjacent sensors in different frequency bands. The overlapping layout of the detection area means that there is a 10% to 15% overlap between the detection end of the previous sensor and the detection start end of the next sensor. This ratio is determined by calculation based on the detection radius of the sensor. This layout can effectively make up for the limitations of the detection range of a single sensor, avoid missed detection of longitudinal cracks in the track due to the spacing between sensors, and ensure the continuity of detection.The phase synchronization control module of the multi-channel signal generator eliminates electromagnetic coupling interference between adjacent sensor units. The multi-channel signal generator is the core component that provides excitation signals to each sensor unit and supports the simultaneous output of excitation signals in multiple independent frequency bands. The phase synchronization control module is a timing control unit built into the generator, which can precisely adjust the phase of the excitation signal of each channel. Electromagnetic coupling interference refers to the mutual influence of the electromagnetic fields generated by adjacent sensors when they are working, which leads to distortion of the eddy current response signal. To eliminate interference, the phase synchronization control module adjusts the phase of the excitation signals of adjacent sensors to be 180 degrees out of phase according to the layout of the sensors, so that the interfering electromagnetic fields generated by them cancel each other out, while keeping the sensitivity of each sensor to cracks unaffected, and ensuring that the acquired eddy current response voltage signal is true and reliable. During the synchronous acquisition of eddy current response voltage signals and associated spatial coordinates, the travel speed of the detection equipment varies with track conditions and power output. To ensure consistent sampling data density per unit track length and avoid insufficient sampling points and loss of crack features due to excessive speed, or data redundancy and excessive storage and computing resources due to excessively slow speed, a dynamic adjustment of the signal sampling frequency is implemented. This includes a speed adaptive compensation mechanism, a closed-loop control logic that automatically adjusts sampling parameters based on real-time travel speed. Its core is maintaining data uniformity through the linkage of speed feedback and sampling rate adjustment. Based on the real-time feedback travel speed value from the speed sensor (a wheel-type incremental encoder mounted on the wheel axle of the detection equipment), the speed sensor captures wheel rotation speed in real time and converts it into the travel speed value of the detection equipment using the wheel circumference. The sampling frequency is 100 Hz to ensure timely response to speed changes. Real-time feedback refers to the transmission of speed data to the control module every 10 milliseconds via a high-speed data bus, providing real-time basis for sampling rate adjustment. The sampling rate of the high-speed data acquisition module is dynamically adjusted within a preset sampling frequency range using an interpolation algorithm. The interpolation algorithm employs linear interpolation, which smoothly calculates the required sampling rate based on the difference between the current speed and the reference speed, avoiding signal distortion caused by sudden changes in the sampling rate. The preset sampling frequency range is calibrated to 10 kHz to 100 kHz based on detection requirements. This range satisfies both redundancy control at low speeds and covers the minimum sampling requirements at high speeds. The core logic of dynamically adjusting the sampling rate is: when the speed increases, the sampling rate is increased proportionally; when the speed decreases, the sampling rate is decreased proportionally, ensuring that the sampling rate is positively correlated with the speed.The signal sampling point density per unit track length is kept constant. The signal sampling point density per unit track length refers to the number of eddy current response voltage signal sampling points per meter of track. It is experimentally calibrated to 1000 points / meter. This density ensures that even a millimeter-sized microcrack corresponds to at least 2 to 3 sampling points, completely preserving the signal change characteristics caused by the crack. The constant density is achieved by adjusting the ratio of the interpolation algorithm. For example, when the speed is increased from 30 km / h to 60 km / h, the sampling rate is increased from 50 kHz to 100 kHz, always maintaining the standard of 1000 sampling points per meter. Simultaneously, the mileage pulse trigger signal of the encoder is synchronized with the acquisition time of the eddy current response voltage signal. The encoder, also known as the mileage encoder, is installed on the driven wheel axle of the detection equipment. For each rotation of the wheel, the encoder outputs a fixed number of mileage pulses, each pulse corresponding to a fixed track length. Synchronizing the mileage pulse trigger signal with the acquisition time means that the pulse signal output by the encoder is used as the acquisition trigger signal for the high-speed data acquisition module. When each pulse arrives, the acquisition module immediately acquires an eddy current response voltage signal, ensuring that each acquired signal sampling point can accurately correspond to the specific position coordinates of the track, avoiding crack location errors caused by signal-position misalignment. The acquired eddy current response voltage signal inevitably contains track vibration noise, mainly from wheel-rail friction and track joint impact during the movement of the detection equipment. Its frequency distribution is complex and can mask the weak signal characteristics corresponding to the crack. Therefore, a wavelet denoising algorithm is needed to preprocess the signal, and this algorithm uses an adaptive threshold optimization strategy to improve the denoising targeting and avoid signal distortion or noise residue caused by a fixed threshold. By analyzing the spectral distribution characteristics of track vibration noise, which refers to the energy distribution pattern of the noise signal in different frequency ranges, the analysis first performs a Fast Fourier Transform on the original acquired signal to convert the time-domain signal into a frequency-domain signal, identifying the frequency bands where noise energy is concentrated and the sensitive frequency bands corresponding to crack signals, providing a basis for subsequent targeted frequency band processing. Dynamic threshold calculation is performed on the coefficients of each frequency band after wavelet decomposition. Wavelet decomposition decomposes the original signal into coefficients of multiple scales according to frequency from low to high. A 5-level decomposition is performed using the db6 wavelet basis function, where levels 1 and 2 are high-frequency coefficients (mainly corresponding to vibration noise), and levels 3 to 5 are low-frequency coefficients (mainly corresponding to crack signals). Dynamic threshold calculation means that for the coefficients of each frequency band, a specific threshold is calculated based on the mean and variance of the energy of that frequency band, rather than using a uniform threshold. For example, the threshold for the high-frequency noise band is set to 30% of the maximum value of the coefficients in that frequency band, and the threshold for the low-frequency signal band is set to 10% of the maximum value, ensuring that the threshold matches the signal characteristics of each frequency band.The high-frequency noise coefficients are nonlinearly compressed based on energy weights. These high-frequency noise coefficients are the high-scale coefficients after wavelet decomposition, and their energy primarily originates from vibration noise. Energy weights refer to the proportion of energy of each high-frequency coefficient to the total energy of that frequency band; a higher weight indicates stronger noise interference. Nonlinear compression employs a soft-thresholding method, setting coefficients with absolute values ​​less than the corresponding frequency band threshold to 0, and retaining coefficients with absolute values ​​greater than the threshold after subtracting the threshold. Additional gradient compression is applied to coefficients with high energy weights to avoid excessive retention of noise components. Low-frequency signal coefficients retain their complete amplitude. These coefficients contain eddy current variations caused by cracks. While their amplitudes are weak, they carry crucial detection information; therefore, they are not compressed or truncated, preserving their original amplitude and phase information to ensure that crack-related details are not lost during subsequent feature extraction. The final reconstructed denoised signal retains the original characteristics of the crack-sensitive frequency band. Reconstruction involves combining the processed frequency band coefficients according to the inverse wavelet transform formula to restore the time-domain signal. Retaining the original characteristics of the crack-sensitive frequency band means that during the reconstruction process, through precise control of the dynamic threshold, the signal amplitude and phase changes of the crack-sensitive frequency band (20 kHz to 50 kHz) are completely preserved, and only abnormal peak noise in this frequency band is removed. The final output denoised signal can clearly present the signal fluctuation corresponding to the crack, and can minimize the interference of vibration noise.

[0026] After eliminating track vibration noise using wavelet denoising algorithms, the denoised eddy current response voltage signal still contains superimposed components of different frequency bands and complex phase information. These intertwined components interfere with the extraction of crack-sensitive features. Therefore, it is necessary to separate key signal components through decoupling operation of a digital lock-in amplifier. This decoupling operation includes two-stage processing: quadrature demodulation and frequency band separation. Quadrature demodulation decomposes the complex signal into two independent components, in-phase and quadrature, by multiplying it with a quadrature reference signal, which can accurately extract the amplitude and phase features of the signal. Frequency band separation, based on quadrature demodulation, divides the signal into intervals according to the signal frequency, separating the components corresponding to different frequency bands to meet the detection requirements of dual-frequency sensors. The two-stage processing involves performing quadrature demodulation first and then frequency band separation to ensure that the processing is orderly and accurate. The first-stage mixer multiplies the denoised eddy current response voltage signal with the local reference signal to generate an in-phase component. The first-stage mixer is the core component of the digital lock-in amplifier responsible for the first-stage signal modulation, featuring low noise and high linearity to ensure distortion-free signal multiplication. The local reference signal is a standard sinusoidal signal generated by a signal generator, with the same frequency and phase as the sensor excitation signal. Its frequency strictly matches the operating frequency of the dual-band sensor (1-10 kHz for low frequency, 100 kHz-1 MHz for high frequency), providing a reference for demodulation. The in-phase component refers to the signal component that is in phase with the local reference signal, reflecting the amplitude changes in the eddy current response signal that are in phase with the excitation signal, and is an important carrier of crack characteristics. In practice, the denoised eddy current response voltage signal is input to one input terminal of the first-stage mixer, and the local reference signal is input to the other. The mixer performs a multiplication operation on the two signals, and the result is filtered by a low-pass filter to obtain a smooth in-phase component, which retains the amplitude characteristics and in-phase phase information of the original signal. The secondary mixer multiplies the phase-shifted reference signal with the voltage signal to generate a quadrature component. The secondary mixer has the same performance as the primary mixer and is used to perform the second-stage modulation. The phase-shifted reference signal is the signal obtained by adjusting the local reference signal by 90 degrees using a phase shifter, forming an orthogonal relationship with the original reference signal (90-degree phase difference). The quadrature component is a signal component with a phase perpendicular to the local reference signal, supplementing the phase characteristics not covered by the in-phase component. The combination of the two can completely restore the complex characteristics of the original signal. In operation, the local reference signal is first phase-shifted by 90 degrees using a high-precision phase shifter, ensuring a phase shift accuracy error of no more than 0.1 degrees. Then, the phase-shifted reference signal is input to the secondary mixer, along with the noise-reduced eddy current response voltage signal. After multiplication, the signal is filtered by a low-pass filter and outputs the quadrature component, which, together with the in-phase component, forms a complete quadrature demodulation result.For signals in different frequency bands in composite excitation, a configurable bandpass filter bank is used to separate the in-phase and quadrature components corresponding to each frequency band. The different frequency band signals in composite excitation refer to the eddy current response signals corresponding to the low-frequency (1-10 kHz) and high-frequency (100 kHz-1 MHz) excitations generated by the dual-band sensor array. The two types of signals are superimposed in the in-phase and quadrature components and need to be further separated. The configurable bandpass filter bank is a module composed of two independent bandpass filters. The center frequency and bandwidth of each filter can be configured by software to adapt to low-frequency and high-frequency signals respectively. Separating the in-phase and quadrature components corresponding to each frequency band means inputting the in-phase and quadrature components obtained by quadrature demodulation into the filter bank to separate the four independent components: low-frequency in-phase, low-frequency quadrature, high-frequency in-phase, and high-frequency quadrature. During implementation, the filter bank parameters are first configured: the center frequency of the low-frequency filter is set to 5 kHz and the bandwidth to 4 kHz, covering the low-frequency range of 1-10 kHz; the center frequency of the high-frequency filter is set to 500 kHz and the bandwidth to 800 kHz, covering the high-frequency range of 100 kHz-1 MHz. Then, the in-phase components are input into the two filters respectively, and the low-frequency in-phase component and the high-frequency in-phase component are output. Similarly, the quadrature components are input into the filters, and the low-frequency quadrature component and the high-frequency quadrature component are output, so as to achieve complete separation of signal components in different frequency bands and provide accurate data for subsequent crack-sensitive characteristic quantity calculation. After completing quadrature demodulation and frequency band separation to obtain in-phase and quadrature components of different frequency bands, in order to further purify the signal and avoid the impact of cross-interference between different frequency bands on feature extraction, the decoupling operation of the digital lock-in amplifier still takes quadrature demodulation and frequency band separation as the core. The logic of this two-stage processing is consistent with the previous one. First, the phase characteristics of the signal are split through quadrature demodulation, and then the frequency range of the signal is divided through frequency band separation to ensure that each step of processing can specifically remove interference and retain effective information. The first-stage mixer multiplies the denoised eddy current response voltage signal with the local reference signal to generate an in-phase component. The definitions of the first-stage mixer, local reference signal, and in-phase component are the same as before, with only the synchronization being enhanced in the operational details: the generation of the local reference signal is strictly synchronized with the sensor excitation signal, and the phase synchronization control module ensures that the phase difference between the two is 0, avoiding distortion of the in-phase component due to phase deviation; the low-pass filter after the mixing operation adopts a Chebyshev Type I filter with an attenuation rate of 60 dB per octave, which can effectively filter out the high-frequency harmonics generated by the multiplication operation, ensuring that the output in-phase component is smooth and free of noise. The secondary mixer multiplies the phase-shifted reference signal with the voltage signal to generate quadrature components. The phase-shifting process uses digital phase-shifting technology, and precise phase adjustment is achieved through an FPGA chip. Compared with analog phase shifters, digital phase shifting can avoid phase deviation caused by temperature drift, ensuring that the phase-shifted reference signal is strictly orthogonal to the original signal. The multiplication operation of the secondary mixer is synchronized with the first-stage mixer in timing, and the low-pass filter parameters after the operation are consistent with those of the first stage, ensuring that the amplitude scales of the quadrature components and the in-phase components are uniform, which facilitates the fusion processing during subsequent characteristic quantity calculations.For signals in different frequency bands in composite excitation, a configurable bandpass filter bank is used to separate the in-phase and quadrature components corresponding to each frequency band. The configurable bandpass filter bank adds a frequency band calibration function. Before separating the signals, a test signal of known frequency is input through a standard signal source to calibrate the center frequency and bandwidth of the filter to ensure filtering accuracy. During the separation process, the output signal of the filter bank is fed back to the control module in real time. If the amplitude of a certain frequency band component is detected to be lower than the sensitivity threshold, the gain of the filter is automatically adjusted to avoid the loss of weak signals. Finally, the separated low-frequency in-phase, low-frequency quadrature, high-frequency in-phase, and high-frequency quadrature components are stored in different data buffers and managed according to frequency band and component type for easy retrieval and calculation as needed. After constructing the three-dimensional feature vector set, in order to achieve a precise transition from feature recognition to parameter quantification, the vector set needs to be input into a dual-cascade evaluation model. This model works collaboratively through two levels of networks to first determine the existence and distribution pattern of cracks, and then quantitatively calculate crack parameters. The first-level convolutional neural network adopts a multi-scale feature fusion structure. The multi-scale feature fusion structure refers to extracting feature information at different scales from the three-dimensional feature vector set by setting convolution kernels of different sizes, and then fusing these features to ensure the comprehensiveness of feature extraction. In practice, the input layer of the convolutional neural network receives a three-dimensional feature vector set (dimensions of spatial location, in-phase components, and orthogonal components). Then, three parallel convolutional layers are set up: the first convolutional layer uses a small kernel (3×3×3) to extract fine-grained features such as crack edges and minor abrupt changes; the second convolutional layer uses a medium kernel (5×5×5) to extract local crack distribution patterns; and the third convolutional layer uses a large kernel (7×7×7) to extract the overall spatial distribution features of the cracks. The output features of the three convolutional layers are fused by a concatenation layer to form a comprehensive feature map containing multi-scale information, providing rich evidence for subsequent identification. The spatial distribution pattern of the three-dimensional feature vector set is extracted through a parallel set of convolutional kernels. This parallel set of convolutional kernels refers to the combination of the three different sizes of convolutional kernels, which work in parallel to extract features simultaneously, avoiding feature loss caused by serial processing. The spatial distribution pattern refers to the arrangement of features corresponding to cracks in the three-dimensional feature vector set in terms of spatial location. For example, the features of longitudinal cracks are continuously distributed along the track extension direction, while the features of transverse cracks are concentrated laterally. During extraction, each convolutional kernel group traverses the three-dimensional feature vector set in a sliding window manner, with the window stride set to 1 to ensure that no feature at any spatial location is missed. After convolution, the nonlinear expressive power of the features is enhanced by an activation function (using the ReLU function), and then the feature dimension is compressed by a pooling layer (using max pooling) to retain key information. The final output is a feature map that clearly reflects the spatial distribution law of cracks.The output includes the probability of crack presence and its classification. The probability of crack presence refers to the confidence level of the model in determining the presence of a crack at the current spatial location, ranging from 0 to 1, with values ​​closer to 1 indicating a higher probability of crack presence. The classification result divides cracks into three categories based on their spatial direction: longitudinal (along the track extension direction), transverse (perpendicular to the track extension direction), and oblique (at an angle of 30-60 degrees to the track extension direction), providing directional reference for subsequent crack parameter calculations. During output, the fused feature map is mapped to a one-dimensional vector through a fully connected layer, and then the probability of each category is calculated using the Softmax function: the first category represents no crack, and the second to fourth categories represent longitudinal, transverse, and oblique cracks, respectively. If the probability of no crack is the highest and exceeds 0.8, the region is determined to be crack-free. If the probability of a certain type of crack is the highest and exceeds 0.7, the probability of that crack presence and its corresponding direction classification result are output, providing input for the second-level network. The second-level physical constraint inversion network embeds the crack depth mapping relationship derived from the electromagnetic field analytical model. This network is an inversion algorithm that integrates physical laws, preventing deviations from engineering reality by embedding constraints from the actual physical model. The electromagnetic field analytical model, based on Maxwell's equations, is a mathematical model describing the interaction between electromagnetic eddy currents and track cracks, accurately reflecting the intrinsic relationship between crack depth, phase shift, and impedance amplitude change rate. The crack depth mapping relationship is a quantitative correspondence obtained by simulating the phase shift and impedance amplitude change rate corresponding to cracks of different depths using the electromagnetic field analytical model. During the embedding process, an track crack simulation model is first constructed using electromagnetic field simulation software, setting multiple simulation scenarios with crack depths ranging from 0.1 mm to 10 mm. The phase shift and impedance amplitude change rate are calculated for each scenario, forming a depth-phase-impedance mapping dataset. This dataset is then used as a physical constraint and embedded into the loss function of the inversion network, ensuring that the network training process consistently adheres to this mapping relationship and that the output crack depth conforms to physical laws. The crack depth output value is forced to satisfy a preset function constraint on the phase shift and impedance amplitude change rate. This preset function constraint is a nonlinear function fitted to a mapping dataset, clearly defining the quantitative relationship between crack depth and these constraints. Forcing the satisfaction of this constraint means penalizing the output result that deviates from this function through a loss function, ensuring that the crack depth output by the network always falls within the function constraint range. During implementation, the inversion network receives the crack presence region and orientation classification results from the first-level output, along with the corresponding phase shift and impedance amplitude change rate data. The hidden layers of the network learn the mapping relationship between features and depth through a multilayer perceptron, while the output layer outputs the predicted crack depth. The loss function consists of two parts: one is the mean square error between the predicted depth and the simulation data, and the other is a penalty term for deviations of the predicted depth from the preset function. The network parameters are optimized through a backpropagation mechanism, ultimately ensuring that the output crack depth both conforms to the data characteristics and strictly satisfies the physical function constraint, thus ensuring the accuracy of the quantitative assessment.

[0027] After the crack depth is calculated by the second-level physical constraint inversion network of the dual-cascade evaluation model, the crack length and inclination angle need to be accurately calculated to fully quantify the key parameters of the crack. The crack length calculation employs a spatial continuity discrimination algorithm. This algorithm is based on the continuous distribution characteristics of crack features in space, calculating the actual length by determining the continuous intervals of feature signals. Its core principle is to eliminate discrete interference signals and retain the continuous features of the true crack. After identifying the crack presence area, which is the spatial location range with a crack presence probability exceeding 0.7 output by the first-level convolutional neural network, the approximate extension direction of this area has been clearly defined through orientation classification, thus defining an effective range for length calculation. After identification, the system automatically extracts the impedance amplitude change rate data corresponding to all spatial locations within this area, eliminating redundant data from crack-free areas and focusing on the core calculation. The impedance amplitude change rate is differentially calculated along the track extension direction, which is the longitudinal direction of the track (train travel direction) and the main measurement direction of crack length. The impedance amplitude change rate is a key characteristic reflecting the degree of influence of the crack on electromagnetic eddy currents; the change rate in the cracked region is much higher than that in the uncracked region. The differential operation calculates the difference in impedance amplitude change rates between adjacent spatial locations to determine the continuity of the feature. The smaller the difference, the more continuous the feature, and the more likely it belongs to the same crack. During the calculation, the difference in impedance amplitude change rate between two adjacent sampling points is calculated point by point according to the spatial coordinate sequence (along the track extension direction), and the continuous intervals where the difference is less than a preset continuity threshold are recorded. The crack length is calculated by summing the lengths of consecutive spatial intervals exceeding the sensitivity threshold. The sensitivity threshold is a critical value for determining the presence of crack characteristics at a certain spatial location. It is calibrated using a large number of standard crack samples and is set to be three times the average rate of change of impedance amplitude in a defect-free area. Exceeding this threshold indicates the presence of a crack at that location. The length of a consecutive spatial interval refers to the spatial range in which multiple consecutive sampling points satisfy the condition that the rate of change of impedance amplitude exceeds the sensitivity threshold and the difference between adjacent points is less than the consecutive threshold. The actual track length corresponding to each sampling point is determined by the mileage pulse of the encoder. During the accumulation process, the system traverses all consecutive intervals, multiplies the number of sampling points in each interval by the track length corresponding to a single sampling point to obtain the length of each consecutive interval, and then sums the lengths of all consecutive intervals belonging to the same crack (if the interval spacing is less than 5 mm, it is determined to be a fracture interval of the same crack and is accumulated together), finally obtaining the actual length of the crack. The crack inclination angle is calculated by the gradient direction detection algorithm, which is a spatial angle of the impedance change vector. The gradient direction detection algorithm determines the crack inclination angle by calculating the direction angle of the impedance change vector in the spatial coordinate system. The impedance change vector is a two-dimensional vector composed of the change of the in-phase component and the change of the orthogonal component, and its direction is related to the crack propagation direction. The spatial angle is the angle between this vector and the trajectory extension direction (longitudinal), which is the crack inclination angle.During calculation, the impedance change gradient at the center of each window within the crack region is first calculated using a 3×3 sliding window, yielding the in-phase and orthogonal component gradients, which together constitute the impedance change vector. Then, the angle between this vector and the longitudinal direction of the track is calculated using the arctangent function, resulting in a preliminary crack inclination angle ranging from 0 to 90 degrees (0 degrees corresponds to a longitudinal crack, and 90 degrees to a transverse crack). Cross-validation is then performed using the crack orientation classification results from the first-level output. These orientation classification results, representing the crack orientation (longitudinal, transverse, or oblique) output by the first-level convolutional neural network, provide a directional reference for inclination angle verification. Cross-validation involves comparing the orientation classification results with the calculated inclination angle to correct inclination angle errors and ensure accurate results. During verification, if the orientation is longitudinal, the corrected inclination angle should be within the range of 0-30 degrees; if it is transverse, the inclination angle should be within the range of 60-90 degrees; and if it is oblique, the inclination angle should be within the range of 30-60 degrees. If the initially calculated inclination angle exceeds the corresponding range, the weight of the gradient calculation is adjusted based on the orientation classification result, and the inclination angle is recalculated until the result is consistent with the orientation classification, ultimately outputting an accurate crack inclination angle. After completing the quantitative calculation of key parameters such as crack depth, length, and inclination angle, in order to provide clear risk guidance for track maintenance, it is necessary to classify crack risk levels and generate structured reports. The risk level classification introduces multi-parameter decision rules. Multi-parameter decision rules are rules that comprehensively consider key parameters such as crack depth and inclination angle, rather than a single parameter, to determine the risk level. This can more comprehensively reflect the impact of cracks on track safety and avoid the one-sidedness of single-parameter judgment. When the crack depth reaches the intermediate risk threshold and the crack angle exceeds the critical angle, the risk level is automatically upgraded. The intermediate risk threshold is a critical crack depth value determined based on the bearing capacity limit of the track material and maintenance standards, and is determined through numerous track strength experiments. For example, the intermediate risk threshold for rails is set at 3 mm; beyond this depth, cracks tend to propagate rapidly. The critical angle refers to the angle at which the crack angle significantly increases its impact on the force transmission of the track; it is set at 45 degrees. Cracks that are oblique or transverse beyond this angle are more likely to cause stress concentration. The automatic upgrade of the risk level means that the risk level is initially divided into three levels: low, medium, and high based on the crack depth (depth < 1 mm is low risk, 1-3 mm is medium risk, and > 3 mm is high risk). If the angle of a medium-risk crack exceeds 45 degrees, it is automatically upgraded to high risk. If the angle of a low-risk crack exceeds 45 degrees and the depth is close to 1 mm, it is upgraded to medium risk, ensuring that the risk assessment accurately reflects the actual safety hazards. When generating a structured report, spatial location coordinates are converted into standard railway mileage information. The structured report is a standardized report containing key information such as crack parameters, risk level, and location information. Spatial location coordinates are the coordinates of the acquisition point associated with the encoder relative to the starting point of the track. Standard railway mileage information is a universal location identifier for the railway system, consisting of line number, mileage, and offset, which facilitates rapid positioning by track maintenance personnel.During conversion, the system incorporates a mileage benchmark database for the target track line. This database stores the standard mileage corresponding to the track's starting point, the mapping relationship between track alignment and mileage, and calculates the corresponding standard railway mileage by substituting the spatial coordinates of the collected points into the mapping relationship. It also marks the left and right rails and their offsets to ensure accurate and traceable location information. Crack parameters and risk levels are encapsulated into a tree-like data structure. This hierarchical structure facilitates data storage, transmission, and parsing. The root node is the track crack detection report; first-level child nodes contain basic information (track number, detection time), crack parameters (depth, length, inclination angle, mileage location), and risk level (level name, judgment criteria); second-level child nodes contain the specific values ​​and units of each parameter, such as depth: 2.5 mm, length: 15 mm, etc. During encapsulation, the system categorizes all data into the tree structure according to a preset hierarchy and adds a data verification code to ensure that data is not lost or tampered with during transmission. Transmission is conducted via a dedicated wireless communication protocol to the ground maintenance management system. This protocol is optimized for track inspection scenarios, employing either the railway-specific GSM-R protocol or a 5G industrial private network protocol. It features strong anti-interference capabilities, low transmission latency (<1 second), and long transmission distance, making it suitable for the mobile inspection scenarios of track inspection vehicles. The ground maintenance management system is the backend system used by the maintenance department to manage track inspection data and issue maintenance tasks. During transmission, the encapsulated tree-structured data is first compressed using a data compression algorithm to reduce bandwidth consumption, and then sent to the ground base station via the wireless communication module on the inspection vehicle. The base station forwards the data to the ground maintenance management system, which then decompresses and verifies the data, records the crack information in the database, and triggers corresponding early warning prompts (high-risk cracks are immediately displayed as pop-up notifications, and medium-risk cracks are included in the maintenance plan), ensuring timely response from maintenance personnel. To ensure the long-term accuracy and stability of the detection system and avoid parameter drift caused by environmental changes, equipment aging, and other factors, this method also includes an online self-calibration module. The online self-calibration module is a closed-loop module that can monitor the detection accuracy in real time and automatically adjust the system parameters. It can maintain the detection accuracy without manual intervention. The core of the module is to achieve self-calibration by updating the benchmark data and optimizing the model parameters.When the detection system enters a defect-free track section, it automatically collects eddy current response signals under the current environmental parameters to update the defect-free reference impedance database. A defect-free track section refers to a section confirmed by historical testing to be free of cracks and damage. The system determines entry into this section based on the impedance amplitude change rate at 500 consecutive sampling points being less than half the sensitivity threshold, and the probability of crack presence output by the first-level convolutional neural network being less than 0.1. Environmental parameters include temperature, humidity, and track surface cleanliness (collected with sensor assistance). These parameters affect the eddy current response signal and must be recorded synchronously. The defect-free reference impedance database stores the characteristics of eddy current response signals under different environmental parameters, providing a benchmark reference for subsequent crack determination. The database is categorized by environmental parameters, and each record contains features such as in-phase components, quadrature components, and impedance amplitude. During updates, the system collects eddy current response signals at a frequency of 100 Hz in the defect-free section, synchronously records the current environmental parameters, calculates the mean and variance of the signal characteristics, and replaces the old data in the database under the same environmental parameters, ensuring that the benchmark data matches the current operating conditions. When the output confidence of the physically constrained inversion network falls below a set threshold, the parameter optimization process for the crack depth mapping relationship is triggered. The output confidence of the physically constrained inversion network is an indicator of the reliability of the network's crack depth prediction. It is obtained by calculating the degree of matching between the predicted value and simulation data and historical measured data, with a value ranging from 0 to 1. The closer to 1, the higher the reliability. The set threshold is experimentally calibrated to 0.8. If it is lower than this threshold, it indicates that the crack depth output by the network may be biased, and the parameters need to be optimized. The parameter optimization process for the crack depth mapping relationship is a process of adjusting the coefficients of the nonlinear mapping function derived from the electromagnetic field analytical model to make the mapping relationship more accurate. When triggered, the system first marks the detection data corresponding to low confidence and extracts the in-phase component, orthogonal component, phase offset, and other features corresponding to the data as optimization samples. The function constraint coefficients are adjusted through the error backpropagation mechanism. The error backpropagation mechanism is an algorithm that transmits the prediction error backward from the network output to the input, thereby adjusting the model parameters. It can accurately locate and optimize the parameters that cause the error. The function constraint coefficients are key parameters of the nonlinear mapping function embedded in the physically constrained inversion network and directly affect the calculation accuracy of the crack depth. During adjustment, the prediction error of low-confidence samples (the difference between the predicted crack depth and the actual verification depth, which is supplemented by subsequent manual verification or high-confidence detection data) is used as the target. The contribution of each function constraint coefficient to the error is calculated layer by layer through the error backpropagation mechanism. The coefficients are adjusted according to the magnitude of the contribution (the larger the contribution, the larger the adjustment). After adjustment, the sample is substituted back to calculate the prediction value. The iteration is repeated until the prediction error is less than 0.05 mm and the output confidence is higher than the set threshold. The parameter optimization is completed to ensure that the output of the physical constraint inversion network is always accurate and reliable.

[0028] This invention synchronously acquires eddy current response voltage signals from a dual-band sensor array using a high-speed data acquisition module, correlates them with the spatial coordinates of the track, and dynamically adjusts the sampling frequency according to the travel speed. Vibration noise is eliminated through wavelet denoising, and the in-phase and quadrature components of each frequency band are decoupled and separated by a digital lock-in amplifier to construct a three-dimensional feature vector set. This is then input into a dual-cascaded evaluation model. The first-level convolutional neural network identifies the existence and spatial distribution pattern of cracks, while the second-level physical constraint inversion network calculates the crack depth, length, and inclination angle. Risk levels are classified according to multi-parameter decision rules, generating a structured report containing standard railway mileage information, which is then transmitted to the ground system. This improves the accuracy and stability of detection and provides a reliable basis for track maintenance.

[0029] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for in-transit nondestructive testing and quantitative assessment of track cracks based on electromagnetic eddy currents, characterized in that: Includes the following steps: S1. A high-speed data acquisition module is used to synchronously acquire the eddy current response voltage signals of each sensor unit, and the encoder is used to associate the spatial position coordinates of the acquisition point relative to the track in real time. At the same time, the signal sampling frequency is dynamically adjusted according to the travel speed fed back by the speed sensor. S2. The wavelet denoising algorithm is used to preprocess the eddy current response voltage signal to eliminate track vibration noise. The preprocessed eddy current response voltage signal is decoupled by a digital lock-in amplifier to separate the in-phase and quadrature components corresponding to each frequency band. Crack sensitive feature quantities are calculated based on the in-phase and quadrature components. The crack sensitive feature quantities are then used to construct a three-dimensional feature vector set according to the spatial position coordinates. S3. Input the three-dimensional feature vector set into the dual-cascade evaluation model. The first stage uses a convolutional neural network to identify the existence and spatial distribution pattern of cracks in the three-dimensional feature vectors. The second stage calculates the crack parameters through a physical constraint inversion network. The crack depth is output by the phase offset and the rate of change of impedance amplitude through a pre-calibrated nonlinear mapping function. The crack length is calculated by the continuous distribution length of the three-dimensional feature vectors in space. The crack inclination angle is determined by the direction angle of the impedance change gradient. The risk level is automatically classified according to the crack depth.

2. The method for in-transit nondestructive testing and quantitative evaluation of track cracks based on electromagnetic eddy currents according to claim 1, characterized in that: The electromagnetic eddy current sensor array connected to the high-speed data acquisition module adopts a dual-band layered layout structure. The operating frequency range of the low-frequency sensor unit covers the subsurface crack detection requirements, and its U-shaped magnetic core is installed at a preset tilt angle in the rail waist area to enhance the magnetic field penetration capability. The operating frequency range of the high-frequency sensor unit covers the surface crack detection requirements, and its rectangular magnetic core symmetrically covers the working surface of the rail head. The detection areas of adjacent sensor units have an overlapping layout in the track extension direction. The phase synchronization control module of the multi-channel signal generator eliminates the electromagnetic coupling interference between adjacent sensor units.

3. The method for in-transit nondestructive testing and quantitative evaluation of track cracks based on electromagnetic eddy currents according to claim 2, characterized in that: The operation of dynamically adjusting the signal sampling frequency includes a speed adaptive compensation mechanism. Based on the real-time feedback of the travel speed value from the speed sensor, the sampling rate of the high-speed data acquisition module is dynamically adjusted within a preset sampling frequency range through an interpolation algorithm, so that the density of signal sampling points within a unit track length remains constant. At the same time, the acquisition time of the encoder's mileage pulse trigger signal and the eddy current response voltage signal is synchronized.

4. The method for in-transit nondestructive testing and quantitative evaluation of track cracks based on electromagnetic eddy currents according to claim 3, characterized in that: The wavelet denoising algorithm adopts an adaptive threshold optimization strategy. By analyzing the spectral distribution characteristics of track vibration noise, dynamic threshold calculation is performed on the coefficients of each frequency band after wavelet decomposition. The high-frequency noise coefficients are nonlinearly compressed according to energy weights, while the low-frequency signal coefficients retain their full amplitude. The final reconstructed denoised signal retains the original characteristics of the crack-sensitive frequency band.

5. The method for in-transit nondestructive testing and quantitative evaluation of track cracks based on electromagnetic eddy currents according to claim 4, characterized in that: The decoupling operation of the digital lock-in amplifier includes two-stage processing of quadrature demodulation and frequency band separation. The first-stage mixer multiplies the noise-reduced eddy current response voltage signal with the local reference signal to generate an in-phase component. The second-stage mixer multiplies the phase-shifted reference signal with the voltage signal to generate a quadrature component. For different frequency band signals in the composite excitation, a configurable bandpass filter bank is used to separate the in-phase and quadrature components corresponding to each frequency band.

6. The method for in-transit nondestructive testing and quantitative evaluation of track cracks based on electromagnetic eddy currents according to claim 5, characterized in that: The decoupling operation of the digital lock-in amplifier includes two-stage processing of quadrature demodulation and frequency band separation. The first-stage mixer multiplies the noise-reduced eddy current response voltage signal with the local reference signal to generate an in-phase component. The second-stage mixer multiplies the phase-shifted reference signal with the voltage signal to generate a quadrature component. For different frequency band signals in the composite excitation, a configurable bandpass filter bank is used to separate the in-phase and quadrature components corresponding to each frequency band.

7. The method for in-transit nondestructive testing and quantitative evaluation of track cracks based on electromagnetic eddy currents according to claim 6, characterized in that: The first-level convolutional neural network of the dual-cascade evaluation model adopts a multi-scale feature fusion structure. It extracts the spatial distribution pattern of the three-dimensional feature vector set through parallel convolutional kernel groups and outputs the crack existence probability and orientation classification results. The second-level physical constraint inversion network embeds the crack depth mapping relationship derived from the electromagnetic field analytical model and forces the crack depth output value to meet the preset function constraints of phase offset and impedance amplitude change rate.

8. The method for in-transit nondestructive testing and quantitative evaluation of track cracks based on electromagnetic eddy currents according to claim 7, characterized in that: The crack length is calculated using a spatial continuity discrimination algorithm. After identifying the crack region, the impedance amplitude change rate is differentially calculated along the track extension direction. The lengths of spatial intervals that continuously exceed the sensitivity threshold are accumulated as the crack length. The crack inclination angle is calculated by a gradient direction detection algorithm to calculate the spatial angle of the impedance change vector, and cross-validated by combining the direction classification results of the first-level output.

9. The method for in-transit nondestructive testing and quantitative evaluation of track cracks based on electromagnetic eddy currents according to claim 8, characterized in that: The risk level classification introduces a multi-parameter decision rule. When the crack depth reaches the intermediate risk threshold and the crack inclination angle exceeds the critical angle, the risk level is automatically upgraded. When the structured report is generated, the spatial location coordinates are converted into standard railway mileage information. Crack parameters and risk levels are encapsulated into a tree data structure and transmitted to the ground engineering management system through a dedicated wireless communication protocol.

10. The method for in-transit nondestructive testing and quantitative evaluation of track cracks based on electromagnetic eddy currents according to claim 9, characterized in that: It also includes an online self-calibration module. When the detection system enters the defect-free track section, it automatically collects the eddy current response signal under the current environmental parameters to update the defect-free reference impedance database. When the output confidence of the physical constraint inversion network is lower than the set threshold, it triggers the parameter optimization process of the crack depth mapping relationship and adjusts the function constraint coefficient through the error backpropagation mechanism.

Images