Rolling bearing monitoring signal degradation stage identification method

Abstract

The invention discloses a rolling bearing monitoring signal degradation stage identification method. An acceleration sensor is used for collecting vibration signals of bearing operation; the signals are sliced to generate samples, and a data set is formed; samples of different parts of the data set are respectively used for training three base learners, the output of the base learners is input into the meta-model for training to obtain a final recognition result, an encoder of each base learner is composed of a convolution block, a full connection layer and a flattening layer, and a classifier of a sub-model is a softmax classifier; and an XGBoost classifier is selected as the meta model. And for the to-be-identified signal, identifying the to-be-identified signal by using the three base learners, fusing the outputs of the base learners, and inputting the fused outputs into the meta learner to obtain a final degradation stage identification result. The method can effectively solve the problem that in the training process of the model, when the number of samples in different stages of bearing degradation is seriously unbalanced, the model cannot be accurately identified.

Description

Technical Field

[0001] The invention belongs to the field of state monitoring and fault diagnosis, and relates to a rolling bearing degradation stage identification method based on integrated learning. Background Art

[0002] Rolling bearings are core components of rotating machinery and are widely used in various industrial equipment. Once a rolling bearing fails, it will cause losses to the company and even cause casualties. Therefore, it is necessary to identify the degradation stage of the rolling bearing so that planned downtime can be carried out before the failure occurs, and predictive maintenance of the equipment can be performed to reduce the losses caused by unplanned downtime. Therefore, accurately identifying the degradation stage of the bearing is of great significance for the intelligent operation and maintenance of the equipment, as well as the subsequent life prediction and early fault identification of the equipment.

[0003] In practical applications, the degradation process of bearings can be divided into three stages: stable operation stage, initial wear stage and severe wear stage (Source: Lei Yaguo-Machinery health prognostics: A systematic review from data acquisition to RUL prediction). The vibration signal of the bearing shows a different distribution in each stage: in the stable operation stage, the vibration signal of the bearing is a stationary signal, while in the latter two stages it is a dynamic time-varying signal. Therefore, traditional feature extraction methods based on expert knowledge, such as FFT transformation, cannot process dynamic time-varying signals, and the characteristic distribution of bearings in the initial wear stage and severe wear stage show linear and nonlinear change trends respectively, making it difficult to use a model that is compatible with both distributions to identify the signal.

[0004] In recent years, with the development of machine learning technology, neural networks have demonstrated powerful nonlinear mapping capabilities and feature extraction capabilities for dynamic data. Therefore, some scholars have begun to use encoders based on machine learning to identify the degradation stage of bearing diagnostic signals. For example, Chen Renxiang et al. used the Source Domain Multi-sample Integrated Geodesic Flow Kernel (SDMIGFK) to identify the life stages of spatial rolling bearings; Dong et al. used the Improved Convolutional Neural Network with Anti-interference (ICNNA) to identify the life stages of spatial rolling bearings; Wu Haonian et al. used the Improved Balanced Distribution Adaptation (IBDA) method to complete the life stage identification of spatial rolling bearings.

[0005] In practical applications, the amount of data generated by bearings in the three operating stages is unbalanced. Most of the operating time of the bearings is in the stable operation stage and the initial wear stage. After entering the severe wear stage, they will soon fail. This results in the sample size of the severe wear stage of the bearing being much smaller than that of the other two stages during the machine learning training process, which affects the training effect of the model and greatly reduces the accuracy of the model's recognition of the severe wear stage of the bearing. Therefore, this method proposes a rolling bearing degradation stage recognition method based on ensemble learning, which can train the model under the condition of unbalanced sample size and achieve good recognition effect. Summary of the invention

[0006] In order to overcome the problem that the sample sizes of different degradation stages are unbalanced during the training of the bearing degradation stage identification model, which reduces the model performance, the present invention proposes a rolling bearing degradation stage identification method based on ensemble learning.

[0007] The technical solution of the present invention:

[0008] A rolling bearing monitoring signal degradation stage identification method is provided. The method adopts an ensemble learning stacking algorithm to establish a two-layer stacking ensemble model, including three base learners and one meta learner, and includes the following steps:

[0009] Step 1: Use a sensor to collect the vibration signal of the rolling bearing throughout its life cycle, and organize and save the vibration signal;

[0010] Step 2: divide the collected full life cycle vibration signal into the bearing degradation stage, which is divided into the stable operation stage, the initial wear stage and the severe wear stage. Use the sliding window sampling method to slice the full life cycle vibration signal, and cut the vibration signals of each stage in the full life cycle vibration signal into data samples of the same shape. Since the sample size of the stable operation stage and the initial wear stage is sufficient, overlapping sampling is not performed on them. Only overlapping sampling is performed on the severe wear stage data to increase the number of samples in the severe wear stage, so as to establish a full life cycle data set for the bearing;

[0011] The base learner is designed to reuse the samples of the severe wear stage with the least sample size without causing overfitting. The data sets used to train the three base learners are expressed as follows:

[0012]

[0013] in, is the collected full life cycle vibration signal and is labeled with the corresponding degradation stage; D1, D2, and D3 represent the data sets of the three base learners respectively; x i Represents the collected vibration signal of the entire life cycle, y i represents the corresponding label, where 0 represents the stable operation stage, 1 represents the initial wear stage, and 2 represents the severe wear stage; n represents the total number of samples; the number of samples collected in the three stages are n1, n2, and n3, respectively, and n1>n2>>n3; in the data set of each base learner, it is randomly divided into a training set and a validation set in a ratio of 5:1;

[0014] Step 3: Use the bearing life cycle data set to train three base learners for the task of dividing the bearing degradation stages; and combine the recognition results of the base learners into a high-dimensional data set and input it into the meta-learner for training to obtain the final result;

[0015] Specifically, the encoders of the base learners used for training are composed of convolution blocks, flattening layers and fully connected layers, and the classifier of the base learners is a softmax classifier; wherein, the convolution blocks and the fully connected layers are used to extract the feature vectors of the full life cycle vibration signal, the flattening layer is used to adjust the extracted feature vectors to one dimension, and the softmax classifier identifies the extracted features; the encoder includes 4 convolution blocks, a flattening layer and 3 fully connected layers, each convolution block is composed of convolution, batch normalization, activation and maximum pooling operations in sequence; wherein, the convolution kernel sizes of the 4 convolution blocks are set to 32×1, 3×1, 3×1, 3×1 in sequence; the number of convolution kernels are 16, 32, 64, 64 respectively; the number of neurons in the fully connected layer is 128, 32, 3 in sequence;

[0016] In order to speed up the convergence of the base learner, the adam optimizer is used to update the parameters. The base learner repeats the above process on the training set to update the parameters until the accuracy of the base learner on the validation set exceeds 95%. Then the update is stopped and the base learner is saved.

[0017] In order to fuse the output recognition results of the three base learners, after the base learner training is completed, the recognition results output by the encoders of different base learners are combined into a higher dimensional data set as the meta-learner training set, which is input into the meta-learner for training. The meta-learner training set is expressed as:

[0018] l n ={l1,l2,…,l i}

[0019] l i ={s 1i ,s 2i ,s 3i}

[0020] Among them, l n represents the meta-learner training dataset, l i is the i-th training sample of the meta-learner, s 1i ,s 2i ,s 3i are the recognition results of the three base learners for the i-th sample;

[0021] The meta-learner uses the XGBoost model to further improve the model accuracy and generalization ability. The training is set to stop when the accuracy of the meta-learner does not increase after ten trainings. The XGBoost training process and the final recognition result are expressed as:

[0022] The meta-learner objective function is recorded as:

[0023]

[0024] Among them, L(φ) is the expression in linear space; is the i-th sample l i The predicted value of Regularization term Ω(f k );∑ k Ω(f k ) represents the complexity of k trees. Through this process, the objective function is iteratively updated to find the optimal solution and save the model.

[0025] Step 4: Save the trained model and identify the degradation stage of the rolling bearing.

[0026] In step 4 of the rolling bearing degradation stage identification method based on ensemble learning, in order to save computer resources and improve the recognition reliability in the actual application of the model, the vibration signal of the bearing is collected for 1 second every 5 minutes, and the degradation stage is identified using the trained model. When the model divides the signal to be identified into the same stage for two consecutive times, the bearing is marked as the degradation stage, the data is recorded and the next round of identification is continued.

[0027] Beneficial effects of the present invention:

[0028] The present invention proposes a rolling bearing degradation stage identification method based on ensemble learning, which uses an ensemble learning stacking algorithm to ensure that the fused model has stronger accuracy and generalization ability, and enhances the degradation stage identification ability of the model. The utilization rate of samples in the severe wear stage is increased, and the problem of unbalanced sample quantity caused by the fact that the number of samples in the severe degradation stage of the bearing is much smaller than that in the other two stages is overcome.

[0029] The method can timely identify the working status of the bearing by processing the operating signal emitted by the rolling bearing in real time, and formulate different maintenance strategies according to different health states. That is, the present invention can be used in the field of bearing predictive maintenance, effectively reducing maintenance costs and improving the prevention level of bearing failures. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flow chart of the present invention.

[0031] Figure 2 : It is the model structure diagram, where (a) is the structure diagram of the base learner, and (b) is the internal structure of the convolution block in the base learner.

[0032] Figure 3 This is the vibration amplitude diagram of the bearing over its entire life.

[0033] Figure 4 This is the RMS value variation curve of bearing No. 2.

[0034] Figure 5 This is the final identification result of the degradation stage of bearing No. 2.

[0035] Figure 6 The results of using the saved model to identify the degradation stages of other bearings, where (a) is bearing No. 1, (b) is bearing No. 3, and (c) is bearing No. 5. DETAILED DESCRIPTION

[0036] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0037] The basic concept of the present invention is: using an acceleration sensor to obtain the full life cycle signal of the bearing to train a neural network model based on machine learning. In the sampling stage, a sliding window is used to slice the original signal to expand the training data. In the model training stage, due to the scarcity of degradation samples of the bearing in the severe wear stage, the training set data is unbalanced. Therefore, an ensemble learning-based method is used to train three independent base learners respectively, and the data of the severe wear stage is reused to alleviate the problem of imbalanced data set samples. The training of each base learner reuses the sample data of the severe wear stage, and then the outputs of the three trained models are integrated into a higher dimensional feature vector set and input into the XGBoost classifier for training. After the training is completed, the final recognition result is output. In the actual application stage, the degradation stage of the bearing is identified using the trained model, and the vibration signal of the bearing is collected for 1 second every 5 minutes, and the degradation stage is identified using the trained model. When the model divides the signal to be identified into the same stage for two consecutive times, the bearing is marked as the degradation stage. This working mode can save computer resources and improve the recognition credibility.

[0038] This embodiment uses the XJTU-SY bearing data set to verify the proposed method. The platform consists of an AC motor, a motor speed controller, a rotating shaft, a support bearing, a hydraulic loading system, and a test bearing. It can carry out accelerated life tests of various rolling bearings or sliding bearings under different working conditions and obtain the full life cycle monitoring data of the test bearing. The adjustable working conditions of the test platform mainly include radial force and speed, where the radial force is generated by the hydraulic loading system and acts on the bearing seat of the test bearing, and the speed is set and adjusted by the speed controller of the AC motor. The test bearing is an LDK UER204 rolling bearing. The experiment designed 3 types of working conditions (speed (r / min) is 2100, 2250, 2400 respectively; radial load (kN) is 12, 11, 10 respectively), with 5 bearings in each type of working condition, totaling 15 bearings. This embodiment uses bearing data under the working condition of a speed of 2100r / min and a radial load of 12kN. Bearing No. 2 under this working condition is used for model training. After the training is completed, bearings No. 1, 3, and 5 are used for testing. The specific steps include:

[0039] 1. Use an acceleration sensor to collect the vibration signal of the bearing throughout its life cycle, organize and save the signal, extract the characteristic value that can characterize the bearing degradation state from the vibration signal, and visualize it. The characteristic value here is expressed by the root mean square value, and its calculation method is as follows:

[0040]

[0041] The training data in this embodiment uses the data generated by bearing No. 3 under the working conditions of a speed of 2100r / min and a radial load of 12kN. The vibration waveform is shown in the figure below: Figure 3 As shown in the figure, the bearing was sampled 161 times in total, and each sampling generated 32768 sampling points. The root mean square value distribution is shown in the figure. Figure 4 As shown in the figure, it can be seen that before the 64th sampling, the RMS value fluctuated steadily, and the bearing was in a stable operation stage. The RMS value of the 64th to 155th samplings fluctuated violently, and its amplitude was much larger than that of the first stage. At this time, the bearing was in a stable wear stage; after the 156th sampling, the RMS value fluctuated more violently and rose rapidly. At this time, the bearing was in a severe wear stage, until the 161st sampling, the bearing completely failed. From the degradation data of the bearing, it can be seen that only 6 samples were taken from the severe wear stage to the complete failure of the bearing, and the amount of data was much less than that of the first two stages. If the samples of all three stages are directly used to train the neural network, the serious imbalance of the data will greatly reduce the model's ability to recognize samples in the severe wear stage.

[0042] 2. Use the sliding window sampling method to slice the vibration signal and generate different samples to establish a bearing full life cycle data set; in this embodiment, since the number of sampling times in the stable operation stage and the initial wear stage of the bearing is large enough, there is no need for overlapping sampling to increase the number of samples. Therefore, for the data generated by each sampling, the window width is set to 2048 and the step size is set to 2048; for the data collected in the severe wear stage, the window width is set to 2048 and the step size is set to 512 to increase the sample size. A total of 360 samples are collected as training data for the severe wear stage.

[0043] 3. Build three base learners for training. The encoder of each base learner consists of a convolution block, a flattening layer, and a fully connected layer. The classifier of the base learner is a softmax classifier. Its encoder includes four convolution blocks, a flattening layer, and three fully connected layers in sequence. Each convolution block consists of a convolution layer, batch normalization, activation, and maximum pooling operations in sequence. The operation of the convolution layer can be expressed as:

[0044]

[0045] in and is the weight and bias matrix of the i-th convolution kernel in the l-th layer, x l (j) represents the jth region of the lth layer. The symbol * is the convolution operation.

[0046] After the sample is extracted by the convolution layer, the distribution of the generated feature vector will be different, so batch normalization is needed to adjust the output feature vector to a normal distribution with a mean of 0 and a variance of 1. However, this will lead to a decrease in the network's expressiveness. To prevent this, two adjustment parameters (scaling and displacement parameters) are added to each neuron. These two parameters are learned through training and are used to inversely transform the transformed activation to enhance the network's expressiveness. In summary, the batch normalization process can be described as:

[0047]

[0048] in and (i) represents the input and output of the batch normalization layer, γ (i) and β (i) are the scale and displacement parameters that need to be learned.

[0049] In order to enable the model to acquire nonlinear expression capabilities, an activation layer is used to activate the standardized feature vector. This embodiment uses the RELU function as the activation function. The activation process can be expressed as:

[0050]

[0051] in is the feature vector after batch normalization, yes The activated value of .

[0052] In order to reduce the size of the model and increase computational efficiency, it is necessary to reduce the parameters to be learned. Therefore, it is necessary to perform maximum pooling on the output after activation to shorten the length of the feature vector and save resources for subsequent operations. The maximum pooling operation can be expressed as:

[0053]

[0054] in represents the value of the jth feature of the i-th sample of the l-th layer neural network, t∈[(j-1)W,jW]. W is the width of the pooling unit, Represents the output value of the pooling layer. In this embodiment, the pooling unit width is set to 3.

[0055] In this embodiment, the convolution kernel sizes in the four convolution blocks are set to 32×1, 3×1, 3×1, and 3×1, respectively; the number of convolution kernels is 16, 32, 64, and 64, respectively; and the number of neurons in the fully connected layer is 128, 32, and 3, respectively.

[0056] 4. Randomly shuffle the samples of the three stages, and then divide the collected samples into data sets, which are used to train three independent base learners. In order to reuse the samples of the severe wear stage with the least sample size without causing overfitting, the data sets used to train the three base learners are expressed as:

[0057]

[0058] in, represents the data set, x i represents the collected samples, y i Denotes the corresponding label, where 0 represents the stable operation stage; 1 represents the initial wear stage; 2 represents the severe wear stage, and n represents the total number of samples. The samples in each data set are divided into a training set and a validation set at a ratio of 5:1.

[0059] 5. Use the adam optimizer to optimize the update process of the model parameters. The model continuously repeats the above process on the training set to update the parameters until the accuracy of the model on the validation set exceeds 95%. Then stop updating and save the model.

[0060] 6. After the base learner training is completed, the encoders of different base learners map the samples to be identified to the feature space to generate feature vector sets, and the feature vector sets of the three models are fused into a higher-dimensional feature vector set. Finally, the stacking algorithm is used to train the XGBoost classifier to identify the fused features. The fusion process and the final recognition result are expressed as follows:

[0061]

[0062] Among them, L(φ) is the expression in linear space; is the i-th sample x i The predicted value of Regularization term Ω(f k );∑ k Ω(f k ) represents the complexity of k trees. In this example, XGBoost sets the number of decision trees k = 100, the maximum depth is 6, and the CART tree structure is adopted. Each node has two child nodes, and each leaf node stores the prediction probability of the corresponding category. This example is a three-classification problem, and the softmax function is used to convert the leaf prediction score into a probability distribution. In this embodiment, the training data comes from bearing No. 2. After the model training is completed, the recognition result of bearing No. 2 is as follows Figure 5 The saved model is used to identify the degradation stages of other bearings, and the identification results are shown in Figure 6As shown in the figure, it can be seen that this method can accurately identify the degradation stage of bearing No. 2 and the other three bearings that have never been seen in the training process. This verifies the effectiveness and generalization of the method based on ensemble learning. This method can reuse the data of the severe wear stage to alleviate the problem of unbalanced data set samples without causing overfitting problems.

[0063] 7. In order to further save computer resources and improve the credibility of model recognition, in the online identification of the degradation stage of rolling bearings, the vibration signal of the bearing is collected for 1 second every 5 minutes, and the degradation stage is identified using the trained model. When the model divides the signal to be identified into the same stage for two consecutive times, the bearing is marked as that degradation stage.

Claims

1. A method for identifying the degradation stage of a rolling bearing monitoring signal, characterized in that: The rolling bearing monitoring signal degradation stage identification method adopts ensemble learning stacking algorithm to establish a two-layer stacking ensemble model, including three base learners and one meta learner. The following steps are involved: Step 1: Use a sensor to collect the vibration signal of the rolling bearing throughout its life cycle, and organize and save the vibration signal; Step 2: divide the collected full life cycle vibration signals into bearing degradation stages, namely stable operation stage, initial wear stage and severe wear stage, slice the full life cycle vibration signals using sliding window sampling method, and cut the vibration signals of each stage in the full life cycle vibration signals into data samples of the same shape; overlap sampling is performed on the severe wear stage data to increase the number of samples in the severe wear stage, so as to establish a full life cycle data set for the bearing; The base learner is designed to reuse the samples of the severe wear stage with the least sample size without causing overfitting. The data set used to train the three base learners is randomly divided into a training set and a validation set in a ratio of 5:1; Step 3: Use the full life cycle data set of the bearing to train three base learners respectively for the task of dividing the bearing degradation stages; and combine the recognition results of the base learners into a high-dimensional data set and input it into the meta-learner for training to obtain a trained model; Step 4, saving the trained model and identifying the degradation stage of the rolling bearing; The vibration signal of the bearing is collected for 1 second every 5 minutes, and the degradation stage is identified using the trained model. When the model divides the signal to be identified into the same stage for two consecutive times, the bearing is marked as the degradation stage, the data is recorded and the next round of identification is carried out.

2. The rolling bearing monitoring signal degradation stage identification method according to claim 1 is characterized in that: The data sets of the three base learners are expressed as: in, is the collected full life cycle vibration signal and is labeled with the corresponding degradation stage; D1, D2, and D3 represent the data sets of the three base learners respectively; x i Represents the collected vibration signal of the entire life cycle, y i Represents the corresponding label, where 0 represents the stable operation stage, 1 represents the initial wear stage, and 2 represents the severe wear stage; n represents the total number of samples; suppose the numbers of samples collected in the stable operation stage, initial wear stage, and severe wear stage are n1, n2, and n3 respectively, and n1>n2>>n3.

3. The rolling bearing monitoring signal degradation stage identification method according to claim 1, characterized in that: In step 3, The encoders of the base learners used for training are composed of convolution blocks, flattening layers and fully connected layers, and the classifier of the base learners is a softmax classifier; among them, the convolution blocks and fully connected layers are used to extract the feature vectors of the full life cycle vibration signals, the flattening layers are used to adjust the extracted feature vectors to one dimension, and the softmax classifier recognizes the extracted features.

4. The rolling bearing monitoring signal degradation stage identification method according to claim 3 is characterized in that: The encoder consists of 4 convolutional blocks, a flattening layer, and 3 fully connected layers. Each convolutional block consists of convolution, batch normalization, activation, and maximum pooling operations in sequence. The convolution kernel sizes of the 4 convolutional blocks are set to 32×1, 3×1, 3×1, and 3×1, respectively; the number of convolution kernels is 16, 32, 64, and 64, respectively; the number of neurons in the fully connected layers is 128, 32, and 3, respectively.

5. The rolling bearing monitoring signal degradation stage identification method according to claim 3 is characterized in that: In order to speed up the convergence of the base learner, the adam optimizer is used to update the parameters. The base learner repeats the above process on the training set to update the parameters until the accuracy of the base learner on the validation set exceeds 95%. Then the update is stopped and the base learner is saved.

6. The rolling bearing monitoring signal degradation stage identification method according to claim 3, characterized in that: In order to fuse the output recognition results of the three base learners, after the base learner training is completed, the recognition results output by the encoders of different base learners are combined into a higher dimensional data set as the meta-learner training set, which is input into the meta-learner for training. The meta-learner training set is expressed as: l n ＝{l1,l2,…,l i} l i ={s 1i ,s 2i ,s 3i } Among them, l n represents the meta-learner training dataset, l i is the i-th training sample of the meta-learner, s 1i ,s 2i ,s 3i They are the recognition results of the three base learners for the i-th sample.

7. The rolling bearing monitoring signal degradation stage identification method according to claim 3, characterized in that: The meta-learner uses the XGBoost model to further improve the model accuracy and generalization ability. The training is set to stop when the accuracy of the meta-learner does not increase after ten trainings. The XGBoost training process and the final recognition result are expressed as: The meta-learner objective function is recorded as: Among them, L(φ) is the expression in linear space; is the i-th sample l i The predicted value of Regularization term Ω(f k );∑ k Ω(f k ) represents the complexity of k trees; through this process, the objective function is iteratively updated to find the optimal solution and save the model.

Images