Bogie frame damage detection method and device and railway vehicle

Abstract

The application discloses a bogie frame damage detection method and device and a railway vehicle, and relates to the technical field of rail transit. The scheme collects acceleration signals of end portions of side beams of each bogie frame and stress data of each stress position through vibration and stress testing of the bogie frame, and constructs a plurality of stress response prediction data sets based on the acceleration signals and the stress data; a stress prediction model is constructed based on the stress response prediction data sets, that is, a mapping relationship between acceleration and dynamic stress of the bogie frame is established by using a neural network technology; in actual application, only target acceleration signals of the end portions of the side beams of the bogie frame need to be collected and input into the stress prediction model, so that target stress time sequence signals of each stress position on the bogie frame can be obtained, thereby facilitating fatigue damage evaluation of the bogie frame, realizing indirect identification and accurate quantification of frame damage, and having the advantages of low long-term monitoring cost, high prediction precision and simple operation process.

Description

Technical Field

[0001] This application relates to the field of rail transit technology, and in particular to a method, device and rail vehicle for detecting bogie frame damage. Background Technology

[0002] As a core load-bearing component of rail vehicles, the bogie frame directly bears the dynamic load between the car body and the track, and its structural strength and fatigue life are crucial to operational safety. To accurately predict the dynamic stress distribution of the bogie frame, three main monitoring methods are currently used: strain gauge-based direct monitoring technology, load inversion method, and vibration-stress frequency transfer function method.

[0003] However, the aforementioned methods all face corresponding challenges in practical long-term applications. Strain gauge monitoring is limited by economic costs and environmental conditions, and is mostly suitable for laboratory or short-term field testing. During long-term service, it faces maintenance difficulties and insufficient data continuity. While the load inversion method can indirectly obtain stress, its accuracy is highly dependent on the accuracy of the simulation model, often resulting in stress prediction errors. Although the vibration-stress transfer function method can reduce the difficulty of sensor deployment, its stress reconstruction algorithm for key components is not yet mature, and its applicability in multi-source excitation and nonlinear systems still requires further theoretical and empirical verification.

[0004] Given the above, how to address the current limitations in the application of long-term monitoring of bogie frame damage, such as high costs, insufficient accuracy due to model dependence, and immature key algorithms, is an urgent problem for technical personnel in this field. Summary of the Invention

[0005] The purpose of this application is to provide a method, device and rail vehicle for detecting bogie frame damage, in order to solve the current application limitations such as high cost of long-term bogie frame damage monitoring, insufficient accuracy due to model dependence and immature key algorithms.

[0006] To address the aforementioned technical problems, this application provides a method for detecting damage to a bogie frame, comprising:

[0007] The target acceleration signal at the end of the side beam of the bogie frame is collected according to a preset cycle;

[0008] The target acceleration signal is input into a pre-constructed stress prediction model to output the target stress time-series signal at each stress location on the bogie frame. The construction process of the stress prediction model includes: performing vibration and dynamic stress tests on the bogie frame, and collecting acceleration signals from the ends of the side beams of each bogie frame and stress data at each stress location on the bogie frame; constructing multiple sets of stress response prediction datasets based on the acceleration signals and stress data; and training a long short-term memory network model based on the stress response prediction datasets to generate the stress prediction model.

[0009] Fatigue damage assessment of the bogie frame is performed based on the target stress time-series signals.

[0010] On the one hand, acceleration signals at the ends of the side beams of each bogie frame and stress data at various stress locations on the bogie frame are collected, including:

[0011] Based on the acceleration sensors and stress plates installed at the ends of the side beams of each bogie frame, the acceleration signals and dynamic hot spot stress data are collected respectively.

[0012] Establish a finite element model of the bogie frame in its preparation state;

[0013] Based on the finite element model and the dynamic hot spot stress data, simulation modal stress calculation is performed to determine the stress locations and corresponding stress data on the bogie frame.

[0014] The stress locations include at least the cap cylinder area, the weld seam of the transverse side beam, the weld seam of the motor hanger, and the weld seam of the transverse beam connecting beam.

[0015] On the other hand, based on the acceleration signals and stress data, multiple sets of stress response prediction datasets are constructed, including:

[0016] The acceleration signals and stress data are subjected to low-pass filtering; wherein the acceleration signals include lateral acceleration signals and vertical acceleration signals.

[0017] The filtered acceleration signals and stress data are resampled according to the Nyquist sampling theorem.

[0018] The resampled acceleration signals and stress data are normalized.

[0019] Based on the normalized acceleration signals and stress data, stress response prediction datasets corresponding to each stress location are established.

[0020] On the other hand, based on the normalized acceleration signals and stress data, stress response prediction datasets corresponding to each stress location are established, including:

[0021] Determine the size of the sliding time window and determine the current position of the sliding time window within all the acceleration signals and stress data;

[0022] Based on the current position, all the acceleration signals in the sliding time window and the stress data at the end of the sliding time window are determined as a set of stress response prediction data;

[0023] Slide the sliding time window forward along the time axis and return to the step of determining the current position of the sliding time window in all the acceleration signals and stress data, until all the acceleration signals and stress data have been traversed to obtain multiple sets of stress response prediction data;

[0024] The stress response prediction dataset is constructed based on the stress response prediction data.

[0025] On the other hand, a long short-term memory network model is trained based on each of the stress response prediction datasets to generate the stress prediction model, including:

[0026] Based on the stress response prediction datasets, training and validation sets are divided, the structure and parameters of the long short-term memory network model are initialized, and the loss function is determined.

[0027] The training set data is input into the long short-term memory network model in batches to obtain predicted values;

[0028] The loss function is called to calculate the error gradient between the predicted value and the corresponding true value, and the parameters of the long short-term memory network model are updated through the backpropagation algorithm and the optimizer. The process returns to the step of inputting the training set data into the long short-term memory network model in batches to obtain the predicted value, until the number of training steps is reached.

[0029] The validation set data is input into the long short-term memory network model to obtain the predicted value corresponding to the validation set data;

[0030] Calculate the mean absolute error, root mean square error, and coefficient of determination based on the predicted and actual values ​​corresponding to the validation set data.

[0031] Based on the mean absolute error, the root mean square error, and the coefficient of determination, determine whether the long short-term memory network model meets the preset requirements;

[0032] If so, the current Long Short-Term Memory network model is determined as the stress prediction model.

[0033] On the other hand, fatigue damage assessment of the bogie frame is performed based on the time-series signals of each of the aforementioned target stresses, including:

[0034] Calculate the damage value within the corresponding time history based on the time sequence signal of each target stress;

[0035] The total damage value of the bogie frame is determined based on the design mileage of the rail vehicle, the damage value, and the mileage corresponding to the damage value.

[0036] On the other hand, after determining the total damage value of the bogie frame, the following is also included:

[0037] Generate a log of the fatigue damage assessment of the bogie frame.

[0038] The total damage value of the steering architecture is uploaded to the cloud platform.

[0039] On the other hand, target acceleration signals at the ends of the side beams of the bogie frame are collected according to a preset cycle, including:

[0040] When the rail vehicle is running, the target acceleration signal is continuously collected from the end of the side beam of the bogie frame.

[0041] To address the aforementioned technical problems, this application also provides a bogie frame damage detection device, comprising:

[0042] The first acquisition module is used to acquire the target acceleration signal at the end of the side beam of the bogie frame according to a preset cycle;

[0043] The prediction module is used to input the target acceleration signal into a pre-built stress prediction model to output the target stress time-series signal at each stress location on the bogie frame. The modules required to build the stress prediction model include: a testing module for performing vibration and dynamic stress tests on the bogie frame; a second acquisition module for acquiring acceleration signals from the ends of the side beams of each bogie frame and stress data at each stress location on the bogie frame; a dataset construction module for constructing multiple sets of stress response prediction datasets based on the acceleration signals and stress data; and a model training module for training a long short-term memory network model based on the stress response prediction datasets to generate the stress prediction model.

[0044] The evaluation module is used to perform fatigue damage assessment on the bogie frame based on the time-series signals of each of the targets.

[0045] On one hand, the second acquisition module includes:

[0046] The acceleration and stress acquisition module is used to acquire acceleration signals and dynamic hot spot stress data respectively based on acceleration sensors and stress plates installed at the ends of the side beams of each bogie frame;

[0047] The finite element model building module is used to build a finite element model of the bogie frame in its preparation state.

[0048] The calculation module is used to perform simulation modal stress calculation based on the finite element model and the dynamic hot spot stress data to determine the stress locations and corresponding stress data on the bogie frame.

[0049] The stress locations include at least the cap cylinder area, the weld seam of the transverse side beam, the weld seam of the motor hanger, and the weld seam of the transverse beam connecting beam.

[0050] On the other hand, the dataset construction module includes:

[0051] A filtering module is used to perform low-pass filtering on each of the acceleration signals and each of the stress data; wherein, the acceleration signals include lateral acceleration signals and vertical acceleration signals;

[0052] The resampling module is used to resample the filtered acceleration signals and stress data according to the Nyquist sampling theorem.

[0053] The normalization module is used to normalize the resampled acceleration signals and stress data.

[0054] The first dataset construction submodule is used to establish the stress response prediction dataset corresponding to each stress location based on the normalized acceleration signals and stress data.

[0055] On the other hand, the first dataset construction submodule includes:

[0056] A sliding time window setting module is used to determine the size of the sliding time window and the current position of the sliding time window in all the acceleration signals and stress data;

[0057] The data selection module is used to determine a set of stress response prediction data based on the current position, including all the acceleration signals in the sliding time window and the stress data at the end of the sliding time window.

[0058] The window sliding module is used to slide the sliding time window in the forward direction of the time axis and return to the step of determining the current position of the sliding time window in all the acceleration signals and stress data, until all the acceleration signals and stress data have been traversed to obtain multiple sets of stress response prediction data;

[0059] The second dataset construction submodule is used to construct the stress response prediction dataset based on the stress response prediction data.

[0060] On the other hand, the model training module includes:

[0061] The dataset partitioning module is used to partition training set data and validation set data based on each stress response prediction dataset, initialize the structure and parameters of the long short-term memory network model, and determine the loss function;

[0062] The batch training module is used to input the training set data into the long short-term memory network model in batches to obtain predicted values;

[0063] The loss calculation module is used to call the loss function to calculate the error gradient between the predicted value and the corresponding true value, and update the parameters of the long short-term memory network model through the backpropagation algorithm and optimizer, and return to the step of inputting the training set data into the long short-term memory network model in batches to obtain the predicted value, until the number of training steps is reached;

[0064] The model validation module is used to input the validation set data into the long short-term memory network model to obtain the predicted value corresponding to the validation set data;

[0065] The indicator calculation module is used to calculate the mean absolute error, root mean square error, and coefficient of determination based on the predicted and actual values ​​corresponding to the validation set data, respectively.

[0066] The judgment module is used to determine whether the long short-term memory network model meets the preset requirements based on the mean absolute error, the root mean square error, and the coefficient of determination; if so, the current long short-term memory network model is determined as the stress prediction model.

[0067] On the other hand, the evaluation module includes:

[0068] The damage value calculation module is used to calculate the damage value within the corresponding time history based on the time series signals of each target stress.

[0069] The total damage value calculation module is used to determine the total damage value of the bogie frame based on the design mileage of the rail vehicle, the damage value, and the mileage corresponding to the damage value.

[0070] On the other hand, it also includes:

[0071] The log generation module is used to generate logs for this fatigue damage assessment of the steering frame.

[0072] The upload module is used to upload the total damage value of the steering architecture to the cloud platform.

[0073] On the other hand, the first acquisition module includes:

[0074] The continuous acquisition module is used to continuously acquire the target acceleration signal at the end of the side beam of the bogie frame when the rail vehicle is running.

[0075] To address the aforementioned technical problems, this application also provides a rail vehicle, including the aforementioned bogie frame damage detection device.

[0076] On the one hand, it also includes: an acceleration sensor; the acceleration sensor is installed at the end of the side beam of the bogie frame of the rail vehicle.

[0077] On the other hand, there are four acceleration sensors, each of which is located at the end of one of the four side beams of the bogie frame.

[0078] On the other hand, it also includes: strain gauges; the strain gauges are disposed at the ends of the side beams of the bogie frame of the rail vehicle.

[0079] On the other hand, it also includes: a communication device for establishing a communication connection between the bogie frame damage detection device and the cloud platform.

[0080] The bogie frame damage detection method provided in this application is based on a bogie frame stress prediction model. It involves conducting vibration and dynamic stress tests on the bogie frame, and collecting acceleration signals from the ends of each bogie frame side beam and stress data from various stress locations on the bogie frame. Multiple stress response prediction datasets are constructed based on these acceleration signals and stress data. A long short-term memory network model is trained based on these datasets to build the stress prediction model, which utilizes neural network technology to establish a mapping relationship between the bogie frame's acceleration and dynamic stress. In practical applications, only the target acceleration signals from the ends of the bogie frame side beams need to be collected and input into the stress prediction model to obtain the target stress time-series signals from various stress locations on the bogie frame. This facilitates fatigue damage assessment of the bogie frame, enabling indirect identification and accurate quantification of frame damage. The method offers advantages such as low long-term monitoring costs, high prediction accuracy, and simple operation.

[0081] In addition, this application also provides a bogie frame damage detection device and a rail vehicle, with the same effect. Attached Figure Description

[0082] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0083] Figure 1 A flowchart illustrating a bogie frame damage detection method provided in this application embodiment;

[0084] Figure 2 A schematic diagram of a sliding time window provided in an embodiment of this application;

[0085] Figure 3 A schematic diagram of a long short-term memory neural network provided in an embodiment of this application;

[0086] Figure 4 A comparison chart of stress time-domain prediction results provided for embodiments of this application;

[0087] Figure 5 A comparison chart of stress amplitude frequency distribution prediction results provided in the embodiments of this application;

[0088] Figure 6 A comparison chart of stress frequency domain prediction results provided in the embodiments of this application;

[0089] Figure 7 A schematic diagram of a bogie frame damage detection device provided in an embodiment of this application;

[0090] Figure 8 This is a schematic diagram of a bogie frame damage detection device provided in an embodiment of this application. Detailed Implementation

[0091] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0092] The core of this application is to provide a bogie frame damage detection method, device, and rail vehicle to solve the current application limitations such as high cost of long-term bogie frame damage monitoring, insufficient accuracy due to model dependence, and immature key algorithms.

[0093] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0094] As a core load-bearing component of subway vehicles, the bogie frame directly bears the dynamic load between the car body and the track. Its structural strength and fatigue life directly affect the safety of vehicle operation. During long-term service, the bogie frame is prone to cracks or even fractures due to fatigue damage accumulation at welds, bolted connections, and areas of geometric abrupt changes, seriously threatening the safety of train operation.

[0095] Currently, fatigue damage monitoring of bogie frames is a crucial aspect of ensuring the operational safety of rail vehicles. The mainstream methods include direct strain gauge monitoring, load inversion, and the vibration-stress frequency transfer function method. These methods each have their own emphasis in terms of principle, application conditions, and implementation effects, collectively forming the technical landscape of this field, but each also faces significant technical bottlenecks and challenges.

[0096] Traditional strain gauge direct monitoring technology measures local micro-strain by attaching resistance strain gauges to the surface of a structure, thereby obtaining the true stress state based on material mechanics relationships. This method provides intuitive data, has mature theory, and possesses irreplaceable accuracy in laboratory calibration and short-term field testing. However, its limitations are particularly pronounced in actual long-term service environments. First, economic cost and maintainability pose significant constraints: a complete bogie monitoring system often requires dozens or even hundreds of measuring points, with high costs for the sensors themselves, the associated data acquisition system, and long-term maintenance. Second, the field environment is extremely harsh; strain gauges and their conductors are exposed to continuous vibration and impact, rain and snow corrosion, oil contamination, and extreme temperature fluctuations, making them prone to detachment, damage, or signal drift, leading to data interruption or distortion. Furthermore, complex wiring can also affect the vehicle's original structure and pose safety hazards. Therefore, while this method can provide the true stress at a "point," it struggles to support continuous, high-reliability monitoring across a "surface" throughout the entire lifespan.

[0097] To overcome the limitations of direct strain gauge placement, the load inversion method offers an indirect approach. Its core idea is to deploy a limited number of strain sensors at key locations on the structure, and inversely deduce the actual load history (such as vertical, lateral, and longitudinal loads) acting on the structure through measured strain responses. Then, combined with a pre-established finite element simulation model of the structure, harmonic response analysis is performed to obtain the frequency response function from load input to stress response at any part of the structure, ultimately calculating the global dynamic stress. The greatest advantage of this method is that it achieves virtual prediction of stress in areas difficult to measure directly through a limited number of measurement points. However, its accuracy is highly dependent on the fidelity of the simulation model. Actual structures exhibit material nonlinearity, contact nonlinearity at joints, residual welding stress, and damage accumulation after long-term service. These factors create an inherent gap between the idealized linear finite element model and the real structure. Model simplification errors, boundary condition assumption deviations, and uncertainties in material parameters are amplified in the chain-like transmission of load inversion and stress calculation, ultimately making it difficult to guarantee the accuracy of stress prediction, especially in areas of localized stress concentration in the structure.

[0098] Another more promising indirect method is the vibration-stress frequency transfer function method. This method attempts to circumvent the challenge of densely deploying strain gauges, instead utilizing widely deployed and more durable accelerometers. Its principle is to establish a transfer function relationship between the vibration acceleration at a specific point on the structure and the stress response at the target point in the frequency domain using experimental data. Once this relationship is established, the stress time history can be indirectly reconstructed through frequency domain transformation simply by monitoring the vibration signal. This method significantly reduces the difficulty of sensor deployment and system cost, and facilitates long-term implementation. However, this technology still faces core challenges in achieving mature application. First, the vibration-stress transfer function is not globally universal; it strongly depends on the dynamic characteristics and load input characteristics of the structure. In actual operation of rail vehicles, the bogie is subjected to complex excitations from multiple sources, across a wide frequency range, and non-stationary, including wheel-rail contact, traction braking, and track excitation. The transfer function may change accordingly, raising questions about its stability. Second, structures exhibit nonlinear dynamic characteristics under long-term service or heavy loads, and the applicability of transfer function algorithms based on linear assumptions in such scenarios urgently requires theoretical breakthroughs and experimental verification. Currently, ensuring the robustness and accuracy of stress reconstruction algorithms for critical components under varying excitations remains a critical technical challenge that needs to be overcome.

[0099] In summary, the three main existing technologies for monitoring fatigue damage in bogie frames are evolving along a path from "direct" to "indirect" and from "local" to "global," but none of them fully address the needs for long-term, reliable, accurate, and cost-effective online monitoring. The strain gauge method is limited by durability and cost; the load inversion method is constrained by model errors; and the vibration-stress transfer function method is hampered by the reliability of the algorithm under complex working conditions.

[0100] With the rapid development of artificial intelligence technologies, especially machine learning and deep learning, using these technologies for stress prediction and damage identification has become a new trend in the field of rail transit research. Machine learning methods, by learning stress variation characteristics from large amounts of historical data, can construct stress prediction models that can adapt to different operating conditions and state changes. This data-driven approach has significant advantages over traditional methods, especially in its performance when dealing with nonlinear, high-dimensional, and complex relationships.

[0101] Based on the above, this application provides a bogie frame damage detection method that, while maintaining the model's performance in handling nonlinearity, high dimensionality, and complex relationships, constructs a high-precision, efficient, and interpretable stress prediction model, thereby further improving the prediction accuracy of bogie frame fatigue damage under full life-cycle service conditions. The method provided in this application is described in detail below:

[0102] Figure 1 This is a flowchart illustrating a bogie frame damage detection method provided in an embodiment of this application. Figure 1As shown, the method includes:

[0103] S10: Collect the target acceleration signal at the end of the side beam of the bogie frame according to the preset cycle.

[0104] The ends of the bogie frame side beams refer to the end areas of the longitudinal main beams on both sides of the bogie, which are critical parts connecting the frame to the wheel axle boxes or primary suspension system. This area directly transmits complex dynamic loads between the wheel and rail. Due to abrupt geometric changes and concentrated welds, it is highly susceptible to stress concentration and fatigue crack initiation, making its structural integrity crucial for train operation safety. Therefore, in practical implementation, it is first necessary to collect target acceleration signals from the ends of the bogie frame side beams according to a preset cycle.

[0105] It should be noted that this embodiment does not limit the preset period. For example, the target acceleration signal at the end of the bogie frame side beam can be collected with a period of 1 second. Furthermore, the stress data for the entire line can be predicted and damage calculated after the rail vehicle has completed operation, depending on the specific implementation. In addition, this embodiment does not limit the specific method of collecting the target acceleration signal. For example, it can be collected based on an acceleration sensor installed at the end of the bogie frame side beam, or it can be collected through other methods, which will not be elaborated in this embodiment.

[0106] S11: Input the target acceleration signal into the pre-built stress prediction model to output the target stress time sequence signal at each stress location on the bogie frame.

[0107] After obtaining the target acceleration signal at the end of the side beam of the bogie frame, the target acceleration signal is input into the pre-built stress prediction model to output the target stress time sequence signal at each stress position on the bogie frame.

[0108] It should be noted that the stress prediction model is a pre-built Long Short-Term Memory (LSTM) network model. An LSTM network model is a special type of recurrent neural network that, by introducing cell states and gating mechanisms (forget gate, input gate, output gate), can effectively capture and process long-term dependencies in time-series data, solving the gradient vanishing problem of traditional recurrent neural networks. It is widely used in time series prediction, natural language processing, and speech recognition. In this embodiment, the input of the stress prediction model is an acceleration signal, and the output is the stress time-series signal at each stress location on the bogie frame. Specifically, the construction process of the stress prediction model involves performing vibration and dynamic stress tests on the bogie frame, collecting acceleration signals from the ends of the side beams of each bogie frame, and collecting stress data at each stress location on the bogie frame; constructing multiple sets of stress response prediction datasets based on the acceleration signals and stress data; and training the LSTM network model based on these datasets to generate the stress prediction model. This embodiment does not limit the specific details of the stress prediction model construction process; it depends on the specific implementation.

[0109] Furthermore, the stress locations on the bogie frame mainly refer to critical areas prone to stress concentration or high stress under wheel-rail dynamic loads. These areas mainly include, but are not limited to, welds at the connections between side beams and cross beams, geometric transitions between hangers and mounting bases, brake beam connection nodes, and the areas around bolt holes. The stress levels at these locations directly affect the fatigue life of the frame and are key areas of focus for structural health monitoring and strength assessment. This embodiment does not limit the specific stress locations involved; they will be determined based on the specific implementation.

[0110] S12: Perform fatigue damage assessment on the bogie frame based on the time-series signals of each target stress.

[0111] Finally, fatigue damage assessment of the bogie frame is performed based on the target stress time-series signals at each stress location. This assessment uses vibration acceleration to evaluate the degree of fatigue damage to prevent damage caused by fatigue fracture of the bogie frame. The specific process of fatigue damage assessment is not limited in this embodiment and will be determined according to the specific implementation.

[0112] In this embodiment, vibration and dynamic stress tests are performed on the bogie frame, and acceleration signals from the ends of the side beams of each bogie frame and stress data at various stress locations on the bogie frame are collected. Based on the acceleration signals and stress data, multiple sets of stress response prediction datasets are constructed. A long short-term memory network model is trained based on each stress response prediction dataset, thus realizing the construction of a stress prediction model. In other words, neural network technology is used to establish the mapping relationship between the acceleration and dynamic stress of the bogie frame. In practical applications, it is only necessary to collect the target acceleration signals from the ends of the side beams of the bogie frame and input them into the stress prediction model to obtain the target stress time-series signals at various stress locations on the bogie frame. This facilitates fatigue damage assessment of the bogie frame, achieving indirect identification and accurate quantification of frame damage. It has the advantages of low long-term monitoring cost, high prediction accuracy, and simple operation.

[0113] In the process of constructing the stress prediction model, in order to acquire acceleration signals and corresponding stress data, in some embodiments, acceleration signals at the ends of the side beams of each bogie frame and stress data at each stress location on the bogie frame are acquired, including:

[0114] S101: Based on the acceleration sensors and stress plates installed at the ends of the side beams of each bogie frame, the acceleration signals and dynamic hot spot stress data are collected respectively.

[0115] S102: Establish a finite element model of the bogie frame in its preparation state.

[0116] S103: Based on the finite element model and dynamic hot spot stress data, simulate modal stress calculation is performed to determine the stress locations and corresponding stress data on the bogie frame.

[0117] Specifically, the bogie frame is first measured. Accelerometers are installed at the ends of the side beams of the bogie frame, and strain gauges are also installed on the bogie frame. During vibration and dynamic stress tests, acceleration signals and dynamic hotspot stress data are collected based on the accelerometers and strain gauges located at the ends of each bogie frame side beam. It should be noted that this embodiment does not limit the specific number of accelerometers. For example, since there are four bogie frame side beam ends, one accelerometer can be installed at each of the four ends, for a total of four accelerometers.

[0118] Furthermore, a precise finite element model of the bogie frame in its prepared state is established. It should be noted that the finite element model of the bogie frame in its prepared state refers to a digital mechanical simulation model that simulates its structural and load-bearing state when all suspension components, braking devices, and other auxiliary equipment are installed and it bears the vehicle's unloaded weight. This model is used to accurately calculate the stress distribution and deformation of the frame under static reference conditions, serving as the foundation for subsequent dynamic loading, fatigue strength analysis, and structural optimization design. Subsequently, based on the finite element model and dynamic hotspot stress data, simulated modal stress calculations are performed to determine the stress locations and corresponding stress data on the bogie frame.

[0119] It is important to note that the stress locations in this embodiment include at least the cap section, the welded areas of the transverse side beams, the welded areas of the motor mount, and the welded areas of the transverse beam connecting beams. Specifically, the cap section refers to the cylindrical structure at both ends of the side beam used to install axle box springs or primary suspension. Because it directly bears the vertical and lateral dynamic loads from the wheelset and has abrupt geometric changes, stress concentration is easily formed at the root fillet or welded area, making it the main point of application for vertical fatigue loads. The welded areas of the transverse side beams are the welded areas connecting the transverse beams and the longitudinal side beams. This area bears the bending and torsional loads of the entire frame. As a material discontinuity area, the weld itself is prone to fatigue cracking under cyclic loads, which has a critical impact on the overall load-bearing stiffness and fatigue life of the frame. The welded areas of the motor mount are the connecting welds of the mounting seats on which the traction motor is suspended from the frame. The motor's operating vibration and starting / braking torque will generate high-frequency alternating stress at this location, and the mount structure is mostly cantilevered, with a complex stress state, making it extremely prone to weld cracking due to fatigue accumulation. The welded joints of the crossbeams are located at the connections between the transverse auxiliary beams and the main beams within the frame. These welds transmit longitudinal traction and braking forces, and bear cyclic tensile and compressive stresses. Stress concentrations are significant at points of abrupt changes in local stiffness, making them weak points in the longitudinal power transmission path of the frame.

[0120] Since the aforementioned stress locations are all structurally located in areas of geometric abrupt change, weld joints, or areas directly subjected to dynamic loads, they continuously endure multi-directional, high-frequency cyclic stresses during long-term service, making them critical risk points where fatigue damage is most likely to initiate and develop. Therefore, stress assessment and monitoring of these locations are necessary to identify potential cracks in advance, providing a direct basis for the safe maintenance and life prediction of the structure.

[0121] In the process of constructing the stress prediction model, in order to build multiple sets of stress response prediction datasets, in some embodiments, multiple sets of stress response prediction datasets are constructed based on each acceleration signal and each stress data, including:

[0122] S111: Perform low-pass filtering on each acceleration signal and each stress data; the acceleration signals include lateral acceleration signals and vertical acceleration signals.

[0123] S112: Resample the filtered acceleration signals and stress data according to the Nyquist sampling theorem.

[0124] S113: Normalize the resampled acceleration signals and stress data.

[0125] S114: Based on the normalized acceleration signals and stress data, establish stress response prediction datasets corresponding to each stress location.

[0126] To construct the dataset, this embodiment requires downsampling and normalization of the data. Specifically, low-pass filtering is applied to each acceleration signal and each stress data point. Low-pass filtering eliminates high-frequency noise in the data, making the signals smoother and reducing the complexity of the model's learning process. This helps improve the stability and convergence speed of the training process, while reducing the risk of overfitting, allowing the model to focus more on the main patterns in the data, thus often improving prediction accuracy. It should be noted that this embodiment does not limit the specific method of low-pass filtering. For example, a 250Hz low-pass filter can be used for the acceleration signals and stress data. Furthermore, the acceleration signals in this embodiment specifically include lateral acceleration signals and vertical acceleration signals.

[0127] Furthermore, the filtered acceleration signals and stress data are resampled according to the Nyquist sampling theorem. It is understood that the Nyquist sampling theorem states that to reconstruct a continuous signal without distortion, the sampling frequency must be greater than twice the highest frequency of the signal. In this embodiment, the filtered data is resampled to ensure that the sampling process satisfies this theorem. By first performing low-pass filtering to limit the signal bandwidth, and then resampling at an appropriate frequency, spectral aliasing can be effectively avoided, and the amount of data can be reduced while retaining the main information, providing a more effective and compliant data foundation for subsequent processing. This embodiment does not limit the specific method of data resampling; for example, the filtered data can be resampled at 500Hz according to the Nyquist sampling theorem.

[0128] Subsequently, the resampled acceleration signals and stress data are normalized to ensure their values ​​fall within the range of [0,1]. This embodiment does not restrict the normalization method; for example, the acceleration signals and stress data can be normalized using the following formula:

[0129] ;

[0130] in, For the normalized data, This is the original data. and These are the maximum and minimum values, respectively.

[0131] Finally, based on the normalized acceleration signals and stress data, stress response prediction datasets corresponding to each stress location are established.

[0132] Figure 2 This is a schematic diagram of a sliding time window provided in an embodiment of this application. Since the acceleration signal and stress data, as input and output data, strictly adhere to the vibration transmission relationship, in some embodiments, such as... Figure 2 As shown, based on the normalized acceleration signals and stress data, stress response prediction datasets corresponding to each stress location are established, including:

[0133] S121: Determine the size of the sliding time window and its current position within all acceleration signals and stress data.

[0134] S122: Based on the current position, determine all acceleration signals in the sliding time window and the stress data at the end of the sliding time window as a set of stress response prediction data.

[0135] S123: Slide the sliding time window forward along the time axis, return to the step of determining the current position of the sliding time window in all acceleration signals and stress data, until all acceleration signals and stress data have been traversed, and multiple sets of stress response prediction data are obtained.

[0136] S124: Construct a stress response prediction dataset based on various stress response prediction data.

[0137] like Figure 2 As shown, both the acceleration signal and stress data are time-series data. To generate the stress response prediction dataset, in this embodiment, the size of the sliding time window (i.e., ...) is first determined. Figure 2 The time series length is determined, and the current position of the sliding time window within all acceleration signals and stress data is identified. Subsequently, based on the current position, all acceleration signals within the sliding time window and the stress data at the end of the sliding time window are combined to form a set of stress response prediction data.

[0138] Furthermore, by sliding the sliding time window forward along the time axis, the process returns to the step of determining the current position of the sliding time window within all acceleration signals and stress data, until all acceleration signals and stress data have been traversed, resulting in multiple sets of stress response prediction data. Finally, a stress response prediction dataset is constructed based on each set of stress response prediction data.

[0139] For example, using the lateral and vertical accelerations at the end of the frame as input signals and the stress data at the cap section as output signals, a corresponding stress response prediction dataset can be constructed in the manner described above. Here, 1y, 1z, 2y, 2z, 3y, 3z, 4y, and 4z represent the lateral and vertical vibration accelerations at the four ends of the frame, respectively, and S represents the stress at the E14 sleeve measuring point. Therefore, the entire system can be considered as an 8-input, 1-output system. Similarly, the construction method for the stress response prediction datasets of the side beam weld, motor hanger weld, and crossbeam connecting beam weld is similar to the above. Based on the obtained stress response prediction datasets for each stress location, multiple multi-input, single-output systems can be constructed.

[0140] To construct the stress prediction model, based on the above embodiments, in some embodiments, a long short-term memory network model is trained based on each stress response prediction dataset to generate the stress prediction model, including:

[0141] S131: Based on the stress response prediction datasets, divide the training set data and validation set data, initialize the structure and parameters of the long short-term memory network model, and determine the loss function.

[0142] S132: Input the training set data in batches into the Long Short-Term Memory network model to obtain the predicted values.

[0143] S133: Call the loss function to calculate the error gradient between the predicted value and the corresponding true value, and update the parameters of the long short-term memory network model through the backpropagation algorithm and optimizer. Return to the step of inputting the training set data into the long short-term memory network model in batches to obtain the predicted value, until the training steps are reached.

[0144] S134: Input the validation set data into the Long Short-Term Memory network model to obtain the predicted values ​​corresponding to the validation set data.

[0145] S135: Calculate the mean absolute error, root mean square error, and coefficient of determination based on the predicted and actual values ​​corresponding to the validation set data.

[0146] S136: Based on the mean absolute error, root mean square error, and coefficient of determination, determine whether the long short-term memory network model meets the preset requirements; if so, determine the current long short-term memory network model as the stress prediction model.

[0147] In this embodiment, the Long Short-Term Memory (LSTM) neural network model uses Mean Squared Error (MSE) as the loss function, selects the Adam optimizer for training, sets the initial learning rate to 0.0001, and sets the batch size for each training iteration to 64. It is understood that the LSM neural network is an improved structure of the traditional recurrent neural network. Figure 3 This is a schematic diagram of a long short-term memory neural network provided in an embodiment of this application. Figure 3As shown, compared to traditional recurrent neural networks, it has several more gating units, but the network structures of the two are similar. Therefore, similar to the parameter design of recurrent neural networks, the influence of time series length, the number of hidden layer neurons, and the number of neural network layers in this application embodiment were analyzed. The time series length range was 10-110 (10 as the interval), the number of hidden layer neurons ranged from 25-125 (25 as the interval), and the number of neural network layers ranged from 1-4 (1 as the interval). After comparison, the optimal combination of long short-term memory neural network parameters was found to be a time series length of 80, a number of hidden layer neurons of 100, and a number of neural network layers of 3.

[0148] In the specific implementation process, firstly, training and validation sets are divided based on each stress response prediction dataset. The structure and parameters of the Long Short-Term Memory (LSTM) network model are initialized, and the loss function MSE is determined. Then, a batch of training data is input into the LTM network model, and the model performs forward propagation to obtain the corresponding predicted values. Further, the error gradient between the predicted and actual values ​​is calculated using the loss function, and the parameters of the LTM network model are updated through backpropagation and an optimizer. The process is repeated, returning to the step of inputting batches of training data into the LTM network model to obtain predicted values, until the preset number of training steps is reached. Subsequently, validation data is input into the LTM network model to obtain the predicted values ​​corresponding to the validation data. The mean absolute error, root mean square error, and coefficient of determination are calculated based on the predicted and actual values ​​corresponding to the validation data, as shown in the following formulas:

[0149] ;

[0150] ;

[0151] ;

[0152] Where MAE is the mean absolute error, RMSE is the root mean square error, and R0 is the root mean square error. 2 As the coefficient of determination, For predicted values, For the true value, This is the mean of the true values.

[0153] Finally, based on the Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination (CDE), it is determined whether the Long Short-Term Memory (LSTM) network model meets the preset requirements. It should be noted that MAE calculates the average absolute error between the predicted and actual values ​​for all samples. MAE is insensitive to outliers and directly reflects the average magnitude of the prediction error. For example, an MAE of 5 means that on average, each prediction deviates from the actual value by 5 units. RMSE calculates the square of the error (amplifying larger errors), then averages and takes the square root to ensure consistency with the original data. RMSE penalizes larger errors more severely, thus better reflecting the accuracy of the prediction. A larger RMSE value indicates a wider distribution of prediction errors or the presence of large errors. R 2 R² measures the model's ability to explain changes in the target variable. The value of R² ranges roughly from 0 to 1 (it can also be negative). The closer R² is to 1, the better the model fits, and the more accurate its predictions are compared to simply using the mean. R² close to 0 means the model's prediction performance is no different from using the mean. A negative value indicates that the model is inferior to simple mean. In this embodiment, there are no restrictions on the preset conditions; they depend on the specific implementation. If it is confirmed that the Long Short-Term Memory (LSTM) network model does not meet the preset requirements, it is retrained. If it is confirmed that the LTM network model meets the preset requirements, the current LTM network model is determined as the stress prediction model, thus completing the construction of the stress prediction model.

[0154] To achieve fatigue damage assessment of the bogie frame, based on the above embodiments, in some embodiments, fatigue damage assessment of the bogie frame is performed based on the time-series signals of each target stress, including:

[0155] S141: Calculate the damage value within the corresponding time history based on the stress time sequence signal of each target.

[0156] S142: Determine the total damage value of the bogie frame based on the design mileage, damage value, and mileage corresponding to the damage value of the rail vehicle.

[0157] Specifically, the damage value within the corresponding time history is first calculated based on the time-series stress signals of each target. Then, the total damage value of the bogie frame is determined based on the design mileage of the rail vehicle, the damage value, and the mileage corresponding to the damage value. The specific formula is as follows:

[0158] ;

[0159] in, This represents the total damage value of the bogie frame. For the design mileage of rail vehicles, The damage value corresponds to the time history. This represents the mileage corresponding to the damage value.

[0160] Figure 4A comparison chart of stress time-domain prediction results provided in the embodiments of this application. For example... Figure 4 As shown, the frequency distribution of stress amplitude was calculated using measured and predicted stress time histories and compared with the results. The predicted frequency distribution of stress amplitude was basically consistent with the calculated results of measured frequency distribution of stress amplitude, and the error was within an acceptable range.

[0161] Figure 5 A comparison chart of the predicted results of stress amplitude frequency distribution from stress rainflow counting provided in the embodiments of this application. Figure 6 A comparison chart of stress frequency domain prediction results provided in the embodiments of this application. (See attached image.) Figure 5 and Figure 6 As shown in the comparison of frequency domain results, the frequency distribution of stress measurement points is between 80-100Hz. It can be seen that the power spectral density variation trends of the predicted and actual values ​​have good consistency. The predicted stress can reflect the true dominant frequency well, and the error is within an acceptable range.

[0162] To effectively record the relevant data from this bogie frame damage assessment, based on the above embodiments, in some embodiments, after determining the total damage value of the bogie frame, the following is also included:

[0163] S151: Generate a log of the fatigue damage assessment of the bogie frame.

[0164] S152: Upload the total damage value of the steering architecture to the cloud platform.

[0165] Specifically, during the fatigue damage assessment of this bogie frame, a detailed assessment log was automatically generated, fully recording the input acceleration signals, output stress data, and calculation steps and intermediate results based on cumulative damage theory, ensuring the transparency and traceability of the assessment process. Subsequently, the total damage value obtained from the assessment was uploaded to the cloud platform in real time via an encrypted interface and stored in a centralized asset health management database for remote access and integrated analysis. Through this process, not only was reliable archiving and efficient sharing of assessment data achieved, but continuous condition monitoring was also supported. This significantly improved the accuracy and timeliness of fatigue damage assessment, providing a solid data foundation for predictive maintenance, thereby effectively reducing the risk of frame failure, optimizing maintenance cycles, and enhancing the overall safety and operational economy of the rail vehicle.

[0166] In addition, in order to predict the stress data of the entire rail vehicle and calculate its damage, when collecting the target acceleration signal at the end of the bogie frame side beam, the target acceleration signal at the end of the bogie frame side beam can be continuously collected while the rail vehicle is running, so as to realize the prediction of the damage of the entire line after the train runs.

[0167] In the above embodiments, the method for detecting bogie frame damage has been described in detail. This application also provides embodiments of a bogie frame damage detection device.

[0168] Figure 7 This is a schematic diagram of a bogie frame damage detection device provided in an embodiment of this application. Figure 7 As shown, the device includes:

[0169] The first acquisition module 10 is used to acquire the target acceleration signal at the end of the side beam of the bogie frame according to a preset cycle;

[0170] The prediction module 11 is used to input the target acceleration signal into a pre-built stress prediction model to output the target stress time-series signal at each stress location on the bogie frame. The modules required to build the stress prediction model include: a testing module for performing vibration and dynamic stress tests on the bogie frame; a second acquisition module for acquiring acceleration signals from the ends of the side beams of each bogie frame and stress data at each stress location on the bogie frame; a dataset construction module for constructing multiple sets of stress response prediction datasets based on the acceleration signals and stress data; and a model training module for training a long short-term memory network model based on the stress response prediction datasets to generate the stress prediction model.

[0171] Evaluation module 12 is used to evaluate the fatigue damage of the bogie frame based on the time-series signals of each target stress.

[0172] On the one hand, the second acquisition module includes:

[0173] The acceleration and stress acquisition module is used to acquire acceleration signals and dynamic hot spot stress data based on acceleration sensors and stress plates installed at the ends of the side beams of each bogie frame.

[0174] The finite element model building module is used to build a finite element model of the bogie frame in its preparation state.

[0175] The calculation module is used to perform simulated modal stress calculations based on the finite element model and dynamic hot spot stress data to determine the stress locations and corresponding stress data on the bogie frame.

[0176] Among them, the stress locations include at least the cap cylinder area, the weld seam of the transverse side beam, the weld seam of the motor hanger, and the weld seam of the transverse beam connecting beam.

[0177] On the other hand, the dataset building module includes:

[0178] The filtering module is used to perform low-pass filtering on various acceleration signals and stress data; the acceleration signals include lateral acceleration signals and vertical acceleration signals.

[0179] The resampling module is used to resample the filtered acceleration signals and stress data according to the Nyquist sampling theorem.

[0180] The normalization module is used to normalize the resampled acceleration signals and stress data.

[0181] The first dataset construction submodule is used to build stress response prediction datasets for each stress location based on the normalized acceleration signals and stress data.

[0182] On the other hand, the first dataset construction submodule includes:

[0183] The sliding time window setting module is used to determine the size of the sliding time window and its current position in all acceleration signals and stress data.

[0184] The data selection module is used to determine a set of stress response prediction data based on the current position, including all acceleration signals in the sliding time window and the stress data at the end of the sliding time window.

[0185] The window sliding module is used to slide the sliding time window forward along the time axis and return to the step of determining the current position of the sliding time window in all acceleration signals and stress data, until all acceleration signals and stress data have been traversed and multiple sets of stress response prediction data are obtained.

[0186] The second dataset construction submodule is used to construct a stress response prediction dataset based on each stress response prediction data.

[0187] On the other hand, the model training module includes:

[0188] The dataset partitioning module is used to partition the training set and validation set based on each stress response prediction dataset, initialize the structure and parameters of the long short-term memory network model, and determine the loss function.

[0189] The batch training module is used to input training set data into the Long Short-Term Memory network model in batches to obtain predicted values;

[0190] The loss calculation module is used to call the loss function to calculate the error gradient between the predicted value and the corresponding true value, and update the parameters of the long short-term memory network model through the backpropagation algorithm and optimizer. It then returns to the step of inputting the training set data into the long short-term memory network model in batches to obtain the predicted value, until the training steps are reached.

[0191] The model validation module is used to input validation set data into the long short-term memory network model to obtain the predicted values ​​corresponding to the validation set data.

[0192] The indicator calculation module is used to calculate the mean absolute error, root mean square error, and coefficient of determination based on the predicted and actual values ​​corresponding to the validation set data, respectively.

[0193] The judgment module is used to determine whether the long short-term memory network model meets the preset requirements based on the mean absolute error, root mean square error, and coefficient of determination; if so, the current long short-term memory network model is determined as the stress prediction model.

[0194] On the other hand, the evaluation module includes:

[0195] The damage value calculation module is used to calculate the damage value within the corresponding time history based on the stress time sequence signal of each target.

[0196] The total damage value calculation module is used to determine the total damage value of the bogie frame based on the design mileage of the rail vehicle, the damage value, and the mileage corresponding to the damage value.

[0197] On the other hand, it also includes:

[0198] The log generation module is used to generate logs for this fatigue damage assessment of the steering frame.

[0199] The upload module is used to upload the total damage value of the steering architecture to the cloud platform.

[0200] On the other hand, the first acquisition module includes:

[0201] The continuous acquisition module is used to continuously acquire target acceleration signals at the ends of the side beams of the bogie frame when the rail vehicle is running.

[0202] Since the embodiments of the apparatus and the embodiments of the method correspond to each other, please refer to the description of the embodiments of the method for the embodiments of the apparatus, which will not be repeated here.

[0203] Figure 8 This is a schematic diagram of a bogie frame damage detection device provided in an embodiment of this application. Figure 8 As shown, the bogie frame damage detection equipment includes:

[0204] Memory 20 is used to store computer programs;

[0205] The processor 21 is used to execute a computer program to implement the steps of the bogie frame damage detection method mentioned in the above embodiments.

[0206] The bogie frame damage detection device provided in this embodiment may include, but is not limited to, smartphones, tablets, laptops, or desktop computers.

[0207] The processor 21 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 21 may be implemented using at least one of the following hardware forms: Digital Signal Processor (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 21 may also include a main processor and a coprocessor. The main processor, also known as the Central Processing Unit (CPU), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 21 may integrate a Graphics Processing Unit (GPU), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 21 may also include an Artificial Intelligence (AI) processor, which handles computational operations related to machine learning.

[0208] The memory 20 may include one or more computer-readable storage media, which may be non-transitory. The memory 20 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In this embodiment, the memory 20 is used to store at least the following computer program 201, which, after being loaded and executed by the processor 21, is capable of implementing the relevant steps of the steering architecture damage detection method disclosed in any of the foregoing embodiments. In addition, the resources stored in the memory 20 may also include an operating system 202 and data 203, and the storage method may be temporary or permanent storage. The operating system 202 may include Windows, Unix, Linux, etc. The data 203 may include, but is not limited to, the data involved in the steering architecture damage detection method.

[0209] In some embodiments, the bogie frame damage detection device may further include a display screen 22, an input / output interface 23, a communication interface 24, a power supply 25, and a communication bus 26.

[0210] Those skilled in the art will understand that Figure 8 The structure shown does not constitute a limitation on the bogie frame damage detection device and may include more or fewer components than shown.

[0211] This application also provides an embodiment corresponding to a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps described in the above method embodiments.

[0212] It is understood that if the methods in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and executes all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0213] Finally, this application also provides a rail vehicle including the bogie frame damage detection device described in the above embodiments. The bogie frame damage detection device includes: a first acquisition module 10, used to acquire target acceleration signals at the ends of the side beams of the bogie frame according to a preset period; a prediction module 11, used to input the target acceleration signals into a pre-built stress prediction model to output target stress time-series signals at each stress location on the bogie frame; wherein, the modules required to build the stress prediction model include: a testing module, used to perform vibration testing and dynamic stress testing on the bogie frame; a second acquisition module, used to acquire acceleration signals at the ends of the side beams of each bogie frame and stress data at each stress location on the bogie frame respectively; a dataset construction module, used to construct multiple sets of stress response prediction datasets according to each acceleration signal and each stress data; a model training module, used to train a long short-term memory network model based on each stress response prediction dataset to generate a stress prediction model; and an evaluation module 12, used to perform fatigue damage evaluation on the bogie frame based on each target stress time-series signal.

[0214] Based on the aforementioned bogie frame damage detection device, vibration and dynamic stress tests are performed on the bogie frame, and acceleration signals from the ends of each bogie frame side beam and stress data from each stress location on the bogie frame are collected. Multiple stress response prediction datasets are constructed based on these acceleration signals and stress data. A long short-term memory network model is trained based on these datasets, thus constructing a stress prediction model. This utilizes neural network technology to establish a mapping relationship between the bogie frame's acceleration and dynamic stress. In practical applications, only the target acceleration signal from the ends of the bogie frame side beams needs to be collected and input into the stress prediction model to obtain the target stress time-series signal from each stress location on the bogie frame. This facilitates fatigue damage assessment of the bogie frame, achieving indirect identification and accurate quantification of frame damage. It offers advantages such as low long-term monitoring costs, high prediction accuracy, and simple operation.

[0215] To acquire acceleration signals from the end of the bogie frame side beam, rail vehicles may also include acceleration sensors; these sensors are located at the end of the bogie frame side beam. Due to the wide variety of acceleration sensors, their working principles and application scenarios differ. Traditional piezoelectric acceleration sensors are based on the piezoelectric effect. When the sensor is subjected to vibration or impact, the mass applies force to the piezoelectric material, generating a charge signal proportional to the acceleration. These sensors typically do not require external power and have advantages such as a wide frequency range and large dynamic range. In contrast, capacitive acceleration sensors sense acceleration by detecting the micro-capacitance changes caused by the movement of a mass. This type offers high accuracy, low power consumption, and good response to static acceleration (such as gravity).

[0216] Another unique type is the thermal convection accelerometer, which has no moving mass inside. Instead, it works by heating gas and detecting the asymmetric changes in the temperature field under acceleration. This structure is extremely resistant to mechanical shock. Piezoresistive accelerometers, based on semiconductor technology, utilize the principle that strain causes a change in resistance; they offer a large output signal and a relatively simple structure.

[0217] With the development of micro-electro-mechanical systems (MEMS) technology, MEMS accelerometers have become mainstream. They utilize integrated circuit technology to integrate micromechanical structures (such as mass blocks and cantilever beams) with signal processing circuitry on a single chip, achieving miniaturization, low cost, and low power consumption. Most MEMS sensors are based on capacitive or piezoresistive principles. Alternatively, high-end accelerometers based on optical principles or servo feedback technology can be selected, offering extremely high sensitivity and stability. This embodiment does not limit the specific type of accelerometer used; one or more of the aforementioned accelerometers can be used, depending on the specific implementation requirements.

[0218] In addition, in order to accurately collect the acceleration signals at the ends of each frame side beam, considering that there are a total of 4 frame side beam ends, a total of 4 acceleration sensors are set, with each acceleration sensor set at the end of one of the 4 bogie frame side beams.

[0219] To collect stress data from the bogie frame, rail vehicles also include strain gauges; these gauges are installed at the ends of the side beams of the bogie frame. As can be understood, a strain gauge is a sensitive element based on metallic or semiconductor materials. Its core principle is the "strain effect," meaning that the resistance of a conductor changes regularly with its mechanical deformation (stretching or compression). It is typically attached to an insulating substrate in a foil structure and must be firmly adhered to the surface of the object being measured, directly converting the minute deformation caused by the object under stress into a change in resistance. The main function of a strain gauge is to indirectly measure various mechanical quantities such as strain, stress, force, torque, and pressure by measuring changes in resistance.

[0220] It should be noted that, in addition to strain gauges, piezoresistive or capacitive MEMS sensors based on microelectromechanical systems (MEMS) technology can also be used; and in scenarios with strong electromagnetic interference or long-distance distributed monitoring, fiber optic grating sensors can also be selected.

[0221] Finally, to ensure that the bogie frame damage detection device can provide real-time feedback on the frame damage to maintenance personnel, the rail vehicle may also include a communication device. This communication device is used to establish a communication connection between the bogie frame damage detection device and the cloud platform. Together, the communication device and the cloud platform constitute the data link and decision-making center connecting the detection equipment and the back-end intelligent analysis center, which is a key infrastructure for realizing remote condition monitoring, in-depth data mining, and predictive maintenance.

[0222] Specifically, the core function of the communication device is to establish a stable and reliable data transmission channel. It is responsible for transmitting the raw data (such as acceleration signals, strain, and vibration signals) or pre-processed characteristic information collected by the bogie damage detection device to the remote location in real time or near real time. Depending on the application environment and requirements, the device may employ various methods such as mobile cellular networks (e.g., 4G / 5G), dedicated wireless networks, satellite communication, or Wi-Fi when the train enters the depot, ensuring effective data transmission even under complex conditions such as high-speed movement and tunnel travel. Furthermore, advanced communication devices also possess data caching, protocol conversion, network self-healing, and a certain level of edge computing capabilities, enabling them to temporarily store data during network interruptions and resume transmission upon recovery, ensuring data integrity and continuity.

[0223] The cloud platform, as the core of the system, bears the heavy responsibility of aggregating, storing, analyzing, and managing massive amounts of data. It receives data uploaded from numerous vehicle communication devices and utilizes its powerful computing resources and storage capabilities to perform long-term data archiving. More importantly, the cloud platform performs in-depth data processing by deploying advanced data analysis algorithms, machine learning models, and expert diagnostic systems. This not only accurately assesses the current state and evolution trend of damage, enabling a shift from scheduled maintenance to condition-based maintenance or predictive maintenance, but also identifies potential design flaws or common problems by aggregating and analyzing data from the entire fleet, providing data support for optimizing design and operational strategies. Simultaneously, the platform provides a visual human-machine interface, sending early warning information, diagnostic reports, and maintenance suggestions to maintenance personnel, ultimately forming an intelligent operation and maintenance closed loop integrating monitoring, analysis, decision-making, and feedback.

[0224] The foregoing provides a detailed description of a bogie frame damage detection method, apparatus, and rail vehicle. The various embodiments are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

[0225] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A method for detecting damage to a bogie frame, characterized in that, include: The target acceleration signal at the end of the side beam of the bogie frame is collected according to a preset cycle; The target acceleration signal is input into a pre-constructed stress prediction model to output the target stress time-series signal at each stress location on the bogie frame. The construction process of the stress prediction model includes: performing vibration and dynamic stress tests on the bogie frame, and collecting acceleration signals from the ends of the side beams of each bogie frame and stress data at each stress location on the bogie frame; constructing multiple sets of stress response prediction datasets based on the acceleration signals and stress data; and training a long short-term memory network model based on the stress response prediction datasets to generate the stress prediction model. Fatigue damage assessment of the bogie frame is performed based on the target stress time-series signals.

2. The method for detecting bogie frame damage according to claim 1, characterized in that, Acceleration signals at the ends of the side beams of each bogie frame and stress data at various stress locations on the bogie frame were collected, including: Based on the acceleration sensors and stress plates installed at the ends of the side beams of each bogie frame, the acceleration signals and dynamic hot spot stress data are collected respectively. Establish a finite element model of the bogie frame in its preparation state; Based on the finite element model and the dynamic hot spot stress data, simulation modal stress calculation is performed to determine the stress locations and corresponding stress data on the bogie frame. The stress locations include at least the cap cylinder area, the weld seam of the transverse side beam, the weld seam of the motor hanger, and the weld seam of the transverse beam connecting beam.

3. The method for detecting bogie frame damage according to claim 1, characterized in that, Based on the acceleration signals and stress data, multiple sets of stress response prediction datasets are constructed, including: The acceleration signals and stress data are subjected to low-pass filtering; wherein the acceleration signals include lateral acceleration signals and vertical acceleration signals. The filtered acceleration signals and stress data are resampled according to the Nyquist sampling theorem. The resampled acceleration signals and stress data are normalized. Based on the normalized acceleration signals and stress data, stress response prediction datasets corresponding to each stress location are established.

4. The method for detecting bogie frame damage according to claim 3, characterized in that, Based on the normalized acceleration signals and stress data, stress response prediction datasets corresponding to each stress location are established, including: Determine the size of the sliding time window and determine the current position of the sliding time window within all the acceleration signals and stress data; Based on the current position, all the acceleration signals in the sliding time window and the stress data at the end of the sliding time window are determined as a set of stress response prediction data; Slide the sliding time window forward along the time axis and return to the step of determining the current position of the sliding time window in all the acceleration signals and stress data, until all the acceleration signals and stress data have been traversed to obtain multiple sets of stress response prediction data; The stress response prediction dataset is constructed based on the stress response prediction data.

5. The method for detecting bogie frame damage according to claim 1, characterized in that, Training a long short-term memory network model based on each of the stress response prediction datasets to generate the stress prediction model includes: Based on the stress response prediction datasets, training and validation sets are divided, the structure and parameters of the long short-term memory network model are initialized, and the loss function is determined. The training set data is input into the long short-term memory network model in batches to obtain predicted values; The loss function is called to calculate the error gradient between the predicted value and the corresponding true value, and the parameters of the long short-term memory network model are updated through the backpropagation algorithm and the optimizer. The process returns to the step of inputting the training set data into the long short-term memory network model in batches to obtain the predicted value, until the number of training steps is reached. The validation set data is input into the long short-term memory network model to obtain the predicted value corresponding to the validation set data; Calculate the mean absolute error, root mean square error, and coefficient of determination based on the predicted and actual values ​​corresponding to the validation set data. Based on the mean absolute error, the root mean square error, and the coefficient of determination, determine whether the long short-term memory network model meets the preset requirements; If so, the current Long Short-Term Memory network model is determined as the stress prediction model.

6. The method for detecting bogie frame damage according to claim 1, characterized in that, Fatigue damage assessment of the bogie frame is performed based on the time-series signals of the target stresses, including: Calculate the damage value within the corresponding time history based on the time sequence signal of each target stress; The total damage value of the bogie frame is determined based on the design mileage of the rail vehicle, the damage value, and the mileage corresponding to the damage value.

7. The bogie frame damage detection method according to claim 6, characterized in that, After determining the total damage value of the bogie frame, the following is also included: Generate a log of the fatigue damage assessment of the bogie frame. The total damage value of the steering architecture is uploaded to the cloud platform.

8. The method for detecting bogie frame damage according to any one of claims 1 to 7, characterized in that, The target acceleration signal at the end of the side beam of the bogie frame is collected according to a preset cycle, including: When the rail vehicle is running, the target acceleration signal is continuously collected from the end of the side beam of the bogie frame.

9. A bogie frame damage detection device, characterized in that, include: The first acquisition module is used to acquire the target acceleration signal at the end of the side beam of the bogie frame according to a preset cycle; The prediction module is used to input the target acceleration signal into a pre-built stress prediction model to output the target stress time-series signal at each stress location on the bogie frame. The modules required to build the stress prediction model include: a testing module for performing vibration and dynamic stress tests on the bogie frame; a second acquisition module for acquiring acceleration signals from the ends of the side beams of each bogie frame and stress data at each stress location on the bogie frame; a dataset construction module for constructing multiple sets of stress response prediction datasets based on the acceleration signals and stress data; and a model training module for training a long short-term memory network model based on the stress response prediction datasets to generate the stress prediction model. The evaluation module is used to perform fatigue damage assessment on the bogie frame based on the time-series signals of each of the targets.

10. The bogie frame damage detection device according to claim 9, characterized in that, The second acquisition module includes: The acceleration and stress acquisition module is used to acquire acceleration signals and dynamic hot spot stress data respectively based on acceleration sensors and stress plates installed at the ends of the side beams of each bogie frame; The finite element model building module is used to build a finite element model of the bogie frame in its preparation state. The calculation module is used to perform simulation modal stress calculation based on the finite element model and the dynamic hot spot stress data to determine the stress locations and corresponding stress data on the bogie frame. The stress locations include at least the cap cylinder area, the weld seam of the transverse side beam, the weld seam of the motor hanger, and the weld seam of the transverse beam connecting beam.

11. The bogie frame damage detection device according to claim 9, characterized in that, The dataset construction module includes: A filtering module is used to perform low-pass filtering on each of the acceleration signals and each of the stress data; wherein, the acceleration signals include lateral acceleration signals and vertical acceleration signals; The resampling module is used to resample the filtered acceleration signals and stress data according to the Nyquist sampling theorem. The normalization module is used to normalize the resampled acceleration signals and stress data. The first dataset construction submodule is used to establish the stress response prediction dataset corresponding to each stress location based on the normalized acceleration signals and stress data.

12. The bogie frame damage detection device according to claim 11, characterized in that, The first dataset construction submodule includes: A sliding time window setting module is used to determine the size of the sliding time window and the current position of the sliding time window in all the acceleration signals and the stress data; The data selection module is used to determine a set of stress response prediction data based on the current position, including all the acceleration signals in the sliding time window and the stress data at the end of the sliding time window. The window sliding module is used to slide the sliding time window in the forward direction of the time axis and return to the step of determining the current position of the sliding time window in all the acceleration signals and stress data, until all the acceleration signals and stress data have been traversed to obtain multiple sets of stress response prediction data; The second dataset construction submodule is used to construct the stress response prediction dataset based on the stress response prediction data.

13. The bogie frame damage detection device according to claim 9, characterized in that, The model training module includes: The dataset partitioning module is used to partition training set data and validation set data based on each stress response prediction dataset, initialize the structure and parameters of the long short-term memory network model, and determine the loss function; The batch training module is used to input the training set data into the long short-term memory network model in batches to obtain predicted values; The loss calculation module is used to call the loss function to calculate the error gradient between the predicted value and the corresponding true value, and update the parameters of the long short-term memory network model through the backpropagation algorithm and optimizer, and return to the step of inputting the training set data into the long short-term memory network model in batches to obtain the predicted value, until the number of training steps is reached; The model validation module is used to input the validation set data into the long short-term memory network model to obtain the predicted value corresponding to the validation set data; The indicator calculation module is used to calculate the mean absolute error, root mean square error, and coefficient of determination based on the predicted and actual values ​​corresponding to the validation set data, respectively. The judgment module is used to determine whether the long short-term memory network model meets the preset requirements based on the mean absolute error, the root mean square error, and the coefficient of determination; if so, the current long short-term memory network model is determined as the stress prediction model.

14. The bogie frame damage detection device according to claim 9, characterized in that, The evaluation module includes: The damage value calculation module is used to calculate the damage value within the corresponding time history based on the time series signal of each target stress. The total damage value calculation module is used to determine the total damage value of the bogie frame based on the design mileage of the rail vehicle, the damage value, and the mileage corresponding to the damage value.

15. The bogie frame damage detection device according to claim 14, characterized in that, Also includes: The log generation module is used to generate logs for this fatigue damage assessment of the bogie frame. The upload module is used to upload the total damage value of the steering architecture to the cloud platform.

16. The bogie frame damage detection device according to any one of claims 9 to 15, characterized in that, The first acquisition module includes: The continuous acquisition module is used to continuously acquire the target acceleration signal at the end of the side beam of the bogie frame when the rail vehicle is running.

17. A rail vehicle, characterized in that, Includes the bogie frame damage detection device as described in any one of claims 9 to 16.

18. The rail vehicle according to claim 17, characterized in that, Also includes: Acceleration sensor; the acceleration sensor is installed at the end of the side beam of the bogie frame of the rail vehicle.

19. The rail vehicle according to claim 18, characterized in that, The number of acceleration sensors is four, and each acceleration sensor is respectively installed at the end of the side beam of the four bogie frame.

20. The rail vehicle according to claim 17, characterized in that, Also includes: Strain gauge; the strain gauge is installed at the end of the side beam of the bogie frame of the rail vehicle.

21. The rail vehicle according to claim 17, characterized in that, Also includes: A communication device is used to establish a communication connection between the bogie frame damage detection device and the cloud platform.

Images