A Method for Predicting the Remaining Useful Life of Bearings Based on Sustainable Storage Pool Computing

Abstract

The present invention provides a method for predicting the remaining useful life of bearings based on sustainable storage pool computing, belonging to the field of engineering prediction and health management systems. The steps are as follows: 1) Perform time-frequency conversion on the original data set; 2) Extract and initialize features from the original data and frequency-domain data; 3) Input the data after feature extraction, the old model, and the stored relevant parameters into the new model and train them; 4) Store the new model and relevant parameters and conduct prediction and evaluation. The method for predicting the remaining useful life of bearings provided by the present invention can alleviate problems such as excessive computing volume, long training time, high requirements for computing devices, inapplicability to applications, and catastrophic forgetting in continuous learning in deep learning.

Description

Technical Field

[0001] The present invention belongs to the field of engineering prediction and health management systems, and relates to a method for predicting the remaining useful life of bearings based on sustainable storage pool computing. Background Art

[0002] With the remarkable development of artificial intelligence and computer computing power, artificial intelligence is gradually being applied to a large number of fields, including the engineering field. There are various problems in the engineering field, such as complex working conditions, a large amount of data, and equipment limitations. It also faces the problem that the data providers are reluctant to open data, making it impossible to immediately apply advanced artificial intelligence technologies. Although transfer learning can solve some of the above problems, it also has some limitations, such as the need for relevant knowledge and non - continuous learning.

[0003] A continuous learning system is a system that combines plasticity (the ability to acquire new knowledge) and stability (the ability to remember old knowledge) as tasks and data change. Therefore, sustainable learning may be a better way to solve this series of problems. A sustainable learning system can improve old models by continuously learning new data, so that they can be applied to all operating conditions. A sustainable learning system can utilize old models and knowledge instead of storing a large amount of old data. When the device computing power is insufficient, it can divide a large amount of data and even provide services by providing a pre - trained continuous learning model to those data providers who are reluctant to disclose data.

[0004] On the other hand, as the core of the prediction and health management system, the accuracy of the remaining useful life prediction has a great impact on its effectiveness. With accurate remaining useful life, engineers can better arrange maintenance, thereby effectively extending the service life of machines, reducing the failure rate, and even preventing catastrophic failures. In addition, due to various reasons such as manufacturing processes and working conditions, the service life of some mechanical components may vary widely, such as bearings. Therefore, the prediction of the remaining useful life is a promising field of continuous learning research.

[0005] However, there are also some limitations in the application of artificial intelligence methods in the prediction of the remaining useful life. For example, neural networks with a forward structure are generally not suitable for dealing with machine learning problems related to time series, especially the prediction of the remaining useful life. Although recurrent neural networks can be used to solve problems related to time series, they still have problems such as overly complex training algorithms, large computational amounts, slow convergence, and difficulty in determining the deep learning structure. In addition, in practical engineering applications, the amount of data from sensors is huge, and a large number of components need to be monitored. Summary of the Invention

[0006] Aiming at the problems existing in the prior art, the present invention provides a method for predicting the remaining useful life of bearings based on sustainable storage pool computing.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] A method for predicting the remaining useful life of bearings based on sustainable storage pool computing, the steps are as follows:

[0009] In the first step, the full life cycle data of the equipment is obtained through experiments, a full life cycle data set is constructed, and the original time domain signal is converted into a frequency domain signal through fast Fourier transform.

[0010] In the second step, feature extraction is performed on the time domain and frequency domain data respectively. The extracted features include the variance, kurtosis, root mean square of the time domain signal, and the mean and variance of the frequency domain signal. Then, all the extracted features are initialized by the absmax method shown in formula (1) to obtain the processed data set.

[0011]

[0012] Among them, x * is the processed data, x is the data before processing, and |x| max is the maximum value of the absolute values of the input of the same type of data.

[0013] In the third step, a new sustainable storage pool computing model is constructed and trained.

[0014] Suppose all the data previously learned by a certain model is uniformly labeled as A, and the newly learned data is labeled as B. According to the old model, related parameters obtained from the learned data A, and the new data B processed in the second step, the steps of continuous learning are realized in a deep storage pool computing model through the method of elastic weight constraint.

[0015] The deep storage pool computing model described above includes a deep storage pool network part and multiple fully connected layers including dropout layers.

[0016] A deep storage pool network part uses a stack of multiple standard storage pool layer networks, where the output of each layer is used as the input of the next layer. At each time step, the state is calculated along the pipeline of the recursive layer, starting from the first pipeline directly provided by the external input and ending at the last pipeline in the storage structure.

[0017] The first layer storage pool is given by formula (2), and other layer storage pools are updated by formula (3). Among them, u(t) is the external input at time step t, and x (i) (t) is the state of the storage layer i at time step t, is the state feedback weight of the i-th layer storage pool, a(i) is the leakage rate of the i-th layer storage pool, W (1) is the connection weight between the input and the first layer storage pool, and W (i) (i > 1) is the connection weight between the (i - 1)-th layer storage pool and the i-th layer storage pool.

[0018]

[0019]

[0020] At each time step t, the overall state of the network is composed of the states of all layer storage pools, i.e., x(t) = x (1) (t), …, x (N) (t). The output y(t) at time step t is calculated according to formula (4), W out is the readout weight matrix adapted to the training set.

[0021] y(t) = W out x(t) = W out (x (1) (t), …, x (N) (t)) (4)

[0022] Considering the limited plasticity of the deep storage pool network, a fully connected layer is added after the deep storage pool layer, and finally a single-node fully connected layer is used to predict the remaining useful life. A dropout layer is added after each fully connected layer, which can effectively improve the stability of the model and reduce the training time of the model.

[0023] In addition, the present invention uses a method of updating the function with elastic weight constraints to achieve the purpose of sustainable learning. The loss function for updating the sustainable storage pool calculation model is as follows:

[0024]

[0025] where λ is the importance parameter of old data, with a default value of 0.9; L B is the loss function without using the sustainable update rule, with a default value of root mean square error; W out，i is the parameter of the readout weight matrix of the current deep storage pool network adapted to the training set, is the parameter of the readout weight matrix of the deep storage pool network of the model trained based on the previous data A; θ fc，i is the parameter of the current fully connected layer, is the parameter of the fully connected layer of the model trained based on the previous data A; F i is the Fisher information matrix. If the learned data is an empty set, all parameters with subscript A are 0.

[0026] Step 4: Store the newly obtained trained model and all parameter matrices of the new model. Use the new model to predict the remaining useful life and conduct an evaluation.

[0027] The beneficial effects of the present invention are as follows:

[0028] By establishing a method for predicting the remaining useful life of bearings based on sustainable storage pool computing, the present invention achieves the effects of alleviating problems such as excessive computational complexity, long training time, high requirements for computing devices, and catastrophic forgetting in continuous learning in deep learning. The method proposed by the present invention is simply designed and easy to use in the deep learning program for predicting the remaining useful life of bearings. Description of the Drawings

[0029] Figure 1 It is a flowchart for implementing a method for predicting the remaining useful life of bearings based on sustainable storage pool computing;

[0030] Figure 2 It is a structural diagram of a standard storage pool layer network;

[0031] Figure 3 It is a structural diagram of a storage pool computing model for predicting the remaining useful life of bearings;

[0032] Figure 4 is a prediction result diagram of a comparative experiment for predicting the remaining useful life of bearings in continuous learning; Figure (a) is the prediction effect diagram of the comparative experiment for continuous learning on all subsets; Figure (b) is the maximum, minimum, and average value diagram after the continuous learning of the comparative experiment. Detailed Implementation Manner

[0033] To make the method problems solved by the present invention, the adopted method solutions, and the achieved method effects clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Additionally, it should be noted that for the sake of description, only parts related to the present invention are shown in the drawings rather than all the content.

[0034] Figure 1This is the implementation flowchart of a method for predicting the remaining useful life of bearings based on sustainable storage pool computing provided by an embodiment of the present invention. The specific implementation experiment uses the most authoritative FEMTO-ST bearing life dataset in the industry, which can be found on the webpage of NASA Prognostics Data Repositor. It is the most authoritative dataset in the field of predicting the remaining useful life of bearings, and the experimental platform is designed and implemented in the AS2M department of the FEMTO-ST Institute. The FEMTO-ST bearing dataset was obtained from an accelerated bearing degradation test. In this implementation, the FEMTO-ST bearing life dataset is divided into seven subsets according to the bearing numbers, and the division is shown in Table II. In addition, the environment for training the experimental model is an Intel Core i5-1135G7 2.42GHz CPU and TensorFlow 2.9.0. The results of this embodiment are the average values of five experiments.

[0035] Table I Dataset Task Division

[0036] Task serial number 1 2 3 4 5 6 7 Bearing serial number 1-1 1-2 1-3 1-4 1-5 1-6 1-7

[0037] The specific implementation experiment process includes:

[0038] First, convert the original time-domain signals of the seven subsets into frequency-domain signals through fast Fourier transform.

[0039] Second, extract features from the time-domain and frequency-domain data respectively and perform initialization processing. Then, merge the time-domain features and frequency-domain features extracted from each subset and perform a time sliding window. The size of the time sliding window selected in this implementation is 30, and the data of the seven subsets after data processing are obtained.

[0040] Third, first construct a sustainable storage pool computing model. The sustainable storage pool computing model consists of a deep storage pool network part and multiple fully connected layers containing dropout layers.

[0041] A standard storage pool layer network is as Figure 2 shown. The structure of the deep storage pool computing model used in this implementation is as Figure 3 shown. In this implementation, the number of layers of the deep storage pool network is five and the number of fully connected layers containing dropout layers is two.

[0042] In addition, this implementation uses the method of elastic weight constraint to update the function. Assume that all the previously learned data is labeled as A, and the newly learned data to be learned is labeled as B. The loss function for updating the sustainable storage pool computing model is as follows:

[0043]

[0044] Among them, λ is the importance parameter of old data, which is fixed at 0.9 in this implementation; L B is the loss function without using the sustainable update rule, and the root mean square error is the default; W out，i is the parameter of the readout weight matrix of the current deep storage pool network adapted to the training set, is the parameter of the readout weight matrix of the deep storage pool network of the model adapted to the training set trained based on the previous data A; θ fc，i is the parameter of the current fully connected layer, is the parameter of the fully connected layer of the model trained based on the previous data A; F i is the Fisher information matrix. If the learned data is an empty set, all parameters with subscript A are 0.

[0045] Then, use the seven subsets after data processing to train the constructed sustainable storage pool calculation model. The input order is to input continuously in ascending order of the task number for the continuous learning task, and the steps are as follows:

[0046] 1) Input the processed task 1 data to train the sustainable storage pool calculation model, and store the trained model 1 and its parameter matrix.

[0047] 2) Input the processed task 2 data, model 1 and its parameter matrix to train the sustainable storage pool calculation model, and store the trained model 2 and its parameter matrix.

[0048] ……

[0049] 7) Input the processed task 7 data, model 6 and its parameter matrix to train the sustainable storage pool calculation model, and store the trained model 7 and its parameter matrix.

[0050] Fourthly, summarize all the stored seven models, and use the seven models to predict the remaining useful life of all subsets, and evaluate the effectiveness of the model and the method.

[0051] Figure 4 is the prediction result graph of the bearing remaining useful life prediction comparison experiment in continuous learning. The left figure shows the performance of the model on all subsets during continuous learning from 1 to 7. The TRMSE in the figure is the root mean square error of the full task set. The right figure shows the maximum, minimum and average values predicted by the model for the full task data after the entire continuous learning process. The training duration of the entire continuous learning process of all four methods in the figure is shown in Table 3.

[0052] Table 2 Training durations of each method

[0053]

[0054] As shown in the implementation results, the method for predicting the remaining effective life of bearings in sustainable storage pool computing can effectively improve problems such as excessive deep learning computation, long training time, high requirements for computing devices, inapplicability to applications, and catastrophic forgetting in continuous learning.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the method solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: modifying the method solutions recorded in the foregoing embodiments, or equivalently replacing some or all of the method features therein, does not cause the essence of the corresponding method solutions to deviate from the scope of the method solutions of the embodiments of the present invention.

Claims

1. A method for predicting the remaining useful life of bearings based on sustainable storage pool computing, characterized in that, it includes the following steps: First step, obtain the full life cycle data of the equipment through experiments, construct a full life cycle data set, and convert the original time domain signal into a frequency domain signal through fast Fourier transform; Second step, extract features from the time domain and frequency domain data respectively, and use the absmax method to initialize all the extracted features to obtain a processed data set; Third step, construct and train a new sustainable storage pool computing model; Assume that all the data previously learned by a certain model is labeled as A, and the new data to be learned is labeled as B; according to the old model, relevant parameters obtained from the learned data A, and the new data B processed in the second step, implement the step of continuous learning in a deep storage pool computing model through the method of elastic weight constraint; The deep storage pool computing model includes a deep storage pool network part and multiple fully connected layers containing dropout layers; A deep storage pool network part uses a stack of multiple standard storage pool layer networks, where the output of each layer is used as the input of the next layer; at each time step, the state is calculated along the pipeline of the recursive layer, starting from the first pipeline directly provided by the external input and ending at the last pipeline in the storage structure; The first-layer storage pool is given by Equation (2), and the other-layer storage pools are updated by Equation (3); where u(t) is the external input at time step t, and x (i) (t) is the state of the storage layer i at time step t, is the state feedback weight of the i-th layer storage pool, a (i) is the leakage rate of the i-th layer storage pool, W (1) is the connection weight between the input and the first-layer storage pool, and W (i) (i > 1) is the connection weight between the (i - 1)-th layer storage pool and the i-th layer storage pool; At each time step t, the overall state of the network consists of the states of all layer storage pools, i.e., x(t) = x (1) (t), …, x (N) (t); the output y(t) at time step t is calculated according to formula (4), and W out is the readout weight matrix adapted to the training set; y(t) = W out x(t) = W out (x (1) (t), …, x (N) (t)) (3) Add a fully connected layer after the deep storage pool layer, and finally use a single-node fully connected layer to predict the remaining useful life; add a dropout layer after each fully connected layer to improve the stability of the model and reduce the model training time; Adopt the method of elastic weight constraint to update the function, and thus achieve the purpose of sustainable learning; the loss function for updating the sustainable storage pool computing model is as follows: Among them, λ is the importance parameter of old data, with a default value of 0.9; L B is the loss function without using the sustainable update rule, with a default value of root mean square error; W out，i is the parameter of the readout weight matrix of the current deep storage pool network adapted to the training set, is the parameter of the readout weight matrix of the model deep storage pool network adapted to the training set trained based on the previous data A; θ fc，i is the parameter of the current fully connected layer, is the parameter of the fully connected layer of the model trained based on the previous data A; F i is the Fisher information matrix; if the learned data is an empty set, all parameters with subscript A are 0; Fourth step, store the newly obtained model after training, and store all the parameter matrices of the new model; use the new model to predict the remaining useful life and evaluate it.

2. A method for predicting the remaining useful life of bearings based on sustainable storage pool computing according to claim 1, characterized in that, the features extracted in the second step include the variance, kurtosis, root mean square of the time domain signal, and the mean and variance of the frequency domain signal.

3. A method for predicting the remaining useful life of bearings based on sustainable storage pool computing according to claim 1, characterized in that, in the second step, formula (1) is used to initialize all the extracted features where x * is the processed data, x is the data before processing, and |x| max is the maximum value of the absolute values of the input data of the same type.

Images