A rack defect detection method that engages an impact feature

Abstract

The application discloses a rack defect detection method of meshing impact characteristics, and belongs to the technical field of rack defect detection. Firstly, the application collects the vibration signal of the train and the rack when the train is meshed with the rack through a vibration sensor, and enhances the vibration signal in the time domain; then, in the gear meshing frequency band, the meshing impact strength and the amplitude feature are extracted based on the energy coupling strength and the amplitude information through the ratio of the abnormal value to the average value, and the corresponding original vector is constructed; then, the eigenvalue of the original vector is extracted, and the strength and amplitude feature matrix is formed; finally, the double-branch defect diagnosis neural network is used to process the feature matrix, and the rack defect diagnosis result is output. Through multi-dimensional feature extraction and neural network analysis, the precision of the rack defect diagnosis is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of toothed rail defect detection technology, and specifically to a method for detecting toothed rail defects based on meshing impact characteristics. Background Technology

[0002] Gear racks are core components in rail transit, mining transportation, and other fields. During long-term service, they are prone to defects such as tooth surface wear and cracks due to load and wear. If not detected in time, they may cause safety accidents such as derailment. Therefore, accurate detection is crucial.

[0003] Current methods for detecting defects in toothed rails primarily rely on vibration signal analysis, but these methods have significant shortcomings. Firstly, in the signal preprocessing stage, existing linear enhancement methods (such as filtering and time-domain averaging) struggle to handle meshing nonlinear vibration signals, and the subtle impact characteristics related to defects are easily masked by environmental noise and background vibration. Secondly, in the feature extraction stage, traditional techniques focus on linear parameters such as amplitude and frequency. For example, they use EMD to decompose the signal to obtain time-frequency domain features, employ wavelet analysis to extract local detail features, or calculate peak values ​​and root mean square (RMS) statistics in the time domain. While these methods can reflect the basic operating state of the toothed rail, these features are easily masked by noise and background vibration, failing to accurately reflect defects and resulting in low diagnostic accuracy. Summary of the Invention

[0004] In view of the above-mentioned shortcomings in the prior art, the present invention provides a tooth rail defect detection method based on meshing impact characteristics, which solves the problem of low diagnostic accuracy in the prior art.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a method for detecting tooth rail defects based on meshing impact characteristics, comprising the following steps:

[0006] Vibration sensors are installed on the rack rail of the section to be tested to collect vibration signals generated when the train passes over the rack rail, and the vibration signals are enhanced in the time domain to obtain the enhanced signal.

[0007] In the gear meshing frequency band of the enhanced signal, the energy coupling strength of any two frequency components is obtained, and the average value is taken to obtain the average energy coupling strength.

[0008] Based on the energy coupling intensity of gear meshing harmonics, the meshing impact intensity characteristics are extracted by the ratio of abnormal to average energy coupling intensity, and the original vector of meshing impact intensity is constructed.

[0009] Based on the amplitude information of the gear meshing frequency band, the meshing impact amplitude features are extracted by the ratio of abnormal to average amplitude, and the original vector of meshing impact amplitude is constructed.

[0010] Feature values ​​are extracted from the original vectors of meshing impact intensity and meshing impact amplitude, respectively, and feature matrices of meshing impact intensity and meshing impact amplitude are constructed.

[0011] A dual-branch defect diagnosis neural network is used to process the meshing impact intensity feature matrix and the meshing impact amplitude feature matrix to obtain the tooth rail defect diagnosis results.

[0012] Furthermore, the process of time-domain enhancement of vibration signals includes:

[0013] The vibration signal is subjected to Hilbert transform to obtain the transformed signal;

[0014] Calculate the transient amplitude based on the transformed signal;

[0015] The enhanced signal is obtained by multiplying the vibration signal by the transient amplitude at each time point.

[0016] Furthermore, the process of obtaining the average energy coupling strength includes:

[0017] A short-time Fourier transform is performed on the enhanced signal to obtain multiple spectra;

[0018] In each spectrum, the frequency bands belonging to the gear meshing frequency range are extracted as the gear meshing frequency bands;

[0019] Select any two frequency components in the gear meshing frequency band and extract the complex amplitude values ​​of the two frequency components in the frequency band;

[0020] Calculate the conjugate complex amplitude at the sum of two frequency components;

[0021] The energy coupling strength is calculated based on the complex amplitude and conjugate complex amplitude of the two frequency components.

[0022] The average energy coupling intensity is obtained by averaging the energy coupling intensities across all energy coupling frequencies in the gear meshing band.

[0023] Furthermore, the formula for calculating the energy coupling strength is:

[0024] ,

[0025] Among them, B(f) i ,f j ) represents the i-th frequency component f i and the j-th frequency component f j The energy coupling strength, f i +f j The conjugate complex amplitude, X(f) i ) represents the i-th frequency component f i The complex amplitude value, X(f) j ) represents the j-th frequency component fj The complex amplitude value.

[0026] Furthermore, the process of constructing the original vector of meshing impact strength includes:

[0027] When two frequency components are simultaneously located in the gear meshing harmonic frequency set, the corresponding energy coupling strength is taken as the meshing energy coupling strength.

[0028] By screening out meshing energy coupling strengths that are greater than the average energy coupling strength, abnormal energy coupling strengths are obtained.

[0029] The ratio of abnormal energy coupling strength to average energy coupling strength is used as the meshing impact strength feature, and the meshing impact strength features of each gear meshing frequency band are used to form the original meshing impact strength vector.

[0030] Furthermore, the process of constructing the original vector of meshing impact amplitude includes:

[0031] The mean amplitude is extracted from the gear meshing frequency band to obtain the average amplitude;

[0032] The amplitude of the frequency component corresponding to the abnormal energy coupling strength is taken as the abnormal amplitude;

[0033] The ratio of abnormal amplitude to average amplitude is used as the characteristic of meshing impact amplitude. The meshing impact amplitude characteristics of each gear in the same meshing frequency band are used to form the original vector of meshing impact amplitude.

[0034] Furthermore, the process of constructing the meshing impact intensity characteristic matrix and the meshing impact amplitude characteristic matrix includes:

[0035] Extract the mean, maximum value, kurtosis, and standard deviation from the original vector of meshing impact strength;

[0036] Extract the mean, maximum value, kurtosis, and standard deviation from the original vector of meshing impact amplitude;

[0037] The mean, maximum value, kurtosis, and standard deviation of each original vector of meshing impact strength are used to construct the meshing impact strength feature matrix;

[0038] The mean, maximum value, kurtosis, and standard deviation of each original vector of meshing impact amplitude are used to construct the meshing impact amplitude characteristic matrix.

[0039] Furthermore, the dual-branch defect diagnosis neural network includes: a first-level intensity feature extraction unit, a first-level amplitude feature extraction unit, an association attention generation unit, a first multiplication unit, a second multiplication unit, a second-level intensity feature extraction unit, a second-level amplitude feature extraction unit, and a fully connected layer;

[0040] The first-level strength feature extraction unit is used to extract features from the meshing impact strength feature matrix to obtain the first-level strength features; the first-level amplitude feature extraction unit is used to extract features from the meshing impact amplitude feature matrix to obtain the first-level amplitude features.

[0041] The associative attention generation unit generates associative attention based on the first-level intensity feature and the first-level amplitude feature; the first multiplication unit multiplies the first-level intensity feature with the associative attention to obtain the first-level intensity enhancement feature; the second multiplication unit multiplies the first-level amplitude feature with the associative attention to obtain the first-level amplitude enhancement feature; the second-level intensity feature extraction unit extracts features from the first-level intensity enhancement feature to obtain the second-level intensity enhancement feature; the second-level amplitude feature extraction unit extracts features from the first-level amplitude enhancement feature to obtain the second-level amplitude enhancement feature; the fully connected layer classifies based on the second-level intensity enhancement feature and the second-level amplitude enhancement feature to obtain the tooth track defect diagnosis result.

[0042] Furthermore, the associative attention generation unit includes: an adder A1, a pointwise convolutional layer, and a sigmoid layer;

[0043] Adder A1 is used to add the first-level intensity enhancement feature and the first-level amplitude enhancement feature element-wise to obtain the intensity-amplitude joint feature; the pointwise convolutional layer is used to perform pointwise convolution on each feature value in the intensity-amplitude joint feature to obtain the mapping feature; the sigmoid layer is used to generate attention based on the mapping feature.

[0044] Furthermore, the first-level intensity feature extraction unit and the first-level amplitude feature extraction unit have the same structure, both including the following connected sequentially: a first 1D convolutional layer, a 1D pooling layer, and a second 1D convolutional layer;

[0045] The second-level intensity feature extraction unit and the second-level amplitude feature extraction unit have the same structure, both including: a first 2D convolutional layer, a second 2D convolutional layer, a third 2D convolutional layer, a Concat layer, a BN layer, and a ReLU layer;

[0046] The first, second, and third 2D convolutional layers are used to extract features at different depths; the Concat layer is used to concatenate features at different depths to obtain concatenated features; the BN layer is used to batch normalize the concatenated features; and the ReLU layer is used to perform non-linear activation on the batch-normalized features.

[0047] The beneficial effects of this invention are as follows:

[0048] 1. After acquiring vibration signals, this invention performs signal enhancement processing. The amplitude of the nonlinear impact region (including defect features) is selectively amplified, while the noise in the stable noise region (low transient amplitude) is relatively suppressed. This achieves targeted enhancement of the "nonlinear impact features" rather than indiscriminate signal amplification, effectively suppressing the interference of environmental noise and background vibration, and solving the problem that weak defect signals are easily masked.

[0049] 2. Traditional technologies rely on linear parameters such as amplitude and frequency, which are easily diluted by the overall vibration signal and cannot accurately reflect defects. This invention, on the one hand, captures the nonlinear characteristics of meshing impact caused by defects by calculating the average energy coupling strength of frequency components within the gear meshing frequency band. This feature is more sensitive to defects. On the other hand, it constructs impact amplitude features based on the ratio of anomalies to average amplitudes, accurately amplifying the amplitude anomalies caused by defects. This dual-dimensional approach ensures that the extracted features are truly correlated with defects, extracting features that are precisely related to defects, accurately reflecting the defects in the gear track, and improving diagnostic accuracy.

[0050] 3. At the diagnostic decision-making level, this invention significantly improves the reliability of results by leveraging a dual-branch neural network. Traditional diagnostic methods often rely on single feature inputs, resulting in limited information dimensions and a high risk of misjudgment and missed diagnosis. The dual-branch defect diagnostic neural network of this invention can process both meshing impact intensity and amplitude feature matrices separately, simultaneously mining defect information from both "intensity" and "amplitude" dimensions. Combined with the powerful feature fusion and pattern recognition capabilities of the neural network, it can more clearly distinguish between normal and defective vibration modes, significantly reducing the misjudgment rate and ultimately outputting more accurate and reliable diagnostic results, thus solving the problem of "low diagnostic accuracy." Attached Figure Description

[0051] Figure 1 A flowchart of a method for detecting defects in toothed rails based on meshing impact characteristics;

[0052] Figure 2 Flowchart of data processing for a dual-branch defect diagnosis neural network;

[0053] Figure 3 A schematic diagram of the structure of the associative attention generation unit;

[0054] Figure 4 This is a schematic diagram of the structure of the first-level intensity feature extraction unit and the first-level amplitude feature extraction unit;

[0055] Figure 5 This is a schematic diagram of the structure of the second-level intensity feature extraction unit and the second-level amplitude feature extraction unit. Detailed Implementation

[0056] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0057] like Figure 1 As shown, a method for detecting toothed rail defects based on meshing impact characteristics includes the following steps:

[0058] Vibration sensors are installed on the rack rail of the section to be tested to collect vibration signals generated when the train passes over the rack rail, and the vibration signals are enhanced in the time domain to obtain the enhanced signal.

[0059] In the gear meshing frequency band of the enhanced signal, the energy coupling strength of any two frequency components is obtained, and the average value is taken to obtain the average energy coupling strength.

[0060] Based on the energy coupling intensity of gear meshing harmonics, the meshing impact intensity characteristics are extracted by the ratio of abnormal to average energy coupling intensity, and the original vector of meshing impact intensity is constructed.

[0061] Based on the amplitude information of the gear meshing frequency band, the meshing impact amplitude features are extracted by the ratio of abnormal to average amplitude, and the original vector of meshing impact amplitude is constructed.

[0062] Feature values ​​are extracted from the original vectors of meshing impact intensity and meshing impact amplitude, respectively, and feature matrices of meshing impact intensity and meshing impact amplitude are constructed.

[0063] A dual-branch defect diagnosis neural network is used to process the meshing impact intensity feature matrix and the meshing impact amplitude feature matrix to obtain the tooth rail defect diagnosis results.

[0064] In this embodiment, the process of time-domain enhancement of the vibration signal includes:

[0065] The vibration signal is transformed by Hilbert transform, resulting in the transformed signal: y(t) = x(t) + j⋅Hilbert{x(t)}, where x(t) is the vibration signal, Hilbert is the Hilbert transform, y(t) is the transformed signal, and j is the imaginary unit;

[0066] Calculate the transient amplitude based on the transformed signal:

[0067] ,

[0068] Where A(t) is the transient amplitude;

[0069] The enhanced signal is obtained by multiplying the vibration signal by the transient amplitude at each time point:

[0070] x en (t)=A(t)⋅x(t),

[0071] Where, x en (t) represents the enhanced signal.

[0072] This invention extracts the transient amplitude of the vibration signal using Hilbert transform and multiplies it with the original signal at time points, enabling adaptive amplification of the transient characteristics of meshing impact in the time domain. This method does not rely on manually set thresholds or filtering bandwidths, and simultaneously preserves the original form of the signal and its transient energy change characteristics. This results in the enhanced signal having a higher amplitude response when the impact occurs, thereby improving the discriminability and signal-to-noise ratio of the meshing impact characteristics.

[0073] In this embodiment, the sensor is fixed to a rigid support component such as a sleeper or support beam below / on the side of the toothed rail by bolts or magnetic attraction, and must be directly facing the meshing area of ​​the toothed rail to be tested (i.e., the support point directly below / on the side when the train wheel teeth mesh with the toothed rail), or on the side of the toothed rail section to be tested (the working surface where the wheels and teeth mesh).

[0074] In this embodiment, the process of obtaining the average energy coupling strength includes:

[0075] A short-time Fourier transform is performed on the enhanced signal to obtain multiple spectra;

[0076] In each spectrum, the frequency bands belonging to the gear meshing frequency range are extracted as the gear meshing frequency bands;

[0077] Select any two frequency components in the gear meshing frequency band and extract the complex amplitude values ​​of the two frequency components in the frequency band;

[0078] Calculate the conjugate complex amplitude at the sum of two frequency components;

[0079] The energy coupling strength is calculated based on the complex amplitude and conjugate complex amplitude of the two frequency components.

[0080] The average energy coupling intensity is obtained by averaging the energy coupling intensities across all energy coupling frequencies in the gear meshing band.

[0081] In this embodiment, the enhanced signal is divided into several frames of fixed length, with overlapping portions between each frame to ensure time-frequency continuity. The sampling frequency is 12800Hz, the length of each frame is set to 2048 points (approximately 0.16s), the frame shift is half the frame length, a Hanning window is used, each frame signal is multiplied by a window function, and then a fast Fourier transform is performed.

[0082] The gear meshing frequency range is: [f g -Δf,nfg ], where f g =z⋅f r f g Let z be the gear meshing frequency, z be the number of teeth, and f be the frequency of gear meshing. r The rotation frequency is n, a positive integer ranging from 3 to 5. Δf is the lower limit offset, with a setting range of 1f. r ~2×f r .

[0083] In this embodiment, the formula for calculating the energy coupling strength is:

[0084] ,

[0085] Among them, B(f) i ,f j ) represents the i-th frequency component f i and the j-th frequency component f j The energy coupling strength, f i +f j The conjugate complex amplitude, X(f) i ) represents the i-th frequency component f i The complex amplitude value, X(f) j ) represents the j-th frequency component f j The complex amplitude value.

[0086] This invention extracts frequency bands within the gear meshing frequency range from each spectrum. By performing a Short-Time Fourier Transform (STFT) on the enhanced signal, this invention can characterize the non-stationary features of the enhanced signal in the frequency domain, thereby accurately extracting the transient meshing response when a train passes over a gear train. Compared to traditional full-frequency domain analysis methods, STFT maintains a dynamic balance between time and frequency resolution, avoiding feature ambiguity caused by the transient characteristics of the signal. In the gear meshing frequency band, the energy coupling strength between any two frequency components is used as a measure of meshing characteristics, and the degree of energy coupling between different frequency components is calculated using a formula. This calculation can reveal the nonlinear interaction effects caused by tooth defects, contact impacts, etc., during gear train meshing, thus reflecting the energy accumulation phenomenon of impact characteristics in the frequency domain.

[0087] In this embodiment, the process of constructing the original vector of meshing impact strength includes:

[0088] When two frequency components are simultaneously located in the gear meshing harmonic frequency set, the corresponding energy coupling strength is taken as the meshing energy coupling strength.

[0089] By screening out meshing energy coupling strengths that are greater than the average energy coupling strength, abnormal energy coupling strengths are obtained.

[0090] The ratio of abnormal energy coupling strength to average energy coupling strength is used as the meshing impact strength feature, and the meshing impact strength features of each gear meshing frequency band are used to form the original meshing impact strength vector.

[0091] The set of gear meshing harmonic frequencies is {f g , 2f g , 3f g , …, nf g The deviation threshold is set to ±2%, meaning that when a frequency component is within the range of 98% to 102% of the harmonic frequency, it is considered that the frequency component is the harmonic frequency.

[0092] This invention uses the energy coupling strength of two frequency components simultaneously located in the gear meshing harmonic frequency set as the meshing energy coupling strength, making the extraction results more focused on the real physical coupling relationship in the gear meshing process. It can accurately identify the nonlinear interactive enhancement features caused by meshing impact or local faults within the gear meshing frequency band, effectively suppressing the interference of background frequency components. By screening the portion with energy coupling strength higher than the average, and constructing the meshing impact strength feature based on its ratio to the average strength, the difference between abnormal energy components and normal energy components is amplified.

[0093] The greater the energy coupling strength, the more intense the nonlinear energy exchange between the frequency components of the toothed rail in the meshing frequency band, which means that the toothed rail is more likely to have defects, and the nonlinear interference of the meshing impact caused by the defects on the vibration signal is more significant, which can more intuitively reflect the abnormal meshing behavior caused by the defects.

[0094] In this embodiment, the process of constructing the original vector of meshing impact amplitude includes:

[0095] The mean amplitude is extracted from the gear meshing frequency band to obtain the average amplitude;

[0096] The amplitude of the frequency component corresponding to the abnormal energy coupling strength is taken as the abnormal amplitude;

[0097] The ratio of abnormal amplitude to average amplitude is used as the characteristic of meshing impact amplitude. The meshing impact amplitude characteristics of each gear in the same meshing frequency band are used to form the original vector of meshing impact amplitude.

[0098] This invention first extracts the average amplitude of the gear meshing frequency band to establish a baseline level for overall energy. Then, using the frequency components corresponding to the abnormal energy coupling strength, its amplitude is selected as the abnormal amplitude, thereby identifying the amplitudes accompanied by obvious nonlinear interaction phenomena in the frequency domain. By calculating the ratio of the abnormal amplitude to the average amplitude, the impact characteristics of the frequency components corresponding to the abnormal energy coupling strength in terms of amplitude are evaluated.

[0099] Energy coupling strength essentially reflects the degree of coupling between two frequency components. When the interaction strength of a pair of frequency components is significantly higher than the average level, it indicates the possible presence of periodic meshing disturbances or transient impacts. These anomalies are signals of changes in gear meshing conditions (such as localized wear, tooth surface defects, or uneven lubrication). "Energy coupling strength" reflects the "coupling strength between two frequency components," but it does not directly represent the energy contribution of that frequency component in the signal spectrum. By mapping amplitude characteristics to "abnormal strength" frequencies, we ensure that only amplitudes reflecting impact energy are extracted, excluding normal meshing background. Extracting the amplitudes corresponding to these frequencies ensures that the original amplitude vector accurately reflects the impact energy distribution, rather than general vibration energy, thereby improving the specificity and reliability of fault diagnosis features.

[0100] In this embodiment, the process of constructing the meshing impact intensity feature matrix and the meshing impact amplitude feature matrix includes:

[0101] Extract the mean, maximum value, kurtosis, and standard deviation from the original vector of meshing impact strength;

[0102] Extract the mean, maximum value, kurtosis, and standard deviation from the original vector of meshing impact amplitude;

[0103] The mean, maximum value, kurtosis, and standard deviation of each original vector of meshing impact strength are used to construct the meshing impact strength feature matrix;

[0104] The mean, maximum value, kurtosis, and standard deviation of each original vector of meshing impact amplitude are used to construct the meshing impact amplitude characteristic matrix.

[0105] This invention performs a short-time Fourier transform on the enhanced signal. Each time window corresponds to a spectrum. Each spectrum generates an original vector of meshing impact intensity and an original vector of meshing impact amplitude. Therefore, there are multiple original vectors of meshing impact intensity and multiple original vectors of meshing impact amplitude.

[0106] The meshing impact intensity characteristic matrix and the meshing impact amplitude characteristic matrix are N×4, where N is the number of rows. The first column of the matrix is ​​the mean, the second column is the maximum value, the third column is the kurtosis, and the fourth column is the standard deviation.

[0107] The mean extracted by this invention reflects the overall level of the feature, indicating the average intensity / amplitude of the defect impact; the maximum value captures extreme impact situations, corresponding to the strongest impact moment caused by the defect; the kurtosis characterizes the steepness of the distribution, revealing whether there are a large number of peak-like defect impacts; and the standard deviation reflects the degree of fluctuation of the feature.

[0108] This invention extracts features from the meshing impact intensity feature matrix and the meshing impact amplitude feature matrix, respectively, and generates attention based on the joint features to enhance the features. The two types of enhanced features are then classified to obtain the tooth track defect diagnosis results.

[0109] like Figure 2 As shown, the dual-branch defect diagnosis neural network includes: a first-level intensity feature extraction unit, a first-level amplitude feature extraction unit, an association attention generation unit, a first multiplication unit, a second multiplication unit, a second-level intensity feature extraction unit, a second-level amplitude feature extraction unit, and a fully connected layer;

[0110] The first-level strength feature extraction unit is used to extract features from the meshing impact strength feature matrix to obtain the first-level strength features;

[0111] The first-level amplitude feature extraction unit is used to extract features from the meshing impact amplitude feature matrix to obtain the first-level amplitude features;

[0112] The associative attention generation unit is used to generate associative attention based on the first-level intensity features and the first-level amplitude features;

[0113] The first multiplication unit is used to multiply the first-level intensity feature with the associated attention to obtain the first-level intensity enhancement feature;

[0114] The second multiplication unit is used to multiply the first-level amplitude feature with the associated attention to obtain the first-level amplitude enhancement feature;

[0115] The second-level intensity feature extraction unit is used to extract features from the first-level intensity enhancement features to obtain the second-level intensity enhancement features;

[0116] The second-level amplitude feature extraction unit is used to extract features from the first-level amplitude enhancement features to obtain the second-level amplitude enhancement features;

[0117] The fully connected layer is used to classify tooth track defects based on the second-level strength enhancement characteristics and the second-level amplitude enhancement characteristics.

[0118] The dual-branch defect diagnosis neural network of this invention extracts branches from the meshing impact intensity feature matrix and the meshing impact amplitude feature matrix separately, and introduces an associative attention mechanism in the intermediate layer to achieve targeted enhancement of the core defect features that are "significantly strong in energy coupling and have abnormally prominent amplitudes." This strengthens the coupling information strongly correlated with the essence of the defect in both types of features, suppresses irrelevant noise interference, and thus significantly improves the discriminativeness of diagnostic features and the robustness to signal changes under complex working conditions. Specifically, the first-level feature extraction unit can fully mine the primary features of intensity and amplitude, and the associative attention generation unit adaptively allocates weights according to the interaction relationship between the two types of features, enabling the network to focus on key feature regions and suppress redundant information; the multiplicative enhancement mechanism realizes dynamic enhancement of features, improving the sensitivity of abnormal responses to the model. The second-level feature extraction unit further extracts the deep information of the enhanced features, forming a more stable and hierarchical representation; finally, the fully connected layer fuses the two high-dimensional features of intensity and amplitude to complete the accurate classification of tooth rail defects.

[0119] like Figure 3 As shown, the associative attention generation unit includes: adder A1, pointwise convolutional layer and sigmoid layer;

[0120] Adder A1 is used to add the first-level intensity enhancement feature and the first-level amplitude enhancement feature element-wise to obtain the intensity-amplitude joint feature; the pointwise convolutional layer is used to perform pointwise convolution on each feature value in the intensity-amplitude joint feature to obtain the mapping feature; the sigmoid layer is used to generate attention based on the mapping feature.

[0121] Adder A1 adds the first-level intensity enhancement feature and the first-level amplitude enhancement feature element-wise, essentially superimposing and fusing the anomalous information of the two types of features. When a feature location simultaneously exhibits a synergistic anomaly of "high intensity" and "large amplitude" (this is the core signal characterization of toothed rail defects, such as the impact caused by a crack leading to simultaneous nonlinear enhancement of intensity and a sudden increase in amplitude energy), the feature value at that location will be significantly higher than in other areas after addition; while at locations with only a single feature anomaly or no anomaly, the feature value will remain at a low level, allowing for rapid localization of the synergistic anomaly features of the two types of features.

[0122] like Figure 4 As shown, the first-level intensity feature extraction unit and the first-level amplitude feature extraction unit have the same structure, both including the following connected sequentially: a first 1D convolutional layer, a 1D pooling layer, and a second 1D convolutional layer.

[0123] The first 1D convolutional layer, the 1D pooling layer, and the second 1D convolutional layer all process the data in the time dimension, that is, they perform convolution and pooling on each column of the matrix. The kernel size of the first 1D convolutional layer and the second 1D convolutional layer is 3, and the stride is set to 1. The kernel size of the 1D pooling layer is 2, and the stride is 2.

[0124] like Figure 5 As shown, the second-level intensity feature extraction unit and the second-level amplitude feature extraction unit have the same structure, both including: a first 2D convolutional layer, a second 2D convolutional layer, a third 2D convolutional layer, a Concat layer, a BN layer, and a ReLU layer;

[0125] The input of the first 2D convolutional layer serves as the input of the second-level intensity feature extraction unit and the second-level amplitude feature extraction unit, and its output is connected to the input of the second 2D convolutional layer and the first input of the Concat layer, respectively.

[0126] The output of the second 2D convolutional layer is connected to the input of the third 2D convolutional layer and the second input of the Concat layer, respectively; the output of the third 2D convolutional layer is connected to the third input of the Concat layer; the output of the Concat layer is connected to the input of the BN layer; the input of the ReLU layer is connected to the output of the BN layer, and its output serves as the output of the second-level intensity feature extraction unit and the second-level amplitude feature extraction unit.

[0127] The kernel size of the first, second, and third 2D convolutional layers is set to 3×3, the stride is set to 1, and the padding is set to Same padding.

[0128] This invention extracts features at different depths through first, second, and third 2D convolutional layers, fuses multi-scale features via a Concat layer, and combines batch normalization of a BN layer with nonlinear activation of a ReLU layer. This achieves multi-dimensional and in-depth extraction and enhancement of gear meshing impact features. The Concat layer preserves complementary information of features from each layer, while the BN and ReLU layers improve the stability and nonlinear expressive power of the feature distribution, respectively. Ultimately, the output features possess both fine-grained details and global correlations, providing a more discriminative and robust input for defect classification and significantly improving the accuracy and generalization ability of gear defect diagnosis.

[0129] In this embodiment, the diagnostic results of the toothed rail defects include: no defects (normal toothed rail meshing state), slight tooth surface wear (slight wear on the tooth surface with little impact on operation), severe tooth surface wear (severe wear on the tooth surface, requiring timely maintenance), tooth root cracks (cracks appear at the tooth root, posing a risk of derailment), and tooth surface spalling (local material spalling off the tooth surface, affecting meshing stability).

[0130] After acquiring vibration signals, this invention performs signal enhancement processing. The amplitude of the nonlinear impact region (including defect features) is selectively amplified, while the noise in the stable noise region (low transient amplitude) is relatively suppressed. This achieves targeted enhancement of the "nonlinear impact features" rather than indiscriminate signal amplification, effectively suppressing interference from environmental noise and background vibration, and solving the problem that weak defect signals are easily masked.

[0131] Traditional techniques rely on linear parameters such as amplitude and frequency, which are easily diluted by the overall vibration signal and cannot accurately reflect defects. This invention addresses this by, on the one hand, calculating the average energy coupling strength of frequency components within the gear meshing frequency band to capture the nonlinear characteristics of meshing impact caused by defects, a feature that is more sensitive to defects; on the other hand, constructing impact amplitude features based on the ratio of anomalies to average amplitudes to accurately amplify amplitude anomalies caused by defects. This dual-dimensional approach ensures that the extracted features are truly correlated with defects, extracting features precisely related to defects, accurately reflecting gear track defects, and improving diagnostic accuracy.

[0132] This invention significantly improves the reliability of diagnostic results at the decision-making level by leveraging a dual-branch neural network. Traditional diagnostic methods often rely on single feature inputs, resulting in limited information dimensions and a high risk of misjudgment and missed diagnosis. The dual-branch defect diagnostic neural network of this invention can process two types of feature matrices—meshing impact intensity and amplitude—separately, mining defect information from both "intensity" and "amplitude" dimensions. Combined with the powerful feature fusion and pattern recognition capabilities of the neural network, it can more clearly distinguish between normal and defective vibration modes, significantly reducing the misjudgment rate and ultimately outputting more accurate and reliable diagnostic results, thus solving the problem of "low diagnostic accuracy."

[0133] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting toothed rail defects based on meshing impact characteristics, characterized in that, Includes the following steps: Vibration sensors are installed on the rack rail of the section to be tested to collect vibration signals generated when the train passes over the rack rail, and the vibration signals are then enhanced in the time domain to obtain the enhanced signal: The vibration signal is subjected to Hilbert transform to obtain the transformed signal; Calculate the transient amplitude based on the transformed signal; The enhanced signal is obtained by multiplying the vibration signal by the transient amplitude at each time point. In the gear meshing frequency band of the enhanced signal, the energy coupling strength of any two frequency components is obtained, and the average value is taken to obtain the average energy coupling strength: ， Among them, B(f) i ,f j ) represents the i-th frequency component f i and the j-th frequency component f j The energy coupling strength, f i +f j The conjugate complex amplitude, X(f) i ) represents the i-th frequency component f i The complex amplitude value, X(f) j ) represents the j-th frequency component f j The complex amplitude value; Based on the energy coupling intensity of gear meshing harmonics, the meshing impact intensity feature is extracted by the ratio of abnormal to average energy coupling intensity, and the original vector of meshing impact intensity is constructed: When two frequency components are simultaneously located in the gear meshing harmonic frequency set, the corresponding energy coupling strength is taken as the meshing energy coupling strength. By screening out meshing energy coupling strengths that are greater than the average energy coupling strength, abnormal energy coupling strengths are obtained. The ratio of abnormal energy coupling strength to average energy coupling strength is used as the meshing impact strength feature, and the meshing impact strength features of each gear meshing frequency band are used to form the original meshing impact strength vector. Based on the amplitude information of the gear meshing frequency band, the meshing impact amplitude features are extracted by the ratio of abnormal to average amplitude, and the original vector of meshing impact amplitude is constructed: The mean amplitude is extracted from the gear meshing frequency band to obtain the average amplitude; The amplitude of the frequency component corresponding to the abnormal energy coupling strength is taken as the abnormal amplitude; The ratio of abnormal amplitude to average amplitude is used as the meshing impact amplitude feature, and the meshing impact amplitude features of each gear meshing frequency band are used to form the original meshing impact amplitude vector. Feature values ​​are extracted from the original vectors of meshing impact intensity and meshing impact amplitude, respectively, and feature matrices of meshing impact intensity and meshing impact amplitude are constructed. A dual-branch defect diagnosis neural network is used to process the meshing impact intensity feature matrix and the meshing impact amplitude feature matrix to obtain the tooth rail defect diagnosis results.

2. The method for detecting tooth rail defects based on meshing impact characteristics according to claim 1, characterized in that, The process of obtaining the average energy coupling strength includes: A short-time Fourier transform is performed on the enhanced signal to obtain multiple spectra; In each spectrum, the frequency bands belonging to the gear meshing frequency range are extracted as the gear meshing frequency bands; Select any two frequency components in the gear meshing frequency band and extract the complex amplitude values ​​of the two frequency components in the frequency band; Calculate the conjugate complex amplitude at the sum of two frequency components; The energy coupling strength is calculated based on the complex amplitude and conjugate complex amplitude of the two frequency components. The average energy coupling intensity is obtained by averaging the energy coupling intensities across all energy coupling frequencies in the gear meshing band.

3. The method for detecting tooth rail defects based on meshing impact characteristics according to claim 1, characterized in that, The process of constructing the meshing impact intensity characteristic matrix and the meshing impact amplitude characteristic matrix includes: Extract the mean, maximum value, kurtosis, and standard deviation from the original vector of meshing impact strength; Extract the mean, maximum value, kurtosis, and standard deviation from the original vector of meshing impact amplitude; The mean, maximum value, kurtosis, and standard deviation of each original vector of meshing impact strength are used to construct the meshing impact strength feature matrix; The mean, maximum value, kurtosis, and standard deviation of each original vector of meshing impact amplitude are used to construct the meshing impact amplitude characteristic matrix.

4. The method for detecting tooth rail defects based on meshing impact characteristics according to claim 1, characterized in that, The dual-branch defect diagnosis neural network includes: a first-level intensity feature extraction unit, a first-level amplitude feature extraction unit, an association attention generation unit, a first multiplication unit, a second multiplication unit, a second-level intensity feature extraction unit, a second-level amplitude feature extraction unit, and a fully connected layer; The first-level strength feature extraction unit is used to extract features from the meshing impact strength feature matrix to obtain the first-level strength features; the first-level amplitude feature extraction unit is used to extract features from the meshing impact amplitude feature matrix to obtain the first-level amplitude features. The associative attention generation unit generates associative attention based on the first-level intensity feature and the first-level amplitude feature; the first multiplication unit multiplies the first-level intensity feature with the associative attention to obtain the first-level intensity enhancement feature; the second multiplication unit multiplies the first-level amplitude feature with the associative attention to obtain the first-level amplitude enhancement feature; the second-level intensity feature extraction unit extracts features from the first-level intensity enhancement feature to obtain the second-level intensity enhancement feature; the second-level amplitude feature extraction unit extracts features from the first-level amplitude enhancement feature to obtain the second-level amplitude enhancement feature; the fully connected layer classifies based on the second-level intensity enhancement feature and the second-level amplitude enhancement feature to obtain the tooth track defect diagnosis result.

5. The method for detecting tooth rail defects based on meshing impact characteristics according to claim 4, characterized in that, The associative attention generation unit includes: adder A1, pointwise convolutional layer and sigmoid layer; Adder A1 is used to add the first-level intensity enhancement feature and the first-level amplitude enhancement feature element-wise to obtain the intensity-amplitude joint feature; the pointwise convolutional layer is used to perform pointwise convolution on each feature value in the intensity-amplitude joint feature to obtain the mapping feature; the sigmoid layer is used to generate attention based on the mapping feature.

6. The method for detecting tooth rail defects based on meshing impact characteristics according to claim 4, characterized in that, The first-level intensity feature extraction unit and the first-level amplitude feature extraction unit have the same structure, both including the following sequentially connected: a first 1D convolutional layer, a 1D pooling layer and a second 1D convolutional layer; The second-level intensity feature extraction unit and the second-level amplitude feature extraction unit have the same structure, both including: a first 2D convolutional layer, a second 2D convolutional layer, a third 2D convolutional layer, a Concat layer, a BN layer, and a ReLU layer; The first, second, and third 2D convolutional layers are used to extract features at different depths; the Concat layer is used to concatenate features at different depths to obtain concatenated features; the BN layer is used to batch normalize the concatenated features; and the ReLU layer is used to perform non-linear activation on the batch-normalized features.

Images