Bearing fault variable working condition diagnosis method based on multi-scale convolutional network and MAML

Abstract

The invention provides a bearing fault variable working condition diagnosis method based on a multi-scale convolutional network and MAML, and relates to the technical field of equipment fault diagnosis. The method comprises the following steps: firstly, introducing fast Fourier transform to pre-process an original time domain vibration signal; secondly, a fault diagnosis model based on a multi-scale convolutional network and MAML is applied; and then, an internal and external circulation updating method based on model-independent element learning is adopted, so that the model can quickly adapt to a new task, and a relatively good prediction effect can be achieved only through a small amount of fine adjustment. According to the method, multi-scale feature extraction and meta-learning are creatively combined, the generalization ability of the model under variable working conditions is remarkably improved, the problem that a traditional fault diagnosis method depends on a single working condition and a large sample size is effectively solved, and the method is particularly suitable for small-sample and multi-working-condition bearing fault diagnosis scenes on an industrial site.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment fault diagnosis, and in particular to a bearing fault variable operating condition diagnosis method based on a multi-scale convolutional network and MAML. Background Art

[0002] With the advancement of modern industry and computer technology, the complexity of mechanical systems is increasing. The widespread use of mechanical equipment has made the safety and reliability of these equipment critical issues. Rotating machinery's rotor components (such as bearings and gears) often face a high risk of failure during long-term operation due to factors such as fatigue stress and wear. Bearing failure, a common type of failure in rotating machinery, accounts for approximately 30% of all failures. For example, severe bearing damage can lead to failures in rotating machinery. Failures not only reduce production efficiency and increase maintenance costs, but in severe cases, can even endanger personnel safety.

[0003] In recent years, deep learning technology, with its powerful data-driven feature learning capabilities, has provided a new solution for breaking through the bottlenecks of traditional fault diagnosis methods. Unlike shallow machine learning models, deep neural networks (convolutional neural networks, long short-term memory neural networks, recurrent neural networks, and so on) use a hierarchical feature representation mechanism constructed through multi-level nonlinear transformations to automatically extract physically meaningful deep fault features from raw time-domain vibration signals. This end-to-end feature learning model effectively avoids the subjective biases of manual feature engineering while fully exploiting the fault-sensitive information implicit in the data through distributed representation learning.

[0004] However, existing technologies still have some shortcomings: First, fault diagnosis models rely heavily on sufficient labeled fault data for model training. However, the non-repeatability of high-risk fault states and sensor deployment limitations mean that the number of available fault samples is far lower than what is required for supervised learning. This is especially true in multi-fault coupling modes, where labeled data is severely scarce. Second, the continuous fluctuations in parameters such as equipment load and speed cause domain shifts between the training set (single operating condition) and the test set (dynamic operating conditions), resulting in performance degradation of models trained on static data. In summary, traditional fault diagnosis methods based on big data or fixed operating conditions face many challenges in practical applications. In particular, when data is limited and operating conditions are constantly changing, achieving efficient diagnosis with small sample sizes becomes an urgent problem that needs to be solved. Summary of the Invention

[0005] (1) Technical problems solved

[0006] In response to the shortcomings of the existing technology, the present invention provides a bearing fault variable operating condition diagnosis method based on multi-scale convolutional networks and MAML, which solves the technical problem of how to achieve efficient diagnosis under small sample conditions.

[0007] (2) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0009] A bearing fault variable operating condition diagnosis method based on multi-scale convolutional network and MAML, comprising:

[0010] collecting time-domain vibration signals of rotating components under different operating conditions and different fault types, performing fast Fourier transform on the time-domain vibration signals, obtaining one-dimensional data based on spectrum information processing, and constructing a fault data set; and dividing the fault data set into a training set and a test set, wherein the training set and the test set correspond to different operating conditions but contain the same fault types;

[0011] Constructing a bearing component fault diagnosis model, wherein the bearing component fault diagnosis model includes a multi-scale feature extraction unit with a residual structure and a fault diagnosis classification unit;

[0012] Based on the training set, the bearing component fault diagnosis model is trained using a model-independent meta-learning (MAML) framework;

[0013] Based on the test set, the trained bearing component fault diagnosis model is tested, and the weight at the time of highest test accuracy is used as the network parameter of the final model;

[0014] The time-domain vibration signal to be diagnosed is used as the input of the final model to obtain the diagnosis results of the bearing fault variable working condition.

[0015] Preferably, the multi-scale feature extraction unit adopts a dual-path parallel convolutional neural network architecture, and the multi-scale feature extraction unit is used to:

[0016] The first feature extraction is performed using several 1×1 convolution kernels, and each convolution layer is processed by batch normalization and ReLU activation function;

[0017] 1×3 and 1×7 convolution kernels are used for the second feature extraction to capture feature information of different scales on the corresponding branches. Each convolution layer is processed by batch normalization and ReLU activation function.

[0018] After concatenating the convolution features of the two branches, a single 1×1 convolution kernel is used to perform feature fusion to obtain the fused features.

[0019] A single 1×1 convolution kernel is applied to the input of the multi-scale feature extraction unit to perform a residual connection with the fused features.

[0020] Preferably, the fault diagnosis classification unit is used to:

[0021] Perform one-dimensional maximum pooling on the high-dimensional fusion features obtained by residual connection and expand them into vector form;

[0022] The expanded vector is fed into the fully connected layer, which then passes through the ReLU activation function, batch normalization, and Dropout layer to extract intermediate features.

[0023] The intermediate features are sent to the output layer, and the softmax function is applied to obtain the predicted probability of each fault type.

[0024] Preferably, the method of training the bearing component fault diagnosis model based on the training set using a model-independent meta-learning (MAML) framework includes:

[0025] After reaching the preset iteration termination condition, the trained bearing component fault diagnosis model is output, otherwise the following steps are repeated:

[0026] In each round of meta-training, several tasks are randomly selected from the training set; each task corresponds to a single working condition, and any task consists of a support set and a query set, and the fault types contained in the two sets are the same;

[0027] Utilizing the multi-scale feature extraction unit to extract features from the one-dimensional data in the support set of the current task to obtain high-dimensional fusion features, the high-dimensional fusion features are fed into the fault diagnosis classification unit, and the model is updated within the task parameters through forward propagation and loss function calculation to obtain model parameters that have been adapted for the current task;

[0028] After traversing all tasks, obtain the common parameters of the model for different tasks.

[0029] Perform forward propagation and loss calculation on the one-dimensional data in the current query set using the common parameters, accumulate the losses of multiple tasks on the query set, and meta-update the initial model parameters across the entire task distribution;

[0030] A batch normalization method is used to stabilize the training process of the multi-scale feature extraction unit and the fault diagnosis classification unit.

[0031] Preferably, the updating of the model's intra-task parameters is expressed as:

[0032]

[0033] Among them, θ′ i For the i-th task T i The model parameters after adaptation; θ is the initial model parameters; α is the inner learning rate, which is used to control the step size of the gradient update in the inner loop; For task T i In its support set on the losses.

[0034] Preferably, the meta-update of the initial model parameters over the entire task distribution is expressed as:

[0035]

[0036] Among them, β is the learning rate of the outer loop, which is used to control the step size of the global model update; It is task T i In its query set The loss on ; k is the number of tasks.

[0037] A bearing fault variable operating condition diagnosis system based on multi-scale convolutional networks and MAML, including:

[0038] a signal acquisition and preprocessing module for collecting time-domain vibration signals of rotating components under different operating conditions and different fault types, performing fast Fourier transform on the time-domain vibration signals, obtaining one-dimensional data based on spectrum information processing, and constructing a fault data set; and dividing the fault data set into a training set and a test set, wherein the training set and the test set correspond to different operating conditions but contain the same fault types;

[0039] A model building module is used to build a bearing component fault diagnosis model, wherein the bearing component fault diagnosis model includes a multi-scale feature extraction unit with a residual structure and a fault diagnosis classification unit;

[0040] A model training module, configured to train the bearing component fault diagnosis model based on the training set using a model-independent meta-learning (MAML) framework;

[0041] A model testing module is used to test the trained bearing component fault diagnosis model based on the test set, and use the weight when the test accuracy is the highest as the network parameter of the final model;

[0042] The result acquisition module is used to take the time domain vibration signal to be diagnosed as the input of the final model to obtain the bearing fault variable working condition diagnosis result.

[0043] A storage medium stores a computer program for diagnosing bearing fault altered operating conditions based on a multi-scale convolutional network and MAML, wherein the computer program enables a computer to execute the bearing fault altered operating condition diagnosis method as described above.

[0044] An electronic device, comprising:

[0045] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, the programs including a method for executing the bearing fault condition diagnosis method as described above.

[0046] (3) Beneficial effects

[0047] This paper provides a bearing fault variable condition diagnosis method based on a multi-scale convolutional network and MAML. Compared with the existing technology, it has the following advantages:

[0048] In the present invention, firstly, the fast Fourier transform is introduced to preprocess the original time domain vibration signal, and the complex time series data is mapped to the frequency domain, so as to effectively extract the main frequency component of the signal and filter out high-frequency noise interference, while avoiding the overfitting problem caused by the expansion of the time domain data dimension, and providing high signal-to-noise ratio input features for subsequent model training; secondly, a fault diagnosis model based on multi-scale convolutional network and MAML is applied, which demonstrates a strong knowledge transfer ability by performing multi-level and multi-band feature extraction on fault signals on multiple tasks, and effectively improves the performance of the model in processing small sample problems under variable working conditions; then, an internal and external loop update method based on model-independent meta-learning is adopted, so that the model can adapt quickly when facing new tasks, and can achieve better prediction effects through only a small amount of fine-tuning. It can quickly and effectively predict under different fault types or new working conditions, while maintaining the overall generalization ability of multiple tasks, thereby improving the cross-task adaptability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 A block diagram of a bearing fault variable operating condition diagnosis method based on a multi-scale convolutional network and MAML provided by an embodiment of the present invention;

[0051] Figure 2 A flowchart of a bearing fault variable operating condition diagnosis method based on a multi-scale convolutional network and MAML provided by an embodiment of the present invention;

[0052] Figure 3 A schematic structural diagram of a bearing component fault diagnosis model provided by an embodiment of the present invention;

[0053] Figure 4A schematic diagram of a vibration signal and its fast Fourier transform under different fault conditions provided by an embodiment of the present invention;

[0054] Figure 5 A schematic diagram of an average fault diagnosis accuracy rate provided by an embodiment of the present invention;

[0055] Figures 6(a) to 6(f) Schematic diagrams of t-SNE visualization under different samples and experimental settings provided by embodiments of the present invention;

[0056] Figures 7(a) to 7(f) Schematic diagrams of confusion matrices under different samples and experimental settings provided by embodiments of the present invention. DETAILED DESCRIPTION

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0058] The embodiment of the present application solves the technical problem of how to achieve efficient diagnosis under small sample conditions by providing a bearing fault variable operating condition diagnosis method based on a multi-scale convolutional network and MAML.

[0059] Model-Agnostic Meta-Learning (MAML), often referred to as "learning to learn," is an important method in the field of machine learning. Its core idea is to improve a model's adaptability to new tasks by learning across multiple tasks. Unlike traditional machine learning methods, meta-learning does not directly learn the solution to a specific task. Instead, it quickly adapts to new tasks by learning shared patterns between tasks, especially when data is limited. MAML is a classic meta-learning method whose core idea is to optimize the model's initial parameters θ so that the model can quickly adapt to new tasks through a small number of gradient updates.

[0060] MAML is characterized by its "model agnosticism," meaning it is not dependent on a specific model architecture and can be widely applied to various deep learning models, such as convolutional neural networks and recurrent neural networks. With sufficiently general initialization parameters, the model can effectively achieve good performance after only a few gradient updates. A multi-scale convolutional neural network is a deep learning architecture that enhances model perception by extracting features at different scales. By introducing multiple convolutional kernels of different sizes or performing multi-scale processing on the input data, it can simultaneously capture local details and global structural information, thereby achieving a more comprehensive understanding of the input data.

[0061] In the present invention, "rotating components" refer to components in machinery that primarily function through rotational motion. Common rotating machinery includes steam turbines, gas turbines, centrifugal compressors, and water pumps. The primary components of these devices include rotors, bearing systems, and stators, with rotational speeds ranging from tens to tens of thousands of revolutions per minute. Furthermore, rotary seals are crucial components for preventing leakage in rotating machinery and are suitable for a variety of applications.

[0062] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0063] Example 1:

[0064] like Figure 1 As shown, an embodiment of the present invention provides a bearing fault variable condition diagnosis method based on a multi-scale convolutional network and MAML, comprising:

[0065] S1. Collecting time-domain vibration signals of rotating components under different operating conditions and different fault types, performing fast Fourier transform on the time-domain vibration signals, obtaining one-dimensional data based on spectrum information processing, and constructing a fault dataset; and dividing the fault dataset into a training set and a test set, wherein the training set and the test set correspond to different operating conditions but contain the same fault types;

[0066] S2. Constructing a bearing component fault diagnosis model, wherein the bearing component fault diagnosis model includes a multi-scale feature extraction unit with a residual structure and a fault diagnosis classification unit;

[0067] S3. Based on the training set, the bearing component fault diagnosis model is trained using a model-independent meta-learning (MAML) framework;

[0068] S4. Testing the trained bearing component fault diagnosis model based on the test set, and using the weight with the highest test accuracy as the network parameter of the final model;

[0069] S5. Using the time-domain vibration signal to be diagnosed as the input of the final model to obtain the diagnosis result of the bearing fault variable working condition.

[0070] The embodiment of the present invention innovatively combines multi-scale feature extraction with meta-learning, significantly improving the generalization ability of the model under variable operating conditions, effectively solving the problem of traditional fault diagnosis methods' dependence on a single operating condition and a large sample size, and is particularly suitable for bearing fault diagnosis scenarios with small samples and multiple operating conditions in industrial sites.

[0071] like Figure 2 As shown, Figure 2 A flowchart of a bearing fault variable operating condition diagnosis method based on multi-scale convolutional networks and MAML is disclosed.

[0072] Next, we will combine Figure 2 The following steps are described in detail:

[0073] In step S1, time domain vibration signals of rotating components with different fault types under different working conditions are collected, fast Fourier transform is performed on the time domain vibration signals, one-dimensional data is obtained based on spectrum information processing, and a fault data set is constructed; and the fault data set is divided into a training set and a test set, wherein the training set and the test set correspond to different working conditions but contain the same fault types.

[0074] In this step, the time domain vibration signal is subjected to a fast Fourier transform, and the formula is as follows:

[0075]

[0076] Where x(n) is the time domain vibration signal, n is the sampling point index, N is the total number of sampling points, x(k) is the frequency domain signal, k is the index of the frequency component, is a complex exponential term used to map the time domain signal to the frequency domain.

[0077] Based on the spectrum information processing, one-dimensional data is obtained to construct a fault data set. Here, this means selecting a frequency domain signal x that contains the richest spectrum information and placing this one-dimensional frequency domain signal x (i.e., the aforementioned one-dimensional data) into the fault data set.

[0078] It should be noted that this step also divides the fault data set into a training set and a test set. The training set and the test set correspond to different working conditions but contain the same fault types.

[0079] In step S2, a bearing component fault diagnosis model is constructed, wherein the bearing component fault diagnosis model includes a multi-scale feature extraction unit with a residual structure and a fault diagnosis classification unit.

[0080] like Figure 3 As shown, Figure 3Disclosed is a structural schematic diagram of a bearing component fault diagnosis model.

[0081] See also Figure 3 The multi-scale feature extraction unit adopts a dual-parallel convolutional neural network architecture, and the multi-scale feature extraction unit is used to:

[0082] First, several 1×1 convolution kernels are used for the first feature extraction. Each convolution layer is processed by batch normalization and ReLU activation function, which can be expressed as:

[0083]

[0084] in, Initial features, * represents convolution operation, ReLU is the activation function, BN represents batch normalization, is the convolution weight of the 1×1 convolution kernel.

[0085] Secondly, 1×3 and 1×7 convolution kernels are used for the second feature extraction to capture feature information of different scales on the corresponding branches. Each convolution layer is processed by batch normalization and ReLU activation function, which is expressed as:

[0086]

[0087] Among them, x1 and x2 are the convolution features output by the first and second branches respectively. is the convolution weight of the 1×3 convolution kernel on branch 1, is the convolution weight of the 1×7 convolution kernel on branch 2.

[0088] Then, the convolution features of the output of the above two branches are concatenated and a single 1×1 convolution kernel is used for feature fusion to obtain the fused feature x fusion .

[0089] At the same time, a single 1×1 convolution kernel (which plays the role of matching dimension) is applied to the input of the multi-scale feature extraction unit to perform residual connection with the fusion feature. The high-dimensional fusion feature of the final output is expressed as:

[0090] x out =x fusion +Conv 1×1 (x)

[0091] Among them, Conv 1×1 It is a convolution operation with a 1×1 convolution kernel.

[0092] See also Figure 3 The main function of the fault diagnosis classification unit is to convert the high-dimensional fusion feature map output by the multi-scale feature extraction unit into the final fault category probability distribution, specifically including:

[0093] First, the high-dimensional fusion features obtained by residual connection are subjected to one-dimensional maximum pooling to compress the time series dimension and expanded into a vector form, which is expressed as:

[0094] f=Flatten(MaxPool(x out ))

[0095] Among them, Flatten is the expansion operation and MaxPool is the maximum pooling.

[0096] Secondly, the expanded vector f is fed into the fully connected layer, which passes through the ReLU activation function, batch normalization, and Dropout layer in turn to extract more compact and discriminative intermediate features:

[0097] h1=Dropout(BN(ReLU(W1f+b1)))

[0098] Among them, W1 and b1 are the weight and bias of the fully connected layer respectively.

[0099] Subsequently, the intermediate feature h1 is sent to the output layer, and the softmax function is applied to obtain the predicted probability of each fault type, which is expressed as:

[0100] y=Softmax(W2h1+b2)

[0101] Among them, W2 and b2 are the weight and bias of the output layer respectively.

[0102] It can be understood that the fault diagnosis classification unit structure can effectively map the time series feature information to the specific fault type probability, providing a basis for subsequent diagnostic decisions.

[0103] In step S3, based on the training set, the bearing component fault diagnosis model is trained using a model-independent meta-learning (MAML) framework.

[0104] This step specifically includes:

[0105] After reaching the preset iteration termination condition (e.g., reaching the preset number of iterations), the trained bearing component fault diagnosis model is output, otherwise the following steps are repeated:

[0106] S100. In each round of meta-training, a number of tasks are randomly selected from the training set; each task corresponds to a single working condition, and any task consists of a support set and a query set, and the fault types contained in the two sets are the same.

[0107] S200. Utilize the multi-scale feature extraction unit to perform feature extraction on the one-dimensional data in the support set of the current task to obtain high-dimensional fusion features, send the high-dimensional fusion features to the fault diagnosis classification unit, update the intra-task parameters of the model through forward propagation and loss function calculation, and obtain the model parameters after adaptation to the current task.

[0108] The intra-task parameter update of the model is expressed as:

[0109]

[0110] Among them, θ′ i For the i-th task T i The model parameters after adaptation; θ is the initial model parameters; α is the inner learning rate, which is used to control the step size of the gradient update in the inner loop; For task T i In its support set on the losses.

[0111] S300: After traversing all tasks, obtain the common parameters of the model for different tasks.

[0112] S400 , using the common parameters to perform forward propagation and loss calculation on the one-dimensional data in the current query set, accumulating the losses of multiple tasks on the query set, and performing meta-update on the initial model parameters over the entire task distribution.

[0113] The meta-update of the initial model parameters across the entire task distribution is expressed as:

[0114]

[0115] Among them, β is the learning rate of the outer loop, which is used to control the step size of the global model update; It is task T i In its query set The loss on ; k is the number of tasks.

[0116] S500: Using a batch normalization method to stabilize the training process of the multi-scale feature extraction unit and the fault diagnosis classification unit.

[0117] In step S4, the trained bearing component fault diagnosis model is tested based on the test set, and the weights with the highest test accuracy are used as network parameters of the final model.

[0118] It should be noted again that the working conditions of the training set and the test set are different, such as the speed, load, and operating environment. However, as mentioned above, the fault categories of the training set tasks are consistent with the fault types required by the test set tasks.

[0119] In step S5, the time domain vibration signal to be diagnosed is used as the input of the final model to obtain the bearing fault variable operating condition diagnosis result.

[0120] At this point, the embodiment of the present invention completes the entire process of the bearing fault variable operating condition diagnosis method based on multi-scale convolutional network and MAML.

[0121] In order to better help understand the advantages of the embodiments of the present invention, the following examples are provided:

[0122] See also Figure 3 The method process includes: 1) collecting vibration data of different fault types under different operating conditions and performing fast Fourier transform on them, selecting the segments with the richest frequency information as model input; 2) dividing the data into training and test sets, each training set has multiple tasks, and each task is divided into a support set and a query set; 3) combining a multi-scale convolutional neural network model with a residual structure to design the proposed method; 4) during the training process, tasks are randomly selected for model training; 5) using the test set to test the trained model. The specific implementation steps of this example are as follows:

[0123] Step 1: Collect vibration data for different fault types under different operating conditions and perform a fast Fourier transform. The vibration data for different fault types comes from the Paderborn University Rotating Components Public Dataset (PU Dataset). The vibration data was measured at drive system speeds of 1500 rpm and 900 rpm, radial forces of 1000 N and 400 N on the test rotating components, and load torques of 0.7 nm and 0.1 nm on the transmission system. The data were measured under four operating conditions: the training set includes three load conditions: Load 1, Load 2, and Load 3, and the test set is the Load 0 load condition, as shown in Table 1.

[0124] Table 1 Working conditions of test set and training set

[0125]

[0126] The original one-dimensional vibration signal is subjected to fast Fourier transform, and the first 1600 data lengths of the spectrum are selected and used as the input of samples.

[0127] Step 2: Divide the data into training and test sets. Each training set contains multiple tasks, and each task is divided into a support set and a query set. The support set sample size for each fault category is 1 (or 5) in two experimental settings, and the query set sample size is 15, 35, or 55 (the same query set sample size is used in both cases). That is, there are 20, 40, and 60 samples for each fault category, each consisting of 1600 fast Fourier transformed points. During training, five tasks are selected from each of the ten categories in the training set to train the model. This means that the experiments are 5-way 1-shot and 5-way 5-shot experiments, as shown in Table 2.

[0128] Table 2 Fault types in the test set and training set

[0129] Label Fault location Fault level Bearing number Sample size 0 normal / K005 20、40、60 1 Outer ring (artificial) 1 KA01 20、40、60 2 outer ring 1 KA04 20、40、60 3 outer ring 2 KA16 20、40、60 4 Inner ring 1 KI04 20、40、60 5 Inner ring 2 KI18 20、40、60 6 Inner ring 3 KI16 20、40、60 7 Outer ring + inner ring 1 KB27 20、40、60 8 Outer ring + inner ring 2 KB23 20、40、60 9 Outer ring + inner ring 3 KB24 20、40、60

[0130] like Figure 4 As shown, Figure 4 A schematic diagram of a vibration signal and its fast Fourier transform provided by an embodiment of the present invention is disclosed. The diagram shows the vibration signal and its fast Fourier transform results under different fault states (i.e., normal state, outer ring fault, inner ring fault, and outer ring + inner ring fault).

[0131] Step 3: A multi-scale convolutional neural network model is combined with a residual structure to design the proposed method.

[0132] The feature extraction module utilizes a dual-parallel convolutional neural network architecture, combined with a residual structure, to process Fourier-transformed signal data. The model first extracts features through multiple 1×1 convolutional layers, each followed by batch normalization and ReLU activation to enhance training stability and nonlinear representation. Subsequently, 1×3 and 1×7 convolutional kernels are used for deeper feature extraction, further capturing feature information at different scales. These kernels are also processed with batch normalization and ReLU activation. After multiple convolution operations, all extracted feature maps are concatenated, increasing the model's feature dimensionality and improving its ability to capture information at different scales. The concatenated feature maps are then fed into a 1×1 convolutional layer, concatenated with the residual, and again processed with batch normalization and ReLU activation. Max pooling is then used to reduce the feature map size, further compressing the data and preventing overfitting.

[0133] Step 4: The working conditions of the training set and the test set are different. The task subset of each iteration is randomly selected, and the number of categories is consistent with the number of categories in the test set.

[0134] For each task Ti , we first start with task T i The support set in Train the task-specific model parameters θ′ i :

[0135]

[0136] Among them, θ′ i For the i-th task T i The model parameters after adaptation; θ is the initial model parameters; α is the inner learning rate, which is used to control the step size of the gradient update in the inner loop; For task T i In its support set on the losses.

[0137] Then, the queryset over all tasks The test loss of the task is calculated and the global parameter θ is updated by minimizing the combination of these test losses. The outer loop update formula is:

[0138]

[0139] Among them, β is the learning rate of the outer loop, which is used to control the step size of the global model update; It is task T i In its query set The loss on ; k is the number of tasks.

[0140] Step 5: Use the support set of the test set to test the trained model to obtain a set of parameters, and then conduct a test experiment on the query set in the test set.

[0141] Each experiment was run 5 times, and the average fault diagnosis accuracy was as follows: Figure 5 shown.

[0142] In addition, Figures 6(a) to (f) show t-SNE (t-distributed stochastic neighbor embedding) visualizations under different samples and experimental settings, where Figure 6(a) corresponds to 20-sample 5-way 1-shot, Figure 6(b) corresponds to 40-sample 5-way 1-shot, Figure 6(c) corresponds to 60-sample 5-way 1-shot, Figure 6(d) corresponds to 20-sample 5-way 5-shot, Figure 6(e) corresponds to 40-sample 5-way 5-shot, and Figure 6(f) corresponds to 60-sample 5-way 5-shot.

[0143] Figures 7(a) to (f) show the confusion matrices under different samples and experimental settings, where Figure 7(a) corresponds to 20-sample 5-way 1-shot, Figure 7(b) corresponds to 40-sample 5-way 1-shot, Figure 7(c) corresponds to 60-sample 5-way 1-shot, Figure 7(d) corresponds to 20-sample 5-way 5-shot, Figure 7(e) corresponds to 40-sample 5-way 5-shot, and Figure 7(f) corresponds to 60-sample 5-way 5-shot.

[0144] Example 2:

[0145] The embodiment of the present invention provides a bearing fault variable operating condition diagnosis system based on a multi-scale convolutional network and MAML, including:

[0146] a signal acquisition and preprocessing module for collecting time-domain vibration signals of rotating components under different operating conditions and different fault types, performing fast Fourier transform on the time-domain vibration signals, obtaining one-dimensional data based on spectrum information processing, and constructing a fault data set; and dividing the fault data set into a training set and a test set, wherein the training set and the test set correspond to different operating conditions but contain the same fault types;

[0147] A model building module is used to build a bearing component fault diagnosis model, wherein the bearing component fault diagnosis model includes a multi-scale feature extraction unit with a residual structure and a fault diagnosis classification unit;

[0148] A model training module, configured to train the bearing component fault diagnosis model based on the training set using a model-independent meta-learning (MAML) framework;

[0149] A model testing module is used to test the trained bearing component fault diagnosis model based on the test set, and use the weight when the test accuracy is the highest as the network parameter of the final model;

[0150] The result acquisition module is used to take the time domain vibration signal to be diagnosed as the input of the final model to obtain the bearing fault variable working condition diagnosis result.

[0151] Example 3:

[0152] An embodiment of the present invention provides a storage medium storing a computer program for diagnosing bearing fault variable operating conditions based on a multi-scale convolutional network and MAML, wherein the computer program enables a computer to execute the bearing fault variable operating condition diagnosis method as described in Example 1.

[0153] Example 4:

[0154] An embodiment of the present invention provides an electronic device, including:

[0155] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, the programs including a method for executing the bearing fault condition diagnosis method as described in Example 1.

[0156] It can be understood that the bearing fault variable operating condition diagnosis system, storage medium and electronic device based on multi-scale convolutional network and MAML provided in the embodiment of the present invention correspond to the bearing fault variable operating condition diagnosis method based on multi-scale convolutional network and MAML provided in the embodiment of the present invention. The explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the bearing fault variable operating condition diagnosis method, and will not be repeated here.

[0157] In summary, compared with the existing technology, the present invention has the following beneficial effects:

[0158] 1. The embodiment of the present invention introduces fast Fourier transform to preprocess the original time-domain vibration signal, maps the complex time series data to the frequency domain, effectively extracts the main frequency component of the signal and filters out high-frequency noise interference, while avoiding the overfitting problem caused by the expansion of the time domain data dimension, and provides high signal-to-noise ratio input features for subsequent model training.

[0159] 2. The embodiment of the present invention applies a fault diagnosis model based on a multi-scale convolutional network and MAML. By performing multi-level and multi-band feature extraction on fault signals on multiple tasks, it demonstrates a strong knowledge transfer capability and effectively improves the performance of the model in handling small sample problems under variable working conditions.

[0160] 3. The embodiment of the present invention adopts an internal and external loop update method based on model-independent meta-learning, so that the model can adapt quickly when facing new tasks, and can achieve better prediction results with only a small amount of fine-tuning. It can quickly and effectively predict different fault types or new working conditions while maintaining the overall generalization ability of multiple tasks, thereby improving the cross-task adaptability of the model.

[0161] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0162] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A bearing fault variable working condition diagnosis method based on multi-scale convolutional network and MAML, characterized by: include: collecting time-domain vibration signals of rotating components under different operating conditions and different fault types, performing fast Fourier transform on the time-domain vibration signals, obtaining one-dimensional data based on spectrum information processing, and constructing a fault data set; and dividing the fault data set into a training set and a test set, wherein the training set and the test set correspond to different operating conditions but contain the same fault types; Constructing a bearing component fault diagnosis model, wherein the bearing component fault diagnosis model includes a multi-scale feature extraction unit with a residual structure and a fault diagnosis classification unit; Based on the training set, the bearing component fault diagnosis model is trained using a model-independent meta-learning (MAML) framework; Based on the test set, the trained bearing component fault diagnosis model is tested, and the weight at the time of highest test accuracy is used as the network parameter of the final model; The time-domain vibration signal to be diagnosed is used as the input of the final model to obtain the diagnosis results of the bearing fault variable working condition.

2. The bearing fault variable operating condition diagnosis method according to claim 1, characterized in that: The multi-scale feature extraction unit adopts a dual-path parallel convolutional neural network architecture, and is used to: The first feature extraction is performed using several 1×1 convolution kernels, and each convolution layer is processed by batch normalization and ReLU activation function; 1×3 and 1×7 convolution kernels are used for the second feature extraction to capture feature information of different scales on the corresponding branches. Each convolution layer is processed by batch normalization and ReLU activation function. After concatenating the convolution features of the two branches, a single 1×1 convolution kernel is used to perform feature fusion to obtain the fused features. A single 1×1 convolution kernel is applied to the input of the multi-scale feature extraction unit to perform a residual connection with the fused features.

3. The bearing fault variable operating condition diagnosis method according to claim 2, characterized in that: The fault diagnosis classification unit is used for: Perform one-dimensional maximum pooling on the high-dimensional fusion features obtained by residual connection and expand them into vector form; The expanded vector is fed into the fully connected layer, which then passes through the ReLU activation function, batch normalization, and Dropout layer to extract intermediate features. The intermediate features are sent to the output layer, and the softmax function is applied to obtain the predicted probability of each fault type.

4. The bearing fault variable operating condition diagnosis method according to claim 1, characterized in that: The method of training the bearing component fault diagnosis model based on the training set using a model-independent meta-learning (MAML) framework includes: After reaching the preset iteration termination condition, the trained bearing component fault diagnosis model is output, otherwise the following steps are repeated: In each round of meta-training, several tasks are randomly selected from the training set; each task corresponds to a single working condition, and any task consists of a support set and a query set, and the fault types contained in the two sets are the same; Utilizing the multi-scale feature extraction unit to extract features from the one-dimensional data in the support set of the current task to obtain high-dimensional fusion features, the high-dimensional fusion features are fed into the fault diagnosis classification unit, and the model is updated within the task parameters through forward propagation and loss function calculation to obtain model parameters that have been adapted for the current task; After traversing all tasks, obtain the common parameters of the model for different tasks. Perform forward propagation and loss calculation on the one-dimensional data in the current query set using the common parameters, accumulate the losses of multiple tasks on the query set, and meta-update the initial model parameters across the entire task distribution; A batch normalization method is used to stabilize the training process of the multi-scale feature extraction unit and the fault diagnosis classification unit.

5. The bearing fault variable operating condition diagnosis method according to claim 4, characterized in that: The intra-task parameter update of the model is expressed as: Among them, θ′ i For the i-th task T i The model parameters after adaptation; θ is the initial model parameters; α is the inner learning rate, which is used to control the step size of the gradient update in the inner loop; For task T i In its support set on the losses.

6. The bearing fault variable operating condition diagnosis method according to claim 4, characterized in that: The meta-update of the initial model parameters across the entire task distribution is expressed as: Among them, β is the learning rate of the outer loop, which is used to control the step size of the global model update; It is task T i In its query set The loss on ; k is the number of tasks.

7. A bearing fault variable working condition diagnosis system based on multi-scale convolutional network and MAML, characterized by: include: a signal acquisition and preprocessing module for collecting time-domain vibration signals of rotating components under different operating conditions and different fault types, performing fast Fourier transform on the time-domain vibration signals, obtaining one-dimensional data based on spectrum information processing, and constructing a fault data set; and dividing the fault data set into a training set and a test set, wherein the training set and the test set correspond to different operating conditions but contain the same fault types; A model building module is used to build a bearing component fault diagnosis model, wherein the bearing component fault diagnosis model includes a multi-scale feature extraction unit with a residual structure and a fault diagnosis classification unit; A model training module, configured to train the bearing component fault diagnosis model based on the training set using a model-independent meta-learning (MAML) framework; A model testing module is used to test the trained bearing component fault diagnosis model based on the test set, and use the weight when the test accuracy is the highest as the network parameter of the final model; The result acquisition module is used to take the time domain vibration signal to be diagnosed as the input of the final model to obtain the bearing fault variable working condition diagnosis result.

8. A storage medium, characterized in that: It stores a computer program for diagnosing bearing fault-variable operating conditions based on a multi-scale convolutional network and MAML, wherein the computer program enables a computer to execute the bearing fault-variable operating condition diagnosis method according to any one of claims 1 to 6.

9. An electronic device, characterized in that: include: one or more processors; Memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, the programs including a method for executing the bearing fault variable operating condition diagnosis method according to any one of claims 1 to 6.

Images