A high-speed railway subgrade retaining structure damage identification system based on wavelet transform
Patent Information
- Application Number
- CN202610113048.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-27
- Publication Date
- 2026-10-13
AI Technical Summary
然而，这些特征矩阵的获取和数据库的建立非常复杂，尤其是在高维数据的情况下，如何合理管理和存储这些特征数据以及如何有效检索匹配的标准数据，都是当前技术的难点
1.本发明通过自适应小波变换对原始振动信号进行处理，能够根据信号的频率成分和突变特征自动选择小波基函数和变换参数。这种自适应选择机制使得系统能够精准地捕捉不同类型的结构损伤特征，尤其对于高频噪声和局部突变信号的有效提取，提升了损伤识别的准确性。通过加速度传感器获取的精细数据和高效的特征提取方法，有效减少了传统方法中的误判和漏判问题；
Smart Images
Figure CN122901074A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of nondestructive testing technology, specifically to a damage identification system for high-speed railway subgrade retaining structures based on wavelet transform. Background Technology
[0002] The subgrade retaining structure of high-speed railways is a crucial component supporting the railway tracks, and its stability and health directly impact the safe operation of the railway. Due to factors such as changes in the external environment, train vibration, and material fatigue, subgrade retaining structures may experience various types of damage over long-term use, such as cracks, holes, and surface spalling. Therefore, effectively monitoring and identifying these damages to ensure the safe operation of railways has become an important research topic in the field of engineering technology.
[0003] Currently, traditional damage detection methods mainly rely on manual inspection or periodic checks with specialized equipment. These methods are not only inefficient and time-consuming, but may also miss some early, subtle damage signals. To improve the accuracy and real-time performance of damage identification, more and more research in recent years has focused on methods based on vibration signal analysis, especially using wavelet transform for frequency domain feature extraction. Due to its excellent time-frequency localization characteristics, wavelet transform has significant advantages in processing complex vibration signals, especially high-frequency noise and abrupt changes.
[0004] However, existing wavelet transform methods are typically fixed, with the selected mother wavelet and the number of transform layers pre-defined. This fixed approach cannot adapt to the complexity of different types of vibration signals, as different vibration signals may have different frequency components and abrupt change characteristics. Therefore, fixed wavelet transform parameters cannot fully extract the multi-level features in the signal, resulting in insufficient accuracy and robustness of the identification results.
[0005] Existing databases typically require storing large amounts of standard frequency domain feature data matrices and classifying them according to different wavelet basis functions, transform scales, and decomposition levels. However, obtaining these feature matrices and establishing the database are very complex, especially with high-dimensional data. How to rationally manage and store this feature data and how to effectively retrieve matching standard data are current technological challenges. Furthermore, the poor scalability and maintainability of these databases increase the difficulties in practical applications.
[0006] To address the aforementioned issues, it is necessary to propose a damage identification system for high-speed railway subgrade retaining structures based on wavelet transform. Summary of the Invention
[0007] The purpose of this invention is to solve the problems existing in the background technology and to propose a high-speed railway subgrade retaining structure damage identification system based on wavelet transform.
[0008] The objective of this invention can be achieved through the following technical solutions: A high-speed railway subgrade retaining structure damage identification system based on wavelet transform includes an acceleration sensor, a preprocessing module, a central computing module, a database module, and a comparison and analysis module.
[0009] The accelerometer collects the original vibration signals of the roadbed retaining structure.
[0010] The accelerometer and the preprocessing module are connected via a wireless signal network; The preprocessing module filters and reduces noise from the acquired raw vibration signal, and selects adaptive wavelet transform parameters based on the main frequency components and abrupt change characteristics of the raw vibration signal.
[0011] The central processing module retrieves adaptive wavelet transform parameters to perform wavelet transform on the original vibration signal, extracting frequency domain features at both the local and global levels to obtain local and global frequency domain feature data matrices. A comprehensive frequency domain feature data matrix is then obtained by fusing the local and global frequency domain feature data matrices.
[0012] The database module stores several standard frequency domain feature data matrices.
[0013] The comparison and analysis module obtains the roadbed retaining structure damage identification label by calculating the matching similarity between the comprehensive frequency domain feature data matrix and the standard frequency domain feature data matrix.
[0014] As a preferred embodiment of the present invention, the acquired original vibration signal is specifically a time series dataset, wherein each data point has a timestamp and an accelerometer location number, and each data point represents the vibration recorded at different time points by the roadbed retaining structure at different locations.
[0015] In a preferred embodiment of the present invention, the preprocessing module filters and reduces noise from the acquired raw vibration signals, establishing a state-space model for the signals: It is assumed that for each accelerometer, all raw vibration signals consist of the current value and the rate of change; it is also assumed that the raw vibration signal at the current time t is determined by the raw vibration signal at the previous time t-1, the rate of change of the raw vibration signal at the previous time t, and the train vibration control input at the previous time t. Kalman filtering is performed based on the state-space model, and the control input in the Kalman filter is calculated based on the train speed and the distance between the train and the accelerometer.
[0016] In a preferred embodiment of the present invention, adaptive wavelet transform parameters are selected based on the main frequency components and abrupt change characteristics of the original vibration signal. The original vibration signal is then transformed from the time domain to the frequency domain using a Fast Fourier Transform (FFT) to obtain the frequency domain signal of the original vibration signal. Spectral analysis is performed on the frequency domain signal of the original vibration signal to obtain low-frequency and high-frequency components, and the energy of the low-frequency and high-frequency components is calculated respectively. The quotient of the energy of the high-frequency component and the energy of the low-frequency component is calculated to obtain the proportion of high-frequency component energy. The maximum and minimum frequencies contained in the frequency domain signal of the original vibration signal are extracted.
[0017] As a preferred embodiment of the present invention, the specific process for analyzing mutation characteristics is as follows: the rate of change of the original vibration signal is calculated, abrupt changes in the signal are detected, and the number of mutation points contained in the original vibration signal is obtained.
[0018] As a preferred embodiment of the present invention, the adaptive wavelet transform parameters include the mother wavelet of the wavelet basis function, the transform scale parameter, and the decomposition series.
[0019] Adaptive mother wavelet selection is based on the energy proportion of high-frequency components and the number of abrupt change points.
[0020] The adaptive transformation scaling parameter and adaptive decomposition level are selected based on the signal sampling frequency of the original vibration signal and the maximum and minimum frequencies contained in the frequency domain signal of the original vibration signal.
[0021] Calculate the reciprocal of the maximum frequency to obtain the minimum scaling parameter in the transform scaling parameters, and calculate the reciprocal of the minimum frequency to obtain the maximum scaling parameter in the transform scaling parameters.
[0022] The signal sampling frequency of the original vibration signal is obtained and divided by the step size scaling factor to obtain the translation parameter in the transformation scale parameter.
[0023] In a preferred embodiment of the present invention, the preset step size ratio is adjusted according to the specific value of the original vibration signal sampling frequency. If the original vibration signal sampling frequency is greater than 50Hz, i.e., a high sampling frequency, the step size ratio is set to a preset value of 1, representing a shift of 1 sampling point each time. If the original vibration signal sampling frequency is less than or equal to 50Hz but greater than 25Hz, i.e., a medium sampling frequency, the step size ratio is set to a preset value of 2, representing a shift of half a sampling point each time. If the original vibration signal sampling frequency is less than or equal to 25Hz, i.e., a low sampling frequency, the step size ratio is set to a preset value of 5 to improve the analysis accuracy.
[0024] In a preferred embodiment of the present invention, the decomposition level is calculated using the signal sampling frequency of the original vibration signal and the maximum and minimum frequencies contained in the frequency domain signal of the original vibration signal. A maximum value for the adaptive decomposition level is then defined.
[0025] As a preferred embodiment of the present invention, wavelet transform is performed on the original vibration signal after filtering and noise reduction, and frequency domain features are extracted at both the local and global levels.
[0026] After filtering and denoising the raw vibration signals collected by each accelerometer, a global frequency domain feature data matrix is obtained by performing wavelet transform on the signals using an adaptive mother wavelet, an adaptive transform scale parameter, and an adaptive decomposition level.
[0027] After filtering and denoising the raw vibration signals acquired by each accelerometer, the signals are divided into several raw vibration signal segments through a preset time window. Wavelet transforms are then applied to each raw vibration signal segment using adaptive mother wavelet, adaptive transform scaling parameters, and adaptive decomposition levels, resulting in several local frequency domain feature data matrices. In both the global and local frequency domain feature data matrices, each row represents a decomposition level index, each column represents a translation position for that decomposition level, and each element represents a wavelet coefficient corresponding to the decomposition level in its row and the translation position in its column.
[0028] As a preferred embodiment of the present invention, to avoid feature loss at signal boundaries, overlapping time windows are used, meaning that each time window overlaps with the previous time window to a certain extent, with the overlap ranging from 10% to 50% within the time window. This ensures the continuity of local features and captures abrupt changes in the signal.
[0029] In a preferred embodiment of the present invention, the local frequency domain feature data matrix and the global frequency domain feature data matrix are fused to obtain a comprehensive frequency domain feature data matrix. The fusion method is a weighted summation. A weight coefficient is assigned to the global frequency domain feature data matrix and each local frequency domain feature data matrix.
[0030] In a preferred embodiment of the present invention, local frequency domain feature data matrices and global frequency domain feature data matrices are fused to form a comprehensive frequency domain feature data matrix. The fusion method involves concatenating matrices arranged in rows. The global frequency domain feature data matrix is used as the first matrix in the concatenated result, and the local frequency domain feature data matrices are used as the second to last matrices in the concatenated result. The order of the local frequency domain feature data matrices in the concatenated result follows the order of the start times of the time window.
[0031] In a preferred embodiment of the present invention, the database module includes first, second, third, and fourth sub-databases, each storing data for a specific mother wavelet. The next level of each sub-database performs secondary classification based on adaptive transform scaling parameters, resulting in second-level sub-classifications based on mother wavelet type and adaptive transform scaling parameters. The next level of each second-level sub-classification performs tertiary classification based on adaptive decomposition levels, resulting in third-level sub-classifications based on mother wavelet type, adaptive transform scaling parameters, and adaptive decomposition levels.
[0032] Each third-level subclass contains the same number of standard frequency domain feature data matrices. Each standard frequency domain feature data matrix has a roadbed retaining structure damage identification label.
[0033] Each mother wavelet, mother wavelet type, adaptive transform scaling parameter, and adaptive decomposition level has a uniquely corresponding standard frequency domain feature data matrix in the database module.
[0034] As a preferred embodiment of the present invention, after receiving the comprehensive frequency domain feature data matrix corresponding to a certain accelerometer, the comparison and analysis module retrieves the standard frequency domain feature data matrix corresponding to its adaptive wavelet transform parameters from the database module.
[0035] Based on the mother wavelet type, locate the first, second, third, or fourth sub-database in the database module; locate the second-level sub-classes in the first, second, third, or fourth sub-databases based on the adaptive transform scaling parameter; locate the third-level sub-classes under a certain second-level sub-class in the first, second, third, or fourth sub-database based on the adaptive decomposition level. Obtain all standard frequency domain feature data matrices under the located third-level sub-classes, and perform similarity calculations by matching the comprehensive frequency domain feature data matrix with all the standard frequency domain feature data matrices.
[0036] In a preferred embodiment of the present invention, the calculation process for matching similarity is as follows: The cosine similarity between the comprehensive frequency domain feature data matrix and each standard frequency domain feature data matrix is calculated to obtain the matching similarity between the comprehensive frequency domain feature data matrix and each standard frequency domain feature data matrix. The standard frequency domain feature data matrix with the highest matching similarity to the comprehensive frequency domain feature data matrix is extracted and marked as the optimally matched standard frequency domain feature data matrix. If the matching similarity between the optimally matched standard frequency domain feature data matrix and the comprehensive frequency domain feature data matrix is greater than a preset threshold, the match is considered successful. The roadbed retaining structure damage identification label corresponding to the optimally matched standard frequency domain feature data matrix is output.
[0037] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention processes the original vibration signal using adaptive wavelet transform, automatically selecting wavelet basis functions and transform parameters based on the signal's frequency components and abrupt change characteristics. This adaptive selection mechanism enables the system to accurately capture different types of structural damage features, especially for the effective extraction of high-frequency noise and local abrupt change signals, thus improving the accuracy of damage identification. The fine data acquired through the accelerometer and the efficient feature extraction method effectively reduce the misjudgment and missed detection problems found in traditional methods. 2. By fusing data through weighted summation or row-by-row concatenation of local and global frequency domain feature data matrices, this invention can comprehensively assess the damage status of roadbed retaining structures by integrating information from different levels. The integrated frequency domain feature data matrix, by calculating its similarity to the standard frequency domain feature data matrix, can efficiently identify the specific type of damage and provide accurate damage identification labels. This process not only improves the system's matching accuracy but also significantly enhances the real-time assessment capability of structural health, helping to promptly identify potential problems and improve the efficiency of maintenance decisions. Attached Figure Description
[0038] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings: Figure 1 This is a system block diagram of a high-speed railway subgrade retaining structure damage identification system based on wavelet transform proposed in the embodiments of the present invention.
[0039] Figure 2 This is a schematic diagram of the wireless signal network proposed in the embodiments of the present invention; Figure 3 This is a schematic diagram of the time-domain signal of the original vibration signal proposed in the embodiments of the present invention; Figure 4 This is a schematic diagram of the frequency domain signal of the original vibration signal proposed in the embodiments of the present invention; Figure 5 This is a three-dimensional wavelet transform grid diagram corresponding to the global hierarchical frequency domain features proposed in the embodiments of the present invention; Figure 6 This is a schematic diagram of one of the database modules proposed in the embodiments of the present invention; Figure 7 This is a second schematic diagram of the database module proposed in the embodiments of the present invention. Detailed Implementation
[0040] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0041] First embodiment: Please see Figure 1 As shown, a high-speed railway subgrade retaining structure damage identification system based on wavelet transform includes an acceleration sensor, a preprocessing module, a central computing module, a database module, and a comparison and analysis module.
[0042] The accelerometer collects the original vibration signals of the roadbed retaining structure.
[0043] The accelerometer and the preprocessing module are connected via a wireless signal network; The preprocessing module filters and reduces noise from the acquired raw vibration signal, and selects adaptive wavelet transform parameters based on the main frequency components and abrupt change characteristics of the raw vibration signal.
[0044] The central processing module retrieves adaptive wavelet transform parameters to perform wavelet transform on the original vibration signal, extracting frequency domain features at both the local and global levels to obtain local and global frequency domain feature data matrices. A comprehensive frequency domain feature data matrix is then obtained by fusing the local and global frequency domain feature data matrices.
[0045] The database module stores several standard frequency domain feature data matrices.
[0046] The comparison and analysis module obtains the roadbed retaining structure damage identification label by calculating the matching similarity between the comprehensive frequency domain feature data matrix and the standard frequency domain feature data matrix.
[0047] Second embodiment: In this embodiment, the acquired raw vibration signal is specifically a time series dataset, where each data point has a timestamp and an accelerometer location number, and each data point represents the vibration recorded at different times by the roadbed retaining structure at different locations.
[0048] Third embodiment: In one implementation, please refer to Figure 2 As shown, the wireless signal network adopts a distributed architecture, with each accelerometer being an independent node in the wireless signal network. Each accelerometer collects its raw vibration signals and sends them to the regional base station, which then sends all the raw vibration signals to the preprocessing module.
[0049] In one embodiment, the aforementioned accelerometer and preprocessing module are connected via a wireless signal network, including LoRa, Zigbee, and NB-IoT.
[0050] It should be noted that LoRa is a low-power wide-area network (LPWAN) technology suitable for long-distance data transmission and small-volume data collection. Its applicable conditions are: monitoring distance greater than 2 kilometers and data traffic greater than 5 bytes / packet; Zigbee is a short-range, low-power wireless communication technology suitable for short-range communication between devices, suitable for low-speed data transmission and environments with dense devices. Its applicable conditions are: monitoring distance less than 100 meters and data traffic less than 5 bytes / packet; NB-IoT is a narrowband communication technology for the Internet of Things (IoT), suitable for applications requiring long-distance transmission, low power consumption, and high coverage. Its applicable conditions are: monitoring distance between 100 meters and 2 kilometers.
[0051] Fourth embodiment: In this embodiment, the specific process of the preprocessing module filtering and denoising the acquired raw vibration signal is as follows: Record all raw vibration signals and define the state-space model of the signals: Assume that for each acceleration sensor i, all raw vibration signals consist of the current value and the rate of change; assume that the raw vibration signal at the current time t consists of the raw vibration signal at the previous time t-1. The rate of change of the original vibration signal at the previous time t Train vibration control input at the previous time t Decide.
[0052] The formula for the state-space model is: in, The predicted system state at the current time t includes the original vibration signal. and the rate of change of the original vibration signal ;in, Let be the time interval from the previous time t-1 to the current time t, where Let be the state transition matrix, describing the changes in the system states. Let be the control matrix, describing the influence of train vibration control input on the system state; where Let Q be the process noise, with a mean of 0 and a covariance of Q.
[0053] It should be noted that the state-space model can predict the system state at the next moment based on the system state at the previous moment, thereby filtering the abnormal data at the next moment, estimating the true system state, and suppressing abnormal fluctuations caused by noise and external interference. By combining state behavior and observation data, the original vibration data can be predicted and updated to obtain more accurate signal estimates.
[0054] In one implementation, Kalman filtering is performed based on a state-space model. The specific process is as follows: First, the predicted system state at the current moment is calculated using the predicted system state from the previous moment. This involves combining the original vibration signal, the rate of change of the original vibration signal, and the train vibration control input state-space model from the previous moment's predicted system state to obtain the predicted system state at the current moment. Then, the Kalman gain is calculated based on the difference between the predicted values of the original vibration signal and its rate of change in the current moment's predicted system state and the actually acquired original vibration signal and its rate of change, balancing the weights of the predicted and observed values. Finally, the original vibration signal in the predicted system state and the actually acquired original vibration signal are weighted and summed based on the Kalman gain, and the filtered result is output. Simultaneously, the covariance Q of the process noise is updated to measure the accuracy of the current estimate and prepare for the next prediction.
[0055] Whenever new raw vibration data is input, the state estimation is continuously optimized through prediction and update steps. The Kalman gain is calculated, and the prediction result is corrected by weighted summation. The process noise covariance of the previous time step and the predicted system state are used as parameters to calculate the predicted system state of the next time step, and Kalman recursive calculation is performed. This process continues, continuously outputting the filtered results.
[0056] It should be noted that the Kalman gain reflects the system's level of confidence in the predicted values and observed data. If the observed data is noisy, the Kalman gain is small, and the system will rely more on the predicted values; conversely, if the observed data is accurate, the Kalman gain is large, and the system will rely more on the observed data.
[0057] In one implementation, train vibration is determined by an empirical formula based on train speed and distance from the acceleration sensor, and its influence on the state at the current time t is corrected by a control matrix. The empirical formula is as follows: Where k is a preset empirical coefficient used to adjust the sensitivity of the model based on experimental data or historical records, representing the basic relationship between train speed and distance on vibration, and its preset value is 1; in, The train speed at acceleration sensor i is measured in m / s; when no train is passing, its value is 0. in, The distance between the accelerometer and the train, measured in meters (m).
[0058] in, This is an empirical index, preset to 1.5, with units of 1 / s, reflecting the degree to which distance attenuates the impact of vibration. Smaller values indicate that the impact of distance attenuates more slowly, while larger values indicate that the impact of vibration attenuates more quickly with greater distance.
[0059] It should be noted that in traditional Kalman filtering, the control input is usually a known quantity, meaning that the external influences of the system (such as external forces, control signals, etc.) are known and are typically obtained through sensors or other measurement methods. However, in the process of acquiring train vibration loads in a high-speed railway environment, the train vibration control input is an unknown variable and is difficult to measure directly. To overcome this problem, this embodiment simulates the impact of train vibration on the system using empirical formulas. Specifically, by using parameters such as the train's speed and the distance between it and the acceleration sensor, combined with experimental data or existing physical models, the influence of train vibration can be inferred, and the train vibration control input can be calculated.
[0060] It should be further explained that the impact of train vibration is estimated using the aforementioned empirical formula, and its influence on the system state is corrected using a control matrix. This process ensures that even if the impact of train vibration on the vibration signal is unknown, the system can dynamically adjust the state estimation based on the relationship between train speed and sensor distance, thereby improving the accuracy of the vibration signal.
[0061] Fourth embodiment: In this embodiment, adaptive wavelet transform parameters are selected based on the main frequency components and abrupt change characteristics of the original vibration signal. The specific process is as follows: In one implementation, please refer to Figure 3 and Figure 4 As shown, the specific process for analyzing the main frequency components is as follows: The original vibration signal is transformed from the time domain to the frequency domain using a Fast Fourier Transform (FFT) to obtain the frequency domain signal of the original vibration signal. Spectral analysis is performed on the frequency domain signal of the original vibration signal, using a preset critical frequency of 25Hz as the high-low frequency boundary to distinguish between low-frequency and high-frequency components, and the energy of the low-frequency and high-frequency components is calculated separately. The quotient of the high-frequency component energy and the low-frequency component energy is calculated to obtain the proportion of high-frequency component energy.
[0062] Extract the maximum and minimum frequencies contained in the frequency domain signal of the original vibration signal.
[0063] It should be noted that a higher proportion of high-frequency components indicates that the original vibration signal has strong energy in the high-frequency band, and there may be abrupt changes or high-frequency noise in the original vibration signal; a lower proportion of high-frequency components indicates that the original vibration signal has strong energy in the low-frequency band, indicating that the signal may contain changes with a longer period (such as the vibration of the entire structure).
[0064] In one implementation, the specific process of analyzing mutation characteristics is as follows: calculate the rate of change of the original vibration signal, detect abrupt changes in the signal, and obtain the number of mutation points contained in the original vibration signal.
[0065] In the original vibration signal, the number of abrupt change points represents the instantaneous impact of damage to the high-speed railway subgrade structure or external disturbances.
[0066] In one implementation, the adaptive wavelet transform parameters include the mother wavelet of the wavelet basis functions, the transform scaling parameter, and the decomposition series.
[0067] The mother wavelet is the fundamental waveform function used to generate wavelet basis functions for wavelet transform through scaling and translation. It generates a complete set of basis functions through transform scales including scaling and translation parameters. Common mother wavelets include Haar wavelets, Morlet wavelets, Daubechies wavelets, Mexican Hat wavelets, and Symlet wavelets.
[0068] The transform scaling parameters, including scaling and translation parameters, jointly determine the frequency resolution and temporal localization capability of the wavelet transform. The scaling parameter controls the scale of the mother wavelet, i.e., the extent to which the wavelet expands in time or frequency, adjusting the wavelet width to analyze signals within different frequency ranges. The translation parameter determines the position of the mother wavelet in the time or frequency domain. Through translation, the mother wavelet slides along the time axis, enabling it to analyze different parts of the signal.
[0069] The influence of the scaling and shifting parameters of the mother wavelet on the wavelet transform can be expressed by the following formula: in, is the mother wavelet function, a is the scaling parameter, and b is the displacement parameter.
[0070] It's important to note that when the mother wavelet is stretched (i.e., its scale increases), its waveform becomes broader, primarily analyzing low-frequency signals. This means it's more sensitive to the overall trend of the signal and can capture slow changes. When the mother wavelet's scale decreases (i.e., it's a compressed wavelet), it becomes narrower, primarily analyzing high-frequency signals. Small-scale wavelets focus more on detailed changes and local abrupt changes in the signal, capturing rapid changes.
[0071] The decomposition level refers to the depth of decomposition during wavelet transform, and is closely related to the number of wavelet transform layers. The decomposition level affects the granularity of low-frequency and high-frequency information extracted at each layer. At each layer of the wavelet transform, the signal is decomposed into two parts: a low-frequency part (approximate part) and a high-frequency part (detail part). Each decomposition extracts the energy of the signal from the original signal and propagates it to a deeper level.
[0072] It's important to note that the higher the decomposition level of the wavelet transform, the finer the signal information and the more details are revealed. Therefore, increasing the decomposition level allows for better capture of subtle changes and noise in the signal. However, as the decomposition level increases, the computational complexity required for the wavelet transform also increases. Each additional decomposition level generates more detailed coefficients, which need to be calculated and stored. For long-term sequences or large-scale data, this increased computational complexity can lead to significantly longer processing times and even excessive consumption of computing resources.
[0073] In one implementation, an adaptive mother wavelet is selected based on the energy proportion of high-frequency components and the number of abrupt change points. The specific process is as follows: Using the adaptive mother wavelet optimization table: Determine the mother wavelet of the wavelet transform.
[0074] In this embodiment, the adaptive transformation scale parameter and the adaptive decomposition level are selected based on the signal sampling frequency of the original vibration signal and the maximum and minimum frequencies contained in the frequency domain signal of the original vibration signal.
[0075] Calculate the reciprocal of the maximum frequency to obtain the minimum scaling parameter aMin in the transform scaling parameters, and calculate the reciprocal of the minimum frequency to obtain the maximum scaling parameter aMax in the transform scaling parameters.
[0076] The original vibration signal sampling frequency fs is obtained and divided by the step size scaling factor to obtain the translation parameter b in the transformation scale parameters.
[0077] In one implementation, the preset step size ratio is adjusted based on the specific value of the original vibration signal's sampling frequency fs. If the original vibration signal's sampling frequency is greater than 50Hz (high sampling frequency), the step size ratio is set to a preset value of 1, meaning that selection 1 indicates shifting by 1 sampling point each time. If the original vibration signal's sampling frequency is less than or equal to 50Hz but greater than 25Hz (medium sampling frequency), the step size ratio is set to a preset value of 2, meaning that shifting by half a sampling point each time. If the original vibration signal's sampling frequency is less than or equal to 25Hz (low sampling frequency), the step size ratio is set to a preset value of 5 to improve analysis accuracy.
[0078] For example, if the maximum frequency of the original vibration signal in the frequency domain is 25Hz and the minimum frequency is 0.5Hz, then the scale parameters a and b are selected as follows: minimum scaling parameter aMin = 0.4; maximum scaling parameter aMax = 2; translation parameter b = 20.
[0079] In one implementation, the decomposition level is calculated using the signal sampling frequency of the original vibration signal and the maximum and minimum frequencies contained in the frequency domain signal of the original vibration signal. This is achieved using a preset formula: Calculate the adaptive decomposition level m. Here, `int` is the integer operator, representing rounding down. The integer part of the calculation result.
[0080] Furthermore, the maximum value of the adaptive decomposition level is limited to 8.
[0081] It should be noted that having too many decomposition levels will significantly increase computational complexity and may introduce noise. Generally, the number of decomposition levels should not exceed eight, because at higher levels, the high-frequency details of the signal are gradually amplified, leading to noise interference.
[0082] Fifth embodiment: Please see Figure 5 As shown, wavelet transform is performed on the original vibration signal after filtering and noise reduction, and frequency domain features are extracted at both the local and global levels.
[0083] After filtering and denoising the raw vibration signals collected by each accelerometer, a global frequency domain feature data matrix is obtained by performing wavelet transform on the signals using an adaptive mother wavelet, an adaptive transform scale parameter, and an adaptive decomposition level.
[0084] After filtering and denoising the raw vibration signals acquired by each accelerometer, the signals are divided into several raw vibration signal segments through a preset time window. Wavelet transforms are then applied to each raw vibration signal segment using adaptive mother wavelet, adaptive transform scaling parameters, and adaptive decomposition levels, resulting in several local frequency domain feature data matrices. In both the global and local frequency domain feature data matrices, each row represents a decomposition level index, each column represents a translation position for that decomposition level, and each element represents a wavelet coefficient corresponding to the decomposition level in its row and the translation position in its column.
[0085] In one implementation, to avoid feature loss at signal boundaries, overlapping time windows are used, where each time window overlaps with the previous one to a certain extent, with the overlap ranging from 10% to 50% within the time window. This ensures the continuity of local features and captures abrupt changes in the signal.
[0086] In one implementation, the local frequency domain feature data matrix and the global frequency domain feature data matrix are fused to obtain a comprehensive frequency domain feature data matrix. The fusion method is weighted summation. A weight coefficient is assigned to the global frequency domain feature data matrix and each local frequency domain feature data matrix. The preset value of the weight coefficient for the local frequency domain feature data matrix is 0.6, and the preset value of the weight coefficient for the global frequency domain feature data matrix is 0.4.
[0087] In one implementation, local frequency domain feature data matrices and global frequency domain feature data matrices are fused to form a comprehensive frequency domain feature data matrix. The fusion method involves concatenating rows of matrices. The global frequency domain feature data matrix is used as the first matrix in the concatenated result, and the local frequency domain feature data matrices are used as the second to last matrices. The order of the local frequency domain feature data matrices in the concatenated result follows the order of the start times of the time window.
[0088] To avoid misunderstandings and ambiguities, it should be clarified that "global" and "local" here refer to a division on a time scale, not a spatial scale. That is, the frequency domain feature extraction at the "global scale" and "local scale" here targets the original vibration signals collected by the same accelerometer across the entire time range and within a local time window. Therefore, the frequency domain feature extraction at the local and global levels here does not refer to performing wavelet transforms separately on the original vibration signals collected by all accelerometers across the entire region and on the original vibration signals collected by some accelerometers in a specific local region.
[0089] Sixth embodiment: Please see Figure 6 and Figure 7 As shown, the database module includes four sub-databases, which store wavelet coefficients for the Haar, Symlet, Daubechies, and Coiflet mother wavelets, respectively. Each sub-database is further subdivided at the next level based on adaptive transform scaling parameters, resulting in second-level sub-classifications based on mother wavelet type and adaptive transform scaling parameters. Each second-level sub-classification is then subdivided at the next level based on the adaptive decomposition level, resulting in third-level sub-classifications based on mother wavelet type, adaptive transform scaling parameters, and adaptive decomposition levels.
[0090] Each third-level subclass contains the same number of standard frequency domain feature data matrices. Each standard frequency domain feature data matrix has a roadbed retaining structure damage identification label.
[0091] Each mother wavelet, mother wavelet type, adaptive transform scaling parameter, and adaptive decomposition level has a uniquely corresponding standard frequency domain feature data matrix in the database module.
[0092] Seventh embodiment: After receiving the comprehensive frequency domain feature data matrix corresponding to a certain accelerometer, the comparison and analysis module retrieves the standard frequency domain feature data matrix corresponding to its adaptive wavelet transform parameters from the database module.
[0093] Based on the mother wavelet type, locate the first, second, third, or fourth sub-database in the database module; locate the second-level sub-classes in the first, second, third, or fourth sub-databases based on the adaptive transform scaling parameter; locate the third-level sub-classes under a certain second-level sub-class in the first, second, third, or fourth sub-database based on the adaptive decomposition level. Obtain all standard frequency domain feature data matrices under the located third-level sub-classes, and perform similarity calculations by matching the comprehensive frequency domain feature data matrix with all the standard frequency domain feature data matrices.
[0094] In one implementation, the matching similarity calculation process is as follows: The cosine similarity between the integrated frequency domain feature data matrix and each standard frequency domain feature data matrix is calculated. The products of elements with the same number of rows and columns in the integrated frequency domain feature data matrix and each standard frequency domain feature data matrix are summed, and then divided by the product of the L2 norms of the integrated frequency domain feature data matrix and the standard frequency domain feature data matrices to obtain the matching similarity between the integrated frequency domain feature data matrix and each standard frequency domain feature data matrix. The standard frequency domain feature data matrix with the highest matching similarity to the integrated frequency domain feature data matrix is extracted and marked as the optimally matched standard frequency domain feature data matrix. If the matching similarity between the optimally matched standard frequency domain feature data matrix and the integrated frequency domain feature data matrix is greater than a preset threshold of 0.9, the match is considered successful. The roadbed retaining structure damage identification label corresponding to the optimally matched standard frequency domain feature data matrix is output.
[0095] Seventh embodiment: The roadbed retaining structure damage identification tags include: Tag 1: Normal service; Tag 2: Surface crack damage; Tag 3: Structural crack damage; Tag 4: Damage caused by uneven settlement; Tag 5: Material fatigue damage; Tag 6: Environmental corrosion damage; Tag 7: Damage caused by failure of anti-slide piles; Tag 8: External impact damage; Tag 9: Damage caused by decreased structural stiffness.
[0096] It should be noted that the significance of the roadbed retaining structure damage identification tags lies in analyzing the collected vibration signal data to identify and determine the working condition and potential damage type of the high-speed railway roadbed in a specific section where an accelerometer is located. These tags help us identify the health status of the roadbed retaining structure based on the frequency domain characteristics of the vibration signals and the actual damage patterns. Through accurate classification of different types of damage, the stability and reliability of the roadbed retaining structure can be effectively assessed, providing data support for subsequent maintenance and repair decisions, thereby ensuring the safe operation and long-term stability of the high-speed railway.
[0097] It should be understood that the terms “comprising” and “including” used in this disclosure and claims indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.
[0098] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the disclosure. As used in this disclosure and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this disclosure and claims means any combination and all possible combinations of one or more of the associated listed items, and includes such combinations; The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.
Claims
1. A damage identification system for high-speed railway subgrade retaining structures based on wavelet transform, comprising an accelerometer, a preprocessing module, a central processing module, a database module, and a comparison and analysis module; wherein the accelerometer collects the original vibration signals of the subgrade retaining structures; and the accelerometer and the preprocessing module are connected via a wireless signal network. Its features are: The preprocessing module filters and reduces noise from the acquired raw vibration signal, and selects adaptive wavelet transform parameters based on the main frequency components and abrupt change characteristics of the raw vibration signal. The central processing module retrieves adaptive wavelet transform parameters to perform wavelet transform on the original vibration signal, and extracts frequency domain features at both the local and global levels to obtain local frequency domain feature data matrices and global frequency domain feature data matrices. A comprehensive frequency domain feature data matrix is obtained by fusing the local and global frequency domain feature data matrices. The database module stores several standard frequency domain feature data matrices; The comparison and analysis module obtains the roadbed retaining structure damage identification label by calculating the matching similarity between the comprehensive frequency domain feature data matrix and the standard frequency domain feature data matrix.
2. The high-speed railway subgrade retaining structure damage identification system based on wavelet transform according to claim 1, characterized in that, The specific process of filtering and denoising the acquired raw vibration signal by the preprocessing module is as follows: State-space model of the signal: Assume that for each acceleration sensor, all original vibration signals consist of the current value and the rate of change; assume that the original vibration signal at the current time t is determined by the original vibration signal at the previous time t-1, the rate of change of the original vibration signal at the previous time t, and the train vibration control input at the previous time t. Kalman filtering is performed based on a state-space model, and the control input in the Kalman filter is calculated based on the train speed and the distance from the train distance acceleration sensor.
3. The high-speed railway subgrade retaining structure damage identification system based on wavelet transform according to claim 1, characterized in that, The specific process for selecting adaptive wavelet transform parameters based on the main frequency components and abrupt change characteristics of the original vibration signal is as follows: The original vibration signal is transformed from the time domain to the frequency domain using a Fast Fourier Transform (FFT) to obtain the frequency domain signal. Spectral analysis is performed on the frequency domain signal to obtain low-frequency and high-frequency components, and the energy of the low-frequency and high-frequency components is calculated respectively. The quotient of the high-frequency and low-frequency component energies is calculated to obtain the proportion of high-frequency component energy. The maximum and minimum frequencies contained in the frequency domain signal of the original vibration signal are extracted. The rate of change of the original vibration signal is calculated to detect abrupt changes in the signal and obtain the number of abrupt change points contained in the original vibration signal. Adaptive mother wavelet selection is based on the energy proportion of high-frequency components and the number of abrupt change points; The adaptive transformation scale parameter and adaptive decomposition level are selected based on the signal sampling frequency of the original vibration signal and the maximum and minimum frequencies contained in the frequency domain signal of the original vibration signal. Calculate the reciprocal of the maximum frequency to obtain the minimum scaling parameter in the transform scale parameters; calculate the reciprocal of the minimum frequency to obtain the maximum scaling parameter in the transform scale parameters. The signal sampling frequency of the original vibration signal is obtained and divided by the step size scaling factor to obtain the translation parameter in the transformation scale parameter. The decomposition level is calculated using the sampling frequency of the original vibration signal and the maximum and minimum frequencies contained in the frequency domain signal of the original vibration signal; and the maximum value of the adaptive decomposition level is limited.
4. The high-speed railway subgrade retaining structure damage identification system based on wavelet transform according to claim 3, characterized in that: The step size ratio is adjusted according to the specific value of the original vibration signal sampling frequency. If the original vibration signal sampling frequency is greater than 50Hz, i.e., a high sampling frequency, the step size ratio is set to a preset value of 1, which means that 1 represents shifting 1 sampling point each time. If the original vibration signal sampling frequency is less than or equal to 50Hz but greater than 25Hz, i.e. a medium sampling frequency, the step size ratio is set to a preset value of 2, which means that half a sampling point is shifted each time. If the original vibration signal sampling frequency is less than or equal to 25Hz, i.e. a low sampling frequency, the step size ratio is set to a preset value of 5 to improve the analysis accuracy.
5. The high-speed railway subgrade retaining structure damage identification system based on wavelet transform according to claim 1, characterized in that, The specific process for extracting frequency domain features at both the local and global levels is as follows: After filtering and denoising the raw vibration signals collected by each accelerometer, the signals are subjected to wavelet transform using adaptive mother wavelet, adaptive transform scale parameter and adaptive decomposition level to obtain the global frequency domain feature data matrix. After filtering and denoising the raw vibration signals collected by each accelerometer, the signals are divided into several raw vibration signal segments through a preset time window. Wavelet transform is then performed on each raw vibration signal segment using adaptive mother wavelet, adaptive transform scale parameter, and adaptive decomposition level to obtain several local frequency domain feature data matrices. In the global frequency domain feature data matrix and the local frequency domain feature data matrix, each row represents a decomposition level index, each column represents a translation position of that decomposition level, and each element represents a wavelet coefficient corresponding to the decomposition level and the translation position corresponding to its own row number and column number. To avoid feature loss at signal boundaries, overlapping time windows are used, meaning that each time window overlaps with the previous time window to a certain extent, with the overlap ranging from 10% to 50% within the time window. The local frequency domain feature data matrix and the global frequency domain feature data matrix are fused to obtain the comprehensive frequency domain feature data matrix.
6. The high-speed railway subgrade retaining structure damage identification system based on wavelet transform according to claim 5, characterized in that, The specific process of data fusion is as follows: Assign a weight coefficient to the global frequency domain feature data matrix and each local frequency domain feature data matrix, and then perform a weighted summation to obtain the comprehensive frequency domain feature data matrix.
7. A high-speed railway subgrade retaining structure damage identification system based on wavelet transform according to claim 5, characterized in that, The specific process of data fusion is as follows: The matrix is concatenated by arranging the data in rows to obtain the comprehensive frequency domain feature data matrix. The global frequency domain feature data matrix is used as the first matrix in the splicing result, and the local frequency domain feature data matrix is used as the second to last matrix in the splicing result; the order of the local frequency domain feature data matrix in the splicing result is the order of the start time of the time window.
8. A high-speed railway subgrade retaining structure damage identification system based on wavelet transform according to claim 1, characterized in that: The database module includes first, second, third, and fourth sub-databases, which store data for each mother wavelet. The next level of each sub-database performs secondary classification based on the adaptive transform scale parameter to obtain a second-level sub-classification based on mother wavelet type and adaptive transform scale parameter. The next level of each second-level sub-classification performs tertiary classification based on the adaptive decomposition level to obtain a third-level sub-classification based on mother wavelet type, adaptive transform scale parameter, and adaptive decomposition level. Each third-level subclass contains the same number of standard frequency domain feature data matrices; each standard frequency domain feature data matrix has a roadbed retaining structure damage identification label. Each mother wavelet, mother wavelet type, adaptive transform scaling parameter, and adaptive decomposition level has a unique set of standard frequency domain feature data matrices in the database module.
9. A high-speed railway subgrade retaining structure damage identification system based on wavelet transform according to claim 1, characterized in that, The specific process for obtaining damage identification tags for roadbed retaining structures is as follows: After receiving the comprehensive frequency domain feature data matrix corresponding to a certain accelerometer, the comparison and analysis module retrieves the standard frequency domain feature data matrix corresponding to its adaptive wavelet transform parameters from the database module. Based on the mother wavelet type, locate the first, second, third, or fourth sub-database in the database module; locate the second-level sub-classification in the first, second, third, or fourth sub-database based on the adaptive transform scaling parameter; locate the third-level sub-classification under a certain second-level sub-classification in the first, second, third, or fourth sub-database based on the adaptive decomposition level; obtain all standard frequency domain feature data matrices under the located third-level sub-classification, and perform matching similarity calculations between the comprehensive frequency domain feature data matrix and all standard frequency domain feature data matrices.
10. A high-speed railway subgrade retaining structure damage identification system based on wavelet transform according to claim 1, characterized in that, The specific process for calculating matching similarity is as follows: Calculate the cosine similarity between the comprehensive frequency domain feature data matrix and each standard frequency domain feature data matrix to obtain the matching similarity between the comprehensive frequency domain feature data matrix and each standard frequency domain feature data matrix; extract the standard frequency domain feature data matrix with the maximum matching similarity to the comprehensive frequency domain feature data matrix and mark it as the optimal matching standard frequency domain feature data matrix; if the matching similarity between the optimal matching standard frequency domain feature data matrix and the comprehensive frequency domain feature data matrix is greater than a preset threshold, it is considered a successful match; output the roadbed retaining structure damage identification label corresponding to the optimal matching standard frequency domain feature data matrix.