An improved smote-based industrial unbalanced data classification method

Abstract

The application discloses an industrial unbalanced data classification method based on an improved SMOTE, and comprises the following steps: firstly, obtaining sample monitoring record data of industrial equipment to obtain an unbalanced data set; then, processing the unbalanced data set by using the improved SMOTE method to obtain a balanced data set; then, training a classifier by taking the balanced data set as a training data set to obtain a trained classifier; finally, obtaining real-time monitoring record data of the industrial equipment, and sending a feature vector of the real-time monitoring record data into the trained classifier to obtain a state of the industrial equipment. The application effectively guarantees the quality of the synthesized minority class samples through a scoring mechanism, and effectively makes the number of the minority class samples relatively balanced through sparsity, so that the classifier can pay attention to the fault data samples, and thus the situation that the classifier classifies the fault data samples as normal data samples due to the too small number of the fault data samples can be avoided as much as possible.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment fault diagnosis technology, and specifically to an industrial imbalance data classification method based on an improved SMOTE. Background Technology

[0002] In complex industrial production environments, industrial equipment inevitably experiences various malfunctions. Most equipment failures are preceded by warning signs, such as vibration signals, lubricating oil temperature signals, and abnormal noise signals before bearing failure. As equipment transitions from a normal state to a malfunction, its status signals undergo a process of development and change. Therefore, by analyzing these changes in status information, it is possible to determine whether a malfunction has occurred and to determine the location, nature, extent, and development trend of the fault—this is the foundation of equipment condition management. To correctly guide the maintenance and repair of industrial equipment and ensure its operational capabilities, industrial managers need to monitor equipment operating status data in real time and effectively, and derive the equipment's status results through analysis and processing of this monitoring data.

[0003] Since equipment operates in normal conditions most of the time in industrial production, with very low failure rates, the data monitored by equipment sensors is mostly normal state data, with very little abnormal state data. This indicates a certain degree of data imbalance in equipment monitoring data. Traditional machine learning models, during training, generally favor the majority class samples, resulting in very low accuracy for the minority class samples. Therefore, they are unsuitable for imbalanced data, making imbalanced data processing a crucial issue that must be considered in equipment fault detection based on industrial big data. Summary of the Invention

[0004] The present invention aims to solve the problem of low accuracy in equipment fault diagnosis caused by the imbalance between normal and fault data, and provides an industrial imbalance data classification method based on an improved SMOTE.

[0005] To solve the above problems, the present invention is achieved through the following technical solution:

[0006] An industrial imbalance data classification method based on an improved SMOTE includes the following steps:

[0007] Step 1: Obtain sample monitoring record data of industrial equipment, extract features from the sample monitoring record data to obtain feature vectors of the sample monitoring record data, and then label the feature vectors of the sample monitoring record data to obtain an imbalanced dataset.

[0008] Step 2: Use the improved SMOTE method to process the imbalanced dataset to obtain a balanced dataset;

[0009] Step 2.1: Define the feature vectors labeled as faulty in the imbalanced dataset as minority class samples, and the feature vectors labeled as normal as majority class samples; set the initial number of minority class samples N. c The number of minority class samples N generated again D The number of nearest neighbor samples k and the number of clusters k ′ ;

[0010] Step 2.2: First, randomly select a minority class sample x from the imbalanced dataset; then calculate the Euclidean distance between the minority class sample x and other minority class samples in the imbalanced dataset, and obtain the k nearest neighbors of the minority class sample x based on the Euclidean distance; finally, perform interpolation on each nearest neighbor of the minority class sample x to obtain a new minority class sample x corresponding to each nearest neighbor. new ;

[0011] Step 2.3: First calculate x for each new minority class sample. new Calculate the Euclidean distances to other minority class samples and all majority class samples in the imbalanced dataset, and obtain a new minority class sample x based on the proximity of the Euclidean distances. new Find the k nearest neighbors; then determine the new minority class sample x. new Is each nearest neighbor a minority or majority class sample? If a new minority class sample x new If a new minority class sample x is a minority class sample, then the weight of that neighbor sample is set to 1; if a new minority class sample x... new If the nearest neighbor sample is a majority class sample, then the weight of that nearest neighbor sample is set to 0; then the weight of the new minority class sample x is calculated. new The sum of the weights of the k nearest neighbors, and the new minority class sample x. new Does the sum of the weights of the k nearest neighbors of x exceed k / 2? If so, retain the new minority class sample x. new Otherwise, discard the new minority class sample x. new ;

[0012] Step 2.4: Repeat steps 2.2-2.3 until a new minority class sample x is retained. new The number reaches the required number N of minority class samples. c and will this N c A new minority class sample x that is retained new As minority class samples, they are incorporated into the imbalanced dataset to obtain the initial balanced dataset;

[0013] Step 2.5: First, perform K-means clustering on the minority class samples in the initial balanced dataset, thereby dividing all minority class samples in the initial balanced dataset into k′ clusters; then, take the mean of all minority class samples in each cluster as the centroid of that cluster; finally, calculate the Euclidean distance from each majority class sample in the initial balanced dataset to the centroid of each cluster, and assign each majority class sample to the cluster with the closest Euclidean distance to its centroid.

[0014] Step 2.6: Calculate the number of majority class samples and the number of minority class samples in each cluster. If the number of majority class samples in a cluster is greater than or equal to the number of minority class samples, then discard the cluster; otherwise, retain the cluster.

[0015] Step 2.7: For each retained cluster, first calculate the Euclidean distance between every two minority class samples within the retained cluster, and construct the Euclidean distance matrix of the minority class samples within the retained cluster accordingly; then add all the off-diagonal elements in the Euclidean distance matrix of the minority class samples within the retained cluster, and divide by the total number of off-diagonal elements to obtain the average distance of the retained cluster; finally, divide the average distance of the retained cluster by the number of minority class samples within the retained cluster to obtain the sparsity of the retained cluster.

[0016] Step 2.8: Calculate the number of minority class samples N(f) generated by each retained cluster:

[0017]

[0018] In the formula, N st (f) represents the sparsity of the retained clusters f, m represents the number of retained clusters f, and N D This represents the number of minority class samples generated again.

[0019] Step 2.9: First, randomly select a minority class sample x′ from the retained clusters; then calculate the Euclidean distance between the minority class sample x′ and other minority class samples in the retained clusters, and obtain the k nearest neighbor samples of the minority class sample x′ based on the Euclidean distance; finally, perform interpolation on each nearest neighbor sample of the minority class sample x′ to obtain the new minority class sample x′ corresponding to each nearest neighbor sample. new ;

[0020] Step 2.10: First calculate each new minority class sample x′ new Calculate the Euclidean distances to other minority class samples and all majority class samples in the preserved cluster, and obtain a new minority class sample x′ based on the proximity of the Euclidean distances. new Find the k nearest neighbors; then determine the new minority class sample x′. new Is each nearest neighbor a minority or majority class sample? If a new minority class sample x′ newIf a new minority class sample x′ is a minority class sample, then the weight of that neighboring sample is set to 1; if a new minority class sample x′ is a minority class sample x′, then the weight of that neighboring sample is set to 1. new If the nearest neighbor sample is a majority class sample, then the weight of that nearest neighbor sample is set to 0; then the weight of the new minority class sample x′ is calculated. new The sum of the weights of the k nearest neighbors, and the new minority class sample x′ new Does the sum of the weights of the k nearest neighbors of x′ exceed k / 2? If so, then retain the new minority class sample x′. new Otherwise, discard the new minority class sample x′. new ;

[0021] Step 2.11: Repeat steps 2.9-2.10 until a new minority class sample x′ is retained. new The number reaches the number of minority class samples N(f) that each retained cluster is responsible for generating;

[0022] Step 2.12: Repeat steps 2.9-2.11 for each retained cluster, and then generate N from all retained clusters. D A new minority class sample x′ is retained. new The minority class samples are incorporated into the initial balanced dataset to obtain the balanced dataset;

[0023] Step 3: Use the balanced dataset as the training dataset to train the classifier and obtain the trained classifier;

[0024] Step 4: Obtain real-time monitoring data of industrial equipment, extract features from the real-time monitoring data to obtain feature vectors, and feed the feature vectors into a trained classifier to obtain the status of the industrial equipment.

[0025] In step 2.1 above,

[0026] The number of minority class samples N generated initially c for:

[0027]

[0028] The number of minority class samples N generated again D for:

[0029] N D =C major -C minor -N c

[0030] In the formula, C major C represents the number of majority class samples in an imbalanced dataset. minor p represents the number of minority class samples in an imbalanced dataset, and p is a hyperparameter that is set manually.

[0031] In step 2.2 above, the formula for the interpolation operation is:

[0032] x new = x + rand·(yx)

[0033] In the formula, rand is a random number between [0, 1], x is a randomly selected minority class sample, y is the nearest neighbor sample of the minority class sample x, and x new This is the new minority class sample corresponding to the nearest neighbor y.

[0034] In steps 1 and 4 above, wavelet transform is used to extract features from the sample monitoring record data and the real-time monitoring record data.

[0035] Given that the traditional KMeans-SMOTE algorithm only combines the K-Means clustering algorithm with the SMOTE algorithm, it does not take into account the possibility that the synthesized samples within a cluster may be of poor quality, nor does it consider the imbalance of minority class samples between clusters, and may lead to an over-concentration of synthesized minority class samples, the improved SMOTE algorithm proposed in this invention effectively ensures the quality of synthesized minority class samples through a scoring mechanism, and effectively balances the number of minority class samples through sparsity, so that the classifier can pay attention to faulty data samples, thereby avoiding the situation where the classifier classifies faulty data samples as normal data samples because there are too few faulty data samples. Attached Figure Description

[0036] Figure 1 This is a flowchart of an industrial imbalance data classification method based on an improved SMOTE.

[0037] Figure 2 Flowchart for improving the SMOTE oversampling algorithm. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific examples.

[0039] To overcome the weaknesses of existing traditional classification algorithms when dealing with imbalanced industrial data, this invention proposes an industrial imbalanced data classification method based on an improved SMOTE, such as... Figure 1 As shown, its aim is to address the technical problem of low recall rate in equipment fault detection caused by the imbalance between normal and fault data samples, while ensuring the accuracy of equipment fault detection. The specific steps include the following:

[0040] Step 1: Obtain sample monitoring record data of industrial equipment, extract features from the sample monitoring record data to obtain feature vectors, and then label the feature vectors to obtain an imbalanced dataset.

[0041] Because the monitoring data recorded by industrial equipment sensors (such as the horizontal vibration amplitude data of bearings or the sound data of motors) generally cannot be directly used in current mainstream classifiers, feature extraction is required before it can be used in a classifier. Therefore, feature extraction processing of the monitoring data is necessary first. In a preferred embodiment of the present invention, wavelet transform is used to extract features from the monitoring data. After feature extraction, the obtained feature vectors are labeled, that is, according to the state (normal and faulty) of the industrial equipment corresponding to the monitoring data, the feature vectors are labeled with corresponding labels, thereby obtaining the initial imbalanced dataset.

[0042] Step 2: Use the improved SMOTE method to process the imbalanced dataset to obtain a balanced dataset.

[0043] The improved SMOTE method uses fault data feature samples from an imbalanced dataset as reference samples to interpolate and synthesize fault data feature samples. A scoring mechanism is used to remove synthesized fault data sample points that are close to normal data feature samples. Once a certain number of synthesized fault data feature samples are obtained, k-means clustering is performed on the fault data feature samples in the initially balanced dataset, retaining clusters with a larger number of fault data feature samples. Then, the number of normal and fault data feature samples within each cluster is calculated, and clusters with an excessive number of normal data feature samples relative to the number of fault data feature samples are removed. Finally, the sparsity of each cluster is calculated, allowing clusters with higher sparsity to generate more fault data feature samples, until the algorithm terminates.

[0044] The improved SMOTE method is illustrated below with a specific example (see [link]). Figure 2 ):

[0045] 1) Initialization: In the imbalanced dataset, feature vectors labeled as faulty are defined as minority class samples, and feature vectors labeled as normal are defined as majority class samples; the initial number of minority class samples is N. c The number of minority class samples N generated again D The number of nearest neighbor samples k and the number of clusters k ′ .

[0046] The number of minority class samples N generated initially c and the number of minority class samples N generated again DAll values ​​can be arbitrarily determined by the individual, and the ultimate goal is to make the number of minority class samples and the number of majority class samples in the dataset as equal as possible. In a preferred embodiment of the present invention,

[0047] The number of minority class samples N generated initially c for:

[0048]

[0049] The number of minority class samples N generated again D for:

[0050] N D =C major -C minor -N c

[0051] In the formula, C major C represents the number of majority class samples in an imbalanced dataset. minor p represents the number of minority class samples in an imbalanced dataset, and p is a hyperparameter that is set manually.

[0052] In a preferred embodiment of the present invention, the number of nearest neighbors k = 5 and the number of clusters k′ = 10.

[0053] 2) Linear interpolation: First, randomly select a minority class sample x from the imbalanced dataset. Then, calculate the Euclidean distance from the minority class sample x to other minority class samples in the imbalanced dataset, and obtain the k nearest neighbors of the minority class sample x (i.e., the k samples with the smallest Euclidean distance to the minority class sample x) based on the Euclidean distance. Finally, perform an interpolation operation on each nearest neighbor of the minority class sample x to obtain a new minority class sample x corresponding to each nearest neighbor. new .

[0054] x new = x + rand·(yx)

[0055] In the formula, rand is a random number between [0, 1], x is a randomly selected minority class sample, y is the nearest neighbor sample of the minority class sample x, and x new This is the new minority class sample corresponding to the nearest neighbor y.

[0056] 3) Scoring operation: First calculate x for each new minority class sample. new Calculate the Euclidean distances to other minority class samples and all majority class samples in the imbalanced dataset, and obtain a new minority class sample x based on the proximity of the Euclidean distances. new Find the k nearest neighbors. Then determine the new minority class sample x. new Is each nearest neighbor a minority or majority class sample? If a new minority class sample x newIf a new minority class sample x is a minority class sample, then the weight of that neighbor sample is set to 1; if a new minority class sample x... new If the nearest neighbor sample is a majority class sample, then the weight of that nearest neighbor sample is set to 0. Then, the weight of the new minority class sample x is calculated. new The sum of the weights of the k nearest neighbors, and the new minority class sample x. new Is the sum of the weights of x's k nearest neighbors greater than k / 2? If so, then x is considered... new If the sample is close to the minority class sample, retain the new minority class sample x. new Otherwise, consider x new Discard the new minority class sample x by moving away from the minority class sample. new .

[0057] 4) Repeat steps 2) and 3) until a new minority class sample x is retained. new The number reaches the required number N of minority class samples. c and will this N c A new minority class sample x that is retained new As minority class samples, they are incorporated into the imbalanced dataset to obtain the initial balanced dataset.

[0058] 4) K-Means clustering: First, perform K-means clustering on the minority class samples in the initial balanced dataset, thereby dividing all minority class samples in the initial balanced dataset into k′ clusters; then, take the mean of all minority class samples in each cluster as the centroid of that cluster; then calculate the Euclidean distance from each majority class sample in the initial balanced dataset to the centroid of each cluster, and assign each majority class sample to the cluster with the closest Euclidean distance to its centroid.

[0059] 6) Retain clusters that meet the requirements: Calculate the number of majority class samples and the number of minority class samples in each cluster. If the number of majority class samples in a cluster is greater than or equal to the number of minority class samples, then discard the cluster; otherwise, retain the cluster.

[0060] 7) Calculate the sparsity of each retained cluster: First, calculate the Euclidean distance between every two minority class samples within the retained cluster, and construct the Euclidean distance matrix for the minority class samples within the retained cluster. Then, sum all the off-diagonal elements in the Euclidean distance matrix for the minority class samples within the retained cluster, and divide by the total number of off-diagonal elements to obtain the average distance of the retained cluster. Finally, divide the average distance of the retained cluster by the number of minority class samples within the retained cluster to obtain the sparsity of the retained cluster.

[0061] 8) Calculate the number of minority class samples N(f) generated by each retained cluster:

[0062]

[0063] In the formula, Nst (f) represents the sparsity of the retained clusters f, m represents the number of retained clusters f, and N D This represents the number of minority class samples generated again.

[0064] 9) Linear interpolation operation: First, randomly select a minority class sample x′ from the retained clusters; then calculate the Euclidean distance from the minority class sample x′ to other minority class samples in the retained clusters, and obtain the k nearest neighbor samples of the minority class sample x′ based on the proximity of the Euclidean distances; finally, perform an interpolation operation on each nearest neighbor sample of the minority class sample x′ to obtain a new minority class sample x′ corresponding to each nearest neighbor sample. new ;

[0065] 10) Scoring operation: First calculate x′ for each new minority class sample. new Calculate the Euclidean distances to other minority class samples and all majority class samples in the preserved cluster, and obtain a new minority class sample x′ based on the proximity of the Euclidean distances. new Find the k nearest neighbors; then determine the new minority class sample x′. new Is each nearest neighbor a minority or majority class sample? If a new minority class sample x′ new If a new minority class sample x′ is a minority class sample, then the weight of that neighboring sample is set to 1; if a new minority class sample x′ is a minority class sample x′, then the weight of that neighboring sample is set to 1. new If the nearest neighbor sample is a majority class sample, then the weight of that nearest neighbor sample is set to 0; then the weight of the new minority class sample x′ is calculated. new The sum of the weights of the k nearest neighbors, and the new minority class sample x′ new Does the sum of the weights of the k nearest neighbors of x′ exceed k / 2? If so, then retain the new minority class sample x′. new Otherwise, discard the new minority class sample x′. new ;

[0066] 11) Repeat steps 9) and 10) until a new minority class sample x′ is retained. new The number reaches the number of minority class samples N(f) that each retained cluster is responsible for generating;

[0067] 12) Repeat steps 9)-11) for each retained cluster, and then generate N from all the retained clusters. D A new minority class sample x′ is retained. new As minority class samples, they are incorporated into the initial balanced dataset to obtain the balanced dataset.

[0068] Step 3: Use the balanced dataset as the training dataset to train the classifier and obtain the trained classifier.

[0069] In a preferred embodiment of the present invention, a random forest model is used as the classifier. The balanced dataset obtained in step 2 is fed into the random forest model and trained to obtain a trained random forest classification model (classifier).

[0070] Step 4: Obtain real-time monitoring data of the industrial equipment. Extract features from the real-time monitoring data using the same feature extraction method as in Step 1 to obtain feature vectors. Then, feed these feature vectors into a trained classifier to determine the status of the industrial equipment, i.e., whether it is in a normal state, a fault state, or a specific fault state. If the industrial equipment is in a faulty operating state, it can promptly alert the manager for handling.

[0071] Experimental results:

[0072] The method of this invention was used to process an industrial imbalanced dataset with an imbalance ratio of 20 (i.e., the number of normal samples is approximately 20 times the number of faulty samples). Twenty-five experiments were conducted to minimize randomness, and the average results were taken. The experimental results are shown in Tables 1 and 2.

[0073] Table 1. Classification results without using the improved SMOTE method

[0074]

[0075]

[0076] Table 2 Classification results using the improved SMOTE method

[0077] Acc F-measure G-means 0.9734 0.6690 0.7636

[0078] In the table, Acc represents classification accuracy, and F-measure and G-means are metrics for evaluating the classification performance of imbalanced datasets. The larger the value, the better the performance.

[0079] The calculation formulas for the three are derived from Table 3 and the following method:

[0080] Table 3 Confusion Matrix

[0081] Classification Predicting the majority class Predicting minority groups Actual majority class TP FN Actual minority FP TN

[0082]

[0083]

[0084]

[0085] Conclusion: The comparison between Tables 1 and 2 shows that after using the improved SMOTE algorithm, the Acc value, i.e., accuracy, hardly decreased, while the F-measure and G-means, metrics for evaluating the classification performance of imbalanced datasets, improved by 8% and 10%, respectively. This indicates that the present invention can effectively improve the classification performance of imbalanced datasets, thereby minimizing the possibility of faulty data being ignored.

[0086] It should be noted that although the embodiments described above are illustrative, they are not intended to limit the invention. Therefore, the invention is not limited to the specific embodiments described above. Any other embodiments obtained by those skilled in the art under the guidance of this invention without departing from its principles are considered to be within the protection scope of this invention.

Claims

1. A classification method for industrial imbalance data based on an improved SMOTE, characterized in that, The steps include the following: Step 1: Obtain sample monitoring record data of industrial equipment, extract features from the sample monitoring record data to obtain feature vectors of the sample monitoring record data, and then label the feature vectors of the sample monitoring record data to obtain an imbalanced dataset. Step 2: Use the improved SMOTE method to process the imbalanced dataset to obtain a balanced dataset; Step 2.1: Define the feature vectors labeled as faulty in the imbalanced dataset as minority class samples, and the feature vectors labeled as normal as majority class samples; set the initial number of minority class samples N. c The number of minority class samples N generated again D The number of nearest neighbor samples k and the number of clusters k′; Step 2.2: First, randomly select a minority class sample x from the imbalanced dataset; then calculate the Euclidean distance between the minority class sample x and other minority class samples in the imbalanced dataset, and obtain the k nearest neighbors of the minority class sample x based on the Euclidean distance; finally, perform interpolation on each nearest neighbor of the minority class sample x to obtain a new minority class sample x corresponding to each nearest neighbor. new ; Step 2.3: First calculate x for each new minority class sample. new Calculate the Euclidean distances to other minority class samples and all majority class samples in the imbalanced dataset, and obtain a new minority class sample x based on the proximity of the Euclidean distances. new Find the k nearest neighbors; then determine the new minority class sample x. new Is each nearest neighbor a minority or majority class sample? If a new minority class sample x new If a new minority class sample x is a minority class sample, then the weight of that neighbor sample is set to 1; if a new minority class sample x... new If the nearest neighbor sample is a majority class sample, then the weight of that nearest neighbor sample is set to 0; then the weight of the new minority class sample x is calculated. new The sum of the weights of the k nearest neighbors, and the new minority class sample x. new Does the sum of the weights of the k nearest neighbors of x exceed k / 2? If so, retain the new minority class sample x. new Otherwise, discard the new minority class sample x. new ; Step 2.4: Repeat steps 2.2-2.3 until a new minority class sample x is retained. new The number reaches the required number N of minority class samples. c and will this N c A new minority class sample x that is retained new As minority class samples, they are incorporated into the imbalanced dataset to obtain the initial balanced dataset; Step 2.5: First, perform K-means clustering on the minority class samples in the initial balanced dataset, thereby dividing all minority class samples in the initial balanced dataset into k′ clusters; then, take the mean of all minority class samples in each cluster as the centroid of that cluster; finally, calculate the Euclidean distance from each majority class sample in the initial balanced dataset to the centroid of each cluster, and assign each majority class sample to the cluster with the closest Euclidean distance to its centroid. Step 2.6: Calculate the number of majority class samples and the number of minority class samples in each cluster. If the number of majority class samples in a cluster is greater than or equal to the number of minority class samples, then discard the cluster; otherwise, retain the cluster. Step 2.7: For each retained cluster, first calculate the Euclidean distance between every two minority class samples within the retained cluster, and construct the Euclidean distance matrix of the minority class samples within the retained cluster accordingly; then add all the off-diagonal elements in the Euclidean distance matrix of the minority class samples within the retained cluster, and divide by the total number of off-diagonal elements to obtain the average distance of the retained cluster; finally, divide the average distance of the retained cluster by the number of minority class samples within the retained cluster to obtain the sparsity of the retained cluster. Step 2.8: Calculate the number of minority class samples N(f) generated by each retained cluster: In the formula, N st (f) represents the sparsity of the retained clusters f, m represents the number of retained clusters f, and N D This represents the number of minority class samples generated again. Step 2.9: First, randomly select a minority class sample x′ from the retained clusters; then calculate the Euclidean distance between the minority class sample x′ and other minority class samples in the retained clusters, and obtain the k nearest neighbor samples of the minority class sample x′ based on the Euclidean distance; finally, perform interpolation on each nearest neighbor sample of the minority class sample x′ to obtain the new minority class sample x′ corresponding to each nearest neighbor sample. new ; Step 2.10: First calculate each new minority class sample x′ new Calculate the Euclidean distances to other minority class samples and all majority class samples in the preserved cluster, and obtain a new minority class sample x′ based on the proximity of the Euclidean distances. new Find the k nearest neighbors; then determine the new minority class sample x′. new Is each nearest neighbor a minority or majority class sample? If a new minority class sample x′ new If a new minority class sample x′ is a minority class sample, then the weight of that neighboring sample is set to 1; if a new minority class sample x′ is a minority class sample x′, then the weight of that neighboring sample is set to 1. new If the nearest neighbor sample is a majority class sample, then the weight of that nearest neighbor sample is set to 0; then the weight of the new minority class sample x′ is calculated. new The sum of the weights of the k nearest neighbors, and the new minority class sample x′ new Does the sum of the weights of the k nearest neighbors of x′ exceed k / 2? If so, then retain the new minority class sample x′. new Otherwise, discard the new minority class sample x′. new ; Step 2.11: Repeat steps 2.9-2.10 until a new minority class sample x′ is retained. new The number reaches the number of minority class samples N(f) that each retained cluster is responsible for generating; Step 2.12: Repeat steps 2.9-2.11 for each retained cluster, and then generate N from all retained clusters. D A new minority class sample x′ is retained. new The minority class samples are incorporated into the initial balanced dataset to obtain the balanced dataset; Step 3: Use the balanced dataset as the training dataset to train the classifier and obtain the trained classifier; Step 4: Obtain real-time monitoring data of industrial equipment, extract features from the real-time monitoring data to obtain feature vectors, and feed the feature vectors into a trained classifier to obtain the status of the industrial equipment.

2. The industrial imbalance data classification method based on improved SMOTE as described in claim 1, characterized in that, In step 2.1, The number of minority class samples N generated initially c for: The number of minority class samples N generated again D for: N D =C major -C minor -N c In the formula, C major C represents the number of majority class samples in an imbalanced dataset. minor p represents the number of minority class samples in an imbalanced dataset, and p is a hyperparameter that is set manually.

3. The industrial imbalance data classification method based on improved SMOTE as described in claim 1, characterized in that, In step 2.2, the formula for the interpolation operation is: x new =x+rand·(y-x) In the formula, rand is a random number between [0,1], x is a randomly selected minority class sample, y is the nearest neighbor sample of the minority class sample x, and x new This is the new minority class sample corresponding to the nearest neighbor y.

4. The industrial imbalance data classification method based on improved SMOTE as described in claim 1, characterized in that, In steps 1 and 4, wavelet transform is used to extract features from the sample monitoring record data and the real-time monitoring record data.

Images