Small sample fault diagnosis method based on multi-scale integrated lightgbm

Abstract

This invention discloses a small-sample fault diagnosis method based on multi-scale ensemble LightGBM, comprising the following steps: acquiring vibration acceleration signals of rotating machinery and extracting samples to obtain training and test sets; constructing a multi-scale ensemble model consisting of a data augmentation module, a label classification module, and an output weighting module; in the data augmentation module, resampling the original samples using a multi-scale sliding window to obtain multiple sub-sample training and sub-sample test sets; in the training phase, training multiple classifiers in the label classification module using the sub-sample training sets; in the testing phase, inputting the sub-sample test sets into the corresponding classifiers, and obtaining the test results of the label classification module on the original samples through ensemble; calculating the frequency domain similarity between the test set and the training set samples using a metric function, and using it as a weight; in the output weighting module, weighting the ensembled test results to obtain the final classification diagnosis result.

Description

Technical Field

[0001] This invention belongs to the field of intelligent fault diagnosis of rotating machinery, and specifically relates to a small-sample fault diagnosis method based on multi-scale integrated LightGBM. Background Technology

[0002] With the continuous advancement of modern industrial technology, mechanical equipment is rapidly developing towards larger scale, greater complexity, and greater intelligence. However, mechanical parts inevitably suffer damage during operation, leading to equipment failure. These failures not only delay production processes but can also cause serious safety accidents, resulting in significant losses for enterprises and society. To ensure the stable and efficient operation of mechanical equipment, intelligent fault diagnosis technology has received considerable attention in recent years as a viable solution and has become an indispensable technology in the field of intelligent manufacturing. As a core component of mechanical equipment, rotating machinery typically operates in harsh environments such as high temperatures and heavy loads, making it more susceptible to damage compared to other parts. Therefore, conducting research on intelligent fault diagnosis for rotating machinery is of significant and far-reaching importance for ensuring the stable and efficient operation of mechanical equipment and reducing or avoiding major accidents.

[0003] Intelligent fault diagnosis of machinery usually requires a large amount of data to establish accurate and reliable machine learning models. However, in actual industrial environments, it is often difficult to obtain sufficient fault data on rotating machinery, mainly due to the following reasons: (1) The degradation of rotating machinery takes months or even years, and the probability of its failure is low, especially for some precision and complex mechanical equipment, such as aircraft engines; (2) When a fault occurs, the mechanical equipment will be required to stop immediately to ensure safety, at which time the sensors can only collect very limited fault data; (3) The labeling of fault data requires high costs, including expenditures on manpower, time and economic resources.

[0004] To address the challenge of scarce fault data, several methods suitable for small-sample fault diagnosis have been developed. Among these, methods based on transfer learning and meta-learning have become a research focus in this field in recent years. For example, the following transfer learning-based small-sample fault diagnosis method utilizes a data-rich source domain to assist model training, transferring universal knowledge from the source domain to the data-scarce target domain, thereby effectively improving the model's fault diagnosis performance in small-sample target domains.

[0005] Shi Jingwen, Hou Liqun. Bearing fault diagnosis based on one-dimensional convolutional attention-gated recurrent network and transfer learning [J]. Vibration and Shock, 2023, 42(03):159-164.

[0006] While transfer-based or meta-learning methods have proven effective in solving small-sample fault diagnosis problems, they still require a sufficient amount of labeled fault data to construct the source domain for model training. Furthermore, the model's diagnostic accuracy in the target domain is affected by the similarity in data distribution between the source and target domains. In real-world industrial environments, obtaining a large amount of fault data similar to the small-sample target domain to construct the source domain is undoubtedly extremely difficult. On the other hand, methods that use manual simulation to obtain fault data also incur significant economic and time costs, and the inevitable differences between artificial and real faults make it difficult to guarantee the model's diagnostic performance for real-world faults. Summary of the Invention

[0007] To address the problems of existing technologies, this invention proposes a small-sample fault diagnosis method for rotating machinery based on multi-scale ensemble LightGBM. This method consists of a data augmentation module, a label classification module, and an output weighting module. In the data augmentation module, a multi-scale sliding window is used to resample the original samples, obtaining a large number of sub-samples with different scale information, thereby expanding the data and avoiding overfitting during model training. In the label classification module, the augmented data is used to train multiple LightGBM classifiers, which are then further integrated into a strong classification model. Through ensemble processing, the diagnostic errors generated by each LightGBM classifier are weakened, enabling the ensemble model to exhibit better diagnostic and generalization performance. In the output weighting module, a weighting strategy based on similarity metrics is used to weight the diagnostic results output by the ensemble model, further improving diagnostic accuracy. This invention does not rely on additional cross-domain knowledge and only requires a small number of labeled samples to train the model, achieving effective small-sample fault diagnosis for rotating machinery.

[0008] The present invention is achieved by at least one of the following technical solutions.

[0009] A small-sample fault diagnosis method based on multi-scale ensemble LightGBM includes the following steps:

[0010] Step 1: Dataset Acquisition and Labeling: Acquire vibration acceleration signals of rotating machinery under a specific experimental condition. Extract data segments from the vibration acceleration signals to form samples, label them, and select several samples to form a training set. The remaining samples constitute the test set. Where, x i and y i Let represent the i-th sample and its corresponding label, k be the number of samples in the training set, and m be the total number of samples;

[0011] Step 2: Construct a few-sample fault diagnosis model based on ensemble learning, including a data augmentation module M. ATag Classification Module M C and output weighting module M W Among them, the data augmentation module M A Used for multi-scale resampling of samples to achieve data augmentation; Label classifier module M C Used for classifying and diagnosing samples, and outputting corresponding category prediction vectors; output weighting module M W This is used to weight the predicted vectors for the above categories in order to improve the accuracy of classification diagnosis;

[0012] Step 3, Data Augmentation: Augment the training set D constructed in Step 1... train and test set D test Input data augmentation module M respectively A The original samples are resampled using a multi-scale sliding window to obtain a large number of subsamples of different scales, and correspondingly, multiple training sets of subsamples with different scale information are constructed. and subsample test set in, and Let T represent the training set and test set of subsamples with the j-th scale information, respectively, and T represent the number of scale categories.

[0013] Step 4, Model Training: In the label classification module M C For each subsample training set obtained in step 3, a set of label classifiers based on Lightweight Gradient Boosting Machine (LightGBM) is trained using the gradient boosting decision tree algorithm. A total of T sets are trained. Each set of label classifiers contains N LightGBM models for binary classification, where N is the number of label categories.

[0014] Step 5, Model Testing: In the label classification module M C Each subsample test set obtained in step 3 is input into the corresponding label classifier. The label classifier outputs the test results. The test results are summed, averaged, and concatenated to obtain T groups of N-dimensional category prediction vectors. The category prediction vectors of each group are further summed and averaged to obtain the integrated category prediction vector, which is the label classification module M. C For test set D test The output of the diagnostic results;

[0015] Step 6, Weight Calculation: Use Fast Fourier Transform to calculate the weights of the training set D obtained in Step 3. train and test set D test The samples in the dataset are transformed from the time domain to the frequency domain to obtain a new frequency domain training set D. train-f and frequency domain test set D test-f , frequency domain training set D train-fThe samples in the dataset are summed and averaged according to their categories to obtain N category centers. The cosine similarity function is used as the metric function to calculate the frequency domain test set D. test-f The similarity between the samples and the centers of each category is used to obtain an N-dimensional metric vector. The metric vector is then normalized to obtain an N-dimensional weight vector.

[0016] Step 7, Output Weighting: In the output weighting module M C The integrated N-dimensional category prediction vectors obtained in steps 5 and 6 are multiplied component by component with the corresponding N-dimensional weight vectors to obtain the weighted category prediction vectors. The category corresponding to the largest vector component is the overall model's prediction for the test set D. test The final diagnosis result.

[0017] Furthermore, in step 1, when extracting samples from the vibration acceleration signal, there is no overlap between the data points of each sample. Additionally, the training set D... train and test set D test The data features have the same distribution and the same label space.

[0018] Furthermore, in step 2, the data augmentation module M A Data augmentation is achieved using a multi-scale sliding window resampling method in the label classification module M. C Includes multiple LightGBM models for binary classification, with output weighting module M. W A similarity measurement strategy is used to weight the category prediction vectors.

[0019] Furthermore, in step 3, when resampling the original samples using a multi-scale sliding window, the window scale is set to the length of the original sample data points. Where T is the number of categories at the scale, and the window's sliding step size is set to 1 to obtain more subsamples, and the categories of the subsamples are consistent with the categories of the original samples.

[0020] Furthermore, in step 4, each group of label classifiers has the same function, performing an N-class classification task. Each LightGBM is used to perform binary classification of a subsample of a certain class. For the subsample training set input to the q-th LightGBM in each group, subsamples belonging to the q-th class are considered positive and labeled 1, while subsamples belonging to the other N-1 classes are considered negative and labeled 0. Each LightGBM is trained using the gradient boosting decision tree algorithm, and its loss function is expressed as:

[0021]

[0022] In the formula, L represents the loss function, and n represents the total number of subsamples used for training. and represents the true value and predicted value of the label for the h-th subsample, respectively. The larger the predicted value, the greater the probability that the subsample belongs to the corresponding positive class.

[0023] Furthermore, in step 5, the subsample test set with the j-th scale information We will use a training set of subsamples with j-th scale information. The trained set of label classifiers is used for testing.

[0024] Further, in step 5, when the subsample test set is input into the label classifier of the corresponding group for testing, the test results of each LightGBM on the subsample will be obtained. By summing and averaging the test results, the test result of each LightGBM on the original sample is obtained. After concatenating the test results of all LightGBMs on the original sample in each label classifier in order, the category prediction vector of each label classifier on the original sample is obtained. After further summing and averaging the category prediction vectors of each group, the integrated category prediction vector is obtained, which is expressed as:

[0025]

[0026] In the formula, P(x) represents the label classification module M. C The ensembled class prediction vector of the output of the original sample x, p N (x) represents the Nth component of the class prediction vector P(x). and n j Let f represent the z-th subsample with the j-th scale information and the total number of subsamples with the j-th scale information, respectively. j,N This represents the Nth LightGBM of the jth scale in the jth label classifier group, where N is the number of label categories, i.e. the number of LightGBMs in each label classifier group, and T is the number of scale categories, i.e. the number of label classifier groups.

[0027] Furthermore, in step 6, the frequency domain training set D train-f and frequency domain test set D test-f The samples in the dataset are the time-domain training set D. train and time-domain test set D test The original sample is represented in the frequency domain by a spectrum, and its data point length is half that of the original sample.

[0028] Furthermore, in step 6, the frequency domain test set D is calculated using the cosine similarity function. test-fThe similarity between the samples and the centers of each category is calculated by concatenating the calculated similarity scores to obtain an N-dimensional metric vector. This metric vector is then normalized using a Softmax function with an adjustment factor, resulting in the following weight vector:

[0029]

[0030] In the formula, W(x) represents the calculated weight vector for the original sample x, w N (x) represents the Nth component of the weight vector W(x), c q Let S(x,c) represent the q-th category center. q S(x,c) represents the cosine similarity between the original sample x and the center of the q-th class. N ) represents the cosine similarity between the original sample x and the center of the Nth category, and λ is an adjustment factor used to control the differences between the components in the normalized weight vector.

[0031] Furthermore, in step 7, the weighted category prediction vector is represented by the following formula:

[0032] P W (x)=[p1(x)w1(x),p2(x)w2(x),…,p N (x)w N (x)] (4)

[0033] In the formula, P W (x) represents the weighted class prediction vector output by the overall model for the original sample x, p N (x) and w N (x) represents the Nth component in the category prediction vector and weight vector of the original sample x, respectively.

[0034] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0035] 1. This invention employs a multi-scale sliding window to resample the original samples, generating a large number of sub-samples to augment the data. Training with these sub-samples effectively avoids overfitting caused by a lack of training data. Furthermore, since the sub-samples originate from the original samples and are numerous, they effectively extract the information contained within the original samples, maximizing the utilization of data resources in small-sample scenarios.

[0036] 2. This invention constructs multiple sets of label classifiers, each set consisting of multiple LightGBM models used for binary classification, to achieve multi-class fault diagnosis capabilities. Simultaneously, these classifier sets are further integrated into a single strong classification model to achieve effective few-sample fault diagnosis. Through integration, the diagnostic errors generated by each classifier set are reduced, enabling the integrated model to possess better few-sample fault diagnosis performance and generalization performance.

[0037] 3. This invention designs a weighting strategy based on frequency domain similarity measurement, which makes full use of the spectral information of vibration acceleration signals and integrates metric learning technology to calculate the weight of each sample, and weights the diagnostic results output by the ensemble model, thereby constructing a data mining system of "time domain training and frequency domain weighting", which further improves the small sample fault diagnosis performance of the ensemble model.

[0038] 4. The LightGBM classifier used in this invention is built based on the gradient boosting decision tree algorithm, which has fewer hyperparameters and higher training efficiency than deep neural networks. Furthermore, this classifier can also be accelerated using GPUs, making it easy to deploy on high-performance GPU servers.

[0039] 5. The method proposed in this invention fully considers the actual situation of industrial applications. It can achieve small-sample fault diagnosis of rotating machinery using only a small amount of labeled data without relying on cross-domain knowledge assistance. While ensuring high accuracy, it avoids the cost of constructing the source domain. Because it is more efficient and convenient, it has certain industrial application value.

[0040] The method proposed in this invention can achieve efficient small-sample fault diagnosis of rotating machinery using only a small amount of labeled data without relying on cross-domain knowledge assistance. While ensuring high accuracy, it avoids the cost of constructing a source domain, thus being more efficient and convenient, and possessing certain industrial application value. This demonstrates the rationality of applying ensemble learning to small-sample fault diagnosis, providing another feasible and efficient technical solution for solving the problem of small-sample fault diagnosis of rotating machinery in industrial practice. Attached Figure Description

[0041] Figure 1 This is a flowchart of the fault diagnosis method based on multi-scale integrated LightGBM in this invention.

[0042] Figure 2 This is a schematic diagram of the model structure of the small-sample fault diagnosis method based on multi-scale integrated LightGBM of the present invention;

[0043] Figure 3 This is a schematic diagram of the weight calculation process of the method of the present invention;

[0044] Figure 4 This is a test accuracy diagram of the method of the present invention under working conditions;

[0045] Figure 5 This is a test accuracy diagram of the method of the present invention under working condition two;

[0046] Figure 6 This is a test accuracy diagram of the method of the present invention under working condition three;

[0047] Figure 7 This is a test accuracy diagram of the method of the present invention under working condition four. Detailed Implementation

[0048] To more clearly illustrate the purpose and technical solution of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for better illustrating the application of this invention, but the technical features involved in the implementation of this invention are not limited thereto.

[0049] like Figure 1 As shown, a small-sample fault diagnosis method based on multi-scale ensemble LightGBM is presented. This method can achieve efficient small-sample fault diagnosis of rotating machinery using only a small number of labeled samples without relying on cross-domain knowledge assistance. The method includes the following steps:

[0050] Step 1: Dataset Acquisition and Labeling. Vibration acceleration signals of the rotating machinery under a specific experimental condition are collected. Several samples are extracted according to a certain data point length and labeled. A small number of these samples (e.g., 1-10 per class) are selected to form the training set. The remaining large number of samples constitute the test set. Where, x i and y i Let represent the i-th sample and its corresponding label, k be the number of samples in the training set, and m be the total number of samples;

[0051] When extracting samples from the vibration acceleration signal, there is no overlap between the data points of each sample. Furthermore, the training set D... train and test set D test The data features have the same distribution and the same label space.

[0052] Step 2: Construct a few-shot fault diagnosis model based on ensemble learning, such as... Figure 2 As shown. The model includes a data augmentation module M. A Tag Classification Module M C and output weighting module M W Among them, the data enhancement module M A Used for multi-scale resampling of samples to achieve data augmentation; Label classifier module M CUsed for classifying and diagnosing samples, and outputting corresponding category prediction vectors; output weighting module M W This is used to weight the aforementioned category prediction vectors to improve the accuracy of classification diagnosis.

[0053] Data Augmentation Module M A The label classification module M is implemented based on a multi-scale sliding window resampling method. C This implementation is based on multiple binary classification LightGBM models, with the output weighting module M. W It is implemented based on the principle of similarity measurement.

[0054] Step 3, Data Augmentation. The training set D constructed in Step 1... train and test set D test Input data augmentation module M respectively A The original samples are resampled using a multi-scale sliding window to obtain a large number of subsamples of different scales, and correspondingly, multiple training sets of subsamples with different scale information are constructed. and subsample test set in, and Let T represent the training set and test set of subsamples with the j-th scale information, respectively, and T represent the number of scale categories.

[0055] Step 4: Model Training. In the label classification module M... C For each sub-sample training set obtained in step 3, a set of label classifiers based on Lightweight Gradient Boosting Machine (LightGBM) is trained using the gradient boosting decision tree algorithm, for a total of T sets. Each set of label classifiers contains N LightGBM models for binary classification, where N is the number of label categories. Therefore, each set of label classifiers has the ability to diagnose faults in N categories.

[0056] Step 5: Model Testing. In the label classification module M... C Each subsample test set obtained in step 3 is input into the corresponding label classifier. The label classifier outputs the test results. The test results are summed, averaged, and concatenated to obtain T groups of N-dimensional category prediction vectors. The category prediction vectors of each group are further summed and averaged to obtain the integrated category prediction vector, which is the label classification module M. C For test set D test The output of the diagnostic results;

[0057] Step 6: Weight calculation. For example... Figure 3 As shown, the training set D obtained in step 3 is transformed using the Fast Fourier Transform technique. train and test set D testThe samples in the dataset are transformed from the time domain to the frequency domain to obtain the frequency domain training set D. train-f and frequency domain test set D test-f The frequency domain training set D train-f The samples in the dataset are summed and averaged according to their categories to obtain N category centers. Then, the cosine similarity function is used as the metric to calculate the frequency domain test set D. test-f The similarity between the samples and the centers of each category is used to obtain an N-dimensional metric vector. The metric vector is then normalized to obtain an N-dimensional weight vector.

[0058] Step 7: Output Weighting. In the output weighting module M... C The integrated N-dimensional category prediction vectors obtained in steps 5 and 6 are multiplied component by component with the corresponding N-dimensional weight vectors to obtain the weighted category prediction vectors. The category corresponding to the largest vector component is the overall model's prediction for the test set D. test The final diagnosis result.

[0059] As one embodiment, in step 3, when resampling the original samples using a multi-scale sliding window, the window scale is set to the length of the original sample data points. Where T is the number of categories at the scale, and the window's sliding step size is set to 1 to obtain more subsamples, and the categories of the subsamples are consistent with the categories of the original samples.

[0060] Further, in step 4, each group of label classifiers functions identically, performing an N-class objective task. Each LightGBM is used to perform binary classification of a subsample belonging to a specific class. For example, the q-th LightGBM in each group is used to perform binary classification of subsamples belonging to the q-th class. For the subsample training set input to the q-th LightGBM, subsamples belonging to the q-th class are considered positive and labeled 1, while subsamples belonging to the other N-1 classes are considered negative and labeled 0. Each LightGBM is trained using the gradient boosting decision tree algorithm, with its loss function expressed as:

[0061]

[0062] In the formula, L represents the loss function, and n represents the total number of subsamples used for training. and represents the true value and predicted value of the label for the h-th subsample, respectively. The larger the predicted value, the greater the probability that the subsample belongs to the corresponding positive class.

[0063] Furthermore, in step 5, the subsample test set with the j-th scale information We will use a training set of subsamples with j-th scale information. The trained set of label classifiers is used for testing.

[0064] Further, in step 5, when the subsample test set is input into the label classifier of the corresponding group for testing, the test results of each LightGBM on the subsample will be obtained. By summing and averaging the test results, the test result of each LightGBM on the original sample is obtained. The test results of all LightGBMs on the original sample in each group of label classifiers are concatenated to obtain the category prediction vector output by each group of label classifiers on the original sample. The category prediction vectors of each group are summed and averaged to obtain the integrated category prediction vector, which is expressed as:

[0065]

[0066] In the formula, P(x) represents the label classification module M. C The ensembled class prediction vector of the output of the original sample x, p N (x) represents the Nth component of the class prediction vector P(x). and n j Let f represent the z-th subsample with the j-th scale information and the total number of subsamples with the j-th scale information, respectively. j,N This represents the Nth LightGBM of the j-th scale in the j-th label classifier group, where N is the number of label categories (i.e., the number of LightGBMs in each label classifier group) and T is the number of scale categories (i.e., the number of label classifier groups).

[0067] Furthermore, in step 6, the frequency domain training set D train-f and frequency domain test set D test-f The samples in the dataset are the time-domain training set D. train and time-domain test set D test The original sample is represented in the frequency domain by a spectrum, and its data point length is half that of the original sample.

[0068] Furthermore, in step 6, the frequency domain test set D is calculated using the cosine similarity function. test-f The similarity between the samples and the center of each category is calculated. By concatenating the calculated similarities, an N-dimensional metric vector is obtained. Subsequently, the metric vector is normalized using a Softmax function with an adjustment factor, resulting in the following weight vector:

[0069]

[0070] In the formula, W(x) represents the calculated weight vector for the original sample x, w N (x) represents the Nth component of the weight vector W(x), c qLet S(x,c) represent the q-th category center. q S(x,c) represents the cosine similarity between the original sample x and the center of the q-th class. N ) represents the cosine similarity between the original sample x and the center of the Nth category, and λ is an adjustment factor used to control the difference between the components in the normalized weight vector, which is usually set to 10.

[0071] Furthermore, in step 7, the weighted category prediction vector can be represented by the following formula:

[0072] P W (x)=[p1(x)w1(x),p2(x)w2(x),…,p N (x)w N (x)] (4)

[0073] In the formula, P W (x) represents the weighted class prediction vector output by the overall model for the original sample x, p N (x) and w N (x) represents the Nth component in the category prediction vector and weight vector of the original sample x, respectively.

[0074] The invention will be further described below with reference to the accompanying drawings and experimental examples:

[0075] To verify the superiority of the proposed method, eight other fault diagnosis methods were compared, including: LightGBM, Random Forest (RF), Multi-granularity Cascaded Forest (gcForest), First-Layer Wide Convolutional Deep Neural Network (WDCNN), WDCNN+Fine-tuning (WDCNN-FT), Prototype Network (PN), Matching Network (MN), and Model-Independent Meta-Learning (MAML). Among these, LightGBM, RF, gcForest, WDCNN, and MWE-LightGBM are all standard supervised learning methods; WDCNN-FT is a transfer learning method using WDCNN as the baseline model; and PN, MN, and MAML are all meta-learning methods, also using WDCNN as the baseline model. Comparing multiple types of methods makes the experimental results more convincing. It should be noted that for transfer learning and meta-learning methods, additional source domain labeled samples will be provided to assist model training.

[0076] As one example, the publicly available University of Paderborn (PU) bearing dataset was used to design related experiments. This dataset consists of 32 rolling bearing datasets, including 6 normal bearing datasets, 12 artificially faulted bearing datasets, and 14 real-world faulted bearing datasets. Artificial faults were generated through electrical discharge machining (EDM), drilling, and electro-engraving, while real-world faults such as pitting and plastic deformation were generated through accelerated bearing life experiments. The dataset uses piezoelectric accelerometers to collect bearing vibration signals under the four operating conditions shown in Table 1, with a sampling frequency of 64 kHz.

[0077] Table 1 Experimental conditions of the PU bearing dataset

[0078]

[0079] As one embodiment, 13 bearings were selected to design a small-sample fault diagnosis task, including 1 normal bearing, 8 artificially faulted bearings, and 4 real faulted bearings. Detailed descriptions of the experimental bearings are shown in Table 2, with three fault locations: inner ring fault, outer ring fault, and combined inner and outer ring fault, and two fault severity levels: Level 1 indicates a fault size less than or equal to 2 mm, and Level 2 indicates a fault size greater than 2 mm. In this embodiment, the normal bearing and the real faulted bearings are considered as the target bearings for the small-sample fault diagnosis, with only a small number of labeled samples available for training. As shown in Table 2, the target bearings in the experiment include five health categories: normal, a real inner ring fault of severity level 1, a real outer ring fault of severity level 1, and a real combined inner and outer ring fault of severity levels 1 and 2. The artificially faulted bearings will be used to provide the source domain for the aforementioned transfer learning and meta-learning methods to assist model training, containing a large number of labeled samples.

[0080] Table 2 Introduction to Experimental Bearings

[0081]

[0082] Non-overlapping sliding windows were used to capture bearing vibration signals and collect samples. 120 samples were collected for each experimental bearing under four different operating conditions, with each sample containing 2048 vibration signal data points. Then, a 5-way K-shot (K=1, 2, 5, 10) small-sample fault diagnosis task was constructed for each of the four operating conditions. Specifically, 1, 2, 5, and 10 samples from each target bearing were randomly selected for training, and the remaining samples were used for testing. For transfer learning and meta-learning methods, an additional 120 samples collected from each artificially faulty bearing were used to construct the source domain. Simultaneously, the model's input samples were transformed from the time domain to the frequency domain to improve the training effect and diagnostic performance of these methods. Table 3 shows the number of samples used for each comparative method. To reduce randomness, 10 repeated experiments were conducted, and the average of the experimental results was used to evaluate the diagnostic performance of each comparative method. The hyperparameter settings of the proposed method are shown in Table 4.

[0083] The results of small-sample fault diagnosis experiments under four operating conditions are as follows: Figure 4 , Figure 5 , Figure 6 and Figure 7 As shown. Experimental results show that LightGBM, RF, and WDCNN achieve low test accuracy in all few-shot fault diagnosis tasks because these methods do not allow the use of source domains or data augmentation strategies, leading to overfitting under few-shot conditions. In contrast, transfer learning and meta-learning methods (such as WDCNN-FT, PN, MN, and MAML) effectively solve the overfitting problem by using source domains to assist model training. For example, in 1-shot and 5-shot tasks, WDCNN-FT achieved average test accuracy of 78.35% and 88.54%, respectively, which is 47.22% and 42.27% higher than the baseline model WDCNN. However, such a comparison is not fair because standard supervised learning methods can only be trained with a small number of labeled samples. Nevertheless, the method proposed in this invention, which is also based on standard supervised learning, achieves the highest test accuracy in all few-shot fault diagnosis tasks. Specifically, in the 1-shot task under condition 1, as shown... Figure 4 As shown, the proposed method achieves a test accuracy of 97.19%, which is 17.81%, 5.13%, 4.28%, and 6.81% higher than WDCNN-FT, PN, MN, and MAML, respectively. Furthermore, since the proposed method does not rely on additional cross-domain knowledge, it avoids the cost of constructing the source domain while maintaining high diagnostic accuracy, making it more efficient and convenient, and significantly superior to other comparative methods.

[0084] Another noteworthy comparative method is gcForest. From Figure 4As can be seen, gcForest's test accuracy in 1-shot and 5-shot tasks is close to that of the proposed method. However, in 1-shot and 2-shot tasks, gcForest's test accuracy is significantly lower than that of the proposed method. Although both methods are based on multi-scale ensembles and employ data augmentation strategies, their classification performance still differs significantly. This may be because gcForest's ensemble is mainly based on the "depth" level, which is prone to overfitting when the number of training samples is extremely scarce (e.g., only 1 or 2 samples per class). Conversely, the proposed method's ensemble is based on the "width" level, and the model does not have a deep structure. Therefore, it only requires a small number of training samples to achieve a good fitting effect, making it more suitable for scenarios with scarce data than gcForest.

[0085] The method proposed in this invention first resamples the original samples using a multi-scale sliding window to augment the data and prevent overfitting during model training when diagnosing small-sample faults. Then, the augmented data is used to train multiple LightGBM classifiers, which are further integrated into a strong classification model. Through ensemble analysis, the diagnostic errors generated by each LightGBM classifier are reduced, allowing the ensemble model to exhibit better diagnostic and generalization performance. Furthermore, during the testing phase, a weighting strategy based on similarity metrics is used to weight the diagnostic results output by the ensemble model, further improving diagnostic accuracy. The results of the above experimental cases fully verify the effectiveness and advancement of the proposed method for diagnosing small-sample faults in rotating machinery.

[0086] It should be noted that although the implementation of the present invention has been described in detail with reference to examples, those skilled in the art should understand that any modifications, substitutions, and improvements made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A small-sample fault diagnosis method based on multi-scale ensemble LightGBM, characterized in that, Includes the following steps: Step 1: Dataset Acquisition and Labeling: Acquire vibration acceleration signals of rotating machinery under a specific experimental condition. Extract data segments from the vibration acceleration signals to form samples, label them, and select several samples to form a training set. The remaining samples constitute the test set. ,in, and They represent the first i Each sample and its corresponding label k This represents the number of samples in the training set. m The total number of samples; Step 2: Construct a few-sample fault diagnosis model based on ensemble learning, including a data augmentation module. Tag classification module and output weighting module Among them, the data enhancement module Used for multi-scale resampling of samples to achieve data augmentation; Label classifier module Used for classifying and diagnosing samples, and outputting corresponding category prediction vectors; output weighting module This is used to weight the predicted vectors for the above categories in order to improve the accuracy of classification diagnosis; Step 3, Data Augmentation: Augment the training set constructed in Step 1. and test set Input data augmentation module respectively The original samples are resampled using a multi-scale sliding window to obtain a large number of subsamples of different scales, and correspondingly, multiple training sets of subsamples with different scale information are constructed. and subsample test set ,in, and They respectively represent having the first j Subsample training set and subsample test set with scale information T The number of categories representing the scale; Step 4, Model Training: In the label classification module For each sub-sample training set obtained in step 3, a set of label classifiers based on Lightweight Gradient Boosting Machine (LightGBM) is trained using the gradient boosting decision tree algorithm, for a total of [number missing] training sessions. T Groups, where each group of label classifiers contains N A LightGBM model for binary classification, N This refers to the number of tag categories; Step 5, Model Testing: In the label classification module Each subsample test set obtained in step 3 is input into the corresponding group's label classifier. The label classifiers of each group output the test results. The test results are then summed, averaged, and concatenated to obtain the final result. T Group N The category prediction vectors of each group are further summed and averaged to obtain the integrated category prediction vector, which is the label classification module. For the test set The output of the diagnostic results; Step 6, Weight Calculation: Use Fast Fourier Transform to calculate the weights of the training set obtained in Step 3. and test set The samples in the dataset are transformed from the time domain to the frequency domain to obtain a new frequency domain training set. and frequency domain test set , frequency domain training set The samples in the sample are summed by category and averaged to obtain N For each category center, the cosine similarity function is used as the metric to calculate the frequency domain test set. The similarity between the samples and the centers of each category is obtained. N The metric vector is then normalized to obtain the final result. N A weight vector of dimension; Step 7, Output Weighting: In the output weighting module The integrated results obtained in steps 5 and 6 N Dimensional category prediction vector and corresponding N The weight vectors are multiplied component by component to obtain a weighted class prediction vector. The class corresponding to the largest vector component is the overall model's prediction for the test set. The final diagnosis result.

2. The small-sample fault diagnosis method based on multi-scale integrated LightGBM according to claim 1, characterized in that, In step 1, when extracting samples from the vibration acceleration signal, there is no overlap between the data points of each sample. Furthermore, the training set... and test set The data features have the same distribution and the same label space.

3. The small-sample fault diagnosis method based on multi-scale integrated LightGBM according to claim 1, characterized in that, In step 2, the data augmentation module Data augmentation is achieved using a multi-scale sliding window resampling method in the label classification module. Includes multiple LightGBM models for binary classification, with output weighting modules. A similarity measurement strategy is used to weight the category prediction vectors.

4. The small-sample fault diagnosis method based on multi-scale integrated LightGBM according to claim 1, characterized in that, In step 3, when resampling the original samples using a multi-scale sliding window, the window scale is set to the length of the original sample data points. ,in T The number of categories is set to a scale, and the window's sliding step is set to 1 to obtain more subsamples, whose categories remain consistent with those of the original samples.

5. The small-sample fault diagnosis method based on multi-scale integrated LightGBM according to claim 1, characterized in that, In step 4, each group of label classifiers has the same function, and they are all used to perform... N The objective task is classification, where each LightGBM is used to perform binary classification for a subsample of a certain class; for each input group... q The sub-sample training set of the LightGBM belongs to the first... q Subsamples belonging to a category will be treated as positive and labeled as 1, while those belonging to other categories will be treated as positive. Subsamples of each category will be treated as the negative class and labeled 0. Each LightGBM is trained using the gradient boosting decision tree algorithm, with the loss function expressed as: In the formula, L Represents the loss function. n This represents the total number of subsamples used for training. and They represent the first i The true and predicted values ​​of the labels for each subsample are compared. The larger the predicted value, the greater the probability that the subsample belongs to the corresponding positive class.

6. The small-sample fault diagnosis method based on multi-scale integrated LightGBM according to claim 1, characterized in that, In step 5, the one with the j Subsample test set of seed scale information Will use by having the first j Subsample training set of seed scale information The trained set of label classifiers is used for testing.

7. The small-sample fault diagnosis method based on multi-scale integrated LightGBM according to claim 1, characterized in that, In step 5, when the subsample test set is input into the label classifier of the corresponding group for testing, the test results of each LightGBM on the subsample are obtained. By summing and averaging the test results, the test result of each LightGBM on the original sample is obtained. After concatenating the test results of all LightGBMs on the original sample in each label classifier in order, the category prediction vector of each label classifier on the original sample is obtained. After further summing and averaging the category prediction vectors of each group, the integrated category prediction vector is obtained, which is represented as follows: In the formula, Represents the tag classification module For the original sample x The output of the integrated class prediction vector, Represents the category prediction vector The N One portion, and These represent the original samples respectively. x In terms of having the first j The first type of scale information z Subsamples and having the first j The total number of subsamples for each scale of information. Indicates the first j The first in the group label classifier N The first LightGBM j Scale information, N This refers to the number of label categories, i.e., the number of LightGBM tags in each label classifier group. T The number of categories is the scale, i.e., the number of groups of label classifiers.

8. The small-sample fault diagnosis method based on multi-scale integrated LightGBM according to claim 1, characterized in that, In step 6, the frequency domain training set and frequency domain test set The samples in the dataset are time-domain training sets. and time-domain test set The original sample is represented in the frequency domain by a spectrum, and its data point length is half that of the original sample.

9. The small-sample fault diagnosis method based on multi-scale integrated LightGBM according to claim 1, characterized in that, In step 6, the cosine similarity function is used to calculate the frequency domain test set. The similarity between the samples and the centers of each category is obtained by concatenating the calculated similarity scores. N The metric vector is normalized using a Softmax function with an adjustment factor, resulting in the following weight vector: In the formula, This indicates the calculated result for the original sample. x The weight vector, Represents the weight vector The N One portion, Indicates the first q Each category center Represents the original sample x With the q The cosine similarity value of each category center is calculated. Represents the original sample x With the N The cosine similarity value of each category center is calculated. This is an adjustment factor used to control the differences among the components in the normalized weight vector.

10. The small-sample fault diagnosis method based on multi-scale integrated LightGBM according to claim 1, characterized in that, In step 7, the weighted category prediction vector is represented by the following formula: In the formula, This indicates that the overall model is based on the original samples. x The output weighted class prediction vector, and These represent the original samples respectively. x The first category prediction vector and the weight vector N Each component.

Images