Device for predicting remaining life of bearing and associated bearing device and method
By measuring the vibration of the bearings and using the neural network for prediction, the problem of difficulty effectively predicting the surplus life of the bearing in the existing technology has been solved, and the quantitative prediction of the surplus life of the bearing is realized, and the productivity is improved.
Patent Information
- Application Number
- CN202410955414.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-07-27
- Filing Date
- 2024-07-17
- Publication Date
- 2025-01-28
AI Technical Summary
The prior art is difficult to effectively predict the remaining life of bearings, resulting in unexpected stopping of the machine and causing productivity losses.
By measuring the vibration of the bearings and using the neural network for prediction, the remaining life of the bearings is determined. The neural network can predict the results of the original vibration measurement and predict it with the characteristics of unknown bearing.
A quantitative prediction of the surplus life of the bearings, reducing the risk of accident stopping of the machine, and increasing productivity.
Smart Images
Figure CN119378349A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to prediction of the remaining life (RL) of a bearing.
[0002] More specifically, the present invention relates to methods and apparatus for predicting RL of a bearing. Background Art
[0003] The RL of a bearing is defined as the remaining duration until the end of the bearing operation, which is defined as the occurrence of a spall in the bearing rings or in the rolling elements of the bearing.
[0004] It is well known that bearing failure can cause unplanned downtime, unforeseen costs and even potential disaster.
[0005] In order to detect a fault in the bearing, the bearing is equipped with at least one sensor for monitoring the bearing to detect a fault in the bearing, such as a crack in one of the rings of the bearing or in one of the rolling elements of the bearing.
[0006] Typically, vibration sensors are used to monitor the condition of bearings. Spectral analysis of the vibration measurements is used to detect the presence of cracks.
[0007] However, when the onset of cracking is detected, the bearing needs to be replaced, which requires the machine in which the faulty bearing is installed to be stopped, which results in a loss of productivity.
[0008] It may not be possible to prevent unexpected machine stops by planning preventive maintenance operations. Summary of the invention
[0009] The present invention therefore aims to predict the RL of a bearing and in particular to optimize the productivity of a machine in which the bearing is installed.
[0010] According to one aspect, a method for predicting the remaining life of a bearing is proposed.
[0011] The method includes:
[0012] - measure bearing vibrations, and
[0013] - Implementing a neural network configured to determine the remaining life of the bearing based on the measured vibrations.
[0014] The device 9 provides a quantitative prediction of the remaining life of the bearing 4 based on the vibration measurement results of the bearing 4.
[0015] The vibration measurements are fed into a neural network without preprocessing to determine the RL of the bearing based on the raw vibration measurements.
[0016] Determine the RL of a bearing without knowing the bearing characteristics.
[0017] According to another aspect, a method for training a neural network configured to determine a remaining life of a bearing based on measured vibrations is presented.
[0018] The method includes:
[0019] - obtaining, at least for a training bearing, at least one training set comprising vibration measurements of the training bearing and a measured remaining life of the training bearing associated with the vibration measurements,
[0020] - implementing a neural network to determine a first output remaining life from the vibration measurements of a training set,
[0021] - performing a first comparison between the measured remaining lifetime of the training set and the first output remaining lifetime, and
[0022] - Tuning the weights of the neural network according to the result of the first comparison.
[0023] Advantageously, the method further comprises:
[0024] - for at least one verification bearing, obtaining at least one verification set, the at least one verification set comprising vibration measurement results of the verification bearing and a measured remaining life of the verification bearing associated with the vibration measurement results,
[0025] - implement a neural network to determine the second output remaining life from the vibration measurements of the validation set,
[0026] - performing a second comparison between the measured remaining lifetime of the validation set and the second output remaining lifetime,
[0027] Wherein, tuning the weight of the neural network further comprises tuning the weight of the neural network according to the result of the second comparison.
[0028] Preferably, each comparison between said remaining lifetime and said output remaining lifetime comprises determining a mean absolute error between said remaining lifetime and said output remaining lifetime.
[0029] According to another aspect, a device for predicting the remaining life of a bearing is provided.
[0030] The device includes:
[0031] - a memory storing a neural network configured to determine a remaining life of the bearing based on the measured vibrations of the bearing, and
[0032] - implementing a device configured to implement a neural network.
[0033] Preferably, the neural network comprises at least one stack of three layers, a dense layer and at least one recurrent layer comprising at least one recurrent unit, the stack of three layers comprising a convolutional layer, a batch normalization layer and a max pooling layer.
[0034] Advantageously, the recurrent unit comprises a long short term memory unit.
[0035] Preferably,
[0036] - the implementing device is further configured to implement a neural network to determine a first output remaining life based on vibration measurements of at least one training set, the training set including vibration measurements of training bearings and measured remaining lives of the training bearings associated with the vibration measurements,
[0037] The device also includes:
[0038] - a comparison device configured to perform a first comparison between the measured remaining lifetime of the training set and the first output remaining lifetime, and
[0039] - A tuning device configured to tune the weights of the neural network depending on the result of the first comparison.
[0040] Advantageously:
[0041] - the implementing device is further configured to implement a neural network to determine a second output remaining life based on vibration measurements of at least one validation set, the validation set including vibration measurements of a validation bearing and a measured remaining life of the validation bearing associated with the vibration measurements,
[0042] - the comparison device is further configured to perform a second comparison between the measured remaining lifetime of the validation set and the second output remaining lifetime, and
[0043] - The tuning device is further configured to tune the weights of the neural network depending on the result of the second comparison.
[0044] According to another aspect, a bearing arrangement is provided.
[0045] The bearing arrangement comprises a bearing, a sensor configured to measure vibrations of the bearing, and a device as defined above connected to the sensor. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Other advantages and features of the present invention will appear, but are in no way limiting, from an examination of the detailed description of embodiments and the accompanying drawings, in which:
[0047] Figure 1 A rotating machine according to the invention is schematically shown;
[0048] Figure 2 An example of a device for predicting the remaining life of a bearing according to the present invention is schematically shown;
[0049] Figure 3 schematically shows an example of the architecture of a neural network according to the invention, and
[0050] Figure 4 An example of a method for training a neural network according to the present invention is shown. DETAILED DESCRIPTION
[0051] refer to Figure 1 , Figure 1 A partial longitudinal cross section of the machine 1 is schematically shown.
[0052] The machine 1 comprises a housing 2 and a shaft 3 supported in the housing 2 by rolling bearings 4, such as roller bearings or ball bearings.
[0053] The rolling bearing 4 is provided with a rotating ring 5 mounted on the shaft 3 and a
[0054] A stationary ring 6 is mounted in a hole of the housing 2. The stationary ring 6 radially surrounds the rotating ring 5. The rotating ring 5 and the stationary ring 6 rotate concentrically relative to each other.
[0055] The rolling bearing 4 is also provided with a row of rolling elements 7, which are radially inserted between the inner and outer raceways of the rotating ring 5 and the stationary ring 6. In the illustrated example, the rolling elements 7 are balls. Alternatively, the rolling bearing may include other types of rolling elements 7, such as rollers. In the illustrated example, the rolling bearing includes a row of rolling elements 7. Alternatively, the rolling bearing may include multiple rows of rolling elements.
[0056] A sensor 8 is installed in the housing 2 to measure the vibration of the bearing 4 .
[0057] The sensor 8 may be mounted on a hole in the housing 2 .
[0058] In a variant, the sensor 8 could be mounted elsewhere on the machine, for example close to the stationary collar 6 or in the vicinity of the housing 2 .
[0059] The sensor 8 transmits an item of data DATA representative of the vibrations of the bearing 4 to means 9 for predicting the remaining life RL of the bearing 4 .
[0060] The bearing 4 , the sensor 8 and the device 9 form a bearing arrangement.
[0061] Figure 2 An example of a device 9 is schematically shown.
[0062] The device 9 comprises a memory 10 storing a neural network 11 for processing items of data DATA transmitted by the sensor 8 , implementation means 12 , tuning means 13 and comparison means 14 .
[0063] The implementation device 12 includes, for example, a processing unit implementing the neural network 11 , an implementation device 12 , a tuning device 13 and a comparison device 14 .
[0064] Figure 3 An example of the architecture of the neural network 11 is schematically shown.
[0065] The neural network 11 can be a hybrid convolutional neural network CNN / recurrent neural network RNN.
[0066] The neural network 11 comprises, for example, a first stack 15 of three layers, a second stack 16 of three layers, a first recurrent layer 17 , a second recurrent layer 18 and a dense layer 19 , which successively process the data items DATA to determine the RL of the bearing 4 .
[0067] Each stack 15, 16 of three layers comprises a convolutional layer 15a, 16a, a batch normalization layer 15b, 16b and a max pooling layer 15c, 16c that successively processes data items.
[0068] The stacks 15 , 16 extract features from the data item DATA and reduce the size of the input signal of the first recurrent layer 17 .
[0069] This reduction in size is a key step that needs to be accomplished to train the neural network 11 in a reasonable amount of time and using reasonable resources.
[0070] The rectified linear unit (ReLu) activation function can be used for convolution. Before applying the activation function, a batch normalization operation is applied after each convolution.
[0071] The first circulation layer 17 includes, for example, three circulation units 17a, 17b, 17c, and the second circulation layer 18 includes, for example, one circulation unit 18a.
[0072] Each recurrent unit 17a, 17b, 17c, 18a includes, for example, a long short-term memory unit (LSTM).
[0073] The first and second recurrent layers 17, 18 aim to capture the temporal dependencies of the data items DATA.
[0074] Dense layer 19 includes ReLu, which gives the prediction of RL until bearing 4 fails.
[0075] Implementing ReLu in dense layer 19 allows neural network 11 to transmit any positive real number of RL of bearing 4.
[0076] Note that the neural network 11 is trained to minimize at least one cost function related to, for example, mean squared error, mean absolute error or binary cross entropy.
[0077] The trained neural network 11 implemented by the implementation means 12 determines the remaining life RL of the bearing 4 from the measured vibrations in the data item DATA transmitted by the sensor 8 .
[0078] The number of three-layer stacks, the number of recurrent layers, and the number of recurrent units in each recurrent layer are determined based on the expected accuracy of the URL transmitted by the neural network 11.
[0079] Figure 4 An example of a method for training a neural network 11 is shown.
[0080] In step 20, a training set is obtained.
[0081] Each training test is obtained from a training bearing and includes vibration measurements of the training bearing and a measured remaining life of the training bearing associated with the vibration measurements.
[0082] For each training set, a training bearing is tested, for example, on a test bench. Vibration measurements are collected as a function of time using an accelerometer located on bearing 4.
[0083] The vibration measurements are processed using spectral analysis of the vibration measurements to determine the occurrence of a rupture, the moment of which is recorded.
[0084] The remaining life RL is equal to the difference between the moment when the vibration starts to be recorded and the moment when the crack occurs.
[0085] The presence of rupture can further be verified by visual control.
[0086] In step 21, the neural network 11 is trained according to the training set.
[0087] For each training set, the implementing means 12 implements the neural network 11 to determine a first output remaining lifetime based on the vibration measurements of the training set.
[0088] In step 22 , for each training set, the comparison means 14 performs a first comparison between the measured remaining lifetime of said training set and a first output remaining lifetime determined by the neural network 11 from the vibration measurements of said training set.
[0089] The comparison device 14 may implement a mean absolute error algorithm.
[0090] The tuning device 13 tunes the weights of the neural network 11 according to the result of the first comparison so as to minimize a cost function related to, for example, the mean absolute error.
[0091] The method may further comprise a validation step 23, 24, 25 of the trained neural network 11 to check whether the accuracy of the RL determined by the neural network 11 is sufficient.
[0092] Assume that a validation set is obtained.
[0093] Each validation set includes vibration measurements of a validation bearing and a measured remaining life of the validation bearing associated with the vibration measurements.
[0094] The validation set can be obtained in the same way as the training set described in step 20.
[0095] At step 23 , for each verification step, the implementation means 12 implements the neural network 11 to determine a second output remaining lifetime based on the vibration measurements of the verification set.
[0096] At step 24 , for each training set, the comparison means 14 performs a second comparison between the measured remaining lifetime of the validation set and a second output remaining lifetime determined by the neural network 11 from the vibration measurements of the validation set.
[0097] The tuning device 13 tunes the weights of the neural network 11 according to the result of the second comparison so as to minimize the cost function.
[0098] The device 9 provides a quantitative prediction of the remaining life of the bearing 4 based on the vibration measurement results of the bearing 4.
Claims
1. A method for predicting the remaining life of a bearing (4), the method comprising: measuring the vibration of the bearing (4), and A neural network (11) is implemented, the neural network being configured to determine the remaining life of the bearing (4) based on the measured vibrations.
2. A method for training a neural network (11) configured to determine the remaining life of a bearing (4) based on measured vibrations, the method comprising: At least for the training bearing, obtaining at least one training set, the at least one training set comprising vibration measurements of the training bearing and a measured remaining life of the training bearing associated with the vibration measurements, implementing a neural network (11) to determine a first output remaining life based on the vibration measurements of the training set, performing a first comparison between the measured remaining lifetime of the training set and the first output remaining lifetime, and The weights of the neural network (11) are tuned according to the result of the first comparison.
3. The method according to claim 2, further comprising: For at least one verification bearing, obtaining at least one verification set, the at least one verification set comprising a vibration measurement result of the verification bearing and a measured remaining life of the verification bearing associated with the vibration measurement result, implementing a neural network (11) to determine a second output remaining life based on the vibration measurements of the validation set, performing a second comparison between the measured remaining lifetime of the validation set and the second output remaining lifetime, Wherein, tuning the weight of the neural network (11) further comprises tuning the weight of the neural network according to the result of the second comparison.
4. The method according to claim 2 or 3, wherein: Each comparison between the remaining lifetime and the output remaining lifetime includes determining a mean absolute error between the remaining lifetime and the output remaining lifetime.
5. A device (9) for predicting the remaining life of a bearing (4), the device comprising: a memory (10) storing a neural network (11) configured to determine the remaining life of the bearing based on the measured vibration of the bearing (4), and An implementation device (12) is configured to implement the neural network (11).
6. The device according to claim 5, wherein: The neural network (11) comprises at least one stack (15, 16) of three layers (15a, 15b, 15c, 16a, 16b, 16c), a dense layer (19) and at least one recurrent layer (17, 18) comprising at least one recurrent unit (17a, 17b, 17c, 18a), the stack (15, 16) of three layers comprising convolutional layers (15a, 16a), batch normalization layers (15b, 16b) and maximum pooling layers (15c, 16c).
7. The device according to claim 6, wherein: The recurrent units (17a, 17b, 17c, 18a) include long short-term memory units.
8. The device according to any one of claims 5 to 7, wherein: The implementation device (12) is further configured to implement the neural network (11) to determine a first output remaining life based on vibration measurement results of at least one training set, the training set including vibration measurement results of training bearings and measured remaining life of the training bearings associated with the vibration measurement results, The device (9) further comprises: a comparison device (14) configured to perform a first comparison between the measured remaining lifetime of the training set and the first output remaining lifetime, and A tuning device (13) is configured to tune the weights of the neural network (11) according to the result of the first comparison.
9. The device according to claim 8, wherein: The implementation device (12) is further configured to implement the neural network (11) to determine a second output remaining life based on vibration measurement results of at least one validation set, the validation set including vibration measurement results of a validation bearing and a measured remaining life of the validation bearing associated with the vibration measurement results, The comparison device (14) is further configured to perform a second comparison between the measured remaining lifetime of the validation set and the second output remaining lifetime, and The tuning device (13) is further configured to tune the weights of the neural network (11) according to the result of the second comparison.
10. A bearing arrangement comprising a bearing (4), a sensor (8) configured to measure vibrations of the bearing (4), and a device (9) according to any one of claims 5 to 9 connected to the sensor.