Rolling stock running gear fault diagnosis model training method, diagnosis method and device

Abstract

The application provides a rolling stock running part fault diagnosis model training method, a diagnosis method and equipment, and relates to the technical field of vehicle fault diagnosis. The method comprises the following steps: establishing a coupling system digital twin model corresponding to a rolling stock-track coupling system according to the dynamic parameters of a target rolling stock and the running monitoring data in a normal healthy state; obtaining experimental monitoring data of the target rolling stock in the normal healthy state and a set fault working condition based on a rolling stock rolling-vibration bench experiment, and obtaining simulation monitoring data of the target rolling stock in the set fault working condition based on the twin model; performing sample segmentation and labeling on the experimental monitoring data and the simulation monitoring data respectively, obtaining source domain and auxiliary domain data to generate a combined domain data set, training an initial diagnosis model to obtain a trained domain generalization model as a rolling stock running part fault diagnosis model of the target rolling stock. The application can improve the accuracy and reliability of rolling stock running part fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle fault diagnosis, in particular to a rolling stock fault diagnosis model training method, a diagnosis method and equipment. BACKGROUND

[0002] The rolling stock is a key subsystem of the rail transit rolling stock, has important functions such as bearing, guiding, traction, braking, and vibration reduction, and includes multiple key components such as wheel sets, axle box bearings, gearboxes, and motors. These rotating components have a harsh service environment and variable operating conditions, are prone to faults such as tread flat spots, wheel polygon, bearing failure, and gear cracks, and have a high frequency, high level, and relatively random cumulative service mileage, which seriously affects the running stability and safety of the train.

[0003] Due to the complexity of the functions and structure of the rolling stock and the harsh and variable operating conditions, it is difficult to use traditional fault diagnosis methods based on mechanism models and signal processing. With the development of artificial intelligence theory, intelligent diagnosis methods based on data driving are used for fault diagnosis of the rolling stock. However, the training of deep neural network models requires sufficient labeled monitoring data as input, but in engineering practice, especially in the field of rail transportation, it is difficult to obtain a large amount of field data with rich fault types and perfect health labels. In addition, the application of deep neural network models requires the assumption that the source domain data of the training model and the target domain data of the model application should be independent and identically distributed, but this assumption is often difficult to establish when there are differences in operating conditions, measurement point locations, and application objects. Therefore, the scarcity of data and the difference in distribution seriously affect the diagnosis accuracy of the deep neural network model, thereby greatly restricting its engineering application. SUMMARY

[0004] The rolling stock fault diagnosis model training method, diagnosis method and equipment provided by the embodiments of the present application solve the problem of how to improve the diagnosis accuracy of the diagnosis model.

[0005] In a first aspect, the embodiments of the present application provide a rolling stock fault diagnosis model training method, comprising:

[0006] According to the dynamic parameters of the target rolling stock and the operating monitoring data of the target rolling stock in the normal healthy state, a coupling system digital twin model corresponding to the rolling stock-track coupling system is established;

[0007] The experimental monitoring data of the target locomotive under the normal health state working condition and the set fault working condition are obtained based on a running part rolling-vibration bench experiment, and the simulation monitoring data of the target locomotive vehicle under the set fault working condition are obtained based on the coupled system digital twin model; wherein, the sampling frequencies of the experimental monitoring data and the simulation monitoring data are the same;

[0008] The experimental monitoring data and the simulation monitoring data are respectively subjected to sample segmentation and labeling to obtain source domain data and auxiliary domain data, and are combined to obtain a combined domain data set;

[0009] The initial diagnostic model is trained according to the combined domain data set to obtain a trained domain generalization model as a running part fault diagnosis model of the target locomotive vehicle.

[0010] In a possible implementation, a coupled system digital twin model corresponding to a locomotive vehicle-track coupled system is established according to the dynamic parameters of a target locomotive vehicle and running monitoring data of the target locomotive vehicle under a normal health state, comprising:

[0011] A dynamic model of the locomotive vehicle-track coupled system is established by using a multi-body dynamics simulation software; wherein, the locomotive vehicle adopts a multi-rigid-body model, and the track adopts a massless model;

[0012] A bearing force element model and a gear transmission system force element model are established by using a system modeling and simulation software;

[0013] The bearing force element model and the gear transmission system force element model are imported into the multi-body dynamics simulation software, and the Park method is used to solve the dynamic model of the locomotive vehicle-track coupled system;

[0014] Virtual-real mapping is established based on the running monitoring data and the simulation data of the dynamic model of the locomotive vehicle-track coupled system, the dynamic model of the locomotive vehicle-track coupled system is subjected to parameter updating through the virtual-real mapping, and a coupled system digital twin model is obtained.

[0015] In a possible implementation, the multi-body dynamics simulation software is Universal Mechanism (UM), and the system modeling and simulation software is MATLAB / Simulink.

[0016] In a possible implementation, the bearing force element model and the gear transmission system force element model are imported into the multi-body dynamics simulation software, comprising:

[0017] The bearing force element model is compiled into a dynamic link library (DLL file) by MATLAB / Simulink, and is imported into UM through the Wizard of external libraries of UM;

[0018] The gear force element model is imported into UM through the gearing module in the special force module.

[0019] In a possible implementation, the parameter updating of the rolling stock-track coupling system dynamics model through virtual-real mapping comprises:

[0020] Suppose that a mapping relationship between the operation monitoring data and the simulation data is:

[0021]

[0022] wherein, is the operation monitoring data, is the simulation data, is a scaling coefficient, is a colored noise;

[0023] The scaling coefficient is estimated by using a least square method;

[0024] Based on the estimated scaling coefficient and the monitoring data, the colored noise is estimated .

[0025] In a possible implementation, the bearing force element model is

[0026]

[0027] wherein, is a number of raceway rows, is a number of rollers, is a contact force between an i-th roller and a j-th raceway, is a contact angle, is an angular position of the i-th roller; The gear transmission system force element model is:

[0028]

[0029]

[0030] wherein, is a first pair of gear teeth when double-tooth engagement occurs, is a second pair of gear teeth when double-tooth engagement occurs, is a Hertz contact stiffness, is a driving tooth bending stiffness, ​​​Bending stiffness of driven tooth, Shear stiffness of driving tooth, Shear stiffness of driven tooth, Axial compression stiffness of driving tooth, Axial compression stiffness of driven tooth, Base body deformation stiffness of driving tooth, Base body deformation stiffness of driven tooth.

[0031] In a possible implementation, the multi-rigid-body model corresponding to the rolling stock includes one car body and two bogies; each bogie includes one frame, two wheelsets, two gearboxes, two traction motors and four axle boxes; the primary suspension includes four coil springs, four primary vertical dampers and four axle box swing arm positioning devices; the secondary suspension includes two air springs, two secondary vertical dampers, two secondary lateral dampers, four anti-snaking dampers, a traction tie rod and an anti-roll torsion bar.

[0032] In a possible implementation, the training of the initial diagnosis model according to the combined domain data set obtains a trained domain generalization model, and the method comprises the following steps:

[0033] The combined domain data set is input into the initial diagnosis model for forward propagation, and cross-entropy loss, adversarial loss, feature difference loss and causal loss of the combined domain data set are calculated;

[0034] In the backward propagation stage, the parameters of the initial diagnosis model are updated by using a stochastic gradient descent algorithm, and when an iteration condition is met, a trained domain generalization model is obtained; wherein the iteration condition is that the number of iterations reaches a set iteration number or the total loss function value is lower than a threshold value; the total loss function is obtained by superimposing the cross-entropy loss, the adversarial loss, the feature difference loss and the causal loss according to preset weights.

[0035] In a possible implementation, the initial diagnosis model comprises a feature extractor, a domain discriminator and a classifier.

[0036] The feature extractor comprises a plurality of convolutional layers, pooling layers, batch normalization layers and fully connected layers; the domain discriminator comprises a plurality of fully connected layers and one softmax layer; and the classifier comprises a plurality of fully connected layers and one sigmoid layer.

[0037] In a possible implementation, before the sample segmentation and labeling of the experimental monitoring data and the simulation monitoring data, the method further comprises the following steps:

[0038] The experimental monitoring data and the simulation monitoring data are randomly oversampled and normalized.

[0039] In a possible implementation, the set fault conditions include: axle box bearing outer ring failure, axle box bearing inner ring failure, axle box bearing roller failure, large gear failure, small gear failure, and tread failure.

[0040] In a second aspect, an embodiment of the present application provides a locomotive vehicle running gear fault diagnosis method, comprising:

[0041] Obtaining real-time vibration monitoring data of a target locomotive vehicle;

[0042] Inputting the real-time vibration monitoring data into a pre-trained running gear fault diagnosis model to obtain a running gear fault label of the target locomotive;

[0043] Determining a running gear fault diagnosis result according to the fault label; wherein the running gear fault diagnosis model is trained based on the locomotive vehicle running gear fault diagnosis model training method in the first aspect or any possible implementation of the first aspect.

[0044] In a third aspect, an embodiment of the present application provides an electronic device, comprising a memory and a processor, the memory stores a computer program, and the processor implements the method in the first aspect or any possible implementation of the first aspect and the method in the second aspect when executing the computer program.

[0045] In the embodiment of the present application, a coupling system digital twin model corresponding to a locomotive vehicle-track coupling system is established according to the dynamic parameters of the target locomotive vehicle and the running monitoring data under the normal healthy state, and by incorporating the actual equipment characteristics and real health running data of the target locomotive vehicle, it is ensured that the digital twin model can fit the physical structure and normal running state of the target locomotive vehicle. Then, based on the running gear rolling-vibration bench test, experimental monitoring data under normal healthy conditions and set fault conditions are obtained, and based on the digital twin model, simulation monitoring data under set fault conditions are obtained, which not only ensures the authenticity of the normal condition data by using the bench test data, but also supplements the set fault condition data which is difficult to obtain in actual engineering through the digital twin model, effectively solving the problem of scarcity of running gear fault data. Then, the experimental monitoring data and the simulation monitoring data are respectively segmented and labeled to obtain source domain data and auxiliary domain data, and then the two are combined to form a combined domain data set. This domain processing method can clearly define the attributes of data from different sources and avoid the blurring of domain features caused by data mixing. Finally, based on the combined domain data set, an initial diagnosis model is trained to obtain a domain generalization model, which can not only make full use of the authenticity of the experimental data to improve the diagnosis accuracy, but also take advantage of the richness of the simulation data to enhance the adaptation ability to various faults, effectively avoiding the problems of low diagnosis accuracy and poor generalization caused by relying only on experimental data or ignoring domain differences, and finally realizing accurate and reliable diagnosis of the running gear fault of the target locomotive vehicle. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a flowchart of a locomotive vehicle running gear fault diagnosis model training method provided by an embodiment of the present application;

[0047] Figure 2 is a digital twin model of a locomotive vehicle-track coupling system including an axle box bearing and a gear transmission system provided by an embodiment of the present application;

[0048] Figure 3 is a schematic diagram of a running gear fault diagnosis model framework provided by an embodiment of the present application;

[0049] Figure 4 is a flowchart of a locomotive vehicle running gear fault diagnosis method provided by an embodiment of the present application;

[0050] Figure 5 is a structural schematic diagram of a locomotive vehicle running gear fault diagnosis device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0051] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0052] Figure 1 is a flowchart of a locomotive vehicle running gear fault diagnosis model training method provided by an embodiment of the present application. As shown in Figure 1 , the method comprises the following steps:

[0053] S101, a coupling system digital twin model corresponding to a locomotive vehicle-track coupling system is established according to the dynamic parameters of a target locomotive vehicle and the running monitoring data of the target locomotive vehicle in a normal healthy state.

[0054] The execution subject of each embodiment of the present application can be a server, a processor, a microprocessor, or other devices with data processing functions. In the actual implementation process, the specific implementation mode of the execution subject can be selected according to actual needs, and the present embodiment does not make special limitations, as long as it is a device with data processing function.

[0055] In the actual implementation process, the coupling system digital twin model is different for different types or models of locomotive vehicles. Therefore, before establishing the coupling system digital twin model corresponding to the locomotive vehicle-track coupling system, not only the monitoring data of the locomotive vehicle running in the normal healthy state is obtained, but also the dynamic parameters of the target locomotive vehicle are obtained.

[0056] According to the dynamic parameters of the target rolling stock, including the vehicle structure size, mass attribute, suspension parameter, etc., and the monitoring data collected when the rolling stock is running in a normal healthy state, such as vibration acceleration signals, a digital twin model of the rolling stock-track coupling system is established. The core of the model is to simulate the dynamic response of key components of the running gear, including axle box bearings and gear transmission systems, in operation.

[0057] In a specific embodiment, as shown in Figure 2 The coupling system digital twin model is a rolling stock-track coupling system digital twin model including axle box bearings and gear transmission systems.

[0058] S102, based on the running gear rolling-vibration bench experiment, obtaining experimental monitoring data of the target rolling stock in a normal healthy state and a set fault working condition, and based on the coupling system digital twin model, obtaining simulation monitoring data of the target rolling stock in the set fault working condition.

[0059] In the specific implementation process, the experimental monitoring data is collected through a full-size or reduced-scale running gear rolling-vibration test bench, covering multiple operating conditions and fault types; the simulation monitoring data is generated by adding typical faults through the digital twin model.

[0060] Among them, the sampling frequency of the experimental monitoring data and the simulation monitoring data is the same, which ensures the spatio-temporal alignment of the simulation and the experimental data, avoids resampling distortion, unifies the feature extraction scale, provides comparable input for the model, and thus improves the generalization ability and robustness of the fault diagnosis model.

[0061] In a specific embodiment, a full-size running gear rolling-vibration test bench is used to carry out running-in experiments under normal healthy state and typical fault state at speeds of 200 km / h, 250 km / h, 300 km / h and 350 km / h. Vibration acceleration sensors are used to collect the responses of the axle box and the gear box. The sampling frequency is set to 20 kHz, and the sampling time is 60 s.

[0062] S103, the experimental monitoring data and the simulation monitoring data are respectively subjected to sample segmentation and labeling to obtain source domain data and auxiliary domain data, and are combined to obtain a combined domain data set.

[0063] The experimental monitoring data and the simulation monitoring data are respectively preprocessed, which includes two main links of sample segmentation and data labeling.

[0064] Among them, the sample segmentation adopts a random overlapping sampling method to divide the continuous monitoring signal into multiple fixed-length sample segments. The data labeling assigns corresponding class labels to the samples according to their corresponding healthy state or fault type.

[0065] After preprocessing, the experimental data is divided into source domain data, and the simulation data is divided into auxiliary domain data, and finally the two are combined to form a combined domain data set for model training.

[0066] S104, training the initial diagnosis model according to the combined domain data set to obtain a trained domain generalization model as a running gear fault diagnosis model of the target locomotive vehicle.

[0067] The combined domain data set is input into the preset initial diagnosis model for training to obtain a model for diagnosing the running gear fault of the target locomotive vehicle, so as to accurately identify and diagnose the actual running state of the target locomotive vehicle.

[0068] The initial diagnosis model is a neural network model. The neural network model has strong automatic feature extraction and non-linear relationship modeling capability, can directly learn the fault essence features from complex vibration signals, and effectively reduces the dependence on expert experience. Its excellent generalization and noise tolerance performance enable it to adapt to variable data and noise interference in actual working conditions, thereby constructing a more accurate and robust running gear fault diagnosis model.

[0069] In this embodiment, a coupling system digital twin model corresponding to the locomotive vehicle-track coupling system is established according to the dynamic parameters of the target locomotive vehicle and the running monitoring data under normal healthy state, and by incorporating the actual equipment characteristics and real health running data of the target locomotive vehicle, it is ensured that the digital twin model can fit the physical structure and normal running state of the target locomotive vehicle. Then, the experimental monitoring data under normal healthy condition and set fault condition are obtained based on the running gear rolling-vibration bench test, and the simulation monitoring data under the set fault condition are obtained based on the digital twin model, which not only ensures the authenticity of the normal condition data by using the bench test data, but also supplements the set fault condition data which is difficult to obtain in actual engineering through the digital twin model, effectively solving the problem of lack of running gear fault data. Then, the experimental monitoring data and the simulation monitoring data are respectively segmented and labeled to obtain source domain data and auxiliary domain data, and then the two are combined to form a combined domain data set. This domain processing method can clearly define the attributes of data from different sources and avoid domain feature ambiguity caused by data mixing. Finally, the initial diagnosis model is trained based on the combined domain data set to obtain a domain generalization model, which can not only make full use of the authenticity of the experimental data to improve the diagnosis accuracy, but also take advantage of the richness of the simulation data to enhance the adaptability to various faults, effectively avoiding the problems of low diagnosis accuracy and poor generalization caused by relying only on experimental data or ignoring domain differences, and finally realizing accurate and reliable diagnosis of the running gear fault of the target locomotive vehicle.

[0070] The above mainly summarizes the data preparation process before the locomotive vehicle running gear fault diagnosis model training, and the data preparation process and model training process are described below.

[0071] In a possible implementation, before the sample segmentation and labeling of the experimental monitoring data and the simulation monitoring data, the method further includes:

[0072] randomly overlapping sampling and normalizing the experimental monitoring data and the simulation monitoring data.

[0073] The randomly overlapping sampling can make full use of limited original monitoring data, increase the number of samples, avoid data waste, ensure the representativeness of the samples to the original data, and effectively alleviate the problem of insufficient samples.

[0074] The normalizing operation can map monitoring data of different magnitudes and different units to a unified data interval, eliminate the interference of data magnitude differences on model training, and avoid training bias of the model due to unbalanced data magnitudes.

[0075] In the embodiment, the randomly overlapping sampling and the normalizing operation are performed before the sample segmentation and labeling of the experimental monitoring data and the simulation monitoring data. The number of samples is increased, the quality of the combined domain data set is improved, high-quality data support is provided for effective training of the initial diagnosis model, and thus the training efficiency of the model and the final fault diagnosis accuracy are improved.

[0076] In a possible implementation, the set fault conditions include: axle box bearing outer ring fault, axle box bearing inner ring fault, axle box bearing roller fault, large gear fault, small gear fault, and tread fault.

[0077] The axle box bearing outer ring fault, the inner ring fault, the roller fault, the large gear fault, the small gear fault, and the tread fault are typical fault conditions in actual operation. By setting the typical fault conditions, the running gear rolling-vibration bench test can be targeted to simulate the main fault types of the target rolling stock running gear, and the system digital twin model can also generate corresponding simulation monitoring data around these fault types, avoiding the problem of missing key fault information in the combined domain data set due to incomplete fault type coverage.

[0078] In the embodiment, the set fault conditions are limited to typical fault conditions, so that the constructed combined domain data set can comprehensively reflect the running states of various common faults of the running gear, and thus the trained domain generalization model has the diagnosis capability for multiple core faults.

[0079] In a possible implementation, the coupled system digital twin model corresponding to the locomotive vehicle-track coupled system is established according to the dynamic parameters of the target locomotive vehicle and the running monitoring data of the target locomotive vehicle in the normal healthy state, and includes:

[0080] A locomotive vehicle-track coupling system dynamics model is established by using a multi-body dynamics simulation software; wherein, the locomotive vehicle adopts a multi-rigid-body model, and the track adopts a massless model;

[0081] A bearing force element model and a gear transmission system force element model are established by using a system modeling and simulation software;

[0082] The bearing force element model and the gear transmission system force element model are imported into the multi-body dynamics simulation software, and the Park method is used to solve the locomotive vehicle-track coupling system dynamics model;

[0083] Based on the operation monitoring data and the simulation data of the locomotive vehicle-track coupling system dynamics model, a virtual-real mapping is established, the parameters of the locomotive vehicle-track coupling system dynamics model are updated through the virtual-real mapping, and a coupling system digital twin model is obtained.

[0084] In a possible implementation manner, the multi-body dynamics simulation software is Universal Mechanism (UM for short). The system modeling and simulation software is MATLAB / Simulink.

[0085] In a possible implementation manner, the bearing force element model and the gear transmission system force element model are imported into the multi-body dynamics simulation software, including:

[0086] The bearing force element model is compiled and output as a dynamic link library (DLL file) by MATLAB / Simulink;

[0087] The UM is imported through the Wizard of external libraries of the UM;

[0088] The gear force element model is imported into the UM through the gearing module in the special force module.

[0089] Before the locomotive vehicle-track coupling system dynamics model is established by using the multi-body dynamics simulation software, detailed dynamics parameters of the target locomotive vehicle are collected, including but not limited to: structure parameters of the car body and the bogie, geometric parameters of the wheelset, type and structure parameters of the axle box bearing, gear meshing parameters (such as transmission ratio, tooth profile parameters) of the gear transmission system, and the like; meanwhile, basic parameters of the track are collected, such as rail type and track gauge.

[0090] A multi-body dynamics simulation tool is used to build a dynamics model of the locomotive vehicle-track coupling system. In the vehicle part, a multi-rigid body model is used to build the vehicle body, bogie, wheelset, etc. as independent rigid bodies without considering their elastic deformation. The connection between the components is simulated through the suspension system (primary suspension and secondary suspension). The track part is represented by an elastic discrete model or a continuous model to reflect the support characteristics and profile features of the rail. Through the dynamics model, the motion state and interaction force of each component of the locomotive vehicle during operation can be calculated.

[0091] For the axle box bearing and gear transmission system of the running gear, a special force element model is built using a system simulation tool. The axle box bearing force element model needs to simulate the contact mechanics characteristics of the rollers and the inner and outer rings to reflect the force changes of the bearing during rotation. The gear transmission system force element model needs to simulate the tooth surface contact force and transmission torque during gear meshing, considering the influence of meshing stiffness, tooth side clearance, etc.

[0092] The axle box bearing force element model and the gear transmission system force element model are imported into the locomotive vehicle-track coupling system dynamics model, and the coupling calculation of the force element model and the dynamics model is realized through interface configuration to form an initial digital twin model containing the details of the core components of the running gear.

[0093] In a specific embodiment, during the process of establishing the dynamics model of the locomotive vehicle-track coupling system by the multi-body dynamics simulation software, the vehicle running speeds include different speeds of 200 km / h, 250 km / h, 300 km / h and 350 km / h, the track line is a straight line, the wheel tread is S1002CN tread, and the rail profile is CN_Rail_60. The track irregularity uses the measured Beijing-Tianjin line track random irregularity. The Park method is used to solve the locomotive vehicle-track coupling system digital twin model including the axle box bearing and the gear transmission system. The integral step is set to 5E-6.

[0094] The locomotive vehicle-track coupling system dynamics model is run to generate simulation data under normal healthy state; the simulation data is compared with the actual normal healthy state operation monitoring data of the target locomotive vehicle, the key parameters in the model are adjusted through a parameter optimization algorithm, the simulation data output by the model is consistent with the actual monitoring data in terms of trend and characteristics, and finally the calibrated coupling system digital twin model is obtained.

[0095] In the embodiment, a locomotive vehicle-track coupling system dynamics model is established by using multi-body dynamics simulation software, which contains a multi-rigid-body locomotive vehicle model and a massless track model, to ensure the basic characterization ability of the model for the locomotive vehicle-track coupling structure. Then, a force element model of the axle box bearing and gear transmission system is constructed by using system modeling and simulation software to ensure the accurate simulation of the mechanical properties of the key components of the running gear. Subsequently, the force element model is imported into the multi-body dynamics simulation software and solved by using the Park method to improve the reliability of the model calculation. Finally, the virtual-real mapping is established based on the operation monitoring data and simulation data to update the model parameters, so that the final coupling system digital twin model can continuously approach the real operation state of the target locomotive vehicle, and the authenticity of the simulation data output by the model is improved.

[0096] In a possible implementation, the parameter updating of the locomotive vehicle-track coupling system dynamics model through virtual-real mapping includes:

[0097] Suppose the mapping relationship between the operation monitoring data and the simulation data is:

[0098]

[0099] wherein, is the operation monitoring data, is the simulation data, is a scaling coefficient, is a colored noise;

[0100] The scaling coefficient is estimated by using the least square method;

[0101] Based on the estimated scaling coefficient and the monitoring data, the colored noise is estimated .

[0102] wherein, it is assumed that there is a mapping relationship between the normal health state operation monitoring data of the target locomotive vehicle and the initial simulation data output by the initial digital twin model, which reflects the amplitude difference and noise interference between the actual monitoring data and the simulation data, that is, the actual monitoring data can be obtained by superimposing noise on the initial simulation data after scaling processing, so as to quantify the correlation between the virtual and real data. The scaling coefficient in the mapping relationship is estimated by using the least square method, aiming to solve the scaling coefficient value that minimizes the error function, so as to eliminate the difference in the amplitude order of the virtual and real data.

[0103] According to the scaling coefficient estimated by using the least square method and the actual operation monitoring data, the noise component in the mapping relationship is back calculated, which reflects the interference factors in the actual operation environment, such as sensor noise, track irregularity interference, etc.

[0104] The estimated scaling factor and noise component are integrated into the locomotive-vehicle-track coupling system dynamics model to adjust the output characteristics of the model. The scaling factor is used to correct the amplitude of the simulation data, so that the model output matches the actual monitoring data in numerical range. The noise component is introduced to make the model output contain the interference characteristics in actual operation.

[0105] In actual implementation, after the updated model is verified, if the fitting degree of its output data and the actual monitoring data meets the preset requirements, the parameter correction of the model is completed, and the coupling system digital twin model is obtained.

[0106] In the embodiment, the mapping relationship formula between the actual monitoring data and the simulation data is determined, the correlation between the two is quantified, the ambiguity of virtual-real mapping is avoided, the scaling factor is estimated by using the least square method, the matching coefficient of the simulation data and the actual monitoring data in amplitude is accurately calculated, the simulation data is ensured to be consistent with the actual data in numerical order, and then the colored noise is estimated based on the estimated scaling factor and the monitoring data, so that the interference factors existing in actual operation can be restored. Such a parameter updating method has clear pertinence, can effectively couple the deviation between the simulation result of the system digital twin model and the actual operation data, and makes the updated model more truly reflect the running state of the target locomotive vehicle.

[0107] In a possible implementation, the bearing force element model is

[0108]

[0109] wherein, is the number of rows of the raceway, is the number of rollers, is the contact force between the i-th roller and the j-th raceway, is the contact angle, is the angular position of the i-th roller; The gear transmission system force element model is:

[0110]

[0111]

[0112] wherein, is the first pair of teeth in double-tooth engagement, is the second pair of teeth in double-tooth engagement, is the Hertz contact stiffness, is the bending stiffness of the driving tooth, is the bending stiffness of the driven tooth, is the shear stiffness of the driving tooth, is the shear stiffness of the driven tooth,​​​ axial compression stiffness of the driving tooth, axial compression stiffness of the driven tooth, matrix deformation stiffness of the driving tooth, matrix deformation stiffness of the driven tooth.

[0113] In the embodiment, key parameters such as the number of raceway rows, the number of rollers, the contact force, and the contact angle are included in the contact force calculation process, which can accurately characterize the mechanical response of the axle box bearing in the running process, and avoids the simulation deviation caused by the missing parameters of the bearing force element model. At the same time, in the gear transmission system force element model, different teeth are distinguished when double teeth are engaged, and various stiffness components such as Hertz contact stiffness and bending stiffness are considered, so that the stiffness characteristics in the gear engagement process can be accurately calculated, and the gear force element model can truly simulate the mechanical behavior of the gear transmission system. The bearing force element model and the gear transmission system force element model improve the simulation accuracy of the locomotive vehicle-track coupling system dynamics model.

[0114] In a possible implementation, as shown in Figure 2 Fig. 1, the multi-rigid-body model corresponding to the locomotive vehicle includes one car body and two bogies (only the front bogie is shown in Figure 2 ). Each bogie includes one frame, two wheelsets, two gearboxes, two traction motors, and four axle boxes; the primary suspension includes four coil springs, four primary vertical dampers, and four axle box swing arm positioning devices; the secondary suspension includes two air springs, two secondary vertical dampers, two secondary lateral dampers, four anti-snaking dampers, a traction tie rod, and a roll-prevention torsion bar.

[0115] In the embodiment, the number of car bodies and bogies in the multi-rigid-body model of the locomotive vehicle, and the number of components such as frames, wheelsets, and gearboxes included in the bogies, as well as the specific components of the primary suspension and the secondary suspension, are defined, which ensures that the multi-rigid-body model can completely reproduce the physical structure of the target locomotive vehicle, avoids the simulation deviation of the dynamic characteristics caused by the omission of key components due to structural simplification, and enables the locomotive vehicle-track coupling system dynamics model to accurately reflect the overall dynamic response of the locomotive vehicle during the calculation process, including the force transmission and motion state between components, thereby improving the simulation accuracy of the actual locomotive vehicle operating state by the entire coupling system digital twin model.

[0116] In a possible implementation, the initial diagnosis model is trained according to the combined domain data set to obtain a trained domain generalization model, including:

[0117] The combined domain data set is input into the initial diagnosis model for forward propagation, and the cross-entropy loss, the adversarial loss, the feature difference loss, and the causal loss of the combined domain data set are calculated;

[0118] The parameters of the initial diagnosis model are updated by using a stochastic gradient descent algorithm in the back propagation stage, and when the iteration condition is met, the trained domain generalization model is obtained. The iteration condition is that the number of iterations reaches the set iteration number or the total loss function value is lower than the threshold value; the total loss function is composed of cross entropy loss, adversarial loss, feature difference loss and causal loss according to the preset weight.

[0119] Specifically, the cross entropy loss is:

[0120]

[0121] Wherein, is the number of samples, is the extracted feature, is the cross entropy.

[0122] Further, the adversarial loss is:

[0123]

[0124] Wherein, is the domain discriminator.

[0125] Further, the feature difference loss includes two parts. The first part is the edge distribution difference between the source domain and the auxiliary domain:

[0126]

[0127] Wherein, and are the number of source domain and auxiliary domain samples, is the kernel function.

[0128] The second part is the conditional distribution difference between the source domain and the auxiliary domain:

[0129]

[0130] Wherein, is the number of health categories.

[0131] Further, the causal loss is:

[0132]

[0133] Wherein, is the cross correlation matrix, and I is the unit matrix.

[0134] Further, the total loss is:

[0135]

[0136] Wherein, the loss weight of each item is , , , .

[0137] In a specific embodiment, during training, the learning rate is set to 0.001, the batch size is 64, and the number of iterations is set to 100, wherein the loss weights are automatically updated with each iteration.

[0138] In a possible implementation, the initial diagnosis model comprises a feature extractor, a domain discriminator, and a classifier.

[0139] The constructed combined domain dataset is input into the initial diagnosis model, and each module of the model processes the data according to a preset procedure. The feature extractor extracts features from the input samples to obtain a feature vector representing the fault characteristics of the samples. Part of the feature vector is input into the classifier to output the predicted result of the fault type corresponding to the sample, and the other part is input into the domain discriminator to output the discrimination result of the data domain (source domain or auxiliary domain) to which the sample belongs.

[0140] Based on the above outputs, four kinds of losses are calculated respectively. The cross-entropy loss is used to measure the difference between the fault type prediction result output by the classification module and the true label of the sample, to ensure the accuracy of the model in identifying the fault category. The adversarial loss is used to quantify the discrimination error of the domain discrimination module in determining the domain to which the sample belongs, to promote the model to learn the cross-domain common features through adversarial training. The feature difference loss is used to evaluate the difference in feature distribution between the source domain and the auxiliary domain samples, to promote the feature distribution of the two domains to be consistent. The causal loss is used to constrain the correlation between the feature vectors, to enhance the causal mapping relationship between the features and the fault types.

[0141] Based on the forward propagation calculation, the total loss obtained by superimposing the above four kinds of losses according to the preset weights is calculated, and the random gradient descent algorithm is used to update the parameters of the initial diagnosis model. By calculating the gradient of the loss function with respect to the parameters of each layer of the model, the parameter values are adjusted in the direction of the gradient descent, and the total loss is gradually reduced. During the model training process, the loss change in the iteration process and the performance indicators of the model on the validation set are continuously monitored. When the preset iteration condition (such as the number of iterations reaching the set threshold, or the model performance not significantly improving in continuous multiple iterations) is met, the training is stopped, and the model obtained at this time is the trained domain generalization model. This model can effectively adapt to the distribution characteristics of data from different sources and has stable diagnosis capability for the running gear faults of target locomotive vehicles.

[0142] In the embodiment, the cross-entropy loss, the adversarial loss, the feature difference loss and the causal loss are calculated in the forward propagation process of the initial diagnosis model, the cross-entropy loss can guarantee the classification accuracy of the model for the fault category, the adversarial loss can reduce the distribution difference between different domain data, the feature difference loss can further align the feature distribution of the source domain and the auxiliary domain, and the causal loss can enhance the causal correlation between the features and the faults, the multi-dimensional loss function comprehensively optimizes the model training target, and combined with the stochastic gradient descent algorithm used in the backward propagation stage, the model parameters can be efficiently updated to ensure that the model is continuously adjusted to adapt to the characteristics of the combined domain data set in the iteration process, and when the iteration condition is met, the domain generalization model can effectively handle the data distribution difference problem and improve the generalization ability and diagnosis precision of the running gear fault diagnosis of the target locomotive vehicle.

[0143] In a possible implementation, the initial diagnosis model comprises a feature extractor, a domain discriminator and a classifier.

[0144] The feature extractor comprises a plurality of convolutional layers, pooling layers, batch normalization layers and fully connected layers; the domain discriminator comprises a plurality of fully connected layers and one softmax layer; and the classifier comprises a plurality of fully connected layers and one sigmoid layer.

[0145] The initial diagnosis model is composed of three core modules, namely the feature extractor, the domain discriminator and the classifier, which cooperatively realize the functions of fault feature extraction, domain difference adaptation and fault type recognition.

[0146] The feature extractor is configured to extract key fault features from the input monitoring data samples, and its structure comprises a plurality of convolutional layers, pooling layers, batch normalization layers and fully connected layers. The convolutional layers capture local features of the input data through sliding windows, and can effectively extract fault-related features such as impact and periodicity in the vibration signal; the pooling layers downsample the feature maps output by the convolutional layers, while retaining key information and reducing data dimension to improve model calculation efficiency; the batch normalization layers standardize the input data of each layer to alleviate the gradient vanishing problem in the model training process and accelerate convergence; and the fully connected layers integrate the local features extracted by the previous modules and output a feature vector containing global fault information.

[0147] The domain discriminator is configured to determine the data domain to which the input feature vector belongs, and its structure comprises a plurality of fully connected layers and one softmax layer. The fully connected layers gradually map the feature vector output by the feature extractor, and capture domain-related features through nonlinear transformation; and the softmax layer outputs the probability distribution of the feature vector belonging to the source domain or the auxiliary domain, which provides a basis for the calculation of the adversarial loss, and through adversarial training with the feature extractor, the feature extractor is prompted to generate domain-independent general features.

[0148] The classifier is used for identifying the fault type, and the structure thereof comprises a plurality of fully connected layers and a sigmoid layer. The fully connected layer further processes the feature vector output by the feature extractor, and strengthens the feature expression related to the fault type. The sigmoid layer outputs the probability value of the sample belonging to each fault type (or normal state), realizes the classification prediction of the multiple fault types, provides the basis for the calculation of the cross-entropy loss, and ensures the accurate identification of the model on the fault type.

[0149] In the embodiment, the modules work cooperatively to provide a reasonable and efficient network structure basis for the initial diagnosis model, and ensure that the model has good feature extraction, domain adaptation and classification capabilities.

[0150] As Figure 3 is a framework schematic diagram of a running gear fault diagnosis model provided by an embodiment of the present application. The running gear fault diagnosis model comprises a feature extractor, a domain discriminator and a classifier.

[0151] Among them, the source domain data and the simulation data (i.e. auxiliary domain data) are input into the feature extractor, the input samples are feature extracted by the feature extractor, and a feature vector capable of representing the fault characteristics of the samples is obtained. Part of the feature vector is input into the classifier, and the fault type prediction result (i.e. the real label) corresponding to the sample is output, and the other part is input into the domain discriminator, and the discrimination result (i.e. the domain label) of the data domain to which the sample belongs is output. The cross-entropy loss, the adversarial loss, the feature difference loss and the causal loss are calculated based on the forward propagation, and the total loss is obtained by superimposing the preset weights.

[0152] In order to verify the diagnosis effect of the running gear fault diagnosis model, real-time vibration monitoring data (i.e. target domain data) of the target locomotive vehicle is obtained, and the model is tested.

[0153] The above embodiment introduces the training process of the running gear fault diagnosis model. In the actual implementation process, the trained domain generalization model is used for running gear fault diagnosis of the diagnosis object, and the predicted fault label is output, so as to determine the running gear state according to the fault label.

[0154] Figure 4 is a flowchart of a locomotive vehicle running gear fault diagnosis method provided by an embodiment of the present application, as Figure 4 shown, comprising the following steps:

[0155] S401, obtaining real-time vibration monitoring data of a target locomotive vehicle.

[0156] S402, inputting the real-time vibration monitoring data into the running gear fault diagnosis model trained in advance to obtain the running gear fault label of the target locomotive vehicle.

[0157] S403, determining the running gear fault diagnosis result according to the fault label.

[0158] The running part fault diagnosis model is trained based on the locomotive running part fault diagnosis model training method provided in any of the preceding embodiments.

[0159] It should be understood that the size of the serial number of each step in the above embodiments does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0160] The following is an apparatus embodiment of the present application. For details not described in detail, reference can be made to the corresponding method embodiments described above.

[0161] Figure 5 The structure schematic diagram of the locomotive running part fault diagnosis apparatus provided by the embodiments of the present application is shown. For the convenience of description, only the parts related to the embodiments of the present application are shown, and the details are as follows:

[0162] As shown in Figure 5 The locomotive running part fault diagnosis apparatus 5 comprises:

[0163] The acquisition module 501 acquires real-time vibration monitoring data of the target locomotive.

[0164] The running part fault diagnosis module 502 is configured to input the real-time vibration monitoring data into the running part fault diagnosis model trained in advance, and obtain a running part fault label of the target locomotive.

[0165] The diagnosis result determination module 503 is configured to determine a running part fault diagnosis result according to the fault label.

[0166] The embodiments of the present application also provide an electronic device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the method in the above method embodiments when executing the computer program. Exemplarily, the electronic device can be a smartphone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc., which are not limited herein.

[0167] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described or recorded in detail in a certain embodiment can be referred to the relevant description of other embodiments. If there is no special description and no logical conflict, the terms and / or descriptions of different embodiments are consistent and can be mutually referenced. The technical features in different embodiments can be combined to form new embodiments according to their inherent logical relationship.

[0168] The above-described embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method of training a rolling stock failure diagnosis model, characterized by, The method comprises the following steps: According to the dynamic parameters of the target locomotive vehicle and the running monitoring data of the target locomotive vehicle in the normal healthy state, a coupling system digital twin model corresponding to the locomotive vehicle-track coupling system is established; Based on the running part rolling-vibration bench test, experimental monitoring data of the target locomotive vehicle in the normal healthy state and in a set fault condition are obtained, and simulation monitoring data of the target locomotive vehicle in the set fault condition are obtained based on the coupling system digital twin model; The experimental monitoring data and the simulation monitoring data are respectively subjected to sample segmentation and labeling to obtain source domain data and auxiliary domain data, and the source domain data and the auxiliary domain data are combined to obtain a combined domain data set; An initial diagnosis model is trained according to the combined domain data set to obtain a trained domain generalization model as a running part fault diagnosis model of the target locomotive vehicle; The method comprises the following steps: The combined domain data set is input into the initial diagnosis model for forward propagation, and the cross-entropy loss, the adversarial loss, the feature difference loss and the causal loss of the combined domain data set are calculated; the feature difference loss comprises the edge distribution difference and the conditional distribution difference between the source domain and the auxiliary domain; In the reverse propagation stage, the parameters of the initial diagnosis model are updated by using a stochastic gradient descent algorithm, and when the iteration condition is met, the trained domain generalization model is obtained; The set fault condition comprises an axle box bearing outer ring fault, an axle box bearing inner ring fault, an axle box bearing roller fault, a large gear fault, a small gear fault and a tread fault.

2. The method of claim 1, wherein, According to the dynamic parameters of the target locomotive vehicle and the running monitoring data of the target locomotive vehicle in the normal healthy state, a coupling system digital twin model corresponding to the locomotive vehicle-track coupling system is established, which comprises the following steps: A multi-body dynamics simulation software is used to establish a locomotive vehicle-track coupling system dynamics model; wherein the locomotive vehicle adopts a multi-rigid body model, and the track adopts a massless model; A bearing force element model and a gear transmission system force element model are established by using system modeling and simulation software; The bearing force element model and the gear transmission system force element model are imported into the multi-body dynamics simulation software, and the Park method is used to solve the locomotive vehicle-track coupling system dynamics model; Based on the running monitoring data and the simulation data of the locomotive vehicle-track coupling system dynamics model, a virtual-real mapping is established, the parameters of the locomotive vehicle-track coupling system dynamics model are updated through the virtual-real mapping, and a coupling system digital twin model is obtained.

3. The rail vehicle bogie fault diagnosis model training method according to claim 2, characterized in that, The method comprises the following steps: The mapping relationship between the running monitoring data and the simulation data is assumed as follows: wherein, is the running monitoring data, is the simulation data, is the scaling factor, is the colored noise; The scaling coefficient is estimated by using the least square method; Estimating colored noise based on estimated scaling factors and monitoring data .

4. The rail vehicle bogie fault diagnosis model training method according to claim 2, characterized in that, The bearing force element model is wherein, is the number of raceway rows, is the number of rollers, is the contact force between the th roller and the th raceway, is the contact angle, is the angular position of the th roller; The gear transmission system force element model is wherein, is the first pair of teeth in double tooth engagement, is the second pair of teeth in double tooth engagement, is the Hertz contact stiffness, is the active tooth bending stiffness, is the driven tooth bending stiffness, is the active tooth shear stiffness, is the driven tooth shear stiffness, is the active tooth axial compression stiffness, is the driven tooth axial compression stiffness, is the active tooth matrix deformation stiffness, is the driven tooth matrix deformation stiffness.

5. The rail vehicle bogie fault diagnosis model training method according to claim 2, characterized in that, The multi-rigid-body model of the locomotive vehicle includes one vehicle body and two bogies; each bogie includes one frame, two wheelsets, two gearboxes, two traction motors and four axleboxes; the primary suspension includes four coil springs, four primary vertical dampers and four axlebox swing arm positioning devices; the secondary suspension includes two air springs, two secondary vertical dampers, two secondary lateral dampers, four anti-snaking dampers, a traction rod and an anti-roll torsion bar.

6. The rail vehicle bogie fault diagnosis model training method according to claim 1, characterized in that, The initial diagnosis model comprises a feature extractor, a domain discriminator and a classifier. The feature extractor comprises multiple convolution layers, pooling layers, batch normalization layers and full connection layers; the domain discriminator comprises multiple full connection layers and one softmax layer; and the classifier comprises multiple full connection layers and one sigmoid layer.

7. The method of claim 1, wherein: Before sample segmentation and labeling are performed on the experimental monitoring data and the simulation monitoring data, the method further comprises: Random overlapping sampling and normalization operations are performed on the experimental monitoring data and the simulation monitoring data.

8. A method of diagnosing a failure of a running gear of a rolling stock, characterized by, The method comprises: obtaining real-time vibration monitoring data of a target locomotive vehicle; inputting the real-time vibration monitoring data into a pre-trained running gear fault diagnosis model to obtain a running gear fault label of the target locomotive vehicle; determining a running gear fault diagnosis result according to the fault label; wherein the running gear fault diagnosis model is trained based on the locomotive vehicle running gear fault diagnosis model training method of any one of claims 1 to 7.

9. An electronic device, comprising: A device comprises a memory and a processor, the memory stores a computer program, and the processor implements the method of any one of claims 1 to 8 when executing the computer program.

Images