An INFO-based method for predicting the remaining useful life of bearings

Abstract

The present invention relates to the technical field of deep learning, and particularly relates to a method for predicting the remaining life of a bearing based on INFO, including preprocessing data, first decomposing vibration data by using the INFO-VMD method, extracting features from the decomposed data to form a feature set; performing feature dimensionality reduction on the feature set; inputting the degraded data of the training set into the INFO-DELM model for training to obtain the training parameters of DELM, setting the DELM training parameters, and then inputting the degraded data of the test set and the training set into the DELM model to obtain prediction data; first filtering the prediction data by using a moving average filter; and then fitting the filtered data by using a polynomial fitting method to obtain the final predicted life. The present invention solves the problem of the accuracy of predicting the remaining life of large-scale mechanical equipment, prevents the occurrence of faults, reduces the accident risk caused by equipment failures, thereby saving maintenance costs and improving production efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning, and particularly to a method for predicting the remaining life of a bearing based on INFO. Background Art

[0002] Deep learning and self-maintenance are one of the technical characteristics of intelligent manufacturing; by continuously learning in practice to enrich its knowledge base, the typical feature of an intelligent manufacturing system is to have a deep self-learning function with a certain logical structure. With the in-depth promotion of intelligent manufacturing technology, the function of bearing life prediction is continuously expanding according to user business needs. Rolling bearings are the most widely used part in rotating machinery and are also one of the key components. Problems in the operation of bearings are likely to lead to mechanical failures, resulting in economic losses such as production stoppages at least, and even serious personal safety accidents. Therefore, the safe and stable operation of bearings is crucial for rotating machinery. Therefore, the research on predicting the remaining life of bearings has important significance and practical value.

[0003] With the continuous development of information technology and manufacturing level, large-scale mechanical equipment is continuously developing towards high-precision, high-efficiency and diverse complex working conditions; large-scale mechanical equipment works under extreme conditions (high temperature, overload), and is prone to failures, causing serious economic losses; predicting the remaining life is an important technical means to ensure the reliability, safety and economy of operating equipment, prevent the occurrence of failures, reduce the accident risk caused by equipment failures, thereby saving maintenance costs and improving production efficiency. Summary of the Invention

[0004] Aiming at the deficiencies of existing algorithms, the present invention proposes an INFO-VMD method for problems such as high complexity of bearing vibration data and difficulty in selecting VMD parameters, and uses INFO to optimize the optimal number of layers k and penalty coefficient α of VMD; aiming at the problems of large feature differences and low prediction accuracy, a prediction model is proposed, which consists of three parts: feature dimensionality reduction, INFO-DELM training model and data fitting.

[0005] The technical solution adopted by the present invention is: a method for predicting the remaining life of a bearing based on INFO includes the following steps:

[0006] Step 1: Preprocess the data, first decompose the vibration data using the INFO-VMD method, and extract features from the decomposed data to form a feature set;

[0007] Further, specifically including:

[0008] S11: Decompose the vibration data using VMD, and extract the minimum envelope entropy of the decomposed modes as the fitness function of the INFO algorithm;

[0009] S12. Initialize the parameters of INFO, and set the maximum number of iterations and population size;

[0010] S13. Use the INFO algorithm to perform iterative optimization on VMD to obtain the optimal number of layers k and penalty coefficient α;

[0011] S14. Set the decomposition parameters of VMD with the optimal number of layers k and penalty coefficient α, decompose the vibration data to obtain k-layer modal components; extract time-domain and frequency-domain features for the modes divided into k layers respectively.

[0012] Step 2. Perform feature dimensionality reduction on the feature set;

[0013] Specifically include:

[0014] S21. Use the method of combining random forest and monotonic algorithm to screen the features, and obtain the optimal feature subset through sorting by importance and monotonicity;

[0015] Furthermore, the importance evaluation index of the random forest is measured by GI, and the GI of each feature is calculated m score, and the formula is:

[0016]

[0017] Among them, m represents the number of features, H represents that there are H classes in the group of data, and P mh represents the proportion of h in the node;

[0018] Furthermore, the formula of the monotonic algorithm is:

[0019]

[0020] Among them, n is the number of acquisition points of the test group, l is the number of objects; i represents the number of groups of features in the test group, and j represents the number of features in a single group of features; is the j-th feature value in the i-th group of features, Pdiff represents the positive difference, and Ndiff represents the negative difference.

[0021] S22. Use principal component analysis to further reduce the dimension of the optimal feature subset to obtain multiple principal component components, and use the first principal component component as the degradation index;

[0022] Step 3. Input the degradation data of the training set into the INFO-DELM model for training to obtain the training parameters of DELM, set the DELM training parameters, and then input the degradation data of the test set and the training set into the DELM model to obtain the prediction data.

[0023] Furthermore, specifically include:

[0024] S31. Initialize the parameters of INFO and DELM, import the degradation index of the test set into the DELM model for training, and confirm the optimal vector through INFO;

[0025] S32. Use the three steps of update rule, vector combination and local search in INFO to find the optimal vector. If the traversal is not completed, continue the update rule until it is completed; in the case where each iteration traversal is completed, enter the next iteration, restart the update rule until all iterations are completed, and update the optimal vector if a better vector appears after the iteration is completed;

[0026] S33. After iteratively optimizing the input layer weights and hidden layer thresholds of DELM through the INFO algorithm, obtain the optimal input layer weights and optimal hidden layer thresholds; set the DELM parameters to the optimal parameters, and import the degradation index of the test set into the DELM model to obtain the prediction data.

[0027] Step 4. First, filter the prediction data using a moving average filter; then fit the filtered data using polynomial fitting to obtain the final predicted life;

[0028] Furthermore, the polynomial expression of the predicted life is:

[0029] y = a1x w + a2x w-1 + … + a w x 1 + c (3)

[0030] where w takes positive integer values, a1, a2, …, a w are polynomial coefficients, and c is a constant.

[0031] Advantages of the present invention:

[0032] 1. Compared with traditional preprocessing methods such as EMD and VMD, the INFO-VMD method has the characteristics of high accuracy, fast convergence and good robustness;

[0033] 2. Compared with classical prediction models, the prediction model of the present invention has higher accuracy. Description of the Drawings

[0034] Figure 1 is the flowchart of the INFO-based bearing remaining life prediction method of the present invention;

[0035] Figure 2 is the INFO-VMD optimization flowchart of the present invention;

[0036] Figure 3 is the INFO-DELM optimization flowchart of the present invention;

[0037] Figure 4 It is a comparison chart of the convergence curves of the INFO algorithm and existing algorithms;

[0038] Figure 5 It is a comparison chart of the convergence curves of the INFO-VMD algorithm and existing algorithms;

[0039] Figure 6 It is a comparison chart of the four principal components of PCA. Specific implementation manners

[0040] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner. Therefore, it only shows the components related to the present invention.

[0041] The dataset "IEEE PHM 2012 Data Challenge" includes test data under 3 working conditions, namely working condition 1: load 4000N, rotation speed 1800 rpm; working condition 2: load 4200N, rotation speed 1650 rpm; working condition 3: load 5000N, rotation speed 1500 rpm. The sampling frequency of all experimental groups in this dataset is 25.6 kHz, recorded every 10 seconds, and the recording time for each time is 0.1 second. There are 3 experimental groups with different working conditions in the dataset, among which 1_1, 1_2, 2_1, 2_2, 3_1 and 3_2 are used as the training set, and 1_3, 1_4, 1_5, 1_6, 1_7, 2_3, 2_4, 2_5, 2_6, 2_7 and 3_3 are used as the test set.

[0042] As Figure 1 shown, a bearing remaining life prediction method based on INFO includes the following steps:

[0043] Step 1: Preprocess the data. First, decompose the vibration data using the INFO-VMD method, and extract features from the decomposed data to form a feature set.

[0044] Extract features from the training set and the test set respectively using the INFO-VMD feature extraction method to form a feature set, where the vector average algorithm (weIghted meaN oF vectOrs, INFO), variational mode decomposition (Variational Mode De-composition, VMD); the feature extraction process is as Figure 2As shown in the figure, the VMD is used to decompose the vibration data, and the minimum envelope entropy is extracted from the decomposed modes as the fitness function of the INFO optimization algorithm; the parameters of INFO are initialized, the maximum number of iterations is set to 9, and the population size is set to 30; the INFO algorithm is used to perform iterative optimization on VMD to obtain the optimal number of layers k and the penalty coefficient α; the decomposition parameters of VMD are set with the optimal number of layers k and the penalty coefficient α, and the vibration data is decomposed to obtain k-layer modal components; the time-domain and frequency-domain features are respectively extracted from the modes divided into k layers.

[0045] In this embodiment, a total of 16 time-domain features and 9 frequency-domain features are extracted; the comparison between INFO and SOA, WHO, and SSA Figure 4 As shown in the figure, the fitness function for comparison uses a unimodal test function. The initialization parameters of the four optimization algorithms are the same, the population size is 30, and the number of iterations is 500; from Figure 4 it can be seen that INFO has a faster convergence speed than SOA, WHO, and SSA, and the iteration speed is greatly improved compared with SOA and SSA.

[0046] INFO-VMD is used to conduct experimental comparisons with GWO-VMD, WOA-VMD, and POA-VMD respectively. Taking the 3_3 dataset as an example, the population size is set to 30, and the maximum number of iterations is set to 30. From Figure 5 it can be seen that INFO-VMD reaches the optimal fitness value at the 5th iteration, which is faster than 6 times of WOA-VMD, much faster than 18 times of GWO-VMD and 24 times of POA-VMD, verifying the fast-convergence characteristic of INFO for VMD.

[0047] Step 2: Perform feature dimensionality reduction on the feature set;

[0048] The feature set is obtained through Step 1. First, the features are screened by combining the random forest and monotonicity algorithm, and the optimal feature subset is obtained by sorting according to importance and increase / decrease;

[0049] Then, the principal component analysis is used to further reduce the dimensionality of the optimal feature subset to obtain multiple principal component components, and the first principal component component is used as the degradation index; the importance evaluation index of the random forest is measured by GI, and now the GI of each feature needs to be calculated m The scoring m represents the number of features, H represents that there are a total of H classes in this group of data, and P mh represents the proportion of h in the node, as shown in formula (1):

[0050]

[0051] The value range of the monotonicity algorithm index is [0, 1]. The closer the index is to 1, the better the representation ability of the feature;

[0052]

[0053] Among them, n is the number of acquisition points of the test group, l is the number of objects, and the value of l is 1; i represents the number of groups of features in the test group, and j represents the number of features in a single group of features; is the j-th feature value in the i-th group of features, Pdiff represents the positive difference, and Ndiff represents the negative difference;

[0054] Such as Figure 6 As shown, the four principal components extracted are shown. The monotonicity indexes of each principal component are 0.1453, 0.0085, 0.0199, and 0.0028 respectively. It can be concluded from the data that the monotonicity of the first principal component is much higher than that of other principal components, and it has a strong degradation trend. The first principal component is used as the degradation index.

[0055] Step 3: Model training. Input the degradation data of the training set into the INFO-DELM model for training. The Deep Extreme Learning Machines (DELM) obtains the best training parameters of DELM. Set DELM with the optimal training parameters, and then input the degradation data of the test set and the training set into the DELM model to obtain the prediction data;

[0056] In this embodiment, the maximum number of iterations of the INFO optimization parameter is set to 20, and the population size is set to 50; Input the degradation data of each group of the training set into the INFO-DELM model for training to obtain the best training parameters of DELM. Set DELM with the optimal training parameters, and then input the degradation index data of the test set and the training set into the DELM model to obtain the prediction data; In order to verify the accuracy of the INFO-DELM model, compare it with three other models (SVR, LSSVM, and ELM), compare the obtained error scores. The comparison verification uses the data of the Bearing3_3 experimental group, and the error comparison is shown in Table 1. It can be seen from Table 1 that the error values of the present invention are the lowest, verifying that the accuracy of the INFO-DELM model is relatively high.

[0057] Such as Figure 3 is the flowchart of the INFO-DELM model, including:

[0058] step1: Initialize the parameters of INFO and DELM. Import the degradation index of the test set into the DELM model for training, and confirm the best vector through INFO;

[0059] Step 2: Use the three steps of update rule, vector combination, and local search in INFO to find the optimal vector. If the traversal is not completed, continue to update the rule until it is completed. In the case where each iteration traversal is completed, enter the next iteration, restart the update rule, and continue until all iterations are completed. If a better vector appears after the iteration is completed, update the optimal vector.

[0060] Step 3: After iteratively optimizing the input layer weights and hidden layer thresholds of DELM through the INFO algorithm, obtain the optimal input layer weights and optimal hidden layer thresholds. Set the DELM parameters to the optimal parameters, and import the test set degradation index into the DELM model to obtain the prediction data.

[0061] Table 1 Comparison of error rates between the model proposed in this example and other models

[0062]

[0063] Step 4: Data fitting prediction. First, filter the prediction data using a moving average filter, and then fit the filtered data using polynomial fitting to obtain the final predicted life. The polynomial expression used is as follows:

[0064] y = a1x w + a2x w-1 + … + a w x 1 + c (3)

[0065] Among them, w takes positive integer values, a1, a2, …, a w are polynomial coefficients, c is a constant. Record the true remaining service life time of all test sets and the predicted remaining service life time respectively. Verify the prediction results through the experimental error formula, and the experimental error formula is as follows:

[0066]

[0067] Among them, ActRUL i represents the true RUL value of the bearing, RUL i is the predicted RUL value of the bearing, and Er i is the error value of the i-th group of experimental groups.

[0068]

[0069]

[0070] Among them, A i is the accuracy score of the i-th group, and the mean percentage calculation formula is as follows:

[0071]

[0072] Where: Er i is the percentage error of the i-th group of experiments.

[0073] The comparison of the prediction results with LSTM and GRU is shown in Table 2. It can be seen from Table 2 that for the experimental groups 1-3, 1-4, 1-5, 1-7, 2-3, 2-6, 2-7, and 3-3 in the 11 groups of data tested by the prediction model of the present invention, the prediction errors are all lower than those of the comparative model. The mean error of the model of the present invention is 12.15%, lower than 22.10% of the LSTM-based method and 32.48% of the GRU-based method. At the same time, the total score of the prediction model of the present invention is 0.62, which is 0.31 higher than the score of the LSTM-based method and 0.36 higher than the score of the GRU-based method. The effectiveness of the present invention for remaining life prediction is verified through experiments, and it has a high prediction accuracy.

[0074] Table 2 Comparison of prediction results

[0075]

[0076] Taking the above ideal embodiments based on the present invention as an inspiration, through the above description, relevant staff can completely make various changes and modifications without departing from the technical idea of the present invention. The technical scope of the present invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A method for predicting the remaining life of a bearing based on INFO, characterized in that, It includes the following steps: Step 1: Decompose the vibration data using the INFO-VMD method, and extract features from the decomposed data to form a feature set; Step 2: Perform feature dimensionality reduction on the feature set; Step 2 specifically includes: S21: Use a method combining random forest and monotonic algorithm to screen features, and obtain an optimal feature subset through sorting by importance and monotonicity; The importance evaluation index of the random forest is measured by GI and the GI m score of each feature is calculated. The formula is: ，(1) Among them, m represents the number of features, H indicates that the group of data has a total of H categories, P mh represents the proportion occupied by h in the node; The formula of the monotonic algorithm is: (2) Among them, n is the number of acquisition points of the test group, l is the number of objects; i represents the number of groups of features in the test group, j represents the number of features in a single group of features; is to represent the i th group of features in the j th feature value, Pdiff represents the positive difference, Ndiff represents the negative difference; S22: Use principal component analysis to further reduce the dimensionality of the optimal feature subset, obtain multiple principal component components, and use the first principal component component as the degradation index; Step 3: Input the degradation data of the training set into the INFO-DELM model for training to obtain the training parameters of DELM, set the DELM training parameters, and then input the degradation data of the test set and the training set into the DELM model to obtain prediction data; Step 3 specifically includes: S31: Initialize the parameters of INFO and DELM, import the degradation index of the test set into the DELM model for training, and confirm the optimal vector through INFO; S32: Use the three steps of update rule, vector combination and local search in INFO to find the optimal vector. If the traversal is not completed, continue the update rule until it is completed; in the case of each iteration traversal completion, enter the next iteration, restart the update rule until all iterations are completed, and update the optimal vector if a better vector appears after the iteration is completed; S33: Iteratively optimize the input layer weights and hidden layer thresholds of DELM through the INFO algorithm to obtain the optimal input layer weights and optimal hidden layer thresholds; set the DELM parameters to the optimal parameters, and import the degradation index of the test set into the DELM model to obtain prediction data; Step 4: First, filter the prediction data using a moving average filter; then fit the filtered data using polynomial fitting to obtain the final predicted life; The polynomial expression of the predicted life is: (3) Among them, w takes positive integer values, a 1 , a 2 , …, a w are polynomial coefficients, c is a constant.

2. The INFO-based bearing remaining life prediction method according to claim 1, wherein, Step 1 specifically includes: S11: Decompose the vibration data using VMD, and extract the minimum envelope entropy of the decomposed modes as the fitness function of the INFO algorithm; S12: Initialize the parameters of INFO, and set the maximum number of iterations and population size; S13: Use the INFO algorithm to iteratively optimize VMD to obtain the optimal number of layers k and penalty coefficient α; S14: Set the decomposition parameters of VMD using the optimal number of layers k and penalty coefficient α, decompose the vibration data to obtain k-layer modal components; extract time-domain and frequency-domain features from the k-layer modes respectively.

Images