Cross-domain fault diagnosis method for engine bearing, electronic equipment and storage medium

Abstract

The invention discloses an engine bearing cross-domain fault diagnosis method, electronic equipment and a storage medium. The diagnosis method comprises the steps that historical vibration signal data sets of the engine bearing under different working conditions are constructed, and vibration signal data under each working condition are divided into a plurality of source domains and a target domain; based on the historical vibration signal data set, training the transfer learning neural network to obtain an engine bearing cross-domain fault diagnosis model; the transfer learning neural network comprises a shared feature extractor, a source specific feature extractor, a source specific classifier and a task classifier; and inputting the historical vibration signal data of the engine bearing under different working conditions into the engine bearing cross-domain fault diagnosis model to generate a diagnosis result. According to the method, the cross-domain diagnosis accuracy on a disclosed bearing data set is improved by more than 20%, and the method is particularly outstanding in noise interference and small sample scenes.

Description

Technical Field

[0001] The present invention belongs to the technical field of engine bearings, and particularly relates to a cross-domain fault diagnosis method, an electronic device, and a storage medium for engine bearings. Background Art

[0002] Early fault diagnosis of rotating machinery is crucial for ensuring the reliability of industrial systems. For example, as a high-load key component, the fault characteristics of an aeroengine bearing are easily affected by factors such as working condition fluctuations, sensor heterogeneity, and equipment aging, resulting in a significant decline in the performance of traditional fault diagnosis models in cross-domain scenarios. Although existing unsupervised domain adaptation (UDA)-based methods can alleviate the problem of data bias in a single domain, there are still two major bottlenecks:

[0003] First, the limitation of single-source domain knowledge transfer. Traditional methods rely on a single source domain for knowledge transfer, and the data distribution of a single source domain is difficult to cover the complex actual scenarios of the target domain (such as multiple working conditions and multiple equipment types), resulting in insufficient transferred knowledge and limited discriminative ability of the model in the target domain.

[0004] Second, the coarse-grained defect of cross-domain distribution alignment. Existing methods mostly use metrics such as maximum mean discrepancy (MMD) for global domain alignment, but their mechanism only focuses on first-order statistics (such as mean difference), ignoring higher-order statistical characteristics (such as variance difference) and fine-grained sub-domain distribution relationships (such as local offsets of similar fault characteristics), resulting in insufficient accuracy of inter-domain feature matching. For example, similar faults may exhibit distribution variance offsets in different domains due to vibration characteristic differences, and traditional alignment methods cannot effectively capture such dynamic changes, restricting the cross-domain transfer efficiency of diagnostic knowledge. Summary of the Invention

[0005] In view of the above deficiencies of the prior art, the present invention provides a cross-domain fault diagnosis method, an electronic device, and a storage medium for engine bearings.

[0006] The object of the present invention is achieved by the following technical solutions:

[0007] In a first aspect, the present invention provides a cross-domain fault diagnosis method for engine bearings, including the following steps:

[0008] Construct a historical vibration signal data set of engine bearings under different working conditions, where the historical vibration signal data set includes vibration signal data of normal bearings, inner ring fault bearings, and outer ring fault bearings. Among them, the vibration signal data under each working condition is divided into multiple source domains and one target domain;

[0009] Based on the historical vibration signal data set, train a transfer learning neural network to obtain a cross-domain fault diagnosis model for engine bearings;

[0010] Input the historical vibration signal data of the engine bearing under different working conditions into the cross-domain fault diagnosis model of the engine bearing to generate a diagnosis result.

[0011] In some embodiments, before training the transfer learning neural network, it is necessary to preprocess the vibration signal data in the historical vibration signal dataset, including:

[0012] Normalize the vibration signal data in the historical vibration signal dataset;

[0013] Segment the vibration signal data after normalization to generate multiple time series samples, and each time series sample includes vibration signal data at a unified number of time points.

[0014] In some embodiments, randomly select 70% of the samples from each source domain in the historical vibration signal dataset as the training set, and the remaining 30% as the test set.

[0015] In some embodiments, based on the historical vibration signal dataset, train the transfer learning neural network to obtain the cross-domain fault diagnosis model of the engine bearing, including:

[0016] First, based on the training data in the historical vibration signal dataset, train the shared feature extractor to extract cross-domain general features;

[0017] Then use the source-specific feature extractor to align the general features of each pair of source domain and target domain to reduce the domain difference;

[0018] Next, use the source-specific classifier to perform classification training on each source domain to optimize the parameters of the source-specific classifier;

[0019] Finally, integrate the results of multiple source-specific classifiers through the task classifier to obtain the cross-domain fault diagnosis model of the engine bearing;

[0020] Among them, the transfer learning neural network includes the shared feature extractor, the source-specific feature extractor, the source-specific classifier, and the task classifier.

[0021] In some embodiments, it further includes: evaluating the cross-domain fault diagnosis model of the engine bearing based on the test data in the historical vibration signal dataset.

[0022] In some embodiments, the different working conditions include different rotational speeds, loads, and environmental conditions.

[0023] In a second aspect, the present invention provides an electronic device, including a memory, a processor, and a computer program stored on the memory and running on the processor. When the processor executes the computer program, the engine bearing cross-domain fault diagnosis method according to any one of the above first aspects is implemented.

[0024] In a third aspect, the present invention provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the engine bearing cross-domain fault diagnosis method according to any one of the above first aspects is implemented.

[0025] The beneficial effects of the present invention are as follows: By constructing a multi-branch parallel network model, the sub-domain distributions of each source domain and the target domain are respectively aligned, and the variance difference representation (VDR) is introduced to enhance the fine-grained alignment ability of the same type of fault features. Further combined with the source domain weighted fusion mechanism, the collaborative migration of multi-source knowledge and high-precision diagnosis are realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0027] Figure 1 It is a flowchart of the engine bearing cross-domain fault diagnosis method provided for an exemplary embodiment;

[0028] Figure 2 It is a schematic structural diagram of the engine bearing cross-domain fault diagnosis model provided for an exemplary embodiment;

[0029] Figure 3 It is a schematic structural diagram of the electronic device provided for an exemplary embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] In order to better understand the technical solutions of the present application, the following will describe the embodiments of the present application in detail with reference to the drawings.

[0031] It should be clear that the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0032] As Figure 1 shown in FIG. -3, the present invention proposes an engine bearing cross-domain fault diagnosis method, including the following steps:

[0033] Step S1: Collect historical vibration signal data of an aero-engine bearing under different working conditions through an acceleration sensor, and construct a historical vibration signal dataset of the engine bearing under different working conditions. The historical vibration signal dataset includes vibration signal data of normal bearings, inner-race fault bearings, and outer-race fault bearings. Among them, the vibration signal data under each working condition is divided into multiple source domains and one target domain. The source domains are used to train the model, and the target domain is used to test the generalization ability of the model.

[0034] In some embodiments, to ensure the diversity and representativeness of the vibration signal data, the collected vibration signal data covers a variety of working conditions, including different rotational speeds, loads, and environmental conditions.

[0035] In some embodiments, before training the transfer learning neural network, it is necessary to preprocess the vibration signal data in the historical vibration signal dataset, including: first, normalize the vibration signal data in the historical vibration signal dataset to eliminate the dimensional differences brought by different sensors and working conditions; then use the sliding time window method with a size of 1024 to segment the normalized vibration signal data to generate multiple time series samples, and each time series sample includes vibration signal data at a unified number (for example, the number is 1024) of time points.

[0036] In some embodiments, the normalization process of the vibration signal data in the historical vibration signal dataset first includes:

[0037] Step 1: Assume that there are i parameters in the collected message data, and each parameter contains j data. Determine the maximum value x of each parameter i,max and the minimum value x i,min .

[0038] Step 2: Calculate the normalized parameter value where x i,j represents the j-th data value among the i parameters.

[0039] In some embodiments, during the preprocessing of the vibration signal data, the data segmentation and sample generation include:

[0040] Step 1: Window sliding: Starting from the first time point, take 1024 consecutive time points as a sample, then slide backward by one time point, and take the next 1024 time points as the next sample, and so on until the entire time series is traversed;

[0041] Step 2: Sample generation: Through the sliding window method, the signals in each source domain and target domain are segmented into multiple samples; each sample contains vibration signal data at 1024 time points and serves as the input to the transfer learning neural network.

[0042] In some embodiments, to ensure the generalization ability of the model, 70% of the samples are randomly selected from each source domain in the historical vibration signal dataset as the training set, and the remaining 30% are used as the test set. This division method ensures the data distribution consistency between the training set and the test set and avoids the overfitting problem.

[0043] Step S2: Based on the historical vibration signal dataset, train the transfer learning neural network to obtain an engine bearing cross-domain fault diagnosis model. The transfer learning neural network includes a shared feature extractor, a source-specific feature extractor, a source-specific classifier, and a task classifier.

[0044] (1) Shared feature extractor: Its function is to project all the original signals from different source domains and the target domain into a common feature space and mine the invariant features between different domains. This shared feature extractor is based on a deep convolutional network architecture and includes multiple convolutional layers, batch normalization layers, ReLU activation functions, and max pooling layers. Through the shared feature extractor, the model can extract cross-domain general features from multiple source domains, laying a foundation for subsequent feature alignment and classification tasks. The specific calculation formula for each convolutional layer is as follows.

[0045] ReLU(x) = max(0, x)

[0046] C i,k = ReLU(X i,k ·W i,k + b i,k )

[0047]

[0048] P i,k = max(B i,k )

[0049] D i,k = Dropout(P i,k , dropout_rate)

[0050] In the formula, ReLU is the activation function, and x is the input value of the ReLU activation function. X i,k is the input of the k-th layer of the i-th module and is also the output of the (k - 1)-th layer (C i,k-1 ) of the i-th module. W i,k is the convolution kernel of the k-th layer of the i-th module. b i,k is the bias term of the k-th layer of the i-th module. C i,k is the output of the convolution layer of the k-th layer of the i-th module. B i,k is the output of the batch normalization layer of the k-th layer of the i-th module. μ i,k and σ i,kThey are the mean and the standard deviation respectively. γ i,k and β i,k are learnable parameters, where γ i,k is the scaling coefficient and β i,k is the offset coefficient. P i,k represents max pooling. D i,k represents the dropout layer, which discards elements in max pooling with the probability of dropout_rate. The entire convolutional layer structure process follows: input → convolution + ReLU → batch normalization → pooling → Dropout → output.

[0051] Integrate the diverse features generated by the shared feature extractor into a unified representation F, providing a robust and comprehensive feature input for different downstream tasks in transfer learning. The formula is as follows.

[0052] A i = max(D i,1 , D i,2 , D i,3 , D i,4 , D i,5 )

[0053] F i = Flatten(A i )

[0054] F = Concatenate(F1, F2,..., F N )

[0055] In the formula, A i is the output of the adaptive max pooling layer of the i-th module, taking the maximum value of each position of each feature vector in D i,k , and obtaining F i by flattening A i into a one-dimensional vector. Finally, all F i are combined to obtain the final output result F.

[0056] (2) Source-specific feature extractor: To avoid the complexity of simultaneously aligning multiple source domains and the target domain, a source-specific feature extractor is designed. This extractor maps each pair of source domain and target domain into a specific feature space and reduces the feature distribution difference between the source domain and the target domain through variance difference representation (VDR). By capturing the variance information in the vibration signal, VDR can more accurately reflect the distribution difference between different domains, thereby improving the model's adaptability to target domain data. The domain adaptation loss function of variance difference representation (VDR) is: In the formula, represents the input sample of the j-th source domain, x t the input sample of the target domain. Among them, G jis the specific feature extractor for the j-th source domain. is to calculate the estimated value of the variance difference representation (VDR). The VDR values of all source domains are summed to obtain the total domain adaptation loss l VDR .

[0057] (3) Source-specific classifier: It is a softmax logistic regression module that receives the output of the source-specific feature extractor and outputs the probability distribution of each sample belonging to different classes. Each source domain has an independent classifier. By minimizing the classification loss function, the model parameters are optimized to improve the prediction accuracy of source domain samples. The sum of the multi-source classification losses can be calculated as: In the formula, represents the true label of the samples in the j-th source domain. C j is the j-th source-specific classifier. is to calculate the cross-entropy loss between the predicted distribution and the true label, and sum the losses of all source domains to obtain l cls . The formula for the total loss is as follows: l total = l cls + λl VDR . Among them, λ is the trade-off parameter between the two objectives. Use the formula to gradually change from 0 to 1 to reduce the influence of parameter sensitivity. p represents the linear change of the training progress from 0 to 1. The network learning rate is set to 0.01, and 15 epochs are performed for one training.

[0058] (4) Task classifier: Its role is to calculate the weight score of the source-specific classifier according to the distribution distance between each pair of source domains and the target domain as and integrate the results of multiple source-specific classifiers to obtain the final diagnosis result In this way, the model can integrate the knowledge of multiple source domains and improve the reliability of the diagnosis result.

[0059] In some embodiments, based on the historical vibration signal dataset, the transfer learning neural network is trained to obtain an engine bearing cross-domain fault diagnosis model, including: first, based on the training data in the historical vibration signal dataset, the shared feature extractor is trained to extract cross-domain general features; then, the source-specific feature extractor is used to align the general features of each pair of source domains and the target domain to reduce the domain difference; then, the source-specific classifier is used to perform classification training on each source domain to optimize the parameters of the source-specific classifier; finally, the results of multiple source-specific classifiers are integrated by the task classifier to obtain the engine bearing cross-domain fault diagnosis model. During the training process, Dropout and Early Stopping techniques are used to prevent overfitting and ensure the generalization ability of the model; specifically as follows:

[0060] Step 1: Model Initial Setup. Initialize the shared feature generator (G_shared) with the input channel number (in_channel = 1) and output dimension (output_size = 2560). Initialize the multi-source classifier (Cs) which contains self.num_source MLP classifiers. The ClassifierMLP has a hidden layer (input_size = 2560 → output_size = args.num_classes). Use an optimizer (defined by _get_optimizer, typically an Adam optimizer with learning rate lr = 0.001 and weight decay weight_decay = 1e-4) and a learning rate scheduler (defined by _get_lr_scheduler, which is StepLR with learning rate lr = 0.001 and a 50% decay every 30 epochs). Set the maximum number of training iterations (args.max_epoch = 100) and configure the dynamic trade-off coefficient (tradeoff = self._get_tradeoff(args.tradeoff, epoch)), with a linear growth strategy that gradually increases from the initial value of 0.1 to 1 for multi-task balance in the loss function. Enable the model training mode (train()) and create a storage directory (T-sne) for feature visualization.

[0061] Step 2: Source Domain and Target Domain Data Loading. Extract the target domain data (target_data) and source domain data (source_data) in batches from the data loader, supporting the multi-source domain mixed training mode (source_combine). Concatenate the multi-source domain data and the target domain data (torch.cat) and input them into the shared feature generator for feature extraction.

[0062] Step 3: Forward Propagation and Feature Extraction. Extract the common features (f_shared) of the multi-source domain and the target domain through the shared feature generator (G_shared) and split the features according to the number of source domains (chunk). Save the extracted features to a file at a specified iteration period (e.g., epoch = 30) for subsequent visualization analysis.

[0063] Step 4: Loss Function Calculation.

[0064] 1) Classification Loss: Calculate the cross-entropy loss between the outputs (y_s) of each source domain classifier and the true labels:

[0065] where represents the true label of the j-th source domain sample, represents the input sample of the j-th source domain sample, F is the shared feature extractor, G jis the specific feature extractor for the j-th source domain, C j is the j-th source-specific classifier.

[0066] 2) Domain adaptation loss: Calculate the difference in feature distributions between the source domain and the target domain through VDR (Domain Difference Regularization): In the formula, represents the input sample of the j-th source domain, x t the input sample of the target domain. Among them, G j is the specific feature extractor for the j-th source domain.

[0067] is the estimated value of calculating the variance difference representation (VDR). Sum the VDR values of all source domains to obtain the total domain adaptation loss l VDR

[0068] 3) Total loss: Combine the classification loss and the domain adaptation loss with weights: L total = L cls + 100·tradeoff·L VDR

[0069] Step 5: Backpropagation and parameter update. Set the maximum gradient norm grad_clip = 5.0 (to prevent gradient explosion), momentum parameters: beta1 = 0.9, beta2 = 0.999, update once for each mini-batch (batch_size = 64), implement L2 regularization through the optimizer's weight_decay = 1e-4, clear the gradients (optimizer.zero_grad()), calculate the gradients through backpropagation (loss.backward()), and update the parameters by the optimizer (optimizer.step()). Statistically calculate the classification accuracy (epoch_acc) and loss components (epoch_loss) for each source domain.

[0070] Step 6: Dynamically adjust the learning rate and monitor the training status. Use ReduceLROnPlateau based on the validation set accuracy (decay when the accuracy stagnates). After each epoch, adjust the learning rate according to the learning rate scheduler (lr_scheduler.step()), the minimum learning rate = 1e-6, and terminate the training in advance when the learning rate is lower than 1e-6. Output the training loss and accuracy of the current epoch, and verify the model performance (self.test()).

[0071] Step 7: Record the best accuracy (best_acc) on the validation set and its corresponding epoch (best_epoch). If the performance of the current epoch is better than the historical best, update the best model state.

[0072] Step 8: Iteratively train until the termination condition is met. Repeat Steps 2 to 7 until the maximum number of training epochs (args.max_epoch) is reached. Finally, output the cross-domain fault diagnosis model of the engine bearing with the optimal domain adaptation ability.

[0073] In some embodiments, it further includes: evaluating the cross-domain fault diagnosis model of the engine bearing based on the test data in the historical vibration signal dataset.

[0074] First, input the test data into the shared feature extractor to extract the cross-domain general features. Then, perform feature alignment on the target domain data through the source-specific feature extractor to reduce the domain difference. Next, use the source-specific classifier to perform classification prediction on the target domain data. Finally, integrate the results of multiple source-specific classifiers through the task classifier to obtain the final diagnosis result. By comparing the prediction result with the true label, evaluate the performance metrics of the model (i.e., the cross-domain fault diagnosis model of the engine bearing), such as diagnostic accuracy, recall rate, and F1 score. Specifically as follows:

[0075] Step 1: Initialize the model evaluation mode, switch the shared feature generator (G_shared) and the classifier (Cs) to the evaluation mode (eval()), and turn off the randomness of Dropout and Batch Normalization. Load the validation set data (dataloaders['val']), and initialize the result record table (results_df) and the accuracy accumulator (acc).

[0076] Step 2: Load and extract features of the target domain data. Extract the target domain data (target_data) and the corresponding true labels (target_labels) from the validation set batch by batch. Extract the target domain features through the shared generator (feat_tgt = G_shared(target_data)).

[0077] Step 3: Predict and fuse probabilities by multiple source classifiers, calculate accuracy and record results, maintain the best model state, terminate testing, and output the model.

[0078] In step S3, input the historical vibration signal data of the engine bearing under different working conditions into the cross-domain fault diagnosis model of the engine bearing to generate a diagnosis result.

[0079] The present invention first proposes a shared feature extractor that projects all original signals into a common feature space to mine invariant features across different domains. Secondly, to avoid the complexity of simultaneously aligning multiple source domains and a target domain while ensuring that related subdomains within the same category are aligned in this feature space, a source-specific feature extractor is designed to map each pair of source domain and target domain into a specific feature space, and the variance difference representation (VDR) method is applied to reduce the feature distribution difference between the source domain and the target domain, improving the model's adaptability to target domain data. Then, a source-specific classifier is designed, which uses a softmax logistic regression module to customize a source-specific classifier for each source domain. This classifier receives the output of the source-specific feature extractor and measures the probability distribution of each sample belonging to different classes. By minimizing the VDR domain adaptation loss, the local distribution difference between each pair of source domain and target domain can be reduced, thereby optimizing the model parameters. At the same time, by minimizing the classification loss, the prediction accuracy of source domain samples can be improved. Finally, the task classifier integrates the results of all source-specific classifiers according to the weight scores of specific source domains to obtain a more reliable diagnostic result.

[0080] Thus, the present invention constructs a multi-branch parallel network model to align the subdomain distributions of each source domain and the target domain respectively, and introduces variance difference representation (VDR) to enhance the fine-grained alignment ability of similar fault features. Further combined with the source domain weighted fusion mechanism, it realizes the collaborative transfer of multi-source knowledge and high-precision diagnosis.

[0081] The cross - domain diagnosis average accuracy of this method has been improved to 98.27% on the aero - engine intermediate bearing fault diagnosis dataset of the School of Astronautics, Harbin Institute of Technology. Four intermediate bearings in different health states are used as the research objects, namely normal, outer - ring fault, and two inner - ring faults. Among them, the size of the outer - ring fault is depth 0.5 mm and length 0.5 mm, and the inner - ring faults include two different fault sizes: depth 0.5 mm, length 0.5 mm and depth 0.5 mm, length 1.0 mm. Four different engine speeds are selected as four different working conditions, which are respectively labeled as A, B, C, and D. Among them, for working condition A, the low - pressure motor speed is 1000 r / min and the high - pressure motor speed is 1200 r / min. For working condition B, the low - pressure motor speed is 3500 r / min and the high - pressure motor speed is 4200 r / min. For working condition C, the low - pressure motor speed is 5000 r / min and the high - pressure motor speed is 6000 r / min. For working condition D, the low - pressure motor speed is 3000 r / min and the high - pressure motor speed is 5400 r / min. Compared with other multi - source domain model methods, the proposed method has significantly improved the diagnostic accuracy for different target domains. The average accuracy of the proposed method on the four target domains (A / B / C / D) reaches 98.67%, 98.28%, 98.19%, and 97.97% respectively, with an overall average of 98.27%. It performs particularly well in the scenarios of noise interference and small samples, providing an innovative solution for the intelligent condition monitoring of aero - engine bearings.

[0082] Figure 3 An example of the physical structure diagram of an electronic device is shown as Figure 3 As shown, the electronic device may include: a processor 710, a communications interface 720, a memory 730, and a communication bus 740. Among them, the processor 710, the communications interface 720, and the memory 730 complete mutual communication through the communication bus 740. The processor 710 can call the logical instructions in the memory 730 to execute the cross - domain fault diagnosis method for engine bearings. The method includes: constructing a historical vibration signal dataset of engine bearings under different working conditions, where the historical vibration signal dataset includes vibration signal data of normal bearings, inner - ring fault bearings, and outer - ring fault bearings. Among them, the vibration signal data under each working condition is divided into multiple source domains and one target domain; training a transfer learning neural network based on the historical vibration signal dataset to obtain a cross - domain fault diagnosis model for engine bearings; where the transfer learning neural network includes a shared feature extractor, a source - specific feature extractor, a source - specific classifier, and a task classifier; inputting the historical vibration signal data of engine bearings under different working conditions into the cross - domain fault diagnosis model for engine bearings to generate a diagnosis result.

[0083] In addition, when the logical instructions in the above-mentioned memory 730 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0084] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the engine bearing cross-domain fault diagnosis method provided by the above-mentioned various methods. The method includes: constructing a historical vibration signal data set of the engine bearing under different working conditions. The historical vibration signal data set includes vibration signal data of normal bearings, inner ring fault bearings, and outer ring fault bearings. Among them, the vibration signal data under each working condition is divided into multiple source domains and one target domain; based on the historical vibration signal data set, training a transfer learning neural network to obtain an engine bearing cross-domain fault diagnosis model; wherein, the transfer learning neural network includes a shared feature extractor, a source-specific feature extractor, a source-specific classifier, and a task classifier; inputting the historical vibration signal data of the engine bearing under different working conditions into the engine bearing cross-domain fault diagnosis model to generate a diagnosis result.

[0085] In another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the engine bearing cross-domain fault diagnosis method provided by the above-mentioned various methods. The method includes: constructing a historical vibration signal data set of the engine bearing under different working conditions, where the historical vibration signal data set includes vibration signal data of normal bearings, inner ring fault bearings, and outer ring fault bearings. Among them, the vibration signal data under each working condition is divided into multiple source domains and one target domain; based on the historical vibration signal data set, training a transfer learning neural network to obtain an engine bearing cross-domain fault diagnosis model; where the transfer learning neural network includes a shared feature extractor, a source-specific feature extractor, a source-specific classifier, and a task classifier; inputting the historical vibration signal data of the engine bearing under different working conditions into the engine bearing cross-domain fault diagnosis model to generate a diagnosis result.

[0086] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative labor.

[0087] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0088] Furthermore, it can be understood that although the operations are described in a specific order in the drawings in the embodiments of the present invention, it should not be understood as requiring these operations to be executed in the specific order shown or in a serial order, or requiring all the operations shown to obtain the desired result. In a specific environment, multitasking and parallel processing may be beneficial.

[0089] The above are only the preferred embodiments of one or more embodiments of this specification, and are not intended to limit one or more embodiments of this specification. Any modification made within the spirit and principle of one or more embodiments of this specification.

Claims

1. A cross-domain fault diagnosis method for engine bearings, characterized in that It includes the following steps: Construct a historical vibration signal dataset of the engine bearing under different working conditions. The historical vibration signal dataset includes vibration signal data of normal bearings, inner ring fault bearings, and outer ring fault bearings. Among them, the vibration signal data under each working condition is divided into multiple source domains and one target domain; Based on the historical vibration signal dataset, train a transfer learning neural network to obtain an engine bearing cross-domain fault diagnosis model. Among them, the transfer learning neural network includes a shared feature extractor, a source-specific feature extractor, a source-specific classifier, and a task classifier; Input the historical vibration signal data of the engine bearing under different working conditions into the engine bearing cross-domain fault diagnosis model to generate a diagnosis result.

2. The engine bearing cross-domain fault diagnosis method according to claim 1, characterized in that Before training the transfer learning neural network, it is necessary to preprocess the vibration signal data in the historical vibration signal dataset, including: Normalize the vibration signal data in the historical vibration signal dataset; Segment the vibration signal data after normalization to generate multiple time series samples. Each time series sample includes vibration signal data with a unified number of time points.

3. The engine bearing cross-domain fault diagnosis method according to claim 1, wherein Randomly select 70% of the samples from each source domain in the historical vibration signal dataset as the training set, and the remaining 30% as the test set.

4. The engine bearing cross-domain fault diagnosis method according to claim 1, wherein Based on the historical vibration signal dataset, train a transfer learning neural network to obtain an engine bearing cross-domain fault diagnosis model, including: First, based on the training data in the historical vibration signal dataset, train the shared feature extractor to extract cross-domain general features; Then use the source-specific feature extractor to align the general features of each pair of source domain and target domain to reduce the domain difference; Then use the source-specific classifier to perform classification training on each source domain to optimize the parameters of the source-specific classifier; Finally, integrate the results of multiple source-specific classifiers through the task classifier to obtain an engine bearing cross-domain fault diagnosis model.

5. The engine bearing cross-domain fault diagnosis method according to claim 1, wherein It also includes: Evaluate the engine bearing cross-domain fault diagnosis model based on the test data in the historical vibration signal dataset.

6. The engine bearing cross-domain fault diagnosis method according to claim 1, characterized in that The different working conditions include different rotational speeds, loads, and environmental conditions.

7. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and running on the processor, wherein, When the processor executes the computer program, it implements the engine bearing cross-domain fault diagnosis method according to any one of claims 1 to 6.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the engine bearing cross-domain fault diagnosis method according to any one of claims 1 to 6.

Images