Steel rail turnout area defect monitoring method based on acoustoelectric fusion technology

Abstract

The invention discloses a steel rail turnout area defect monitoring method based on an acoustoelectric fusion technology. The method comprises the following steps: deploying an ultrasonic guided wave sensor array, an acoustic emission sensor array and a carrier coupling device in a turnout area; obtaining a first characteristic parameter set representing internal macroscopic defects through ultrasonic guided waves; obtaining a second characteristic parameter set representing dynamic defect initiation and expansion through acoustic emission; acquiring a third characteristic parameter set representing the electrical continuity and the surface contact state through carrier monitoring; performing time alignment, spatial registration and data level and feature level fusion on the three feature sets, constructing a multi-modal fusion feature vector, and inputting a pre-training model to output defect types, positions and severity; omnibearing sensing from inside to surface and from static state to dynamic state is realized, the problems of weak micro crack recognition, large noise interference, incapability of sensing internal damage, low recognition rate, high false alarm rate and the like in a single technology are solved, and the defect recognition precision and the monitoring reliability are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of non-destructive testing and structural health monitoring technology for rail transit, and particularly relates to a method for monitoring defects in the steel rail turnout area based on acoustic-electric fusion technology. Background Technology

[0002] In recent years, with the rapid development of railway transportation towards high speed and heavy load, the safe operation of track structures faces increasingly severe challenges. As a key component of track lines, turnouts play a crucial role in train track changing. Their complex structure, stress concentration, and harsh service environment make them highly susceptible to the initiation and spread of defects. Defects in steel turnout areas are diverse, including fatigue cracks, rail wear, spalling, loose bolts, and fractures. Failure to detect and address these defects in a timely manner will seriously threaten train safety. Therefore, conducting efficient, accurate, and continuous structural health monitoring of turnout areas has become an important research direction in the field of rail transit operation and maintenance, attracting significant attention from domestic and international research institutions and engineering departments.

[0003] Currently, defect monitoring in rail turnout areas primarily employs detection technologies based on single physical principles. Among these, ultrasonic guided wave technology excites and receives guided wave signals within the rail, utilizing phenomena such as reflection, transmission, and mode conversion resulting from the interaction of waves with defects during propagation to achieve long-distance, wide-area defect detection. It is commonly used for damage identification in internal and concealed areas. Acoustic emission technology, on the other hand, collects transient elastic wave signals released during material deformation or crack propagation under stress, enabling real-time monitoring of dynamic defects with high sensitivity and real-time response capabilities. Furthermore, carrier wave monitoring technology, based on electrical principles, utilizes changes in the electrical characteristics of track circuits, such as impedance, current, and voltage, to assess track connectivity, surface contact conditions, and defects like rail breaks, and is widely used in railway signaling systems.

[0004] However, all of the aforementioned individual monitoring technologies have significant limitations in application. Ultrasonic guided wave technology has weak ability to identify microcracks and shallow surface defects, and its signal stability is difficult to guarantee due to factors such as rail surface deposits and temperature changes. While acoustic emission technology is sensitive to dynamic defects, its signals are easily interfered with by environmental noise such as train operation and track bed vibration, and it is difficult to accurately locate and quantitatively assess defects. Carrier wave monitoring technology mainly reflects the electrical connectivity of track circuits, cannot detect internal structural damage, and has almost no detection capability for non-electrical contact defects. Because the monitoring "blind spots" of various technologies are intertwined, a single method cannot comprehensively cover the complex and diverse defect types in the turnout area, resulting in problems such as low recognition rate, high false alarm rate, and information fragmentation in practical applications, making it difficult to meet the urgent needs of modern railways for high-reliability and high-precision structural health monitoring. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a method for monitoring defects in the steel rail turnout area based on acoustic-electric fusion technology. This method overcomes the problems of existing single monitoring technologies, such as weak ability to identify micro-cracks, large environmental noise interference, inability to detect internal structural damage, low recognition rate, and high false alarm rate. By fusing multi-modal information, this method achieves complementary advantages, comprehensively covers multiple defect types, and improves the accuracy of defect identification and the reliability of monitoring.

[0006] Technical solution: The method for monitoring defects in the steel rail turnout area based on acoustic-electric fusion technology described in this invention includes the following steps:

[0007] S1. Deploy an ultrasonic guided wave monitoring sensor array, an acoustic emission sensor array, and a carrier coupling device in the area of ​​the rail turnout to be monitored.

[0008] S2. Excite and receive ultrasonic guided wave signals into the rail through the ultrasonic guided wave monitoring sensor array, analyze the ultrasonic guided wave signals, and extract a first set of characteristic parameters that characterize the macroscopic defects and structural discontinuities inside the rail.

[0009] S3. Passively receive the transient elastic wave signal generated by the rail during the stress or defect propagation process through the acoustic emission sensor array, analyze the transient elastic wave signal, and extract the second feature parameter set characterizing the dynamic defect initiation and propagation process;

[0010] S4. A carrier signal of a preset frequency is injected into the rail through the carrier coupling device, and the changes in the transmission characteristics of the carrier signal in the rail are detected. The carrier signal is analyzed to extract a third set of characteristic parameters that characterize the changes in the electrical continuity and surface contact state of the rail.

[0011] S5. Perform time alignment and spatial registration on the first feature parameter set, the second feature parameter set, and the third feature parameter set, and perform data-level fusion and feature-level fusion to construct a unified multimodal fusion feature vector;

[0012] S6. Input the multimodal fusion feature vector into the pre-trained defect identification and classification model, and output a comprehensive diagnostic result of the type, location and severity of defects in the rail turnout area from the defect identification and classification model.

[0013] The method described in this invention effectively overcomes many limitations of existing single monitoring methods in monitoring defects in the rail turnout area through acoustic-electric fusion technology. Specifically, step S2 utilizes an ultrasonic guided wave monitoring sensor array to actively excite and analyze ultrasonic guided wave signals, which can accurately capture macroscopic defects and structural discontinuities inside the rail, compensating for the insufficient perception of deep internal damage by passive monitoring. Step S3 passively receives transient elastic wave signals through an acoustic emission sensor array, capturing the initiation and propagation process of dynamic defects in real time, significantly enhancing the ability to identify early damage such as microcracks. Simultaneously, combined with the carrier coupling device in step S4 for sensitive detection of changes in the electrical continuity and surface contact state of the rail, it effectively suppresses the impact of environmental noise on monitoring under complex operating conditions. To mitigate interference from the test results, step S5 constructs a unified multimodal fusion feature vector with complementary information through temporal alignment and spatial registration of multimodal features, as well as data-level and feature-level fusion, overcoming the shortcomings of low recognition rate and high false alarm rate of single parameters. Finally, step S6 inputs the fused features into the pre-trained model to achieve accurate comprehensive diagnosis of defect type, location, and severity, comprehensively covering multiple types of defects from macroscopic structural discontinuities to microscopic crack propagation, and from internal damage to surface state changes, significantly improving the accuracy of defect identification in the steel rail turnout area and the overall reliability of the monitoring system.

[0014] Preferably, the rail turnout area mentioned in step S1 includes at least one or more of the following: the tip of the switch rail, the frog center, the section where the main rail and the switch rail are in close contact, the rail web area, the rail bottom area, and the rail electrical connection section; the deployment positions of the ultrasonic guided wave monitoring sensor array, the acoustic emission sensor array, and the carrier coupling device are set according to the analysis results of the stress concentration points, common damage locations, and monitoring blind spots of the turnout structure.

[0015] The preferred deployment scheme extends the monitoring coverage to key areas such as the tip of the switch rail, the frog, the close-fitting section, the rail web, the rail base, and the electrical connection section. It also precisely sets the placement of the ultrasonic guided wave sensor array, acoustic emission sensor array, and carrier coupling device by combining the analysis results of stress concentration points, common damage locations, and monitoring blind spots in the turnout structure. This eliminates potential monitoring blind spots associated with a single sensor layout at the source, achieving effective coverage of all critical parts of the rail turnout area. This setup not only ensures that the acoustic-electric fusion monitoring technology can simultaneously collect comprehensive damage information from multiple sources, including macroscopic structural discontinuities, dynamic defect initiation and propagation, electrical continuity, and surface contact states, but also lays an accurate spatial foundation for subsequent temporal alignment and spatial registration of multimodal features. This significantly improves the accuracy and overall perception capability of the monitoring system in locating damage in complex turnout structures.

[0016] Preferably, the first set of characteristic parameters in step S2 includes at least one or more of the following: amplitude attenuation, transit time, wave velocity variation, frequency component variation, and mode conversion characteristics of the ultrasonic guided wave signal.

[0017] The preferred scheme extracts at least one or more of the following features from ultrasonic guided wave signals: amplitude attenuation, transit time, wave velocity variation, frequency component variation, and mode transition characteristics. These features serve as the first set of characteristic parameters, enabling the monitoring system to comprehensively characterize macroscopic defects and structural discontinuities within the rail from multiple physical dimensions. Amplitude attenuation and wave velocity variation directly reflect the absorption of guided wave propagation energy by defects and the geometrical alteration of the propagation path. Transit time provides precise distance information for defect location, while frequency component variation and mode transition characteristics further reveal complex structural information such as defect morphology, size, and cross-sectional changes. The comprehensive extraction of multiple parameters effectively avoids the problem of single parameters being susceptible to interference or providing incomplete information in complex turnout structures. This provides rich and complementary structural health status characterization for subsequent multimodal fusion, significantly enhancing the accuracy of macroscopic defect identification and the quantitative description capability of structural discontinuities.

[0018] Preferably, the second set of characteristic parameters described in step S3 includes at least one or more of the impact number, energy, amplitude, rise time, duration, and frequency spectrum characteristics of the acoustic emission event.

[0019] The preferred approach extracts at least one or more of the following as a second set of feature parameters from the acoustic emission signal: impact number, energy, amplitude, rise time, duration, and frequency spectrum features. This enables the monitoring system to finely characterize the initiation and propagation process of dynamic defects from multiple dimensions, including the time domain, frequency domain, and energy distribution. Specifically, the impact number and energy reflect the intensity and activity level of the defect, amplitude and rise time characterize the type and severity of the defect source, duration reveals the duration of defect propagation, and frequency spectrum features further distinguish different defect modes and their evolution stages. The comprehensive extraction of multiple parameters effectively overcomes the limitations of single acoustic emission parameters in complex noise environments, which are prone to misjudgment or incomplete information. This provides rich and dynamically responsive defect evolution information for subsequent multimodal fusion, significantly improving the early identification capability and dynamic tracking accuracy of the initiation and propagation process of micro-cracks.

[0020] Preferably, the third set of characteristic parameters in step S4 includes at least one or more of the following: amplitude attenuation, phase shift, harmonic distortion rate, and impedance spectrum characteristics of the carrier signal.

[0021] The preferred approach extracts at least one or more of the following as a third set of characteristic parameters from the carrier signal: amplitude attenuation, phase shift, harmonic distortion rate, and impedance spectrum characteristics. This enables the monitoring system to sensitively characterize changes in the electrical continuity and surface contact state of the rail from an electrical properties perspective. Amplitude attenuation and phase shift directly reflect changes in contact resistance and interface degradation along the carrier transmission path; harmonic distortion rate reveals signal distortion caused by nonlinear contact or local defects; and impedance spectrum characteristics further provide information on interface characteristics over a wide bandwidth. The comprehensive extraction of multiple parameters effectively compensates for the shortcomings of traditional structural health monitoring methods in sensing the electrical connection section and surface contact state. This provides crucial contact interface state information for subsequent multimodal fusion, significantly enhancing the monitoring sensitivity and diagnostic reliability for contact degradation and surface damage in key areas such as the rail web, rail base, and electrical connection section.

[0022] Preferably, the time alignment and spatial registration in step S5 includes: synchronously acquiring ultrasonic guided wave signals, transient elastic wave signals and carrier signals based on a unified time base, and correcting the acquisition time, sampling window and corresponding monitoring position of different sensor channels so that the characteristic parameters of different physical domains are mapped to the same time window and the same spatial monitoring unit.

[0023] The preferred scheme synchronously acquires ultrasonic guided wave signals, transient elastic wave signals, and carrier signals based on a unified time base, and precisely corrects the acquisition time, sampling window, and corresponding monitoring position of different sensor channels. This ensures that the first, second, and third feature parameter sets from different physical domains can accurately correspond within the same time window and the same spatial monitoring unit. This process effectively eliminates the information misalignment problem caused by differences in acquisition timing and spatial position deviations between multi-source sensors, providing a reliable data foundation with temporal consistency and spatial alignment for subsequent data-level fusion and feature-level fusion. It ensures that multimodal information can achieve true complementary advantages during the fusion process, thereby significantly improving the accuracy of defect location calibration by the fused feature vector and the continuity and consistency of the comprehensive diagnostic results in the time dimension.

[0024] Preferably, the data-level fusion and feature-level fusion in step S5 include: after normalizing the synchronously acquired ultrasonic guided wave signal, transient elastic wave signal and carrier signal, extracting time-domain features, frequency-domain features and time-frequency-domain features in each technical domain respectively, and then concatenating, weighting or extracting deep features through a neural network to form a unified multimodal fusion feature vector.

[0025] The preferred scheme normalizes the synchronously acquired ultrasonic guided wave signals, transient elastic wave signals, and carrier signals to eliminate differences in the dimensions and amplitude ranges of data from each physical domain. Based on this, time-domain, frequency-domain, and time-frequency-domain features are extracted separately within each technical domain, fully preserving the sensitivity of a single monitoring method to specific defect types. Then, through cross-domain feature vector concatenation, weighting, or deep feature extraction from neural networks, multi-level fusion from raw data to high-level semantic representation is achieved. This fusion strategy balances the preservation of the integrity of raw information by data-level fusion with the ability to mine high-dimensional abstract features by feature-level fusion. It enables deep interaction and complementary enhancement of structural health information from the three physical domains of ultrasonic guided waves, acoustic emission, and carrier coupling, effectively solving the problem of one-sided information and insufficient discriminative power of single-domain features in complex turnout environments. Finally, it constructs a rich and highly compact multimodal fusion feature vector, providing more robust and discriminative input support for subsequent defect identification and classification models.

[0026] Preferably, the defect identification and classification model described in step S6 is trained based on historical monitoring data and known defect samples, and the defect identification and classification model is constructed using one or more of the following: support vector machine, random forest, gradient boosting machine, convolutional neural network, recurrent neural network, and deep neural network.

[0027] The preferred approach utilizes historical monitoring data and known defect samples to construct a defect identification and classification model using one or more of the following: Support Vector Machine (SVM), Random Forest, Gradient Boosting Machine (GPG), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Deep Neural Network (DNN). This allows the model to fully learn the complex mapping relationships between multimodal fusion feature vectors and various defect types, locations, and severity levels. The flexible selection of multiple model architectures allows for adaptation to the interpretability advantages of shallow machine learning methods based on the actual data scale and feature distribution characteristics. It also leverages the powerful nonlinear representation capabilities of deep neural networks to automatically mine deep feature correlations, thereby effectively improving the comprehensive diagnostic accuracy and generalization ability for various defects in the rail turnout area. This training mechanism ensures that the model, based on fully absorbing historical prior knowledge, can make accurate and stable identification and classification responses to newly acquired multimodal fusion features, providing the system with reliable adaptive diagnostic capabilities.

[0028] Preferably, the types of defects mentioned in step S6 include at least one or more of the following: internal cracks, surface peeling, bolt hole cracks, rail head crushing, rail surface wear, structural discontinuities, and poor electrical connections; the severity is graded and evaluated based on the defect evolution trend corresponding to the fusion features, the signal change amplitude, and the model output confidence level.

[0029] The preferred scheme includes at least one or more of the following defect types: internal cracks, surface peeling, bolt hole cracks, rail head crushing, rail surface wear, structural discontinuities, and poor electrical connections. This allows the monitoring system to comprehensively cover all typical failure modes in the rail turnout area, from internal structural damage to surface contact deterioration, and from material fatigue to electrical performance degradation. Simultaneously, it classifies and assesses the severity based on the defect evolution trend, signal variation amplitude, and model output confidence level corresponding to the fused features, thereby quantifying the development trend and hazard level of defects beyond simply identifying their types. This setup fully utilizes the rich information contained in multimodal fusion features, organically combining ultrasonic guided wave perception of macroscopic structural anomalies, acoustic emission tracking of dynamic crack propagation, and carrier coupling monitoring of electrical contact status. This achieves high-precision identification of defect types and refined classification of severity, providing a comprehensive and reliable diagnostic basis for differentiated early warning and precise maintenance decisions for rail turnouts.

[0030] Preferably, the ultrasonic guided wave monitoring sensor array, acoustic emission sensor array, and carrier coupling device are synchronously acquired through a unified trigger signal, and noise suppression, bandpass filtering, and outlier removal are performed on each channel before acquisition.

[0031] The preferred scheme synchronously acquires data from the ultrasonic guided wave monitoring sensor array, acoustic emission sensor array, and carrier coupling device using a unified trigger signal. This ensures precise alignment of the three physical domain signals along the time axis, fundamentally eliminating information misalignment caused by inconsistent acquisition start times. Simultaneously, noise suppression, bandpass filtering, and outlier removal are performed on each channel before acquisition, effectively filtering out environmental noise, electromagnetic interference, and sensor anomalies that contaminate the original signal quality. This setup provides a timely and high-signal-to-noise ratio raw data foundation for subsequent multimodal feature extraction and significantly reduces the risk of misjudgment due to inconsistent data quality during fusion. From the data acquisition source, it ensures the stability and reliability of the acoustic-electric fusion monitoring system under complex turnout conditions.

[0032] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: 1. By integrating ultrasonic guided wave, acoustic emission, and carrier wave monitoring technologies, this invention leverages their respective advantages in sensing internal macroscopic defects, dynamic micro-defects, and surface electrical states, achieving comprehensive perception from the inside to the surface and from static to dynamic states. This effectively overcomes the problems of weak micro-crack identification capability, large environmental noise interference, inability to detect internal structural damage, low recognition rate, and high false alarm rate inherent in single technologies, significantly improving defect identification accuracy and monitoring reliability; 2. By performing time alignment and spatial registration of the three physical domain signals and deep fusion at the data and feature levels, a unified multimodal fusion feature vector is constructed. Combined with an intelligent classification model, it can accurately distinguish various defect types such as cracks, peeling, wear, and poor electrical connections, effectively reducing false alarms and missed alarms; 3. By using a unified trigger signal to achieve synchronous acquisition of ultrasonic guided wave, acoustic emission, and carrier wave signals, and by integrating sensor arrays with stress concentration points and damage locations in the turnout structure, multiple detection methods can be completed in one go, eliminating the need for multiple operations over time, and significantly improving the comprehensive monitoring efficiency and automation level of the rail turnout area. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0034] Figure 2 This is a schematic diagram of the deployment of the acoustic-electric fusion monitoring system of the present invention in the steel rail turnout area;

[0035] Figure 3 This is a schematic diagram comparing the effects of the method of the present invention with those of a single technical method in principle. Detailed Implementation

[0036] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0037] This invention provides a method for monitoring defects in the steel rail turnout area based on acoustic-electric fusion technology, such as... Figure 1 As shown, its core process includes four main stages: multi-sensor synchronous data acquisition, multi-modal feature extraction, acoustic-electric feature fusion, and intelligent diagnostic decision-making.

[0038] like Figure 2 As shown, in a typical turnout monitoring scenario, the implementation process is as follows:

[0039] S1: System optimization and deployment.

[0040] Based on the mechanical characteristics and common defect patterns of turnouts, sensor arrays are optimally arranged at key stress-bearing locations and stress concentration areas, such as the switch tip, frog center, and the close-fitting section between the stock rail and switch rail. Specifically, ultrasonic guided wave transducer pairs are installed on the rail web and rail base to excite and receive specific modes of ultrasonic guided waves (such as A0 or S0 modes); high-sensitivity acoustic emission sensors are installed near areas where microcracks or peeling may occur; and carrier coupling devices are installed at the electrical connection sections of the rails or in areas where contact status needs to be monitored. All sensors are synchronously triggered and acquire data through a unified time base signal, ensuring strict temporal alignment and spatial registration of multi-source data.

[0041] S2: Synchronous acquisition and feature extraction of multimodal data.

[0042] Ultrasonic guided wave monitoring: The ultrasonic excitation unit generates an encoded excitation signal (such as linear frequency modulation or phase encoding), which is amplified and drives the transmitting transducer. The receiving transducer captures the signal propagating along the rail, which is then pre-amplified, filtered, and converted from analog to digital before undergoing pulse compression and mode separation. The first set of characteristic parameters is extracted from the processed signal, such as: the amplitude A_UGW of the echo along a specific path, the time difference of arrival Δt_UGW, the signal energy E_UGW, the center frequency f_c_UGW, and the dispersion curve offset.

[0043] Acoustic emission monitoring: The acoustic emission acquisition unit operates in continuous or triggered mode, acquiring sensor signals in real time. Acoustic emission events are detected using either a threshold method or a short-time energy method. For each valid event, a second set of feature parameters is extracted, such as: event impact count (Count_AE), absolute energy (E_AE), peak amplitude (Amplitude_AE), rise time (RiseTime_AE), duration (Duration_AE), and the dominant frequency (Freq_AE) obtained through Fast Fourier Transform (FFT).

[0044] Carrier monitoring: The carrier signal generation unit generates a stable high-frequency sinusoidal signal, which is injected into the rail through a coupling circuit. The receiving unit detects the current or voltage signal flowing through the monitoring circuit. By analyzing the amplitude, phase, and harmonic components of the received signal, a third set of characteristic parameters is extracted, such as: signal transmission attenuation Att_Carrier, phase angle Phase_Carrier, total harmonic distortion rate THD_Carrier, and impedance amplitude |Z| and phase angle Φ obtained by scanning within a certain frequency range.

[0045] S3: Data fusion and high-dimensional feature construction.

[0046] The three sets of feature parameters acquired within the same time window are standardized to eliminate the influence of dimensions. Then, feature-level fusion is performed; for example, a fused feature vector F_fusion is constructed: F_fusion = [A_UGW, Δt_UGW, E_UGW, f_c_UGW, ..., Count_AE, E_AE, Amplitude_AE, ..., Att_Carrier, Phase_Carrier, THD_Carrier, ...]

[0047] This vector integrates information from different physical domains, including acoustic and electrical domains. Furthermore, this vector can be input into a feature selection module or autoencoder to reduce redundancy and extract more discriminative deep fusion features.

[0048] S4: Machine learning-based intelligent defect diagnosis.

[0049] The processed fused feature vector is input into a pre-trained defect identification and classification model. This model can be trained using algorithms such as Support Vector Machine (SVM), Random Forest, Gradient Boosting Machine (GBDT), or Deep Neural Network (DNN). The training data should contain a large number of labeled samples, covering multimodal fusion features under "defect-free" conditions and various typical defect states (such as transverse cracks, longitudinal cracks, core damage, peeling, bolt loosening, etc.). The model ultimately outputs a structured diagnostic result, for example: {Defect type: 'railhead surface peeling', Confidence: 0.92, Suspected location: approximately 2.5m from sensor A at the railhead, Severity index: 0.7}.

[0050] To illustrate the expected beneficial effects of the acoustic-electric fusion method described in this invention, Figure 3 A schematic diagram comparing its principles with those of a single monitoring technology method is provided.

[0051] like Figure 3 As shown in (a) and (b), traditional single-technology monitoring (such as ultrasonic guided waves only or acoustic emission only) has inherent monitoring blind spots (as shown in the shaded area in the figure) when dealing with complex turnout defects due to its own physical principles. This may result in insensitivity to certain types of defects or susceptibility to interference.

[0052] like Figure 3As shown in (c), the acoustic-electric fusion method employed in this invention achieves synchronous acquisition and information fusion of multiple physical quantities (sound waves, elastic waves, and electrical signals) through the coordinated deployment of multiple sensors. Features from different technologies are fused and analyzed at the data processing center, which is expected to mutually verify and complement each other's shortcomings. For example, for a suspected defect, ultrasonic signals may indicate structural discontinuity, acoustic emission signals may indicate dynamic activity, and carrier signals may indicate changes in electrical characteristics. After integrating these relevant features, the fusion algorithm is expected to more accurately and reliably determine the defect type (such as 'railhead surface peeling') and assess its state, thereby theoretically reducing blind spots and improving recognition accuracy.

[0053] therefore, Figure 3 This illustration demonstrates the potential advantages of the present invention, which, through multi-source information fusion, can bring compared to traditional methods at the technical principle level.

Claims

1. A method for monitoring defects in the steel rail turnout area based on acoustic-electric fusion technology, characterized in that, Includes the following steps: S1. Deploy an ultrasonic guided wave monitoring sensor array, an acoustic emission sensor array, and a carrier coupling device in the area of ​​the rail turnout to be monitored. S2. Excite and receive ultrasonic guided wave signals into the rail through the ultrasonic guided wave monitoring sensor array, analyze the ultrasonic guided wave signals, and extract a first set of characteristic parameters that characterize the macroscopic defects and structural discontinuities inside the rail. S3. Passively receive the transient elastic wave signal generated by the rail during the stress or defect propagation process through the acoustic emission sensor array, analyze the transient elastic wave signal, and extract the second feature parameter set characterizing the dynamic defect initiation and propagation process; S4. A carrier signal of a preset frequency is injected into the rail through the carrier coupling device, and the changes in the transmission characteristics of the carrier signal in the rail are detected. The carrier signal is analyzed to extract a third set of characteristic parameters that characterize the changes in the electrical continuity and surface contact state of the rail. S5. Perform time alignment and spatial registration on the first feature parameter set, the second feature parameter set, and the third feature parameter set, and perform data-level fusion and feature-level fusion to construct a unified multimodal fusion feature vector; S6. Input the multimodal fusion feature vector into the pre-trained defect identification and classification model, and output a comprehensive diagnostic result of the type, location and severity of defects in the rail turnout area from the defect identification and classification model.

2. The method according to claim 1, characterized in that, The steel rail turnout area mentioned in step S1 includes at least one or more of the following: the tip of the switch rail, the frog core, the section where the main rail and the switch rail are closely connected, the rail web area, the rail bottom area, and the rail electrical connection section; the deployment positions of the ultrasonic guided wave monitoring sensor array, the acoustic emission sensor array, and the carrier coupling device are set according to the analysis results of the stress concentration points, common damage locations, and monitoring blind spots of the turnout structure.

3. The method according to claim 1, characterized in that, The first set of characteristic parameters mentioned in step S2 includes at least one or more of the following: amplitude attenuation, transit time, wave velocity variation, frequency component variation, and mode conversion characteristics of the ultrasonic guided wave signal.

4. The method according to claim 1, characterized in that, The second set of characteristic parameters mentioned in step S3 includes at least one or more of the impact number, energy, amplitude, rise time, duration, and frequency spectrum characteristics of the acoustic emission event.

5. The method according to claim 1, characterized in that, The third set of characteristic parameters mentioned in step S4 includes at least one or more of the following: amplitude attenuation, phase shift, harmonic distortion rate, and impedance spectrum characteristics of the carrier signal.

6. The method according to claim 1, characterized in that, The time alignment and spatial registration described in step S5 include: synchronously acquiring ultrasonic guided wave signals, transient elastic wave signals and carrier signals based on a unified time base, and correcting the acquisition time, sampling window and corresponding monitoring position of different sensor channels so that the characteristic parameters of different physical domains are mapped to the same time window and the same spatial monitoring unit.

7. The method according to claim 1, characterized in that, The data-level fusion and feature-level fusion described in step S5 include: after normalizing the synchronously acquired ultrasonic guided wave signal, transient elastic wave signal and carrier signal, firstly extracting time-domain features, frequency-domain features and time-frequency-domain features in each technical domain, and then concatenating, weighting or extracting deep features through a neural network to form a unified multimodal fusion feature vector.

8. The method according to claim 1, characterized in that, The defect identification and classification model described in step S6 is trained based on historical monitoring data and known defect samples. The defect identification and classification model is constructed using one or more of the following: support vector machine, random forest, gradient boosting machine, convolutional neural network, recurrent neural network, and deep neural network.

9. The method according to claim 1, characterized in that, The types of defects mentioned in step S6 include at least one or more of the following: internal cracks, surface peeling, bolt hole cracks, rail head crushing, rail surface wear, structural discontinuities, and poor electrical connections; the severity is graded and evaluated based on the defect evolution trend, signal change amplitude, and model output confidence level corresponding to the fusion features.

10. The method according to claim 1, characterized in that, The ultrasonic guided wave monitoring sensor array, acoustic emission sensor array, and carrier coupling device are synchronously acquired through a unified trigger signal, and noise suppression, bandpass filtering, and outlier removal are performed on each channel before acquisition.

Images