A small-sample training method and diagnostic method for bearing fault diagnosis model
By improving the small-sample training method of the bearing fault diagnosis model, and combining pre-training, meta-training and fine-tuning of public datasets, the problem of sample scarcity in bearing fault diagnosis is solved, and high-precision and highly adaptable fault diagnosis results are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-17
- Publication Date
- 2026-04-03
AI Technical Summary
Existing bearing fault diagnosis methods based on deep neural networks suffer from low model reliability due to scarce samples, making them difficult to apply effectively in practical engineering.
A bearing fault diagnosis model with a one-dimensional residual network and encoder-decoder structure is adopted. It combines pre-training, meta-training and fine-tuning of public datasets, utilizes multi-head self-attention layers to improve feature extraction capabilities, and establishes a dataset through data cleaning and augmentation to achieve small-sample training.
Under conditions of scarce sample size, a bearing fault diagnosis with high generalization, high flexibility and high reliability was achieved, improving the fault diagnosis accuracy and model adaptability.
Smart Images
Figure CN117194990B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fault diagnosis technology, specifically to a small-sample training method and a diagnostic method for a bearing fault diagnosis model. Background Technology
[0002] As a crucial component in mechanical equipment, the operating condition of bearings directly impacts the equipment's performance, reliability, and lifespan. Currently, bearing fault diagnosis methods based on vibration signals are the most common, characterized by rich information, clear fault characteristics, strong correlation, non-invasiveness, and real-time performance.
[0003] Among them, fault diagnosis based on deep neural network models has shown significant advantages in the field of fault diagnosis due to its better spatial representation ability and adaptability. Typically, a highly reliable deep neural network requires a large amount of training data. However, in reality, due to privacy concerns regarding mechanical equipment faults and the cost of fault labeling, it is often impossible to obtain sufficient bearing defect samples, resulting in low model reliability and difficulty in practical engineering implementation.
[0004] Currently, existing few-shot learning methods in the field of bearing fault diagnosis can be categorized into sample generation-based methods, transfer-based methods, and meta-learning-based methods. Sample generation-based methods expand the original sample set by synthesizing samples, but they rely excessively on the completeness of sample features, resulting in weak generalization performance. Transfer-based methods utilize existing knowledge data to improve performance on the target task, but they cannot adapt well to the target domain when the difference between the source and target domains is too large, lacking flexibility in inter-domain adaptation. Meta-learning-based methods quickly adapt to tasks through multi-task learning, but their small model capacity results in accuracy that cannot meet production requirements, and the model cannot achieve good diagnostic results under conditions of scarce samples. Summary of the Invention
[0005] To address the aforementioned problems in existing technologies, this invention provides a small-sample training method for bearing fault diagnosis models, and proposes a bearing fault diagnosis model that can achieve high generalization, high flexibility, and high reliability even under conditions of scarce sample numbers.
[0006] In a first aspect, embodiments of this application provide a small-sample training method for a bearing fault diagnosis model, including the following steps:
[0007] Step A: Collect the bearing's operating parameters and establish a dataset, which includes a public dataset and a small sample local dataset.
[0008] Step B involves establishing a bearing fault diagnosis model, which includes a one-dimensional residual network and an encoder-decoder structure arranged in parallel. The one-dimensional residual network is based on the existing ResNet50 and improved for bearing features. As a feature extractor, it improves the sampling layer, adds a multi-head self-attention layer, and corresponding parameters.
[0009] Step C involves pre-training the established bearing fault diagnosis model based on the collected public dataset to ensure that the trained bearing fault diagnosis model has good generalization ability.
[0010] Step D: Meta-training of the pre-trained bearing fault diagnosis model based on the collected small sample local dataset.
[0011] Step E involves fine-tuning the trained bearing fault diagnosis model by further adjusting and improving the network parameters to obtain a well-trained bearing fault diagnosis model. This ensures that the trained bearing fault diagnosis model has higher reliability in practical applications.
[0012] In another alternative to the first aspect, in step A, after establishing the dataset, the public dataset and the small sample local dataset are divided into a training set and a corresponding test set in a ratio of 8:2. The training set is a dataset of unlabeled bearing faults, and the test set is a dataset of labeled bearing faults.
[0013] In one alternative of the first aspect, in step B, the one-dimensional residual network includes five modules connected in sequence. Module 1 includes a sampling layer and an average pooling layer connected in sequence. Modules 2, 3, and 4 each include a multi-layer batch normalization layer. Module 5 includes a batch normalization layer and a multi-head self-attention layer connected in sequence.
[0014] In another alternative to the first aspect, step C specifically involves:
[0015] Step C1 involves inputting the public dataset into a one-dimensional residual network for processing to obtain signal features. The public dataset uses the existing bearing dataset CWRU.
[0016] Step C2 involves inputting the public dataset into the encoder-decoder structure for processing to obtain text description features;
[0017] Step C3: Based on the obtained signal features and text description features, calculate the pre-training loss function, which is a cosine loss function, until the pre-training loss function converges to a preset value to obtain the pre-trained bearing fault diagnosis model.
[0018] In another alternative to the first aspect, step D specifically involves:
[0019] Step D1: Use the pre-trained bearing fault diagnosis model to map the small local dataset from a one-dimensional feature space to a multi-dimensional feature space to obtain the bearing fault diagnosis results diagnosed by the model.
[0020] Step D2: Establish a meta-training loss function until the meta-training loss function converges to a preset value to obtain the meta-trained bearing fault diagnosis model.
[0021] In another alternative to the first aspect, in step A, after collecting the bearing's operating parameters, the collected bearing operating parameters are further cleaned and augmented to improve the data value.
[0022] In another alternative to the first aspect, the data cleaning includes outlier handling and missing value handling.
[0023] In another alternative to the first aspect, the data enhancement includes sliding window enhancement and resampling enhancement.
[0024] In another alternative to the first aspect, the operating parameters of the bearing include sampling frequency, sampling time, and vibration signal.
[0025] Secondly, embodiments of this application provide a bearing fault diagnosis method, which uses a bearing fault diagnosis model trained by a small-sample training method based on the first aspect of this application or any implementation thereof to diagnose faulty bearings.
[0026] The beneficial technical effects of this invention include:
[0027] Overcoming the problem of low fault diagnosis accuracy caused by the scarcity of bearing fault samples in actual engineering, the bearing fault diagnosis model is pre-trained, meta-trained, and fine-tuned based on the collected public dataset and small local dataset. This results in a bearing fault diagnosis model with good generalization ability, strong extraction capability, high reliability, and high fault diagnosis accuracy. Attached Figure Description
[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0029] Figure 1 This is a flowchart of a small-sample training method for a bearing fault diagnosis model in this invention;
[0030] Figure 2 This is a schematic diagram of a small-sample training method for a bearing fault diagnosis model in this invention;
[0031] Figure 3 This is a block diagram of a bearing fault diagnosis model in this invention;
[0032] Figure 4 This is a network architecture block diagram of the one-dimensional residual network in this invention;
[0033] Figure 5 This is a schematic diagram of the network architecture of the one-dimensional residual network in this invention. Detailed Implementation
[0034] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.
[0035] In the following description, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. The following description provides multiple embodiments of this application, which can be substituted or combined with each other. Therefore, this application can also be considered to include all possible combinations of the same and / or different embodiments described. Thus, if one embodiment includes features A, B, and C, and another embodiment includes features B and D, then this application should also be considered to include embodiments containing one or more other possible combinations of A, B, C, and D, even if such embodiments are not explicitly described in the following text.
[0036] The following description provides examples and does not limit the scope, applicability, or examples set forth in the claims. Changes may be made to the function and arrangement of the described elements without departing from the scope of this application. Various processes or components may be appropriately omitted, substituted, or added to the examples. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Furthermore, features described with respect to some examples may be combined into other examples.
[0037] Example 1:
[0038] Reference Figure 1 A method for training a bearing fault diagnosis model using small samples, comprising the following steps:
[0039] Step A: Collect the bearing's operating parameters and establish a dataset, which includes a public dataset and a small sample local dataset.
[0040] Furthermore, step A specifically includes:
[0041] Step A1: Install a vibration sensor on the bearing base. According to the acquisition requirements, acquire the bearing's operating parameters, including sampling frequency, sampling time, filter, and vibration signal. Acquiring the vibration signal involves starting the acquisition device, acquiring the vibration signal data, and importing the acquired vibration signal data into the computer.
[0042] Step A2 involves cleaning the collected bearing operating parameter data. The data cleaning includes outlier handling, missing value handling, and duplicate value handling. Missing and outlier values are set to zero, and duplicate values are deleted.
[0043] Step A3 involves performing data augmentation on the collected bearing operating parameter data, including sliding window enhancement and resampling enhancement.
[0044] Step A4: Divide the dataset into training and test sets in an 8:2 ratio.
[0045] Reference Figure 2 Step B: Establish a bearing fault diagnosis model. The bearing fault diagnosis model includes a one-dimensional residual network and an encoder-decoder structure set up in parallel.
[0046] The one-dimensional residual network is an improvement on the existing ResNet50 for bearing features. As a feature extractor, it improves the sampling layer, adds a multi-head self-attention layer and corresponding parameters.
[0047] The one-dimensional residual network consists of five modules connected in sequence. Module 1 consists of a sampling layer and an average pooling layer connected in sequence. Modules 2, 3, and 4 each consist of multiple batch normalization layers. Module 5 consists of a batch normalization layer and a multi-head self-attention layer connected in sequence.
[0048] Specifically, in the first batch normalization layer, the data is processed in two parallel paths. The first path of data sequentially passes through a convolutional layer, a batch normalization layer, an activation function layer, another convolutional layer, another batch normalization layer, an activation function layer, an average pooling layer, another convolutional layer, and an average pooling layer. The second path of data sequentially passes through a convolutional layer and a batch normalization layer. After these two parallel passes, the data is summed and finally passed through the activation function layer. After passing through the batch normalization layer, the number of channels is 4 / S times the original number, and the channel width is 1 / S times the original number, where S is the multiplier.
[0049] The second batch normalization layer processes data in parallel, with two paths. The first path passes through a convolutional layer, a batch normalization layer, an activation function layer, another convolutional layer, another batch normalization layer, an activation function layer, an average pooling layer, another convolutional layer, and an average pooling layer. The second path of data is left unprocessed. The data is then summed after these two parallel passes and finally passed through the activation function layer. After the second batch normalization layer, the number of channels and channel width remain unchanged.
[0050] Furthermore, the activation function in the activation function layer is ReLU.
[0051] In the multi-head self-attention layer, the input multi-dimensional data is first flattened into one-dimensional data, and then multi-head projection, attention calculation and multi-head splicing operations are sequentially performed on the one-dimensional data, with a number of heads of 8.
[0052] Specifically, module one consists of three sampling layers and one average pooling layer connected in sequence. Each of the three sampling layers includes a convolutional layer, a batch normalization layer, and an activation function layer connected in sequence. The activation function is ReLU. The convolutional layer in the first sampling layer has a kernel size of 1*3, a stride of 2, and 32 channels. The convolutional layer in the second sampling layer has a kernel size of 1*3, a stride of 2, and 32 channels. The convolutional layer in the third sampling layer has a kernel size of 1*3, a stride of 2, and 64 channels. The convolutional layer in the fourth average pooling layer has a kernel size of 2 and a stride of 2.
[0053] Module 2 consists of three batch normalization layers connected in sequence. The first batch normalization layer has 64 channels, a channel width of 128, and a scaling factor of 1. The parameters of the latter two batch normalization layers are the same, with 256 channels and a channel width of 128.
[0054] Module 3 consists of four batch normalization layers connected in sequence. The first batch normalization layer has 256 channels, a channel width of 128, and a scaling factor of 2. The parameters of the last three batch normalization layers are the same: 512 channels and 64 channels.
[0055] Module 4 consists of five batch normalization layers connected in sequence. The first batch normalization layer has 512 channels, a channel width of 64, and a scaling factor of 2. The parameters of the next four second batch normalization layers are the same: 1024 channels and a channel width of 32.
[0056] Module 5 consists of three batch normalization layers and one multi-head self-attention layer connected in sequence. The first batch normalization layer has 1024 channels, a channel width of 32, and a scaling factor of 2. The parameters of the next two second batch normalization layers are the same, with 2048 channels and a channel width of 16. The fourth multi-head self-attention layer has 2048 channels and an output width of N.
[0057] Step C: Pre-train the established bearing fault diagnosis model based on the collected public dataset, which is the Case Western Reserve University Bearing Dataset (CWRU).
[0058] Step C specifically involves:
[0059] Step C1: Input the public dataset into a one-dimensional residual network for processing to obtain signal features;
[0060] Step C2 involves inputting the public dataset into the encoder-decoder structure for processing, transforming the signal features into a specific textual description of the current bearing fault, thereby obtaining the textual description features.
[0061] Step C3: Based on the obtained signal features and text description features, calculate the pre-training loss function, which is a cosine loss function, until the pre-training loss function converges to a preset value to obtain the pre-trained bearing fault diagnosis model.
[0062] Step D: Meta-training of the pre-trained bearing fault diagnosis model based on the collected small sample local dataset.
[0063] Step D specifically involves:
[0064] Step D1 involves using the pre-trained bearing fault diagnosis model to map the small local dataset from a one-dimensional feature space to a multi-dimensional feature space, thereby obtaining the bearing fault diagnosis results identified by the model. The mapping formula is:
[0065]
[0066]
[0067] k represents the samples in the training set, k ′ N represents the samples in the corresponding test set. k Represents the total number of samples in the dataset, (x i ,y i ) represents data x i and its corresponding label y i d represents the cosine distance metric function, and the mapping formula can be used to determine the type of bearing fault. Each training and test set is divided into a support set and a query set in an 8:2 ratio to support data queries and ensure data integrity.
[0068] Step D2: Establish a meta-training loss function until the meta-training loss function converges to a preset value to obtain the meta-trained bearing fault diagnosis model.
[0069] Step E involves fine-tuning the meta-trained bearing fault diagnosis model by mapping the input samples to a low-dimensional feature space and calculating the fine-tuning loss function, which is the cross-entropy loss function, to obtain the trained bearing fault diagnosis model.
[0070] The method of this invention uses a publicly available bearing vibration signal dataset for pre-training, which ensures the model's accuracy while giving it good generalization ability. The pre-trained model uses a two-stream model with a one-dimensional residual network ResNet50 as the feature extractor, which ensures the model's capacity while accelerating the model's convergence speed. Meta-training is performed on local data, allowing the model to flexibly adapt to the local data distribution. Fine-tuning techniques are used to further adjust the network parameters, making the model more reliable.
[0071] This invention utilizes publicly available external bearing datasets as auxiliary training through pre-training, meta-training, and fine-tuning steps, exhibiting high feature extraction capability and high reliability, and thus has great application prospects.
[0072] Example 2:
[0073] A bearing fault diagnosis method is characterized by diagnosing faulty bearings using a bearing fault diagnosis model trained based on a small-sample training method provided in Embodiment 1.
[0074] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope of the present invention.
Claims
1. A small-sample training method for a bearing fault diagnosis model, characterized in that, Including the following steps: A. Collect bearing operating parameters and establish a dataset, which includes a public dataset and a small sample local dataset; B. Establish a bearing fault diagnosis model, which includes a one-dimensional residual network and an encoder-decoder structure set in parallel. C. Pre-train the established bearing fault diagnosis model based on the collected public dataset; D, Meta-training of the pre-trained bearing fault diagnosis model based on the collected small sample local dataset; E. Fine-tune the trained bearing fault diagnosis model to obtain a well-trained bearing fault diagnosis model.
2. The small-sample training method for a bearing fault diagnosis model according to claim 1, characterized in that, In step A, after establishing the dataset, the public dataset and the small sample local dataset are divided into a training set and a corresponding test set in a ratio of 8:
2. The training set is the dataset with unlabeled bearing faults, and the test set is the dataset with labeled bearing faults.
3. The small-sample training method for a bearing fault diagnosis model according to claim 2, characterized in that, In step B, the one-dimensional residual network includes five modules connected in sequence. Module 1 includes a sampling layer and an average pooling layer connected in sequence. Modules 2, 3, and 4 each include multiple batch normalization layers. Module 5 includes a batch normalization layer and a multi-head self-attention layer connected in sequence.
4. The small-sample training method for a bearing fault diagnosis model according to claim 3, characterized in that, Step C specifically involves: C1 involves inputting the public dataset into a one-dimensional residual network for processing to obtain signal features; C2, inputs the public dataset into the encoder-decoder structure for processing to obtain text description features; C3. Based on the obtained signal features and text description features, calculate the pre-training loss function, which is a cosine loss function, until the pre-training loss function converges to a preset value to obtain the pre-trained bearing fault diagnosis model.
5. The small-sample training method for a bearing fault diagnosis model according to claim 4, characterized in that, Step D specifically involves: D1 uses a pre-trained bearing fault diagnosis model to map a small local dataset from a one-dimensional feature space to a multi-dimensional feature space in order to obtain the bearing fault diagnosis results diagnosed by the model. D2. Establish the meta-training loss function until the meta-training loss function converges to the preset value to obtain the meta-trained bearing fault diagnosis model.
6. The small-sample training method for a bearing fault diagnosis model according to claim 1, characterized in that, In step A, after collecting the bearing's operating parameters, the collected bearing operating parameters are also cleaned and augmented.
7. The small-sample training method for a bearing fault diagnosis model according to claim 6, characterized in that, The data cleaning includes outlier handling and missing value handling.
8. The small-sample training method for a bearing fault diagnosis model according to claim 6, characterized in that, The data enhancements include sliding window enhancement and resampling enhancement.
9. The small-sample training method for a bearing fault diagnosis model according to claim 6, characterized in that, The bearing's operating parameters include sampling frequency, sampling time, and vibration signal.
10. A bearing fault diagnosis method, characterized in that, A bearing fault diagnosis model trained using the small-sample training method of any one of claims 1 to 9 is used to diagnose faulty bearings.