A method, device, electronic terminal and storage medium for evaluating damage status of high-speed rail gearbox

Abstract

The present invention discloses a method, device, electronic terminal and storage medium for evaluating the damage state of a high-speed rail gearbox in the technical field of high-speed rail wheel-rail operation and maintenance. The method includes: acquiring the vibration signal of the high-speed rail gearbox to be identified; inputting the vibration signal into a pre-trained evaluation model for the damage state of the high-speed rail gearbox; respectively extracting the coarse-grained domain generalization features and the fine-grained domain generalization features in the vibration signal according to a static convolution module and a dynamic convolution module, and constructing dynamic domain generalization features; inputting the dynamic domain generalization features into a damage state discriminator to obtain multi-dimensional prediction probability data of multiple damage states, and determining the damage state evaluation result according to the multi-dimensional prediction probability data corresponding to multiple damage states. The present invention can solve the technical problem that due to the incomplete extraction of the domain generalization features of the signal, some weak but important domain generalization information is missing, resulting in a decrease in the accuracy of the damage state evaluation and making it difficult to ensure the safety and efficiency of high-speed rail operation.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of high - speed rail wheel - rail, and particularly to a method, device, electronic terminal and storage medium for evaluating the damage state of a high - speed rail gearbox. Background Technique

[0002] As a core component of the transmission system of high - speed trains, the high - speed rail gearbox is listed as a key component during high - speed rail operation. Once the damage to the high - speed rail gearbox is delayed in processing or mispredicted, it may have a serious impact on the safety and operation efficiency of high - speed trains.

[0003] Currently, the damage state assessment platform of high - speed rail gearboxes using big data technology is increasingly being emphasized in the industry. Among them, by combining technologies such as machine learning and artificial intelligence, specific algorithms and models are built to be able to intelligently evaluate the damage state of high - speed rail gearboxes, significantly improving the efficiency of maintenance and the safety of high - speed rail operation.

[0004] However, some existing methods for evaluating the damage state of high - speed rail gearboxes based on deep learning have two defects in practical engineering applications. On the one hand, the operating environment of high - speed railways is complex and changeable, resulting in different distribution characteristics of the state data of high - speed train gearboxes collected in different environments. This inconsistency in data distribution makes it difficult to extract the state characteristics of high - speed train gearboxes. On the other hand, when some existing models identify the damage state of high - speed rail gearboxes through domain - generalization learning, the domain - generalization features of the extracted signals are not comprehensive, resulting in the loss of some weak but important domain - generalization information, leading to a decrease in the accuracy of damage state assessment and making it difficult to ensure the safety and efficiency of high - speed rail operation.

[0005] Therefore, there is an urgent need for a method, device, electronic terminal and storage medium for evaluating the damage state of a high - speed rail gearbox to solve the above - mentioned technical problems. Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a method for evaluating the damage state of a high - speed rail gearbox, which can solve the technical problems that when some existing models identify the damage state of high - speed rail gearboxes through domain - generalization learning, the domain - generalization features of the extracted signals are not comprehensive, resulting in the loss of some weak but important domain - generalization information, leading to a decrease in the accuracy of damage state assessment and making it difficult to ensure the safety and efficiency of high - speed rail operation.

[0007] To achieve the above - mentioned purpose, the present invention is implemented by the following technical solutions:

[0008] In the first aspect, the present invention provides a method for evaluating the damage state of a high - speed rail gearbox, including:

[0009] Obtain the vibration signal of the high - speed rail gearbox to be identified;

[0010] Input the vibration signal into a pre-trained high-speed rail gearbox damage state evaluation model, where the high-speed rail gearbox damage state evaluation model includes a dynamic domain generalization feature extractor and a damage state discriminator, and the dynamic domain generalization feature extractor includes a static convolution module and a dynamic convolution module;

[0011] Based on the high-speed rail gearbox damage state evaluation model, extract the coarse-grained domain generalization features and fine-grained domain generalization features in the vibration signal according to the static convolution module and the dynamic convolution module respectively, and construct dynamic domain generalization features according to the coarse-grained domain generalization features and the fine-grained domain generalization features;

[0012] Input the dynamic domain generalization features into the damage state discriminator to obtain multi-dimensional prediction probability data of multiple damage states, and determine the damage state evaluation result according to the multi-dimensional prediction probability data corresponding to multiple damage states.

[0013] Furthermore, the high-speed rail gearbox damage state evaluation model further includes a dynamic coefficient discriminator, and the training process of the high-speed rail gearbox damage state evaluation model includes:

[0014] Segment the training data under different working conditions without overlap according to a specified length standard to form independent subsets, assign training class labels to the independent subsets with known damage types, and assign training domain labels to the training data under different working conditions to obtain a training data set;

[0015] Input the training data set into the dynamic domain generalization feature extractor, extract the coarse-grained domain generalization features and fine-grained domain generalization features in the training data, and construct dynamic domain generalization features according to the coarse-grained domain generalization features and the fine-grained domain generalization features;

[0016] Based on the dynamic convolution module and the dynamic coefficient discriminator, establish a dynamic coefficient discriminant loss function to counteract and remove the domain information in the fine-grained domain generalization features. Based on the damage state discriminator and the dynamic domain generalization feature extractor, establish a dynamic domain generalization feature discriminant loss function to counteract and remove the domain information in the dynamic domain generalization features to complete the training.

[0017] Furthermore, extracting the coarse-grained domain generalization features and fine-grained domain generalization features in the vibration signal according to the static convolution module and the dynamic convolution module respectively, and constructing the dynamic domain generalization features according to the coarse-grained domain generalization features and the fine-grained domain generalization features includes:

[0018] Extract the coarse-grained domain generalization feature information at the one-dimensional level through the static convolution module, and the expression is:

[0019] ,

[0020] where, represents the vibration signal, and represent the convolution kernel and bias in the one-dimensional convolution operation respectively, and represent the convolution kernel and bias in the static convolution module respectively, represents the non-linear activation function, represents the coarse-grained features output by the static convolution module;

[0021] Extract fine-grained domain generalization feature information at the one-dimensional level through the dynamic convolution module, and the expression is:

[0022] ,

[0023] ,

[0024] ,

[0025] wherein, and represent the convolution kernel and bias in the dynamic convolution module respectively, represents the fine-grained features output by the dynamic convolution module, are four convolution kernels with the same structure but different parameters included in the dynamic convolution module, represents the masking operation, represents the dynamic coefficient generated by the meta-regulator in the dynamic convolution module, represents the meta-regulator model parameters, and the symbol represents the structural reorganization of the convolution kernel;

[0026] Construct dynamic domain generalization features based on the coarse-grained domain generalization features and fine-grained domain generalization features, and the expression includes:

[0027] ,

[0028] wherein, represents the dynamic domain generalization features.

[0029] Furthermore, after constructing the dynamic domain generalization features:

[0030] Perturb the training samples in the training dataset based on the dynamic coefficients generated by the meta-regulator in the dynamic convolution module, and randomly exchange the dynamic coefficients. The expression is:

[0031] ,

[0032] ,

[0033] wherein, and represent two randomly selected training samples respectively and the set of dynamic coefficients generated by the neuron regulator, represents the training dataset, represents the dynamic coefficient perturbation operation, and the symbol represents the swap operation;

[0034] Calculate the dynamic domain generalization feature loss of the same training sample before and after perturbation. The expression is:

[0035] ,

[0036] In the formula, is the perturbation loss between the dynamic domain generalization feature output before perturbation and the dynamic domain generalization feature output after perturbation of the same training sample, represents the total number of training samples, represents the dynamic domain generalization feature of the th training sample before perturbation, and the dynamic coefficient is , represents the dynamic domain generalization feature of the th training sample after perturbation, and the dynamic coefficient is , represents calculating the L2 norm value;

[0037] Minimize the perturbation loss according to the backpropagation and optimization algorithms to extract the discriminative fine-grained features between different training samples.

[0038] Furthermore, the dynamic coefficient discriminator includes a dynamic coefficient classifier and a dynamic coefficient domain discriminator. The dynamic coefficient discriminant loss function includes:

[0039] ,

[0040] In the formula, represents the total number of training samples with training class labels assigned in the training dataset, represents calculating the cross-entropy loss, represents the probability prediction function, represents the logarithmic function, represents the dynamic coefficient classifier, represents the dynamic coefficient of the i-th training sample, represents the training class label, represents the total number of training samples in the training dataset, represents the dynamic coefficient domain discriminator, represents the training domain label, is the dynamic coefficient discriminant loss;

[0041] Adversarial training is performed between the dynamic convolution module and the dynamic coefficient domain discriminator to remove the domain information extracted by the dynamic convolution module.

[0042] Further, the damage state discriminator includes a dynamic domain generalization feature classifier and a dynamic domain generalization feature domain discriminator, and the dynamic domain generalization feature discrimination loss function includes:

[0043] ,

[0044] In the formula, represents the total number of training samples with training class labels assigned in the training dataset, represents calculating the cross-entropy loss, represents the probability prediction function, represents the logarithmic function, represents the dynamic domain generalization feature classifier, represents the dynamic domain generalization feature of the i-th sample, represents the training class label, represents the total number of training samples in the training dataset, represents the dynamic domain generalization feature domain discriminator, represents the training domain label, is the dynamic domain generalization feature discrimination loss;

[0045] Adversarial training is performed between the dynamic domain generalization feature extractor and the dynamic domain generalization feature domain discriminator to remove the domain information in the dynamic domain generalization features.

[0046] Further, the optimization algorithm adopts one of the adaptive moment estimation algorithm, the stochastic gradient descent algorithm, and the root mean square propagation algorithm.

[0047] In a second aspect, the present invention provides a high-speed rail gearbox damage state evaluation device, including:

[0048] An acquisition module, configured to acquire the vibration signal of the high-speed rail gearbox to be identified;

[0049] An input module, configured to input the vibration signal into a pre-trained high-speed rail gearbox damage state evaluation model, where the high-speed rail gearbox damage state evaluation model includes a dynamic domain generalization feature extractor and a damage state discriminator, and the dynamic domain generalization feature extractor includes a static convolution module and a dynamic convolution module;

[0050] An extraction module, configured to, based on the high-speed rail gearbox damage state evaluation model, respectively extract the coarse-grained domain generalization features and the fine-grained domain generalization features in the vibration signal according to the static convolution module and the dynamic convolution module, and construct dynamic domain generalization features according to the coarse-grained domain generalization features and the fine-grained domain generalization features;

[0051] An identification module, configured to input the dynamic domain generalization features into a damage state discriminator, obtain multi-dimensional prediction probability data of multiple damage states, and determine a damage state evaluation result according to the multi-dimensional prediction probability data corresponding to the multiple damage states.

[0052] In a third aspect, the present invention provides an electronic terminal, including a processor and a memory connected to the processor. A computer program is stored in the memory. When the computer program is executed by the processor, the steps of the method described in any one of the above are executed.

[0053] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in any one of the above are implemented.

[0054] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0055] By inputting the vibration signal to be identified into a pre-constructed high-speed rail gearbox damage state evaluation model, the dynamic domain generalization features of the vibration signal to be identified are extracted by a dynamic domain generalization feature extractor. The dynamic domain generalization features can fully reflect the domain generalization information in the vibration signal to be identified, prevent incomplete extraction of the domain generalization features of the signal and the loss of some weak but important domain generalization information, and help improve the performance output of the subsequent classifier, realizing effective online diagnosis of the high-speed rail gearbox.

[0056] By calculating and optimizing the dynamic coefficient discrimination loss and the dynamic domain generalization feature discrimination loss, the domain generalization features between different samples under different working conditions are learned, overcoming the defect that traditional deep learning damage state evaluation methods are difficult to extract cross-working condition state features, and effectively improving the accuracy of high-speed rail gearbox damage state evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is a schematic flowchart of a method for evaluating the damage state of a high-speed rail gearbox provided in Embodiment 1 of the present invention;

[0058] Figure 2 is a schematic structural diagram of a dynamic convolution module of a method for evaluating the damage state of a high-speed rail gearbox provided in Embodiment 1 of the present invention;

[0059] Figure 3 is a schematic structural diagram of a high-speed rail gearbox damage state evaluation model of a method for evaluating the damage state of a high-speed rail gearbox provided in Embodiment 1 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0060] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations on the technical solution of the present application. Without conflict, the technical features in the embodiments of the present application and the embodiments can be combined with each other.

[0061] The term "and / or" in this article is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this article generally represents an "or" relationship between the associated objects before and after.

[0062] Non-overlapping segmentation: Non-overlapping segmentation of vibration signals generally means that in the signal processing process, the signal is divided into multiple continuous segments without overlapping parts between each segment. It allows for higher resolution in the frequency domain while reducing signal distortion caused by window functions.

[0063] In practical applications, which method to choose depends on the characteristics of the signal and the goals of the analysis. If non-overlapping segmentation is required, the maximum overlap discrete wavelet transform (MODWT) can be considered. These methods can handle signals of any length and can avoid a series of problems caused by overlap.

[0064] Embodiment 1:

[0065] Figure 1 It is a flowchart of the method for evaluating the damage state of the high-speed rail gearbox in Embodiment 1 of the present invention. This flowchart only shows the logical sequence of the method described in this embodiment. On the premise of non-conflict, in other possible embodiments of the present invention, the steps shown or described may be completed in a different order from Figure 1 that shown.

[0066] The method for evaluating the damage state of the high-speed rail gearbox provided in this embodiment can be applied to a terminal and can be executed by a device for evaluating the damage state of the high-speed rail gearbox. This device can be implemented in software and / or hardware, and this device can be integrated in the terminal. For example: any smart phone, tablet computer or computer device with a communication function. See Figure 1 、 Figure 2 and Figure 3 shown, the method of this embodiment specifically includes the following steps:

[0067] Step 1: Obtain the vibration signal of the high-speed rail gearbox to be identified, and use a sensor to collect the vibration signals of the high-speed rail wheel-rail under different working conditions.

[0068] Step 2: Input the vibration signal into a pre-trained high-speed rail gearbox damage state assessment model. The high-speed rail gearbox damage state assessment model includes a dynamic domain generalization feature extractor and a damage state discriminator. The dynamic domain generalization feature extractor includes a static convolution module and a dynamic convolution module.

[0069] Step 3: Based on the high-speed rail gearbox damage state assessment model, extract the coarse-grained domain generalization features and fine-grained domain generalization features in the vibration signal according to the static convolution module and the dynamic convolution module respectively. Construct the dynamic domain generalization features according to the coarse-grained domain generalization features and the fine-grained domain generalization features:

[0070] Extract the coarse-grained domain generalization feature information through the static convolution module at the one-dimensional level. The expression is:

[0071] ,

[0072] where represents the vibration signal, and represent the convolution kernel and bias in the one-dimensional convolution operation respectively, and represent the convolution kernel and bias in the static convolution module respectively, represents the non-linear activation function, represents the coarse-grained features output by the static convolution module;

[0073] Extract the fine-grained domain generalization feature information through the dynamic convolution module at the one-dimensional level. The expression is:

[0074] ,

[0075] ,

[0076] ,

[0077] where and represent the convolution kernel and bias in the dynamic convolution module respectively, represents the fine-grained features output by the dynamic convolution module, are four convolution kernels with the same structure but different parameters included in the dynamic convolution module, represents the masking operation, represents the dynamic coefficient generated by the meta-regulator in the dynamic convolution module, represents the meta-regulator model parameters, and the symbol represents the structural reorganization of the convolution kernel;

[0078] Construct a dynamic domain generalization feature based on the coarse-grained domain generalization feature and the fine-grained domain generalization feature. The expression includes:

[0079] ,

[0080] where, represents the dynamic domain generalization feature;

[0081] In this embodiment, after constructing the dynamic domain generalization feature:

[0082] Perturb the training samples in the training dataset based on the dynamic coefficients generated by the meta-regulator in the dynamic convolution module, and randomly exchange the dynamic coefficients. The expression is:

[0083] ,

[0084] ,

[0085] where, and respectively represent two randomly selected training samples and and the sets of dynamic coefficients generated by the meta-regulator, represents the training dataset, represents the dynamic coefficient perturbation operation, and the symbol represents the exchange operation;

[0086] Calculate the loss of the dynamic domain generalization feature output by the same training sample before and after perturbation. The expression is:

[0087] ,

[0088] In the formula, is the perturbation loss between the dynamic domain generalization feature output by the same training sample before perturbation and the dynamic domain generalization feature output after perturbation, represents the total number of training samples, represents the th training sample before perturbation, and the dynamic coefficient at this time is , represents the th training sample after perturbation, and the dynamic coefficient at this time is , represents calculating the two-norm value;

[0089] Minimize the perturbation loss according to the backpropagation and optimization algorithms to extract discriminative fine-grained features between different training samples, where "discriminative" means that in this application, for the same sample, after being perturbed by the dynamic coefficients of different samples, the dynamic domain generalization features extracted by the dynamic domain generalization feature extractor are different but with small differences, so as to enhance the generalization of the model.

[0090] Step 4: Input the dynamic domain generalization features into the damage state discriminator to obtain multi-dimensional prediction probability data of multiple damage states, and determine the damage state evaluation result according to the multi-dimensional prediction probability data corresponding to multiple damage states;

[0091] Among them, the damage state discriminator includes a dynamic domain generalization feature classifier and a dynamic domain generalization feature domain discriminator, both of which belong to a common classifier in the neural network model and include a fully connected layer and a SoftMax function layer. When the dynamic domain generalization features are input into the damage state discriminator, the dynamic domain generalization feature classifier obtains multi-dimensional prediction probability data of the sample corresponding to multiple damage states, and the dynamic domain generalization feature domain discriminator obtains multi-dimensional prediction probability data of the sample corresponding to different domains. An explanation on how to "obtain multi-dimensional prediction probability data of multiple damage states according to the damage state discriminator": After extracting the dynamic domain generalization features corresponding to the sample, input them into the "dynamic domain generalization feature classifier " under the damage state discriminator, which includes a fully connected layer and a SoftMax function layer. The dynamic domain generalization features are processed by the fully connected layer to output low-dimensional features corresponding to the number of damage states, and are converted into probabilities through the SoftMax function layer, and finally multi-dimensional prediction probability data of multiple damage states are obtained. The damage state refers to the known damage type.

[0092] The high-speed rail gearbox damage state evaluation model further includes a dynamic coefficient discriminator. The training process of the high-speed rail gearbox damage state evaluation model includes:

[0093] Segment the training data under different working conditions without overlap according to the specified length standard to form independent subsets, assign training class labels to the independent subsets with known damage types, and assign training domain labels to the training data under different working conditions to obtain a training data set;

[0094] Among them, the training class label refers to the label given to different known damage types in the independent subset. For the same known damage type, its training class label is the same, while for different known damage types, their training class labels are different; the training domain label refers to the label given to the training data collected under different working conditions. For the training data collected under the same working condition, their training domain labels are the same, while for the training data collected under different working conditions, their training domain labels are different;

[0095] Input the training dataset into the dynamic domain generalization feature extractor to extract the coarse-grained domain generalization features and fine-grained domain generalization features in the training data, and construct dynamic domain generalization features based on the coarse-grained domain generalization features and fine-grained domain generalization features.

[0096] Based on the dynamic convolution module and the dynamic coefficient discriminator, establish a dynamic coefficient discriminant loss function to counteract and remove the domain information in the fine-grained domain generalization features. Based on the damage state discriminator and the dynamic domain generalization feature extractor, establish a dynamic domain generalization feature discriminant loss function to counteract and remove the domain information in the dynamic domain generalization features to complete the training.

[0097] Among them, the dynamic coefficient discriminator includes a dynamic coefficient classifier and a dynamic coefficient domain discriminator, and the dynamic coefficient discriminant loss function includes:

[0098] ,

[0099] In the formula, represents the total number of training samples with training class labels assigned in the training dataset, represents calculating the cross-entropy loss, represents the probability prediction function, represents the logarithmic function, represents the dynamic coefficient classifier, represents the dynamic coefficient of the i-th training sample, represents the training class label, represents the total number of training samples in the training dataset, represents the dynamic coefficient domain discriminator, represents the training domain label, is the dynamic coefficient discriminant loss;

[0100] Through the confrontation between the dynamic convolution module and the dynamic coefficient domain discriminator to remove the domain information extracted in the dynamic convolution module. Specifically, the dynamic convolution module generates dynamic coefficients, and the dynamic coefficient domain discriminator continuously learns to better distinguish the dynamic coefficient domain labels, that is, the training domain labels corresponding to the training samples. And according to the backpropagation and optimization algorithm, minimize the dynamic coefficient discriminant loss, so that the dynamic coefficients generated by the dynamic convolution module during the model optimization process do not contain domain information, so that the dynamic coefficient domain discriminator cannot correctly distinguish the dynamic coefficient domain labels, that is, the training domain labels corresponding to the training samples. Through the minimization process of the dynamic coefficient discriminant loss by the backpropagation and optimization algorithm, the mutual confrontation between the dynamic convolution module and the dynamic coefficient domain discriminator is realized, and the domain information extracted in the dynamic convolution module is removed.

[0101] Among them, the damage state discriminator includes a dynamic domain generalization feature classifier and a dynamic domain generalization feature domain discriminator, and the dynamic domain generalization feature discrimination loss function includes:

[0102] ,

[0103] In the formula, represents the total number of training samples assigned with training class labels in the training dataset, represents calculating the cross-entropy loss, represents the probability prediction function, represents the logarithmic function, represents the dynamic domain generalization feature classifier, represents the dynamic domain generalization feature of the i-th sample, represents the training class label, represents the total number of training samples in the training dataset, represents the dynamic domain generalization feature domain discriminator, represents the training domain label, is the dynamic domain generalization feature discrimination loss;

[0104] The dynamic domain generalization feature extractor and the dynamic domain generalization feature domain discriminator are antagonistic to remove the domain information in the dynamic domain generalization feature. Specifically, the dynamic domain generalization feature extractor extracts the dynamic domain generalization feature, while the dynamic domain generalization feature domain discriminator continuously learns to better distinguish the dynamic coefficient domain label, that is, the training domain label corresponding to the training sample. And according to the backpropagation and optimization algorithm, the dynamic domain generalization feature discrimination loss is minimized, so that the dynamic domain generalization feature extracted by the dynamic domain generalization feature extractor during the model optimization process does not contain domain information, thus making the dynamic domain generalization feature domain discriminator unable to correctly distinguish the dynamic coefficient domain label, that is, the training domain label corresponding to the training sample. Through the minimization process of the dynamic domain generalization feature discrimination loss by the backpropagation and optimization algorithm, the mutual antagonism between the dynamic domain generalization feature extractor and the dynamic domain generalization feature domain discriminator is realized, and the domain information extracted by the dynamic domain generalization feature extractor is removed.

[0105] Furthermore, the optimization algorithm adopts one of the adaptive moment estimation algorithm, the stochastic gradient descent algorithm, and the root mean square propagation algorithm, all of which belong to the prior art and will not be elaborated here.

[0106] The so-called backpropagation refers to:

[0107] Taking the gradient of the calculated loss, and then the optimization algorithm updates the network parameters according to the obtained gradient, which is expressed by the formula as follows:

[0108] ,

[0109] ,

[0110] ,

[0111] ,

[0112] ,

[0113] In the formula, represents the calculated gradient, , , and represent the network parameters of the dynamic domain generalization feature extractor, the dynamic coefficient classifier in the dynamic coefficient discriminator and the dynamic coefficient domain discriminator, and the damage state classifier and domain discriminator in the damage state discriminator, is a preset learning rate parameter, , are weight parameters.

[0114] The following further describes this embodiment in conjunction with the accompanying drawings and experimental cases:

[0115] 1. Experimental data

[0116] Taking the bearing damage data set collected on a self-made wheel set bearing failure test bench as an example, this test bench can simulate the structural relationship and motion relationship of a high-speed rail gearbox and collect vibration data under three different working conditions. There are 8 bearing health states: normal (N), inner race fault (IF), ball fault (BF), outer race fault (OF), inner race and ball compound fault (IBF), outer race and inner race compound fault (OIF), outer race and ball compound fault (OBF), outer race, inner race and ball compound fault (OIBF). As shown in Table 1, under different working conditions (B1, B2, B3, and B4), data for each health state is collected and non-overlappingly segmented into 100 samples, and the sample length is 4096 points.

[0117] Table 1 Sample information under different working conditions

[0118]

[0119] 2. Method verification

[0120] To verify the effectiveness and superiority of the proposed invention, this experimental case implemented damage state assessment tasks under 8 different domain generalization diagnosis tasks. Each domain generalization diagnosis task can be expressed as B1B3B4→B4, where the bold B1 represents a training set with damage state labels under one working condition, B3 and B4 represent training sets without damage state labels under the other two working conditions, and B2 represents a test set under another working condition. To verify the effectiveness of a dynamic domain generalization feature-driven high-speed rail gearbox damage state assessment method proposed by the present invention, comparisons were made with five other advanced diagnostic methods, including M1: the conventional domain adversarial neural network DANN, M2: replacing the residual block in M1 with the dynamic residual module in the dynamic domain generalization feature extractor disclosed in this application, M3: a new hybrid generalization network - the inner and outer domain generalization network IEDGNet, M4: an advanced causal disentanglement domain generalization method CDDG, M5: a contrast-assisted domain-specific removal network CDSRN, which extracts transferable features from the perspective of domain-specific removal. In the present invention, the initial learning rate is 0.001, the batch size is 20, and the training period is 200. Among them, to accelerate the model convergence speed, an exponential decay strategy is adopted.

[0121] Compared with other high-speed rail gearbox damage assessment methods, the recognition accuracy of the method of the present invention has achieved good results in most domain generalization diagnosis tasks, and the comparison results are shown in Table 2.

[0122] Table 2 Experimental results

[0123]

[0124] Example two:

[0125] The second embodiment of the present invention provides a high-speed rail gearbox damage state assessment device, including:

[0126] An acquisition module for acquiring the vibration signal of the high-speed rail gearbox to be identified;

[0127] An input module for inputting the vibration signal into a pre-trained high-speed rail gearbox damage state assessment model, where the high-speed rail gearbox damage state assessment model includes a dynamic domain generalization feature extractor and a damage state discriminator, and the dynamic domain generalization feature extractor includes a static convolution module and a dynamic convolution module;

[0128] An extraction module for respectively extracting the coarse-grained domain generalization feature and the fine-grained domain generalization feature in the vibration signal based on the high-speed rail gearbox damage state assessment model, and constructing a dynamic domain generalization feature according to the coarse-grained domain generalization feature and the fine-grained domain generalization feature;

[0129] An identification module, configured to input the dynamic domain generalization features into a damage state discriminator, obtain multi-dimensional prediction probability data for multiple damage states, and determine a damage state evaluation result based on the multi-dimensional prediction probability data corresponding to the multiple damage states.

[0130] The high-speed rail gearbox damage state evaluation device provided in the second embodiment of the present invention can execute the high-speed rail gearbox damage state evaluation method provided in the first embodiment of the present invention, and has corresponding functional modules and beneficial effects for executing the method.

[0131] Embodiment 3:

[0132] The third embodiment of the present invention further provides an electronic terminal, including a processor and a memory connected to the processor. A computer program is stored in the memory, and the processor is configured to operate according to the instructions to execute the steps of the method described in Embodiment 1.

[0133] The electronic terminal provided in the third embodiment of the present invention can execute the high-speed rail gearbox damage state evaluation method provided in the first embodiment of the present invention, and has corresponding functional modules and beneficial effects for executing the method.

[0134] Embodiment 4:

[0135] The fourth embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the method described in Embodiment 1, and has corresponding functional modules and beneficial effects for executing the method.

[0136] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a device, or a computer program product. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0137] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (apparatus), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate for implementing in the process Figure 1 one process or multiple processes and / or blocks Figure 1means for the functions specified in one or more boxes.

[0138] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including an instruction means that implements the functions specified in one Figure 1 process or more processes and / or boxes Figure 1 or more boxes.

[0139] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one Figure 1 process or more processes and / or boxes Figure 1 or more boxes.

[0140] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art of this technology, without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A method for assessing the damage status of a high-speed railway gearbox, characterized in that: include: Acquire a vibration signal of a high-speed rail gearbox to be identified; Inputting the vibration signal into a pre-trained high-speed railway gearbox damage state assessment model, wherein the high-speed railway gearbox damage state assessment model includes a dynamic domain generalized feature extractor and a damage state discriminator, and the dynamic domain generalized feature extractor includes a static convolution module and a dynamic convolution module; Based on the high-speed railway gearbox damage state assessment model, the coarse-grained domain generalization features and the fine-grained domain generalization features in the vibration signal are extracted according to the static convolution module and the dynamic convolution module respectively, and the dynamic domain generalization features are constructed according to the coarse-grained domain generalization features and the fine-grained domain generalization features; Inputting the dynamic domain generalized features into a damage state discriminator to obtain multidimensional prediction probability data of multiple damage states, and determining a damage state assessment result according to the multidimensional prediction probability data corresponding to the multiple damage states; According to the static convolution module and the dynamic convolution module, respectively extracting the coarse-grained domain generalization features and the fine-grained domain generalization features in the vibration signal, and constructing the dynamic domain generalization features according to the coarse-grained domain generalization features and the fine-grained domain generalization features includes: The static convolution module is used to extract the coarse-grained domain generalization feature information in one dimension, and the expression is: ， in, Indicates vibration signal, and Respectively represent the convolution kernel and bias in the one-dimensional convolution operation, and Represent the convolution kernel and bias in the static convolution module respectively, represents a nonlinear activation function, Represents the coarse-grained features output by the static convolution module; The dynamic convolution module is used to extract fine-grained domain generalization feature information at the one-dimensional level, and the expression is: ， ， ， in, and Represent the convolution kernel and bias in the dynamic convolution module respectively, represents the fine-grained features output by the dynamic convolution module, The dynamic convolution module includes four convolution kernels with the same structure but different parameters. Represents a mask operation, represents the dynamic coefficients generated by the metaregulator in the dynamic convolution module, represents the meta-regulator model parameter, symbol Represents the structural reorganization of the convolution kernel; According to the coarse-grained domain generalization feature and the fine-grained domain generalization feature, a dynamic domain generalization feature is constructed, and the expression includes: ， in, Represents dynamic domain generalization features.

2. A high-speed railway gearbox damage status assessment method according to claim 1, characterized in that: The high-speed railway gearbox damage state assessment model further includes a dynamic coefficient discriminator, and the training process of the high-speed railway gearbox damage state assessment model includes: The training data under different working conditions are divided into non-overlapping parts according to a specified length standard to form independent subsets, and training class labels are assigned to the independent subsets with known damage types, and training domain labels are assigned to the training data under the different working conditions to obtain a training data set; Inputting the training data set into a dynamic domain generalization feature extractor, extracting coarse-grained domain generalization features and fine-grained domain generalization features in the training data, and constructing dynamic domain generalization features according to the coarse-grained domain generalization features and the fine-grained domain generalization features; Based on the dynamic convolution module and the dynamic coefficient discriminator, a dynamic coefficient discrimination loss function is established to combat the removal of domain information in the fine-grained domain generalization feature. Based on the damage state discriminator and the dynamic domain generalization feature extractor, a dynamic domain generalization feature discrimination loss function is established to combat the removal of domain information in the dynamic domain generalization feature to complete the training.

3. A high-speed railway gearbox damage status assessment method according to claim 1, characterized in that: After constructing the dynamic domain generalization features: Based on the dynamic coefficients generated by the meta-regulator in the dynamic convolution module, the training samples in the training data set are perturbed and the dynamic coefficients are randomly exchanged. The expression is: ， ， in, and Represents two random training samples respectively and The dynamic coefficient set generated by the meta-regulator, represents the training data set, represents the dynamic coefficient perturbation operation, symbol Indicates a swap operation; Calculate the dynamic domain generalization feature loss of the same training sample output before and after perturbation, expressed as: ， In the formula, is the perturbation loss between the dynamic domain generalization features output before perturbation and the dynamic domain generalization features output after perturbation for the same training sample, represents the total number of training samples, Indicates the first The dynamic domain generalization feature of training samples, the dynamic coefficient is , After the disturbance The dynamic domain generalization feature of training samples, the dynamic coefficient is , Indicates the calculation of the bi-norm value; The perturbation loss is minimized according to the back-propagation and optimization algorithm to extract fine-grained features that are discriminative between different training samples.

4. A high-speed railway gearbox damage status assessment method according to claim 2, characterized in that: The dynamic coefficient discriminator includes a dynamic coefficient classifier and a dynamic coefficient domain discriminator, and the dynamic coefficient discrimination loss function includes: ， In the formula, represents the total number of training samples assigned training class labels in the training dataset, represents the calculation of cross entropy loss, represents the probability prediction function, represents the logarithmic function, represents the dynamic coefficient classifier, represents the dynamic coefficient of the i-th training sample, represents the training class label, represents the total number of training samples in the training data set, represents the dynamic coefficient domain discriminator, represents the training domain label, is the dynamic coefficient discrimination loss; The dynamic convolution module is used to counter the dynamic coefficient domain discriminator to remove the domain information extracted in the dynamic convolution module.

5. A high-speed railway gearbox damage status assessment method according to claim 2, characterized in that: The damage state discriminator includes a dynamic domain generalized feature classifier and a dynamic domain generalized feature domain discriminator, and the dynamic domain generalized feature discrimination loss function includes: ， In the formula, represents the total number of training samples assigned training class labels in the training dataset, represents the calculation of cross entropy loss, represents the probability prediction function, represents the logarithmic function, represents the dynamic domain generalization feature classifier, represents the dynamic domain generalization feature of the i-th sample, represents the training class label, represents the total number of training samples in the training data set, represents the dynamic domain generalization feature domain discriminator, represents the training domain label, It is the discriminative loss of dynamic domain generalization features; The domain information in the dynamic domain generalization feature is removed by the dynamic domain generalization feature extractor and the dynamic domain generalization feature domain discriminator.

6. A high-speed railway gearbox damage status assessment method according to claim 3, characterized in that: The optimization algorithm adopts one of an adaptive moment estimation algorithm, a stochastic gradient descent algorithm and a root mean square transfer algorithm.

7. A high-speed railway gearbox damage status assessment device, characterized in that: include: An acquisition module, used to obtain the vibration signal of the high-speed rail gearbox to be identified; An input module, used for inputting the vibration signal into a pre-trained high-speed railway gearbox damage state assessment model, wherein the high-speed railway gearbox damage state assessment model comprises a dynamic domain generalized feature extractor and a damage state discriminator, and the dynamic domain generalized feature extractor comprises a static convolution module and a dynamic convolution module; An extraction module is used to extract the coarse-grained domain generalization features and the fine-grained domain generalization features in the vibration signal based on the high-speed railway gearbox damage state assessment model according to the static convolution module and the dynamic convolution module, and construct the dynamic domain generalization features according to the coarse-grained domain generalization features and the fine-grained domain generalization features; An identification module, used for inputting the dynamic domain generalization feature into a damage state discriminator to obtain multi-dimensional prediction probability data of multiple damage states, and determining a damage state assessment result according to the multi-dimensional prediction probability data corresponding to the multiple damage states; According to the static convolution module and the dynamic convolution module, respectively extracting the coarse-grained domain generalization features and the fine-grained domain generalization features in the vibration signal, and constructing the dynamic domain generalization features according to the coarse-grained domain generalization features and the fine-grained domain generalization features includes: The static convolution module is used to extract the coarse-grained domain generalization feature information in one dimension, and the expression is: ， in, Indicates vibration signal, and Respectively represent the convolution kernel and bias in the one-dimensional convolution operation, and Represent the convolution kernel and bias in the static convolution module respectively, represents a nonlinear activation function, Represents the coarse-grained features output by the static convolution module; The dynamic convolution module is used to extract fine-grained domain generalization feature information at the one-dimensional level, and the expression is: ， ， ， in, and Represent the convolution kernel and bias in the dynamic convolution module respectively, represents the fine-grained features output by the dynamic convolution module, The dynamic convolution module includes four convolution kernels with the same structure but different parameters. Represents a mask operation, represents the dynamic coefficients generated by the metaregulator in the dynamic convolution module, represents the meta-regulator model parameter, symbol Represents the structural reorganization of the convolution kernel; According to the coarse-grained domain generalization feature and the fine-grained domain generalization feature, a dynamic domain generalization feature is constructed, and the expression includes: ， in, Represents dynamic domain generalization features.

8. An electronic terminal, characterized in that: The method comprises a processor and a memory connected to the processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 6 are executed.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images