A cross-working-condition planetary gearbox fault diagnosis method

Abstract

A large amount of labeled source domain fault data and some unlabeled target domain fault data are acquired, and the label space of the target domain fault data is a subset of the label space of the source domain fault data; wavelet packet transformation is performed on the collected data to obtain wavelet coefficients; a diagnosis model is constructed, including a feature extractor, a domain discriminator and a classifier, and the feature extractor is used to extract features of the source domain fault data and the target domain fault data respectively; the source domain fault data is subjected to weighted processing and then subjected to adversarial training with the target domain fault data through the domain discriminator, an adversarial loss is calculated, a classification loss is calculated through a classification result of the classifier, and parameters of the diagnosis model are optimized through the adversarial loss and the classification loss; data to be detected of the target domain are input into the optimized diagnosis model, and the diagnosis model outputs a fault diagnosis result; compared with existing planetary gear box fault diagnosis methods, the planetary gear box cross-condition fault diagnosis can be realized under the condition that the number of fault categories of the target domain is unknown.

Description

Technical Field

[0001] The present invention relates to the technical field of simulation analysis, and in particular to a cross-operating-condition planetary gearbox fault diagnosis method. Background Art

[0002] Gearboxes are widely used, playing a crucial role in aerospace, power generation, transportation, and other fields. However, due to their complex structure and operating conditions, gearboxes are prone to failures such as tooth fatigue pitting, tooth root cracks, and even tooth or shaft breakage. These failures can affect the equipment's operating accuracy, transmission efficiency, and service life, resulting in economic losses for users. Therefore, gearbox fault diagnosis is crucial.

[0003] Common gearbox failures include gear damage, bearing damage, shaft breakage, oil leakage, and high oil temperatures. Bearings are the most critical components in a gearbox, and their failure often leads to catastrophic damage. Furthermore, during operation, the bearings are subjected to repeated alternating loads between the rings and rolling element surfaces. Due to issues with installation, lubrication, and maintenance, defects such as pitting, cracks, and surface spalling can develop, leading to bearing failure and consequent damage to the gear pair and gearbox.

[0004] Currently, train component fault inspection and maintenance primarily relies on a "judgment by experience + manual intervention" model. Specifically, operational personnel develop business rules for component failures based on their experience, and maintenance personnel conduct troubleshooting and repair of reported faulty components. However, this maintenance approach fails to effectively and effectively detect gearboxes online, preventing early identification of gearbox failures. This leads to delayed repairs and increased operating costs. Furthermore, when using machine learning for fault detection, data on some target faults is limited, making it difficult to train the machine learning model on these target types. Summary of the Invention

[0005] In response to the above problems, the present invention proposes a cross-operating-condition planetary gearbox fault diagnosis method, which specifically includes the following steps:

[0006] S1. Obtain a large amount of labeled source domain fault data and some unlabeled target domain fault data, where the label space of the target domain fault data is a subset of the label space of the source domain fault data.

[0007] S2, performing wavelet packet transform on the collected data, obtaining wavelet packet coefficients and using them as input of the diagnostic model;

[0008] S3. Construct a diagnostic model, including a feature extractor, a domain discriminator, and a classifier. Use the feature extractor to extract features of the source domain fault data and the target domain fault data respectively. After weighting the source domain fault data, perform adversarial training on it with the target domain fault data through the domain discriminator, calculate the adversarial loss, and calculate the classification loss based on the classification results of the classifier. Optimize the parameters of the diagnostic model using the adversarial loss and classification loss.

[0009] S4. Perform wavelet packet transform on the data to be detected in the target domain, input the wavelet packet coefficients into the optimized diagnosis model, and the diagnosis model outputs the fault diagnosis result.

[0010] Furthermore, weighted processing of the fault data in the source domain includes:

[0011]

[0012] Among them, K s Indicates the total number of fault categories in the source domain; n s Indicates the number of source domain samples; n t Indicates the number of samples in the target domain; n k Indicates the number of pseudo labels of the kth category of target domain samples; Indicates the weight of the k-th category of the source domain obtained during the i-th mini-batch during training; Indicates the i-th mini-batch, used to calculate the weight The intermediate variable of ; α is a hyperparameter that controls the weight update rate.

[0013] Furthermore, the number of pseudo labels of the kth category of the target domain sample n k Expressed as:

[0014]

[0015] in, Indicates when When it is established, its value is 1, otherwise it is 0; represents the class prototype of the source domain; k represents the kth fault category of the source domain.

[0016] Furthermore, in the i-th mini-batch, the calculation process of the class prototype of the source domain includes:

[0017]

[0018] Among them, x k represents the samples belonging to the kth class in the source domain; |S k | represents the k-th sample set S in the source domain k The number of samples in G(x k ) represents the sample xk Features extracted by feature extractor G; Represents the average value of the current k-th category feature; α is a hyperparameter that controls the update rate; Represents the prototype of the kth class calculated in the i-th mini-batch.

[0019] Furthermore, the calculation process of adversarial loss includes:

[0020]

[0021] Among them, L d (G,D) denotes adversarial loss, G denotes feature extractor, and D denotes domain discriminator; represents the weight used when weighting the fault data in the source domain; represents the k-th sample in the source domain; x t represents the target domain sample.

[0022] Furthermore, the calculation process of classification loss is expressed as:

[0023]

[0024] Among them, L c (G, F) represents classification loss; E represents expectation; x, y~p s (x,y) represents the source domain sample x and label y with probability p s The joint distribution of (x, y), G represents the feature extractor, F represents the classifier; K s Represents the total number of fault categories in the source domain; 1{k=y} means that the result is 1 only when k=y, and the result is 0 in other cases.

[0025] Furthermore, the feature extractor includes three cascaded convolution modules, the last convolution module is followed by two fully connected layers in sequence, and each convolution module includes a convolution layer, a batch normalization layer, an activation function, and a pooling layer in sequence.

[0026] Furthermore, the domain discriminator includes three cascaded fully connected layers. The features extracted by the feature extractor are weighted and used as the input of the domain discriminator. The domain discriminator identifies whether the weighted features belong to the target domain or the source domain, and the adversarial loss is back-propagated on the feature extractors that extract source domain fault data and target domain fault data to optimize the network parameters of the feature extractors that extract source domain fault data and target domain fault data.

[0027] Furthermore, the classifier includes a fully connected layer that weights the features extracted by the feature extractor and uses them as input. The classifier then categorizes the fault data. The classification loss is only backpropagated through the feature extractor that extracts the source domain fault data, optimizing the network parameters of the feature extractor. Compared to existing planetary gearbox fault diagnosis methods, this invention can perform cross-operating-condition planetary gearbox fault diagnosis even when the number of target domain fault categories is unknown.

[0028] Compared with the existing planetary gearbox fault diagnosis method, the present invention can realize cross-operating condition fault diagnosis of planetary gearbox when the number of fault categories in the target domain is unknown. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A schematic diagram of a fault diagnosis method disclosed in the present invention;

[0030] Figure 2 Schematic diagram of the neural network used in the present invention. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0032] The present invention provides a cross-operating-condition planetary gearbox fault diagnosis method, which specifically includes the following steps:

[0033] S1. Obtain a large amount of labeled source domain fault data and some unlabeled target domain fault data, where the label space of the target domain fault data is a subset of the label space of the source domain fault data.

[0034] S2, performing wavelet packet transform on the collected data and using the wavelet packet coefficients as input of the diagnostic model;

[0035] S3. Construct a diagnostic model, including a feature extractor, a domain discriminator, and a classifier. Use the feature extractor to extract features of the source domain fault data and the target domain fault data respectively. After weighting the source domain fault data, perform adversarial training with the target domain fault data using the domain discriminator, calculate the adversarial loss, and calculate the classification loss based on the classification results of the classifier. Optimize the parameters of the diagnostic model using the adversarial loss and classification loss.

[0036] S4. Perform wavelet packet transform on the data to be detected in the target domain, input the wavelet coefficients into the optimized diagnosis model, and the diagnosis model outputs the fault diagnosis result.

[0037] The present invention includes two feature extractors: one for extracting features from source-domain fault data and the other for extracting features from target-domain fault data. The feature extractor employed in this invention comprises three cascaded convolutional modules, with the last convolutional module followed by two fully connected layers. Each convolutional module sequentially comprises a convolutional layer, a batch normalization layer, an activation function, and a pooling layer. Table 1 provides the feature extractor parameters for this embodiment.

[0038] The domain discriminator employed in this invention comprises three cascaded fully connected layers. The features extracted by the feature extractor are weighted and used as input to the domain discriminator. The domain discriminator identifies whether the weighted features belong to the target domain or the source domain. An adversarial loss is then applied to the feature extractors that extract source and target domain fault data, backpropagating the network parameters of the feature extractors to optimize the network parameters. Table 1 provides the domain discriminator parameters for this embodiment.

[0039] The classifier used in this invention includes a fully connected layer. The features extracted by the feature extractor are weighted and used as input to the classifier. The classifier classifies the fault data type, and the classification loss is only backpropagated to the feature extractor that extracts the source domain fault data, optimizing the network parameters of the feature extractor that extracts the source domain fault data. This embodiment provides the classifier parameters shown in Table 1.

[0040] In this application, it is necessary to ensure that the fault type of the source domain fault data is greater than the fault type of the target domain fault data. The diagnostic model needs to be retrained for different target domain fault data, but the labels of the target domain fault data do not need to be collected during the training process.

[0041] In this embodiment, under one working condition, a large amount of labeled fault data is obtained as source domain samples; under another working condition, some unlabeled fault data is obtained as target domain samples; in particular, the number of source domain fault categories is greater than the number of target domain fault categories.

[0042] For the collected data, wavelet packet transform is performed every 4096 points, and 64×64 wavelet packet coefficients are obtained as the input of the diagnosis model.

[0043] The training set data of the source domain and the target domain are input into the network of the present invention, the source domain fault data is weighted, and adversarial training is performed with the target domain data. The adversarial loss and classification loss are calculated, and the network parameters are optimized. In the specific implementation, the neural network parameters used are shown in Table 1.

[0044] Table 1 Neural network parameters

[0045]

[0046] During training, the weight calculation of the source domain depends on the source domain fault class prototypes and the pseudo labels of the target domain data. The source domain class prototype calculation formula is:

[0047]

[0048] Among them, x i 、y i represents the samples belonging to the kth class in the source domain; |S k | represents the k-th sample set S in the source domain k The number of samples in G(x i ) represents the sample x i Features extracted by feature extractor G; Represents the average value of the current k-th category feature; α is a hyperparameter that controls the update rate; Represents the prototype of the kth class calculated at the nth iteration.

[0049] The target domain pseudo label calculation formula is:

[0050]

[0051] in, Indicates when When it is established, its value is 1, otherwise it is 0; Represents the class prototype of the source domain; k represents the kth class of the source domain.

[0052] The process of calculating weights includes:

[0053]

[0054] Among them, K s Indicates the total number of fault categories in the source domain; n s Indicates the number of samples in the source domain; n t Indicates the number of samples in the target domain; n k represents the number of pseudo labels of the kth category of the target domain sample; α is a hyperparameter that controls the weight update rate; Indicates the weight of the k-th category of the source domain obtained during the i-th mini-batch during training; Indicates the i-th mini-batch, used to calculate the weight Intermediate variables such as Figure 2 , this embodiment calculates In the process, k mini-batches are used, each mini-batch corresponds to a fault type in the source domain fault data, the prototype of each category is found from the source domain fault data and the number of fault data of each category is obtained, the momentum update of each class is performed through the mini-batch to obtain the weight corresponding to each class, the feature vector extracted by the feature extractor is weighted by the weight and then input into the domain discriminator and classifier for domain identification and classification respectively, during the training process, the parameters of the domain discriminator, classifier and feature extractor are updated by the loss function until the value of the loss function is less than the set threshold or the maximum number of training times is reached, where the training goal of the classifier is to enable the classifier to be able to The category of source domain fault data is identified within the range, and the classifier and source domain feature extractor are optimized during the back propagation of the classification loss; the domain discriminator is used to distinguish source domain fault data from target domain fault data. The training goal is to make the discriminator unable to distinguish whether the fault data comes from the target domain or the source domain. The parameters of the domain discriminator and the feature extractors of the target domain and the source domain are optimized during the back propagation process. After the optimization is completed, the target domain feature extractor extracts the features of the target domain fault data and inputs them into the classifier for fault type identification. The label of the target domain fault data is not required in this application. Even if the fault type of the target domain fault data is unclear, the present application can also be used to predict the fault type of the target data.

[0055] By calculating the adversarial loss and classification loss, the sum of these two losses is back-propagated to optimize the parameters of the diagnostic model until the two losses are minimized. In actual application, a threshold can be set. When the loss is less than the set threshold, the iteration can be stopped to complete the parameter optimization. The process of calculating the adversarial loss includes:

[0056]

[0057] Among them, L d (G,D) denotes adversarial loss, G denotes feature extractor, and D denotes domain discriminator; represents the weight used when weighting the fault data in the source domain; represents the k-th sample in the source domain; x t represents the target domain sample.

[0058] The classification loss is:

[0059]

[0060] Among them, L c (G, F) represents classification loss; E represents expectation; x, y~p s (x,y) represents the sample-label pair (x,y) of the source domain, and its obedience probability is p sThe joint distribution of (x, y), G represents the feature extractor, F represents the classifier; K s Represents the total number of fault categories in the source domain; 1{k=y} means that the result is 1 only when k=y, and the result is 0 in other cases.

[0061] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A cross-operating-condition planetary gearbox fault diagnosis method, characterized in that: The specific steps include: S1. Obtain a large amount of labeled source domain fault data and some unlabeled target domain fault data, where the label space of the target domain fault data is a subset of the label space of the source domain fault data. S2, performing wavelet packet transform on the collected data to obtain wavelet packet coefficients, and using the wavelet packet coefficients as input of the diagnostic model; S3. Construct a diagnostic model, including a feature extractor, a domain discriminator, and a classifier. The feature extractor is used to extract features of the source domain fault data and the target domain fault data respectively. After weighted processing of the source domain fault data, adversarial training is performed with the target domain fault data through the domain discriminator to calculate the adversarial loss. The classification loss is calculated based on the classification results of the classifier. The parameters of the diagnosis model are optimized based on the adversarial loss and the classification loss. The process of weighted processing of the source domain fault data includes: Among them, K s Indicates the total number of fault categories in the source domain; n s Indicates the number of source domain samples; n t Indicates the number of samples in the target domain; n k Indicates the number of pseudo labels of the kth category of target domain samples; Indicates the weight of the k-th category of the source domain obtained during the i-th mini-batch during training; Indicates the i-th mini-batch, used to calculate the weight The intermediate variable of ; α is the hyperparameter that controls the weight update rate; The number of pseudo labels n for the kth category of target domain samples k Expressed as: in, Indicates when When it is established, its value is 1, otherwise it is 0; represents the class prototype of the source domain; k represents the kth fault category of the source domain; In the i-th mini-batch, the calculation process of the class prototype of the source domain includes: Among them, x k represents the samples belonging to the kth class in the source domain; |S k | represents the k-th sample set S in the source domain k The number of samples in G(x k ) represents the sample x k Features extracted by feature extractor G; Represents the average value of the current k-th category feature; α is a hyperparameter that controls the update rate; Represents the prototype of the kth class calculated in the i-th mini-batch; S4. Perform wavelet packet transform on the data to be detected in the target domain, input the wavelet coefficients into the optimized diagnosis model, and the diagnosis model outputs the fault diagnosis result.

2. A cross-operating-condition planetary gearbox fault diagnosis method according to claim 1, characterized in that: The calculation process of adversarial loss includes: Among them, L d (G,D) denotes adversarial loss, G denotes feature extractor, and D denotes domain discriminator; represents the weight used when weighting the fault data in the source domain; represents the k-th sample in the source domain; x t represents the target domain sample.

3. The cross-operating-condition planetary gearbox fault diagnosis method according to claim 1, characterized in that: The calculation process of classification loss is expressed as: Among them, L c (G,F) represents classification loss; Ε represents expectation; x, y~p s (x,y) represents the source domain sample x and label y with probability p s The joint distribution of (x, y), G represents the feature extractor, F represents the classifier; K s Represents the total number of fault categories in the source domain; 1{k=y} means that the result is 1 only when k=y, and the result is 0 in other cases.

4. A cross-operating-condition planetary gearbox fault diagnosis method according to claim 1, 2 or 3, characterized in that: The feature extractor consists of three cascaded convolutional modules. The last convolutional module is followed by two fully connected layers. Each convolutional module consists of a convolutional layer, a batch normalization layer, an activation function, and a pooling layer.

5. A cross-operating-condition planetary gearbox fault diagnosis method according to claim 1 or 2, characterized in that: The domain discriminator consists of three cascaded fully connected layers. The features extracted by the feature extractor are weighted and used as the input of the domain discriminator. The domain discriminator identifies whether the weighted features belong to the target domain or the source domain. The adversarial loss is back-propagated on the feature extractors that extract source domain fault data and target domain fault data, and the network parameters of the feature extractors that extract source domain fault data and target domain fault data are optimized.

6. A cross-operating-condition planetary gearbox fault diagnosis method according to claim 1 or 3, characterized in that: The classifier includes a fully connected layer. The features extracted by the feature extractor are weighted and used as the input of the classifier. The classifier classifies the type of fault data, and the classification loss is only backpropagated to the feature extractor that extracts the source domain fault data, optimizing the network parameters of the feature extractor that extracts the source domain fault data.

Images