A fault sensitive band adaptive enhancement method for train bogie multi-measuring point vibration signals

Abstract

This invention discloses an adaptive enhancement method for fault-sensitive frequency bands of multi-measurement vibration signals from train bogies. It employs a multi-window candidate fault frequency drift trajectory constraint to smoothly update wavelet parameters and constructs an adaptive graph gain modulation mechanism based on a triplet of measurement point-candidate fault frequency trajectory-wavelet frequency band. This enhances high-reliability fault-sensitive frequency bands while suppressing low-reliability pseudo-candidate frequencies and noise bands, thus stably preserving weak fault impact characteristics under complex operating conditions. This provides a reliable signal representation for subsequent bogie fault diagnosis, condition identification, or fault source analysis. It addresses the problems in existing technologies, such as difficulty in stably selecting fault-sensitive frequency bands under conditions of multi-source excitation coupling, changing operating conditions, strong noise interference, and differences in multi-measurement point responses. These problems include the ease with which weak fault impact characteristics are submerged by noise, the difficulty of adapting fixed wavelet parameters to frequency band drift, and the coarse-level fusion of multi-measurement point information.

Description

Technical Field

[0001] This invention belongs to the field of train bogie condition monitoring and vibration signal processing technology, specifically relating to a fault-sensitive frequency band adaptive enhancement method for multi-point vibration signals of train bogies. Background Technology

[0002] The bogie is a critical load-bearing and transmission component in train operation, and its operating status directly affects the safety and stability of train operation. The bogie transmission system typically includes components such as traction motors, gearboxes, axle boxes, and wheelsets. During train operation, it is subjected to the combined effects of multiple excitation sources, including wheel-rail excitation, traction load variations, gear meshing impact, axle box support vibration, and external random noise. When a critical bogie component experiences an early failure, the fault impact often manifests as a vibration response with weak amplitude, short duration, and unstable frequency band distribution, easily overwhelmed by low-frequency structural vibration, high-frequency random noise, and coupled excitation from other components. Therefore, how to adaptively select the fault-sensitive frequency band under complex operating conditions and strong noise, and how to stably enhance the weak fault impact components, are important issues in train bogie vibration signal processing and fault diagnosis.

[0003] In existing methods for fault diagnosis of rotating machinery, spectral coherence, cyclic spectral coherence, enhanced envelope spectrum, and weighted envelope spectrum methods have been used to extract fault features from components such as bearings and gears. These methods typically highlight fault features by identifying candidate fault frequencies, selecting information frequency bands, and generating enhanced or weighted envelope spectra. However, these methods are mostly geared towards single measurement points or relatively stable operating conditions, focusing on candidate frequency identification and envelope spectrum enhancement, but lacking tracking and constraint on the drift trajectory of candidate fault frequencies within a continuous window. In the multi-source coupling and variable operating conditions of train bogies, single-window candidate frequencies or single-band evaluation indicators are easily affected by noise spurious peaks, occasional impacts, and structural resonance interference, leading to unstable selection of fault-sensitive frequency bands.

[0004] Existing adaptive frequency band selection and spectrum reconstruction methods typically evaluate frequency bands using metrics such as spectral kurtosis, impulse factor, sparsity, and envelope entropy, and select the optimal resonant frequency band or target characteristic frequency band for enhanced reconstruction. While these methods can mitigate the feature aliasing problem caused by fixed frequency band division to some extent, their frequency band selection often relies on the statistical characteristics of the current window or the current signal, making it difficult to consistently cover the real fault-sensitive region that drifts with changing operating conditions. Furthermore, when noise spurious peaks or occasional impacts are strong, the frequency band selection result may abruptly change, thus affecting the stable preservation of weak fault impact characteristics.

[0005] Wavelet transform, wavelet packet decomposition, and wavelet graph convolution methods possess good time-frequency localization capabilities, making them suitable for processing non-stationary impact signals. However, the center frequency, scale, and bandwidth of traditional wavelet methods largely depend on preset parameters, empirical selection, or optimization using a single evaluation index. If the wavelet bandwidth is too narrow, it is easy to miss the drifted fault impact component; if the bandwidth is too wide, it will introduce wheel-rail excitation, structural vibration, and random noise. Therefore, existing wavelet enhancement methods still struggle to balance fault impact preservation and noise suppression under complex bogie operating conditions.

[0006] Furthermore, train bogies typically house multiple vibration measurement points, including traction motors, gearboxes, and axle boxes. Existing multi-sensor fault diagnosis methods often employ multi-channel signal splicing, feature fusion, classifier recognition, or classification result voting. Some methods also utilize graph neural networks to model sensor relationships. These methods primarily target fault classification or state recognition, with fusion levels concentrated at the channel, feature, or classification result levels. They lack reliability evaluation for fine-grained objects such as "a specific measurement point, a candidate fault frequency trajectory, or a wavelet frequency band," and do not directly utilize graph learning results for wavelet coefficient enhancement or suppression. Therefore, in bogie scenarios with significant differences in response among multiple measurement points, simple splicing or channel-level fusion easily introduces invalid measurement points, noise bands, and spurious candidate frequencies.

[0007] In summary, although existing technologies have made some progress in candidate fault frequency extraction, enhanced envelope spectrum, adaptive frequency band reconstruction, wavelet time-frequency analysis, and multi-sensor fusion diagnosis, they still have the following shortcomings: First, they lack stability constraints on the continuous window drift trajectory of candidate fault frequencies, making them susceptible to noise spurious peaks and occasional impacts; second, they lack a wavelet parameter smoothing update mechanism driven by the candidate fault frequency drift trajectory, making it difficult to adapt to frequency band drift under complex operating conditions; and third, they lack a wavelet coefficient-level gain modulation mechanism oriented towards the measurement point-candidate fault frequency trajectory-wavelet frequency band triplet, making it difficult to finely distinguish the reliability of different measurement points, different candidate frequencies, and different frequency bands.

[0008] Therefore, it is necessary to propose an adaptive enhancement method for fault-sensitive frequency bands of multi-point vibration signals of train bogies. This method uses a multi-window candidate fault frequency drift trajectory to constrain wavelet parameter updates and constructs an adaptive graph gain modulation mechanism based on the measurement point-candidate fault frequency trajectory-wavelet frequency band triplet. This enhances the high-reliability fault-sensitive frequency band and suppresses the low-reliability pseudo-candidate frequencies and noise frequency bands, thereby stably preserving the weak fault impact characteristics under complex operating conditions and providing a reliable signal representation for subsequent bogie fault diagnosis, condition identification, or fault source analysis. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides an adaptive enhancement method for fault-sensitive frequency bands of vibration signals from multiple measurement points on train bogies. The aim is to solve the problems in existing technologies, such as the difficulty in stably selecting fault-sensitive frequency bands under conditions of multi-source excitation coupling, changing operating conditions, strong noise interference, and differences in responses from multiple measurement points; the easy submersion of weak fault impact characteristics by noise; the difficulty of adapting fixed wavelet parameters to frequency band drift; and the coarse level of multi-measurement point information fusion.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] An adaptive enhancement method for fault-sensitive frequency bands of vibration signals from multiple measurement points on train bogies includes the following steps:

[0012] Step 1: Collect vibration signals from multiple measurement points of the train bogie transmission system, and perform windowing and preprocessing; extract candidate fault frequencies from each measurement point and each time window;

[0013] Step 2: Correlate the candidate fault frequencies in the continuous window to form CFF drift trajectories, and perform wavelet parameter smoothing updates;

[0014] Step 3: Based on the updated wavelet parameters, perform wavelet decomposition on the vibration signals at each measuring point; obtain the wavelet coefficient components corresponding to different measuring points, different candidate fault frequency trajectories, and different wavelet frequency bands.

[0015] Step 4: Construct triplet nodes for measurement points, candidate fault frequency trajectories, and wavelet bands, and establish a TCB-AGM triplet adaptive graph based on the triplet nodes. Learn the wavelet gain of the triplet adaptive graph to obtain the wavelet coefficient gain corresponding to each triplet node, which directly applies to the corresponding measurement point, the corresponding candidate fault frequency trajectory, and the corresponding wavelet band.

[0016] Step 5: Perform inverse wavelet transform and multi-measurement point fusion reconstruction on the modulated wavelet coefficients to obtain the enhanced fault-sensitive vibration signal, the enhanced signal of each measurement point, the fault-sensitive wavelet time-frequency representation, and the reliability map of the measurement point-candidate fault frequency trajectory-wavelet frequency band triplet or the fault-sensitive frequency band enhancement map.

[0017] Furthermore, the processing steps in step 1 include:

[0018] Step 1.1: Collect data from the train bogie transmission system Vibration signal at each measuring point The signals from each measuring point are windowed to obtain the first... Multiple measurement points within a time window Preprocessing of the signals within each window is denoted as follows: ;

[0019] Step 1.2: Extract the vibration signal from each measuring point and each window to extract the candidate fault frequency set, denoted as: ;in, Indicates the first The measuring point at the ... The first one extracted from the window Candidate fault frequencies, The total number of candidate fault frequencies;

[0020] Step 1.3: Construct the candidate frequency evaluation function ,like At that time, the corresponding frequency This serves as the candidate fault frequency within the window.

[0021] Furthermore, step 2 includes the following processing steps:

[0022] Step 2.1: For the candidate fault frequencies extracted within the continuous window, perform trajectory association to form the candidate fault frequency drift trajectory;

[0023] Step 2.2: Calculate the stability for each CFF drift trajectory Based on stability Construct the parameter update gate function: , The preset trajectory stability threshold is used to determine whether the trajectory is allowed to participate in wavelet parameter updates;

[0024] Step 2.3: Smoothly update the wavelet center frequency, bandwidth, and scale search interval based on the CFF drift trajectory.

[0025] Furthermore, the update process in step 2.3 is as follows:

[0026] Step 2.3.1: For the first A stable CFF drift trajectory is obtained, and the trajectory is smoothly updated by combining the center frequency of the previous window, the current CFF frequency, and the trajectory stability.

[0027] ;

[0028] in, For the first The first window The wavelet center frequency corresponding to each trajectory The wavelet center frequency of the previous window, For the frequency corresponding to the CFF trajectory of the current window, Update the gating coefficients for the parameters. For smoothing coefficients;

[0029] Step 2.3.2: The wavelet bandwidth is determined by the harmonic interval, sideband width, and CFF drift uncertainty, denoted as:

[0030] ;

[0031] in, For the first The wavelet bandwidth corresponding to each CFF trajectory. Basic broadband, The width of the sideband distribution. For harmonic intervals or harmonic extension range, For CFF drift uncertainty, This is the adjustment coefficient;

[0032] Step 2.3.3: Determine the wavelet scale search interval based on the center frequency and bandwidth, where the center frequency is... Mother wavelet,

[0033] No. The wavelet scale search interval corresponding to each CFF trajectory is represented as follows:

[0034] ;

[0035] in, The center frequency of the mother wavelet.

[0036] Furthermore, step 3 uses a mother wavelet with time-frequency localization capability for wavelet decomposition, including but not limited to Morlet wavelet, complex Morlet wavelet, and generalized Morse wavelet.

[0037] Furthermore, step 4 includes the following processing steps:

[0038] Step 4.1: Construct a triplet node by combining each measurement point, CFF drift trajectory, and wavelet frequency band: ;

[0039] All triplet nodes are combined into a node set: ;

[0040] The feature vector of each node is defined as: ;

[0041] in, For the first One measuring point, For the first A CFF drift trajectory, For the first Each wavelet frequency band; the eigenvectors contain, in order, local peak significance, harmonic integrity, sideband matching degree, CFF trajectory stability, CFF drift uncertainty, cross-measurement point consistency, measurement point location prior, and corresponding wavelet frequency band energy or impact index.

[0042] Step 4.2: Construct the triplet adaptive graph, denoted as: ;in, For a set of nodes, Let be the set of edges. It is an adjacency matrix;

[0043] The graph-side relationships include the measurement point response relationship, frequency band adjacency relationship, harmonic / sideband structure relationship between candidate fault frequencies, prior relationship of mechanical connection of measurement points, and trajectory continuation relationship of candidate fault frequencies.

[0044] Step 4.3: Adjacency Matrix It consists of both prior adjacency and data adaptive adjacency:

[0045] ;

[0046] in: It is the a priori adjacency matrix composed of mechanical connection relationships, frequency band adjacency relationships, and harmonic / sideband structure relationships. This is the data adjacency matrix learned from node feature similarity and response consistency. The fusion coefficient of prior and data-driven adjacency;

[0047] The adaptive adjacency representation of data is as follows:

[0048] ;

[0049] in: This is the index for the triple node. , For the first Node characteristics For learnable mapping functions, For nodes The set of neighbors;

[0050] Step 4.4: Update the graph nodes of the triplet graph to obtain the wavelet coefficient gain of each triplet node, denoted as: ;

[0051] in, , indicating the first The measuring point, the first CFF trajectory, number The reliability gain of each wavelet band. For low-reliability nodes, its The smaller the value, the more the corresponding frequency band is suppressed; for high-reliability nodes, its... The larger the value, the stronger the fault-sensitive wavelet component is; This represents the transpose of the learnable weight vector of the gain mapping layer. Indicates the first After the layer graph node is updated, the first The measuring point, the first CFF trajectory, number The implicit feature representation of each wavelet frequency band corresponding to the triplet node. This represents the bias term of the gain mapping layer.

[0052] Furthermore, step 5 includes the following processing steps:

[0053] Step 5.1: Modulate the wavelet coefficients corresponding to each measurement point, each CFF trajectory, and each frequency band based on the wavelet coefficient gain output by the triplet adaptive graph;

[0054] Step 5.2: Perform inverse wavelet transform on the modulated wavelet coefficients to obtain the... The enhanced signals from multiple measurement points are fused based on the reliability of each measurement point to obtain a multi-measurement point fault-sensitive reconstructed signal.

[0055] The measurement point fusion weights are obtained by aggregated triplet gains. ,

[0056] The enhanced fault-sensitive signal over the entire time range is finally obtained, denoted as:

[0057] ;

[0058] in, This indicates a reconstruction based on overlapping and addition.

[0059] The final output is an enhanced fault-sensitive vibration signal. Fault-sensitive signals enhanced at each measuring point Fault-sensitive wavelet time-frequency representation Measurement point—CFF trajectory—frequency band triplet reliability diagram And fault-sensitive frequency band enhancement maps or frequency band reliability maps.

[0060] Furthermore, step 5.1 employs a bidirectional modulation method:

[0061]

[0062] in, To enhance the coefficient, The inhibition coefficient; These are wavelet coefficient components. When... At higher levels, the corresponding frequency band is enhanced; when At lower levels, the corresponding frequency band is suppressed.

[0063] Beneficial effects:

[0064] (1) This invention evaluates the stability of candidate fault frequency drift trajectories and generates parameter update gating quantities based on frequency continuity, amplitude variation continuity, harmonic structure continuity, and multi-measurement point consistency. For candidate fault frequencies that can form stable drift trajectories, they are included in wavelet parameter updates; for candidate fault frequencies that only suddenly appear in a single window or single measurement point, wavelet parameter updates are not directly driven. Based on the candidate fault frequency drift trajectories and parameter update gating quantities, this invention uses the CFF-TWPU module to smoothly update the wavelet center frequency, bandwidth, and scale search interval, enabling the wavelet enhancement frequency band to adaptively adjust with the drift of the real fault-sensitive frequency band.

[0065] (2) Based on the updated wavelet parameters, this invention performs wavelet decomposition on the vibration signals of each measuring point to obtain wavelet coefficient components corresponding to different measuring points, different candidate fault frequency trajectories, and different wavelet frequency bands. To address the differences in response at multiple measuring points of the train bogie, this invention constructs a triplet node structure of measuring point—candidate fault frequency trajectory—wavelet frequency band, and establishes a TCB-AGM triplet adaptive graph based on the triplet node structure. This graph incorporates the measuring point response relationship, frequency band adjacency relationship, harmonic / sideband structure relationship between candidate fault frequencies, prior relationship of measuring point mechanical connections, and continuation relationship of candidate fault frequency trajectories into the graph structure modeling.

[0066] (3) This invention constructs the adjacency matrix of the triplet adaptive graph by combining the prior adjacency matrix and the data adaptive adjacency matrix, and obtains the wavelet coefficient gain corresponding to each triplet node by updating the graph nodes. The wavelet coefficient gain is not used to directly output the fault category, but directly acts on the wavelet coefficients of the corresponding measurement point, the corresponding candidate fault frequency trajectory and the corresponding wavelet frequency band, so as to enhance the sensitive frequency band of high reliability faults and suppress the low reliability pseudo candidate frequency, noise frequency band or inconsistent frequency band across measurement points.

[0067] (4) This invention performs inverse wavelet transform and multi-measurement point fusion reconstruction on the modulated wavelet coefficients to obtain the enhanced fault-sensitive vibration signal, the enhanced signal of each measurement point, the fault-sensitive wavelet time-frequency representation, and the reliability map of the measurement point-candidate fault frequency trajectory-wavelet frequency band triplet or the enhanced map of the fault-sensitive frequency band. The above results can be used as input for subsequent train bogie fault diagnosis, condition identification or fault source analysis. The core of this invention is not to perform candidate fault frequency identification, envelope spectrum generation, wavelet transform or multi-measurement point classification fusion alone, but to: on the one hand, to solve the problem of fixed wavelet parameter failure caused by fault-sensitive frequency band drift under complex working conditions by constraining the smooth update of wavelet parameters through the multi-window candidate fault frequency drift trajectory; on the other hand, to solve the problem of how to differentiate the enhancement of different measurement points, different candidate frequencies and different frequency bands under the condition of different response of multiple measurement points by outputting the wavelet coefficient gain through the adaptive map of measurement point-candidate fault frequency trajectory-wavelet frequency band triplet. Attached Figure Description

[0068] Figure 1 This is the overall logic flowchart of the fault-sensitive frequency band adaptive enhancement method for multi-point vibration signals of train bogies;

[0069] Figure 2 This is a detailed structural diagram of the TCB-AGM module in this invention;

[0070] Figure 3 This is the TCB-AGM gain network diagram in this invention. Detailed Implementation

[0071] This embodiment uses multi-point vibration signals from a train bogie transmission system as the processing object. It employs a multi-window candidate fault frequency drift trajectory constraint to smoothly update wavelet parameters, and combines this with adaptive gain modulation of the measurement point-candidate fault frequency trajectory-wavelet frequency band triplet to achieve adaptive enhancement of fault-sensitive frequency bands and suppression of low-reliability noise frequency bands. The overall process is as follows: Figure 1 As shown, the detailed structure of the TCB-AGM module is as follows: Figure 2 As shown, the TCB-AGM graph gain network is as follows: Figure 3 As shown.

[0072] Combined with appendix Figure 1-3 This embodiment provides a detailed description of an adaptive enhancement method for fault-sensitive frequency bands of vibration signals from multiple measurement points on train bogies. The specific steps of this method are as follows:

[0073] Step 1: Preprocessing of multi-point signals and extraction of candidate fault frequencies.

[0074] Step 1.1: Vibration sensors are installed at multiple key locations in the train bogie transmission system to collect vibration signals from multiple measurement points. Measurement points include, but are not limited to, traction motor measurement points, gearbox measurement points, left axle box measurement points, and right axle box measurement points. Assume there are a total of... The measuring point, the first The vibration signals collected from each measuring point are represented as follows:

[0075]

[0076] The signals from each measuring point are windowed to obtain the first... Multiple measurement points within a time window:

[0077]

[0078] in, Indicates the window length. Indicates the window sliding step size. For each window, the signal is detrended, normalized, filtered, or subjected to speed / operating condition calibration to obtain the preprocessed signal, denoted as: .

[0079] Step 1.2: Extract a set of candidate fault frequencies from the vibration signal at each measuring point and within each window. The candidate fault frequencies can be obtained by cyclostationary characteristics, spectral coherence characteristics, envelope spectrum stable peak values, time-frequency energy peak values, or cross-window frequency stability analysis.

[0080] The process of extracting candidate fault frequencies specifically includes: performing cyclostationary characteristic analysis, spectral coherence characteristic analysis, envelope spectrum peak analysis, and time-frequency energy distribution analysis on the preprocessed vibration signal at each measuring point and within each window to obtain candidate frequency points that can reflect the periodic impact components. For cyclostationary characteristics, frequency components with obvious peaks in the cyclic frequency direction can be extracted by calculating the cyclic spectrum or cyclic autocorrelation results of the signal; for spectral coherence characteristics, possible candidate fault frequencies can be determined based on the local peak positions in the spectral coherence plane; for envelope spectrum stable peaks, envelope demodulation can be performed on the signal within the window to extract frequencies in the envelope spectrum that satisfy the significance of local peaks and harmonic structures; for time-frequency energy peaks, frequency components with concentrated energy in the local time-frequency region can be extracted based on wavelet transform, short-time Fourier transform, or other time-frequency analysis results; for cross-window frequency stability analysis, frequency points with similar frequency positions, continuous peak changes, and similar harmonic structures in adjacent windows can be used as candidate fault frequencies. Subsequently, a candidate frequency evaluation function is used. A comprehensive evaluation of the above frequency points is conducted when... At that time, the corresponding frequency This serves as the candidate fault frequency within the window.

[0081] No. The measuring point, the first The set of candidate fault frequencies for each window is represented as:

[0082]

[0083] in, Indicates the first The measuring point at the ... The first one extracted from the window Candidate fault frequencies, This represents the total number of candidate fault frequencies.

[0084] Step 1.3: Construct the candidate frequency evaluation function:

[0085]

[0086]

[0087] in, To demonstrate the significance of frequency peaks, To demonstrate the saliency of the harmonic structure, It reflects local frequency band energy or envelope energy. These are the weighting coefficients for the three items. Considering that strong energy noise and structural resonance in the bogie vibration signal can easily interfere with candidate frequency extraction, this embodiment sets the weight of harmonic structure significance higher than the weight of local peak significance, and the weight of local peak significance higher than the weight of local energy. Specifically, the weights are:

[0088]

[0089] Right now:

[0090]

[0091] For each frequency point to be evaluated Based on the significance of its frequency peak Significance of harmonic structure and local frequency band energy or envelope energy Calculate the comprehensive evaluation value of candidate frequencies .when At that time, it is assumed that the frequency point simultaneously satisfies the peak prominence, harmonic structure, and local energy conditions, and the corresponding frequency is... As the candidate fault frequency within this window; when At that time, this frequency point will not be considered as a candidate fault frequency. The preset candidate frequency evaluation threshold is used to determine whether the frequency point to be evaluated meets the candidate fault frequency screening conditions.

[0092] Step 2: Multi-window CFF drift trajectory construction and wavelet parameter smoothing update.

[0093] Step 2.1: For the candidate fault frequencies extracted within the continuous window, perform trajectory association to form the candidate fault frequency drift trajectory.

[0094] Definition of the first The candidate fault frequency drift trajectories are as follows:

[0095]

[0096] in, Indicates the first The trajectory in the first Candidate fault frequency in each window.

[0097] When associating candidate frequencies in adjacent windows, construct the trajectory association distance:

[0098]

[0099]

[0100] in, The candidate frequency interval between adjacent windows. The variation of the peak amplitude of candidate frequencies in adjacent windows. To ensure the continuity of the harmonic structure of candidate frequencies in adjacent windows, These are the correlation weights. Considering that the essence of CFF drift trajectory is that the frequency changes continuously with the window, and that the actual fault frequency usually has a certain harmonic structure, while the amplitude is easily affected by load, speed, wheel-rail impact, and measurement point location, in this embodiment, we take:

[0101]

[0102] Right now:

[0103]

[0104] When satisfied At that time, it was believed Compared to the previous window Belonging to the same CFF drift trajectory, among which, The preset trajectory association distance threshold is used to determine whether the candidate fault frequencies in adjacent windows meet the trajectory continuous association condition.

[0105] Step 2.2: Calculate the stability of each CFF drift trajectory to determine whether the trajectory is allowed to participate in wavelet parameter updates.

[0106] Definition of the first The stability of the CFF drift trajectory is:

[0107]

[0108] in, For frequency continuity, For the continuity of amplitude variation, For the continuity of harmonic structure, To ensure consistency across multiple measurement points, These are the weighting coefficients.

[0109] Frequency continuity is expressed as:

[0110]

[0111] The consistency of multiple measurement points can be expressed as:

[0112]

[0113] in, This is a frequency continuity scaling parameter used to control the degree of influence of the frequency continuity evaluation value on the difference in candidate fault frequencies between adjacent windows. This represents the mean of the candidate frequencies corresponding to this trajectory across multiple measurement points. This indicates the permissible frequency deviation range.

[0114] Construct the parameter update gate function:

[0115]

[0116] in, A preset trajectory stability threshold is used to determine whether the trajectory is allowed to participate in wavelet parameter updates. When a candidate fault frequency only occurs suddenly in a single window or at a single measurement point and cannot form a stable trajectory, its... Lower, corresponding It is relatively small and does not directly drive wavelet parameter updates.

[0117] Step 2.3: Smooth update of wavelet parameters constrained by CFF drift trajectory. This invention uses the CFF-TWPU module to smoothly update the wavelet center frequency, bandwidth, and scale search interval based on the CFF drift trajectory.

[0118] Step 2.3.1: For the first The stable CFF drift trajectory, the first The wavelet center frequency of each window is not directly assigned as the current candidate fault frequency, but is instead smoothly updated by combining the center frequency of the previous window, the current CFF frequency, and the trajectory stability.

[0119]

[0120] in, For the first The first window The wavelet center frequency corresponding to each trajectory The wavelet center frequency of the previous window, For the frequency corresponding to the CFF trajectory of the current window, Update the gating coefficients for the parameters. This is the smoothing coefficient. When the CFF drift trajectory is stable, The wavelet center frequency is relatively large, and it updates slowly following the fault frequency; when the CFF drift trajectory is unstable... The smaller the frequency, the better. The wavelet center frequency remains the same as the previous window parameter, avoiding frequency band shift caused by noise spurious peaks.

[0121] Step 2.3.2: The wavelet bandwidth is determined by the harmonic interval, sideband width and CFF drift uncertainty.

[0122]

[0123] in, For the first The wavelet bandwidth corresponding to each CFF trajectory. Basic broadband, The width of the sideband distribution. For harmonic intervals or harmonic extension range, For CFF drift uncertainty, This is the adjustment coefficient.

[0124] Drift uncertainty can be defined as the most recent Variance of CFF drift trajectory within a window:

[0125]

[0126] When the CFF trajectory is stable, The smaller the wavelet bandwidth, the less noise enters; when there is a slight drift in the CFF, Increase the bandwidth appropriately to cover the fault-sensitive frequency band after drift; when the CFF trajectory changes abruptly and the stability is low, adjust the gating coefficient. Limit its impact on bandwidth updates.

[0127] Step 2.3.3: Determine the wavelet scale search interval based on the center frequency and bandwidth. For a center frequency of... The mother wavelet, whose scale and center frequency satisfy the following relationship:

[0128]

[0129] in, The center frequency of the mother wavelet. For scale parameters, The center frequency of the wavelet.

[0130] Therefore, the first The wavelet scale search interval corresponding to each CFF trajectory can be represented as:

[0131]

[0132] This interval changes smoothly with the CFF drift trajectory, thereby enabling dynamic tracking of the fault-sensitive frequency band.

[0133] Step 3: Wavelet decomposition based on updated parameters. Based on the updated wavelet parameters obtained above, perform wavelet decomposition on the signals at each measurement point. Morlet wavelet, complex Morlet wavelet, generalized Morse wavelet, or other mother wavelets with time-frequency localization capabilities can be used.

[0134] Taking Morlet wavelets as an example:

[0135]

[0136] No. The measuring point, the first The first window, the The wavelet coefficients corresponding to the CFF trajectories are:

[0137]

[0138] in, , Indicates time translation, This represents the conjugate of the mother wavelet. Several wavelet frequency bands are formed based on wavelet scale or frequency division:

[0139]

[0140] The corresponding wavelet coefficient components are obtained: .in This indicates the wavelet frequency band or scale band.

[0141] Step 4: TCB-AGM triplet adaptive graph construction and gain learning.

[0142] Step 4.1: This invention uses the TCB-AGM module to establish an adaptive graph of the measurement point-CFF drift trajectory-wavelet frequency band triplet.

[0143] Each measurement point, CFF drift trajectory, and wavelet frequency band combination is used to construct a triplet node:

[0144]

[0145] in, For the first One measuring point, For the first A CFF drift trajectory, For the first A wavelet frequency band.

[0146] All triplet nodes are combined into a node set:

[0147]

[0148] The feature vector of each node is defined as:

[0149]

[0150] Each of the primitive directions represents: local peak significance, harmonic integrity, sideband matching degree, CFF trajectory stability, CFF drift uncertainty, cross-measurement point consistency, measurement point location prior, and corresponding wavelet band energy or impulsiveness index. The measurement point location prior is... It can be determined based on the mechanical connection relationship between the measuring point and candidate fault sources such as traction motor, gearbox, and axle box, the propagation path length, or the empirical response attenuation relationship.

[0151] Step 4.2: Construction of the triplet adaptive graph.

[0152]

[0153] in, For a set of nodes, Let be the set of edges. It is an adjacency matrix.

[0154] Graph edge relationships include the following categories:

[0155] (1) Response edge of the same CFF drift trajectory at different measurement points

[0156] When two nodes have the same CFF drift trajectory and the same wavelet frequency band, but come from different measurement points, establish a measurement point response edge: This edge is used to represent the consistency of the response of the same candidate fault frequency trajectory across different measurement points.

[0157] (2) Frequency band adjacency edge between adjacent frequency bands within the same measurement point

[0158] When two nodes originate from the same measurement point and the same CFF trajectory, but their wavelet frequency bands are adjacent, establish a frequency band adjacency edge: This edge is used to represent the continuity of the fault-sensitive frequency band in the scale or frequency direction.

[0159] (3) Harmonic / sideband structure between CFF trajectories

[0160] When two CFF trajectories satisfy a harmonic, octave, or sideband relationship, a CFF structure edge is established: ;

[0161] For example, when the following conditions are met: or: At that time, it was assumed that there was a harmonic or sideband relationship between the two CFF trajectories.

[0162] (4) Priority edge of mechanical connection of measuring point

[0163] Based on the bogie drivetrain structure, establish the mechanical connection between the traction motor, gearbox, and axle box: This edge is used to represent the propagation correlation of vibration response between different measurement points.

[0164] (5) CFF trajectory continuation edge

[0165] For triplet nodes belonging to the same CFF drift trajectory in adjacent windows, establish trajectory continuation edges: This edge is used to express the continuity of candidate fault frequencies across time windows.

[0166] Step 4.3: Adaptive Graph Adjacency Weight Update. Adjacency Matrix It consists of both prior adjacency and data adaptive adjacency:

[0167]

[0168] in: It is the a priori adjacency matrix composed of mechanical connection relationships, frequency band adjacency relationships, and harmonic / sideband structure relationships. This is the data adjacency matrix learned from node feature similarity and response consistency. is the fusion coefficient of prior and data-driven adjacency.

[0169] Data adaptive adjacency can be represented as:

[0170]

[0171] in: This is the index for the triple node. , For the first Node characteristics For learnable mapping functions, For nodes The neighbor set. This design enables the graph structure to reflect both the bogie's mechanical connections and frequency band structure priors, and to adaptively adjust the edge weights based on the vibration response of multiple measurement points in the current window.

[0172] Step 4.4: Triple Adaptive Graph Wavelet Gain Learning. The graph nodes of the triplet graph are updated to obtain the wavelet coefficient gain for each triplet node. This invention uses the TCB-AGM graph gain network.

[0173] No. Layer graph update can be represented as:

[0174]

[0175] in, , For the first Layer node representation, For learnable parameter matrices, It is a non-linear activation function.

[0176] The final output is the wavelet coefficient gain for each triplet node:

[0177]

[0178] in, , indicating the first The measuring point, the first CFF trajectory, number The reliability gain of each wavelet band. For low-reliability nodes, its The smaller the value, the corresponding frequency band is suppressed; for high-reliability nodes, its... The larger the value, the stronger the fault-sensitive wavelet component is; This represents the transpose of the learnable weight vector of the gain mapping layer. Indicates the first After the layer graph node is updated, the first The measuring point, the first CFF trajectory, number The implicit feature representation of each wavelet frequency band corresponding to the triplet node. This represents the bias term of the gain mapping layer.

[0179] Step 5: Wavelet coefficient modulation, multi-point reconstruction, and result output

[0180] Step 5.1: Modulate the wavelet coefficients corresponding to each measurement point, each CFF trajectory, and each frequency band based on the wavelet coefficient gain output by the triplet adaptive graph.

[0181] To suppress noise bands, a two-way modulation method is used:

[0182]

[0183] in, To enhance the coefficient, The inhibition coefficient; These are wavelet coefficient components. When... At higher levels, the corresponding frequency band is enhanced; when At lower levels, the corresponding frequency band is suppressed.

[0184] Step 5.2: Perform inverse wavelet transform on the modulated wavelet coefficients to obtain the... Enhanced signal at each measurement point:

[0185]

[0186] in, This represents the inverse wavelet transform.

[0187] Furthermore, to avoid invalid or noisy measurement points affecting the final reconstruction result due to equal-weighted fusion of enhanced signals from different measurement points, this invention calculates the measurement point fusion weight based on the aggregation result of the triplet gain corresponding to each measurement point, and fuses the enhanced signals of multiple measurement points to obtain a multi-measurement point fault-sensitive reconstruction signal:

[0188]

[0189] in, For the first The first window The fusion weight of each measurement point is used to characterize the contribution of the enhanced signal of that measurement point to the fault-sensitive reconstructed signal of multiple measurement points.

[0190] The measurement point fusion weights can be obtained by triplet gain aggregation:

[0191]

[0192] When the overall triplet gain of a certain measurement point is high across multiple CFF trajectories and multiple wavelet frequency bands, it indicates that the measurement point contains a relatively reliable fault-sensitive component within the current window, and therefore its fusion weight is large; conversely, when the overall triplet gain of a certain measurement point is low, it indicates that the measurement point may be subject to strong noise interference or have an insignificant fault response, and therefore its fusion weight is small.

[0193] After obtaining the multi-point fault-sensitive reconstructed signals for each window, the results of each window are then concatenated into an enhanced fault-sensitive signal over the entire time range by overlapping and adding the signals.

[0194]

[0195] in, This indicates that the fault-sensitive reconstructed signals from multiple measurement points in adjacent windows are overlapped and added together to obtain an enhanced fault-sensitive signal over the entire time range.

[0196] This embodiment ultimately outputs an enhanced fault-sensitive vibration signal. Fault-sensitive signals enhanced at each measuring point Fault-sensitive wavelet time-frequency representation Measurement point—CFF trajectory—frequency band triplet reliability diagram The outputs include fault-sensitive frequency band enhancement maps or frequency band reliability maps. These outputs can be used as inputs for subsequent train bogie fault diagnosis, condition identification, or fault source analysis.

Claims

1. A fault-sensitive frequency band adaptive enhancement method for multi-point vibration signals of train bogies, characterized in that, Includes the following steps: Step 1: Collect vibration signals from multiple measurement points of the train bogie transmission system, and perform windowing and preprocessing; extract candidate fault frequencies from each measurement point and each time window; Step 2: Correlate the candidate fault frequencies in the continuous window to form CFF drift trajectories, and perform wavelet parameter smoothing updates; Step 3: Based on the updated wavelet parameters, perform wavelet decomposition on the vibration signals at each measuring point; obtain the wavelet coefficient components corresponding to different measuring points, different candidate fault frequency trajectories, and different wavelet frequency bands. Step 4: Construct triplet nodes for measurement points, candidate fault frequency trajectories, and wavelet bands, and establish a TCB-AGM triplet adaptive graph based on the triplet nodes. Learn the wavelet gain of the triplet adaptive graph to obtain the wavelet coefficient gain corresponding to each triplet node, which directly applies to the corresponding measurement point, the corresponding candidate fault frequency trajectory, and the corresponding wavelet band. Step 5: Perform inverse wavelet transform and multi-measurement point fusion reconstruction on the modulated wavelet coefficients to obtain the enhanced fault-sensitive vibration signal, the enhanced signal of each measurement point, the fault-sensitive wavelet time-frequency representation, and the reliability map of the measurement point-candidate fault frequency trajectory-wavelet frequency band triplet or the fault-sensitive frequency band enhancement map.

2. The fault-sensitive frequency band adaptive enhancement method for multi-point vibration signals of train bogies according to claim 1, characterized in that, The processing steps in step 1 include: Step 1.1: Collect data from the train bogie transmission system Vibration signal at each measuring point The signals from each measuring point are windowed to obtain the first... Multiple measurement points within a time window Preprocessing of the signals within each window is denoted as follows: ; Step 1.2: Extract the vibration signal from each measuring point and each window to extract the candidate fault frequency set, denoted as: ;in, Indicates the first The measuring point at the ... The first one extracted from the window Candidate fault frequencies, The total number of candidate fault frequencies; Step 1.3: Construct the candidate frequency evaluation function ,like At that time, the corresponding frequency This serves as the candidate fault frequency within the window.

3. The fault-sensitive frequency band adaptive enhancement method for multi-point vibration signals of train bogies according to claim 1, characterized in that, Step 2 includes the following processing steps: Step 2.1: For the candidate fault frequencies extracted within the continuous window, perform trajectory association to form the candidate fault frequency drift trajectory; Step 2.2: Calculate the stability for each CFF drift trajectory Based on stability Construct the parameter update gate function: , The preset trajectory stability threshold is used to determine whether the trajectory is allowed to participate in wavelet parameter updates; Step 2.3: Smoothly update the wavelet center frequency, bandwidth, and scale search interval based on the CFF drift trajectory.

4. The fault-sensitive frequency band adaptive enhancement method for multi-point vibration signals of train bogies according to claim 3, the update process in step 2.3 is as follows: Step 2.3.1: For the first A stable CFF drift trajectory is generated and smoothly updated by combining the center frequency of the previous window, the current CFF frequency, and the trajectory stability. ； in, For the first The first window The wavelet center frequency corresponding to each trajectory The wavelet center frequency of the previous window, For the frequency corresponding to the CFF trajectory of the current window, Update the gating coefficients for the parameters. For smoothing coefficients; Step 2.3.2: The wavelet bandwidth is determined by the harmonic interval, sideband width, and CFF drift uncertainty, denoted as: ； in, For the first The wavelet bandwidth corresponding to each CFF trajectory. Basic broadband, The width of the sideband distribution. For harmonic intervals or harmonic extension range, For CFF drift uncertainty, This is the adjustment coefficient; Step 2.3.3: Determine the wavelet scale search interval based on the center frequency and bandwidth, where the center frequency is... Mother wavelet, No. The wavelet scale search interval corresponding to each CFF trajectory is represented as follows: ； in, The center frequency of the mother wavelet.

5. The fault-sensitive frequency band adaptive enhancement method for multi-point vibration signals of train bogies according to claim 1, characterized in that, Step 3 uses a mother wavelet with time-frequency localization capability for wavelet decomposition, including but not limited to Morlet wavelet, complex Morlet wavelet, and generalized Morse wavelet.

6. The fault-sensitive frequency band adaptive enhancement method for multi-point vibration signals of train bogies according to claim 1, characterized in that, Step 4 includes the following processing steps: Step 4.1: Construct a triplet node by combining each measurement point, CFF drift trajectory, and wavelet frequency band: ; All triplet nodes are combined into a node set: ; The feature vector of each node is defined as: ; in, For the first One measuring point, For the first A CFF drift trajectory, For the first Each wavelet frequency band; the eigenvectors contain, in order, local peak significance, harmonic integrity, sideband matching degree, CFF trajectory stability, CFF drift uncertainty, cross-measurement point consistency, measurement point location prior, and corresponding wavelet frequency band energy or impact index. Step 4.2: Construct the triplet adaptive graph, denoted as: ;in, For a set of nodes, Let be the set of edges. It is an adjacency matrix; The graph-side relationships include the measurement point response relationship, frequency band adjacency relationship, harmonic / sideband structure relationship between candidate fault frequencies, prior relationship of mechanical connection of measurement points, and trajectory continuation relationship of candidate fault frequencies. Step 4.3: Adjacency Matrix It consists of both prior adjacency and data adaptive adjacency: ； in: It is the a priori adjacency matrix composed of mechanical connection relationships, frequency band adjacency relationships, and harmonic / sideband structure relationships. This is the data adjacency matrix learned from node feature similarity and response consistency. The fusion coefficient of prior and data-driven adjacency; The adaptive adjacency representation of data is as follows: ； in: This is the index for the triple node. , For the first Node characteristics For learnable mapping functions, For nodes The set of neighbors; Step 4.4: Update the graph nodes of the triplet graph to obtain the wavelet coefficient gain of each triplet node, denoted as: ;in, , indicating the first The measuring point, the first CFF trajectory, number The reliability gain of each wavelet band. For low-reliability nodes, its The smaller the value, the more the corresponding frequency band is suppressed; for high-reliability nodes, its... The larger the value, the stronger the fault-sensitive wavelet component is; This represents the transpose of the learnable weight vector of the gain mapping layer. Indicates the first After the layer graph node is updated, the first The measuring point, the first CFF trajectory, number The implicit feature representation of each wavelet frequency band corresponding to a triplet node. This represents the bias term of the gain mapping layer.

7. The fault-sensitive frequency band adaptive enhancement method for multi-point vibration signals of train bogies according to claim 1, characterized in that, Step 5 includes the following processing steps: Step 5.1: Modulate the wavelet coefficients corresponding to each measurement point, each CFF trajectory, and each frequency band based on the wavelet coefficient gain output by the triplet adaptive graph; Step 5.2: Perform inverse wavelet transform on the modulated wavelet coefficients to obtain the... The enhanced signals from multiple measurement points are fused based on the reliability of each measurement point to obtain a multi-measurement point fault-sensitive reconstructed signal. The measurement point fusion weights are obtained by aggregated triplet gains. , The enhanced fault-sensitive signal over the entire time range is finally obtained, denoted as: ； in, This indicates a reconstructed structure based on overlapping and addition. The final output is an enhanced fault-sensitive vibration signal. Fault-sensitive signals enhanced at each measuring point Fault-sensitive wavelet time-frequency representation Measurement point—CFF trajectory—frequency band triplet reliability diagram And fault-sensitive frequency band enhancement maps or frequency band reliability maps.

8. The fault-sensitive frequency band adaptive enhancement method for multi-point vibration signals of train bogies according to claim 7, characterized in that, Step 5.1 uses bidirectional modulation: ； in, To enhance the coefficient, The inhibition coefficient; These are wavelet coefficient components. When At higher levels, the corresponding frequency band is enhanced; when At lower levels, the corresponding frequency band is suppressed.

Images