Industrial machine fault detection method, device and storage medium based on data partitioning hybrid resampling

Abstract

The present invention relates to an industrial machine fault detection method, device and storage medium based on data partitioning hybrid resampling. The present invention first uses a hybrid sampling technology to balance the category distribution of current sensor data. Small boundary samples are used to generate new small samples. The undersampling technology based on the clustering algorithm balances the current category distribution by removing large samples while taking into account the data distribution of non-fault samples. Then, a training set with a relatively balanced category distribution is constructed, and a random forest classification model is constructed, which can be used for industrial machine fault detection. The data partitioning hybrid resampling technology in the present invention can not only alleviate the problem of easy overfitting in the oversampling mechanism, but also effectively alleviate the problem of information loss in undersampling. At the same time, compared with random undersampling, the undersampling technology based on the clustering algorithm can effectively retain the characteristics of the data distribution by considering the data distribution of non-defective samples.

Description

Technical Field

[0001] The present invention relates to an industrial machine fault detection method, device and storage medium based on data partitioning, hybrid resampling. Background Art

[0002] Classification has become a hot topic in the field of machine learning and has been used in data analysis and prediction in many fields. The evaluation of class imbalanced datasets is a complex task. In binary classification problems, class imbalance refers to the situation where the number of samples in one category is significantly less than the number of samples in another category. Among them, samples with a small number are called small samples, and samples with a large number are called large samples. If traditional classifiers are used to process unbalanced datasets, the performance of the model will often be biased towards large samples. Therefore, how to improve the recognition rate of the model for small samples without reducing the classification performance of the model for large samples is the goal of the class imbalance learning field.

[0003] The detection process of industrial machine faults can use data classification models in machine learning. The data is collected from sensors on the machine. Usually, there are fewer fault samples than non-fault samples. Therefore, using classification models to predict industrial machine faults is a class imbalance classification problem. The goal of the machine fault detection model is to effectively identify fault data in the sensor. An effective classification model needs to improve the recognition rate of fault sensor data.

[0004] The use of class imbalance classification technology can effectively identify the faults of industrial machines. There are already many algorithms for solving the class imbalance problem in fault identification, mainly including three categories: data level, algorithm level and hybrid method. Among them, the data level method is divided into three categories: oversampling, undersampling and hybrid sampling methods. Oversampling includes random oversampling technology and oversampling technology based on generating new small samples. By increasing the number of small samples to balance the category distribution of the current data, however, incorrect or redundant small samples will be formed, which is easy to cause overfitting of the classification model. The undersampling method achieves a balanced category distribution by removing large samples, which is easy to lose useful information of large category samples. The hybrid sampling method uses both oversampling and undersampling mechanisms, so it can improve the generalization ability of the classification model and reduce the probability of overfitting of the classification model. Summary of the invention

[0005] In order to make up for the shortcomings of the existing industrial machine fault detection based on unbalanced classification methods, the present invention aims to provide an industrial machine fault detection method, device and storage medium based on data partitioning and hybrid resampling.

[0006] In one aspect of the present invention, a method for industrial machine fault detection based on data partitioning hybrid resampling is provided, comprising the following steps:

[0007] Step 1) Sensor data partitioning based on data distribution:

[0008] Receive sensor data related to industrial machines, divide the data into non-fault class set Max and fault class set Min, and further divide the sensor data related to industrial machines into small sample sets D in the boundary area based on the characteristics of data distribution. bor+ , small noise sample set D noise+ , sample set D of major safety areas safe- .

[0009] Step 2) De-noise the data:

[0010] Remove the noise small sample set D from the fault class set Min noise+ , we get Min1 and |Min1|, where Min1=Min-D noise+ , |Min1| is the number of samples in Min1.

[0011] Step 3) Oversampling based on boundary areas:

[0012] Use the small sample set D in the boundary area bor+ Generate a new sensor small sample set D new+ , D new+ Used to characterize newly generated defect data to balance the distribution of category data. Its quantity is |D new+ |.

[0013] Step 4) Undersampling based on data distribution: The sample set D of the large category of safe areas obtained in step 1 safe- Perform cluster-based undersampling to obtain D' safe- .

[0014] Step 5) Construct a training data set: Based on the data set obtained in steps 2-4, construct a training data set D = Min1∪D for the machine defect detection model new+ ∪D' safe- .

[0015] Step 6) Training the industrial machine fault detection model: Use the training data set D obtained in step 5 to train the random forest model M. The identification of industrial machine faults can be converted into a binary classification problem, and the sensor data is divided into two categories. Among them, a small number of sensor data belongs to a small class or a fault class, and a large number of sensor data belongs to a large class or a non-fault class. By using the training random forest model M to classify the sensor data related to the industrial machine, if the category is a small class, it is judged as a fault class, thereby identifying the fault in the industrial machine.

[0016] Furthermore, in step 1), the small sample set D in the boundary areabor+ , small noise sample set D noise+ , sample set D of major safety areas safe- It is obtained by using the 5 nearest neighbors method.

[0017] Furthermore, step 3) is specifically:

[0018] First, we need to calculate the number of small samples that need to be generated in the boundary area |D new+ |;

[0019] Then, the SMOTE algorithm is used to oversample the small samples in the boundary area, and the newly generated small sample set is D new+ .

[0020] Furthermore, |D new+ The calculation formula of | is as follows:

[0021] |D new+ |=1 / 2*f*|D safe- |-|Min1|

[0022] Where f is the post-balance rate, |D safe- | is a sample set of large-scale safe areas D safe- The number of samples.

[0023] Furthermore, step 4) is specifically:

[0024] First, we first analyze the large class of safe area sample set D safe- Perform clustering to obtain several clusters;

[0025] Then random undersampling is performed within each cluster.

[0026] Finally, the under-sampled clusters are merged to obtain the large cluster set D' after sampling safe- Among them, D' safe- The number of samples in |D' safe- |,|D' safe- |=(|Min1|+|D new+ |) / f.

[0027] Furthermore, the clustering adopts k-means clustering.

[0028] Furthermore, the number of non-faulty sensor data retained in each cluster is (|Min1|+|D new+ |) / (f*k), where k represents the data class generated by k-means clustering.

[0029] Another aspect of the present invention provides an industrial machine fault detection device based on data partitioning hybrid resampling, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned industrial machine fault detection method based on data partitioning hybrid resampling when executing the program.

[0030] According to another aspect of the present invention, the storage medium stores a computer program, and the computer program is used to execute the above-mentioned industrial machine fault detection method based on data partitioning, hybrid resampling.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The present invention first uses a hybrid sampling technology to balance the category distribution of current sensor data. On the one hand, a small boundary sample with rich classification information is used to generate a new small sample. Compared with the random oversampling technology, this method does not need to copy the original small sample of the fault class, and can generate new fault class sample information. On the other hand, the undersampling technology based on the clustering algorithm balances the current category distribution by removing large samples while taking into account the data distribution of non-fault class samples. Then, a training set with a relatively balanced category distribution is constructed, and a random forest classification model is constructed, which can be used for industrial machine fault detection. Compared with the existing imbalance processing mechanism, the hybrid resampling technology based on data partitioning can not only alleviate the problem of easy overfitting in the oversampling mechanism, but also effectively alleviate the problem of information loss in undersampling. At the same time, compared with random undersampling, the undersampling technology based on the clustering algorithm can effectively retain the characteristics of the data distribution by considering the data distribution of non-defect class samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a schematic diagram of an industrial machine fault detection method based on data partitioning and hybrid resampling proposed by the present invention.

[0034] Figure 2 It is a structural diagram of the device of the present invention. DETAILED DESCRIPTION

[0035] The present invention uses a hybrid sampling technique to effectively solve the class imbalance problem in sensor data. When using binary imbalance classification technology to detect industrial machine faults, the fault class is a small sample and the non-fault class is a large sample. On the one hand, an oversampling technique based on boundary sensor samples is used to increase the number of fault class samples in the training set; on the other hand, considering the data distribution of sensor samples, a clustering-based undersampling technique is used to remove non-fault class samples. It mainly includes the following steps:

[0036] Step 1) Sensor data partitioning based on data distribution:

[0037] Receive sensor data related to industrial machines and divide the data into two sets: non-fault set Max and fault set Min. Based on the characteristics of data distribution, further divide the sensor data related to industrial machines into small sample sets D in the boundary area. bor+ , small noise sample set D noise+ , sample set D of major safety areas safe- Among them, the division of sensor data adopts the 5 nearest neighbor method, that is, the data distribution characteristics of each sensor sample are determined based on the category of the samples in the 5 nearest neighbors of each sensor data.

[0038] Step 2) De-noise the data:

[0039] Remove the noise fault class sample set D obtained in step 1 from the industrial machine sensor small class dataset Min noise+ , we get Min1 and |Min1|. Among them, Min1=Min-D noise+ , |Min1| is the number of sensor samples in Min1. In the model of industrial machine fault detection, noise data is of no value. Instead, it will reduce the performance of the model and increase the probability of model overfitting. Therefore, it is necessary to remove the fault noise data from the fault class set Min. Among them, the recognition of the noise data set is based on the 5 nearest neighbors method. If all the 5 nearest neighbors of a small sample are large class samples, then the small sample is in the noise area, so D noise+ .

[0040] Step 3) Oversampling based on boundary areas:

[0041] Use the small sample set D of the boundary area obtained in step 1 bor+ Generate a new small sample of sensors and get D new+ and |D new+ |. Among them, D new+ It is the newly generated defect data, which can be used to balance the distribution of category data. new+ | for D new+ The number of industrial machine sensor samples in the classification. Boundary samples are samples distributed on the classification decision boundary. Compared with samples in other areas, boundary samples are closer to the decision boundary and have richer classification information, which is more important for the construction of classification models. If the number of boundary samples is large enough, a classification model with a high recognition rate can be trained. Therefore, new small samples can be generated using the boundary area to obtain more defective sensor data in the boundary area. The details are as follows: First, calculate the number of defective samples to be generated | D new+ |, where |D new+ |=1 / 2*f*|D safe-|-|Min1|. f is the imbalance rate that needs to be achieved after mixed resampling, which is called the post-balance rate. |Min1| is the number of small sample fault class data after denoising obtained in step 2. |D safe- | is the number of non-faulty data in the safe area obtained in step 1. Then, the SMOTE algorithm is used to generate new small samples based on the non-faulty sensor samples in the boundary area to balance the category distribution. The newly generated small sample set is D new+ .

[0042] Step 4) Undersampling based on data distribution:

[0043] For the large category safe area sample set D obtained in step 1 safe- Perform cluster-based undersampling to obtain D' safe- , can further balance the category distribution of sensor data. Using clustering-based algorithms and undersampling within clusters can effectively avoid the problem of large sample information loss in random undersampling. Specifically: First, use the k-means clustering algorithm to cluster the large sample set D in the safe area. safe- Divide it into several clusters. Then, perform random undersampling within the cluster. Finally, merge the undersampled clusters to get D' safe- Among them, D' safe- The number of samples in |D' safe- |=(|Min1|+|D new+ |) / f, the number of non-faulty sensor data retained in each cluster is (|Min1|+|Dnew+|) / (f*k).

[0044] Step 5) Build the training dataset:

[0045] Based on the data set obtained in steps 2-4, a training data set D for the machine defect detection model is constructed. The class imbalance in D is equal to f, that is, the number of small samples in D divided by the number of large samples is equal to f. The filtered small sample set Min1 obtained in step 2 and the small sample set D generated based on the boundary small samples obtained in step 3 new+ And the large sample set D' of the safe area based on cluster undersampling obtained in step 4 safe- Together they constitute the training set for the industrial machine fault detection model. That is, D = Min1∪D new+ ∪D' safe- .

[0046] Step 6) Training the industrial machine fault detection model: Using the training set D obtained in step 5, the industrial machine fault detection model M is trained based on the random forest model. M is an integrated classification model, which has higher generalization ability than the single classifier model. M can be used to predict the category of industrial-related sensor data, and if it is determined to be a small category, the fault in the industrial machine can be identified. Since the training set used to train M has a relatively balanced category distribution, M has a higher recognition rate for fault-type data.

[0047] Preferably, in step 1), since the probability of industrial machine failure is relatively low, if a data classification model is used to predict industrial machine failure, a class imbalance processing mechanism needs to be used to improve the probability of the classification model identifying faulty data. A mixed sampling technique based on data partitioning is used, and the data can be partitioned using the proportion of large sample categories in the 5 nearest neighbors to obtain 3 regions. The 5 nearest neighbors of the data with small noise are all large samples; the 5 nearest neighbors of the large class of safe area samples are all large samples; and the small samples in the 5 nearest neighbors of the large sample are regarded as small boundary samples.

[0048] Preferably, in step 2), in order to improve the robustness of the machine fault detection model based on the classification algorithm, the noise data in the training set is removed. The noise data does not contain useful information about the model training and will interfere with the training of the classification model.

[0049] Preferably, in step 3), new small samples are generated using small samples in the boundary area, and oversampling by the SMOTE algorithm can improve the classification performance of the model for fault class samples.

[0050] Preferably, in step 4), if the five nearest neighbors of a large sample are all large samples, the large sample is in a safe area. First, the samples in the area are clustered, and then random undersampling is performed in each cluster. The undersampling model based on the clustering mechanism can not only reduce the bias of the model towards large samples, but also ensure that the distribution of the large category data after oversampling is highly consistent with the initial data distribution.

[0051] Preferably, in step 5), the small sample set Min1 is integrated, and the small sample set D generated based on the boundary small samples is new+ And the large sample set D' of the safe area after cluster undersampling safe- , a training set D is obtained for training the industrial machine fault detection model, and the category distribution in D is approximately balanced, and the imbalance rate is equal to the predefined post-balance rate f.

[0052] Preferably, in step 6), a random forest model can be trained based on D, and the prediction of the sensor sample category by the integrated classification model realizes industrial machine fault detection.

[0053] Example:

[0054] This embodiment proposes an industrial machine fault detection method based on data partitioning and hybrid resampling, such as Figure 1 As shown, the model includes a data partitioning mechanism based on the distribution of industrial machine sensor data, a faulty sensor data filtering mechanism, an oversampling mechanism based on the boundary sensor data area, an undersampling mechanism based on the sensor data distribution, the construction of a training data set, and an industrial machine fault detection model.

[0055] First, the sensor data related to industrial machines are divided into large samples and small samples. Among them, the small samples are used to describe fault data. Then, the data partitioning technology based on data distribution is used to divide the data regions with different characteristics, so as to obtain the small sample set D in the boundary area. bor+ , small noise sample set D noise+ , sample set D of major safety areas safe- The sample sets in different areas have different data distribution characteristics. The 5 nearest neighbor method is used to identify samples in different areas. If all the neighbor samples of a small sample are large samples, then the small sample is a noise sample. At the same time, the large samples in the 5 nearest neighbors of the small sample are placed in the boundary area. On the contrary, if all the neighbor samples of a small sample are small samples, then the small sample is in the safe area. In order to improve the robustness of the fault detection model and reduce the probability of overfitting, the sample set D in the small noise area is removed from the initial data set. noise+ Then, we use the small samples in the boundary area to oversample. Specifically, first, we need to evaluate the number of small samples that need to be generated. Then, we use the SMOTE algorithm to oversample the small samples in the boundary area. The newly generated small sample set is D new+ .

[0056] At the same time, the undersampling technique based on the clustering algorithm is used to remove large samples. By randomly undersampling within the cluster, this technique balances the category distribution while maintaining the category distribution of the initial large sample. Then, a training dataset D = Min1∪D with balanced category distribution is constructed. new+ ∪D' safe- Finally, the random forest classification model M is trained using the sensor samples related to industrial machines in D. M is used to classify the sensor data related to industrial machines. If the category is a small category, it is judged as a fault category, thereby identifying the fault in the industrial machine.

[0057] The embodiments of the present invention can be applied to network devices. The embodiments can be implemented by software, hardware, or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of the device in which it is located reading the corresponding computer program instructions in the non-volatile memory into the memory and running them, wherein the computer program is used to execute the method determined by the above steps 1) to 6). From the hardware level, if Figure 2 As shown, it is a hardware structure diagram of the industrial machine fault detection method and device based on data partitioning and mixed resampling of the present invention, except Figure 2 In addition to the processor, network interface, memory and non-volatile memory shown, the device may also include other hardware to facilitate expansion at the hardware level. On the other hand, the present application also provides a computer-readable storage medium, the storage medium storing a computer program, the computer program being used to execute the method determined in the above steps 1) to 6).

[0058] As for the embodiments, since they basically correspond to the method embodiments, the relevant parts can refer to the partial description of the method embodiments. The above-described device embodiments are only illustrative, and those skilled in the art can understand and implement them without creative work.

[0059] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary techniques in the art that are not disclosed in the present application. The specification and examples are intended to be exemplary only.

[0060] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity, or apparatus.

[0061] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. An industrial machine fault detection method based on data partitioning and hybrid resampling, characterized by: The following steps are involved: Step 1) Sensor data partitioning based on data distribution: Receive sensor data related to industrial machines, divide the data into non-fault class set Max and fault class set Min, and further divide the sensor data related to industrial machines into small sample sets D in the boundary area based on the characteristics of data distribution. bor+ , small noise sample set D noise+ , sample set D of major safety areas safe- ; Step 2) De-noise the data: Remove the noise small sample set D from the fault class set Min noise+ , we get Min1 and |Min1|, where Min1=Min-D noise+ , |Min1| is the number of samples in Min1; Step 3) Oversampling based on boundary areas: Use the small sample set D in the boundary area bor+ Generate a new sensor small sample set D new+ , D new+ Used to characterize newly generated defect data to balance the distribution of category data. Its quantity is |D new+ |; Step 4) Undersampling based on data distribution: The large class of safe area sample set D obtained in step 1 safe- Perform cluster-based undersampling to obtain D' safe- ; Step 5) Construct a training data set: Based on the data set obtained in steps 2-4, construct a training data set D = Min1∪D for the machine defect detection model new+ ∪D' safe- ; Step 6) Training an industrial machine fault detection model: Using the training data set D obtained in step 5, train a random forest model M; by using the trained random forest model M to classify the sensor data related to the industrial machine, if the category is a small category, it is judged as a fault class, thereby identifying the fault in the industrial machine.

2. The industrial machine fault detection method based on data partitioning and hybrid resampling according to claim 1 is characterized in that: In step 1), the small sample set D in the boundary area bor+ , small noise sample set D noise+ , sample set D of major safety areas safe- It is obtained by using the 5 nearest neighbors method.

3. The industrial machine fault detection method based on data partitioning and hybrid resampling according to claim 1 is characterized in that: Step 3) is specifically: First, we need to calculate the number of small samples that need to be generated in the boundary area |D new+ |; Then, the SMOTE algorithm is used to oversample the small samples in the boundary area, and the newly generated small sample set is D new+ .

4. The industrial machine fault detection method based on data partitioning and hybrid resampling according to claim 3 is characterized in that: |D new+ The calculation formula of | is as follows: |D new+ |=1 / 2*f*|D safe- |-|Min1| Where f is the post-balance rate, |D safe- | is a sample set of large-scale safe areas D safe- The number of samples.

5. The industrial machine fault detection method based on data partitioning and hybrid resampling according to claim 1 is characterized in that: Step 4) is specifically: First, we first analyze the large class of safe area sample set D safe- Perform clustering to obtain several clusters; Then random undersampling is performed within each cluster; Finally, the under-sampled clusters are merged to obtain the large cluster set D' after sampling safe- ; where D' safe- The number of samples in |D' safe- |,|D' safe- |=(|Min1|+|D new+ |) / f.

6. The industrial machine fault detection method based on data partitioning and hybrid resampling according to claim 5 is characterized in that: The clustering adopts k-means clustering.

7. The industrial machine fault detection method based on data partitioning and hybrid resampling according to claim 6 is characterized in that: The number of non-faulty sensor data retained in each cluster is (|Min1|+|D new+ |) / (f*k), where k represents the data class generated by k-means clustering.

8. Industrial machine fault detection device based on data partitioning and hybrid resampling, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, an industrial machine fault detection method based on data partitioning, hybrid resampling, as described in any one of claims 1 to 7 is implemented.

9. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and the computer program is used to execute the industrial machine fault detection method based on data partitioning, hybrid resampling as described in any one of claims 1 to 7.

Images