A method for predicting the service life of a spindle bearing

Abstract

The present invention relates to the field of bearing life prediction, and particularly to a method for predicting the life of a spindle bearing. The method includes: obtaining the parameters of the bearing at different historical moments, where the parameters include multiple sub-parameters; constructing a long short-term memory network model, calculating the learning rate of the long short-term memory network model, and training the long short-term memory network model with the historical parameters to obtain a prediction model for predicting the bearing life. The present invention adaptively obtains the learning rate, thereby improving the accuracy of the prediction result of the long short-term memory network model.

Description

Technical Field

[0001] The present invention relates to the field of bearing life prediction, and particularly to a method for predicting the life of a spindle bearing. Background Art

[0002] Bearings are common mechanical equipment components, and the quality of bearings is also a key factor affecting the industrial production and manufacturing process. In a high-speed and high-load operating environment, spindle bearings are prone to wear or failure, which in turn affects the operating efficiency and accuracy of the equipment. Therefore, it is necessary to predict the life of the bearings in order to replace or repair them before the bearings fail, thereby avoiding downtime due to failures, improving production efficiency, and equipment reliability. Long short-term memory network is a type of time-recursive neural network used to process time series data. Its unique gating mechanism enables it to effectively capture and learn long-term dependencies in the data, so it is applied in the field of time series prediction.

[0003] The Chinese patent application document with the publication number CN111581888A discloses a method for constructing a prediction model for the remaining service life of a wind turbine bearing. The method is implemented by the following steps: Step 1, collect bearing vibration frequency data; Step 2, preprocess the data; Step 3, construct a long short-term memory network and configure network parameters, specify training options; Step 4, train the network; Step 5, verify the prediction model. By providing a method for constructing a prediction model for the remaining service life of a wind turbine bearing based on a long short-term memory network, the prediction of the remaining service life of the wind turbine bearing is realized.

[0004] When using a long short-term memory network model to predict the service life of a bearing, first, a long short-term memory network model needs to be constructed, then the long short-term network model is trained, and finally, the trained long short-term network model is used to predict the life of the bearing. In the existing training process, a fixed learning rate is usually adopted. However, if the learning rate is too large, it may cause the model to oscillate near the optimal value and even fail to converge; if the learning rate is too small, it will lead to a slow training process and be easily trapped in a local optimum. Eventually, the accuracy of the prediction results of the long short-term memory network model is relatively low. Summary of the Invention

[0005] In order to solve the problem that the accuracy of the predicted bearing life is relatively low due to training the long short-term network model with a fixed learning rate, the present invention provides a method for predicting the life of a spindle bearing.

[0006] The present invention provides a method for predicting the life of a spindle bearing, adopting the following technical solutions:

[0007] Obtain the parameters of the bearing at different historical moments, where the parameters include multiple sub-parameters;

[0008] Build a long short-term memory network model, calculate the learning rate of the long short-term memory network model, and use historical parameters to train the long short-term memory network model to obtain a prediction model for predicting the bearing life;

[0009] Among them, the calculation method of the learning rate is as follows: divide the historical moments to obtain the time periods in which each historical moment is located; calculate the first influence factor and the second influence factor of the sub-parameters at each historical moment. The first influence factor represents the average deviation of the sub-parameters within the time period corresponding to the historical moment, and the second influence factor represents the fluctuation range of the sub-parameters within the time period corresponding to the historical moment; perform weighted summation on the first influence factor and the second influence factor to obtain the influence degree of the sub-parameters at each historical moment; perform weighted summation on the influence degrees of multiple sub-parameters at each historical moment to obtain the influence degree corresponding to the historical moment, and calculate the learning rate at each historical moment. The expression is:

[0010]

[0011] In the formula, represents the learning rate corresponding to the t-th historical moment, represents the preset initial learning rate, represents the influence degree of the t-th historical moment.

[0012] During the process of training the long short-term memory network model, adaptively adjust the learning rate according to the abnormality degree of different historical moments. When the abnormality degree is high, reduce the learning rate to avoid the overreaction of the model to abnormal data. When the abnormality degree is low, increase the learning rate to accelerate the model convergence, thereby improving the accuracy of the prediction result of the long short-term memory network model.

[0013] Preferably, the method for dividing the historical moments to obtain the time periods in which each historical moment is located is: taking any historical moment as the center, extending multiple moments forward and backward along the time direction to obtain the time period corresponding to the historical moment.

[0014] By dividing the corresponding time periods for the historical moments, it is convenient to analyze the parameters within the corresponding time periods, thereby facilitating the judgment of whether the parameters at the corresponding historical moments are in an abnormal state.

[0015] Preferably, the method further includes: calculating the attention weight of the sub-parameters at each historical moment. The expression is:

[0016]

[0017] In the formula, represents the attention weight of the sub-parameters at the t-th historical moment, represents the sub-parameters at the t-th historical moment, represents the mean of all sub-parameters before the t-th historical moment, and norm represents the normalization function.

[0018] Through the attention weights, it is possible to preliminarily understand the possibility that the sub-parameters in the corresponding historical moment are in an abnormal state, providing a theoretical basis for updating the parameters in the long short-term memory network model.

[0019] Preferably, the method further includes: decomposing the sub-parameters of the historical moment by using the STL algorithm to obtain the trend component of each moment.

[0020] Preferably, the expression of the first influence factor is:

[0021]

[0022] In the formula, represents the first influence factor of the sub-parameters at the t-th historical moment, represents the normalization function, represents the attention weight of the sub-parameters at the t-th historical moment, , respectively represent the minimum value and the maximum value of the trend component, represents the trend component at the t-th historical moment, represents the sub-parameters at the (t + i)-th historical moment, represents the mean of all sub-parameters before the t-th historical moment, represents the t The time length between the sub-parameters at the i-th historical moment and the sub-parameters at the t-th historical moment, and n represents half of the number of remaining moments in the time period where the t-th historical moment is located.

[0023] The first influence factor is calculated through multiple factors, improving the accuracy of the calculation result. The first influence factor comprehensively reflects the possibility that the parameters in the corresponding historical moment are abnormal.

[0024] Preferably, the method further includes: taking the variance of all sub-parameters before each historical moment as the overall variance corresponding to that historical moment, and taking the variance of the sub-parameters at each moment in the time period where each historical moment is located as the local variance corresponding to that historical moment.

[0025] Preferably, the expression of the second influence factor is:

[0026]

[0027] In the formula, represents the second influence factor corresponding to the t-th historical moment, represents the overall variance corresponding to the t-th historical moment, It represents the local variance corresponding to the t-th historical moment, and β represents the preset weight.

[0028] The second influence factor is obtained by comparing the overall variance and the local variance, and the ratio method is more suitable for capturing significant fluctuations in parameters.

[0029] Preferably, the expression for the influence degree of the sub-parameter is:

[0030]

[0031] In the formula, It represents the influence degree of the sub-parameter at the t-th historical moment, It represents the first influence factor of the sub-parameter at the t-th historical moment, It represents the second influence factor corresponding to the t-th historical moment.

[0032] The influence degree is obtained by weighted fusion of the first influence factor and the second influence factor, and the importance of the sub-parameter at the corresponding historical moment is comprehensively reflected by the influence degree.

[0033] Preferably, the expression for the influence degree of the historical moment is:

[0034]

[0035] In the formula, It represents the influence degree at the t-th historical moment, It represents the influence degree of the first sub-parameter at the t-th historical moment, It represents the influence degree of the second sub-parameter at the t-th historical moment.

[0036] Preferably, the multiple sub-parameters are respectively the vibration frequency and the temperature.

[0037] The temperature change can reflect the lubrication state, friction condition and potential thermal damage of the bearing, and can evaluate the health state of the bearing. The vibration frequency can reflect the state information of the machine, such as imbalance, misalignment and bearing faults. By considering both the temperature and the vibration frequency simultaneously, the operating state of the bearing can be evaluated more comprehensively, avoiding the limitations brought by a single data source.

[0038] The present invention has the following technical effects:

[0039] During the process of training the long short-term memory network model, the learning rate is adaptively adjusted according to the degree of abnormality at different historical moments. When the degree of abnormality is high, the learning rate is reduced to avoid the overreaction of the model to abnormal data. When the degree of abnormality is low, the learning rate is increased to accelerate the convergence of the model, thereby improving the accuracy of the prediction result of the long short-term memory network model. Brief Description of the Drawings

[0040] Figure 1 It is a flowchart of a method for predicting the life of a spindle bearing according to the present invention.

[0041] Figure 2 It is a schematic diagram of parameter weights according to the present invention.

[0042] Figure 3 It is a comparison chart of the prediction results of the long short-term memory network model according to the present invention and the prior art. Detailed implementation manners

[0043] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts fall within the protection scope of the present invention.

[0044] An embodiment of the present invention discloses a method for predicting the life of a spindle bearing. Referring to Figure 1 , the method includes the following steps:

[0045] S1: Obtain the parameters of the bearing at different historical moments, where the parameters include multiple sub-parameters.

[0046] Obtain the parameters of the bearing at different moments in the working state. The multiple sub-parameters are respectively temperature data and vibration frequency data. When the bearing works abnormally, there is relatively large friction inside the bearing or friction occurs between the bearing and external components, which causes the temperature of the bearing to rise. At the same time, when the bearing is abnormal, it will shake. Therefore, the working state of the bearing can be reflected by the bearing temperature and vibration frequency. Fill the missing values of the obtained parameters using an interpolation method (such as linear interpolation), and perform smoothing processing on the temperature data and vibration frequency data respectively using a moving average or low-pass filter.

[0047] Exemplarily, the frequency of collecting the historical temperature data and historical vibration frequency data of the bearing is once every 1 second. The unit of the historical temperature data is °C, and the unit of the historical vibration frequency data is Hz. Perform standardization or normalization processing on the collected temperature and vibration frequency data.

[0048] S2: Divide the historical moments to obtain the time periods in which each historical moment is located.

[0049] Centering on any historical moment, extend a plurality of moments in the time direction before and after to obtain the time period of the corresponding historical moment.

[0050] Exemplarily, the historical moments are , ,..., . For the moment among them, extend 5 moments in the time direction before and after to obtain Time period of a moment ( …, …, ). Similarly, for the moment, extend 5 moments forward and backward along the time direction to obtain the time period of the moment ( …, …, ).

[0051] S3: Calculate the first influence factor and the second influence factor of the sub-parameters at each historical moment.

[0052] The first influence factor represents the average deviation of the sub-parameters within the time period where the corresponding historical moment is located, and the second influence factor represents the fluctuation range of the sub-parameters within the time period where the corresponding historical moment is located.

[0053] S31: Calculate the attention weight of the sub-parameters at each historical moment. The expression is:

[0054]

[0055] In the formula, represents the attention weight of the sub-parameters at the t-th historical moment, represents the sub-parameters at the t-th historical moment, represents the mean value of all sub-parameters before the t-th historical moment, and norm represents the normalization function. Exemplarily, for the 10th historical moment it is the value of , …, the mean value of the sub-parameters at the moments.

[0056] Combined with Figure 2 as shown, adjust the attention weight according to the difference between the sub-parameters at the current t-th historical moment and the average value. The greater the difference, the greater the attention weight, making it easier to capture abnormal situations when training the long short-term memory network model; the smaller the difference, the smaller the attention weight, reducing the attention to normal situations and the amount of computation. The attention weight can reflect the abnormal situation of this sub-parameter data to a certain extent. Specifically, the larger its value, the greater the possibility of abnormality of the sub-parameters at the corresponding moment, and vice versa, the smaller its value, the smaller the possibility of abnormality of the sub-parameters at the corresponding moment.

[0057] Identifying abnormal trends in historical temperature data or vibration frequency data depends not only on the temperature data or vibration frequency data at a single time point, but also on the temperature data or vibration frequency data at surrounding time points. By comparing the difference between the current temperature data (or vibration frequency data) and the historical average temperature data (or vibration frequency data), anomalies at a single time point can be captured. It can be understood that each of the temperature data and the vibration frequency data corresponds to a weight of concern.

[0058] S32: Use the STL algorithm to decompose the sub-parameters at historical moments to obtain the trend components at each moment, and calculate the first influence factor.

[0059] Use the STL algorithm to decompose the temperature data at historical moments to obtain the trend components of the temperature data at each moment, and use the STL algorithm to decompose the vibration frequency data at historical moments to obtain the trend components of the vibration frequency data at each moment. It should be noted that the STL algorithm is a prior art, and the calculation method of the trend components will not be elaborated here.

[0060] The expression of the first influence factor is:

[0061]

[0062] In the formula, represents the first influence factor of the sub-parameter at the t-th historical moment, represents the normalization function, represents the weight of concern of the sub-parameter at the t-th historical moment, and represent the minimum and maximum values of the trend component respectively, represents the trend component at the t-th historical moment, represents the sub-parameter at the (t + i)-th historical moment, represents the mean value of all sub-parameters before the t-th historical moment, represents the t The time length between the sub-parameter at the (t + i)-th historical moment and the sub-parameter at the t-th historical moment, and n represents half of the number of remaining moments in the time period where the t-th historical moment is located.

[0063] Exemplarily, for the time period where the moment is located is ( ,..., ,..., ), the number of moments (number of data points) in the time period is 11, and at this time the value of n is 5.

[0064] Among them, It represents the difference between the sub-parameters corresponding to each moment within the time period where the t-th historical moment is located and the overall sub-parameter. The larger its value, the greater the probability that the sub-parameters at the t-th historical moment and its surroundings are abnormal, further indicating that the bearing is in an abnormal working state.

[0065] It represents the normalized value of the difference between the trend component corresponding to the sub-parameter at the t-th historical moment and the maximum value of the trend component, reflecting the relative size of the trend component corresponding to the sub-parameter at the t-th historical moment in the overall trend component. The larger this difference, the more likely the sub-parameter at the t-th historical moment is abnormal.

[0066] To sum up, the first influence factor represents the abnormality possibility of the sub-parameter at the corresponding historical moment. The larger its value, the greater the possibility that the sub-parameter at the corresponding historical moment is abnormal, further indicating the greater possibility that the bearing is in an abnormal working state at the corresponding moment.

[0067] S33: Calculate the second influence factor.

[0068] Take the variance of the sub-parameters at all moments before each historical moment as the overall variance corresponding to the corresponding historical moment, and take the variance of the sub-parameters at each moment within the time period where each historical moment is located as the local variance corresponding to the corresponding historical moment.

[0069] Exemplarily, for the 10th historical moment for example, take , , …, the variance of the sub-parameters at the moments as the overall variance of the historical moment , and take the variance of the sub-parameters at each historical moment within the time period ( , …, , …, , …, ) where is located as the local variance of the historical moment

[0070] The expression of the second influence factor is:

[0071]

[0072] In the formula, represents the second influence factor corresponding to the t-th historical moment, represents the overall variance corresponding to the t-th historical moment, represents the local variance corresponding to the t-th historical moment, and β represents a preset weight, and the value of the weight is set artificially according to the situation. Exemplarily, the value of the weight is 0.5.

[0073] The second influence factor indicates the difference between the local variance and the overall variance of the sub-parameters within the time period in which the t-th historical moment is located. The larger its value, the greater the change in the values of the sub-parameters near the t-th historical moment, and the greater the likelihood that the bearing is abnormal.

[0074] It can be understood that for the t-th historical moment, the temperature data corresponds to a first influence factor and a second influence factor, and the vibration frequency data corresponds to a first influence factor and a second influence factor.

[0075] S4: Perform weighted summation on the first influence factor and the second influence factor to obtain the influence degree of the sub-parameters at each historical moment.

[0076] The expression for the influence degree of the sub-parameters is:

[0077]

[0078] In the formula, represents the influence degree of the sub-parameters at the t-th historical moment, represents the first influence factor of the sub-parameters at the t-th historical moment, represents the second influence factor corresponding to the t-th historical moment.

[0079] The abnormal situation of the sub-parameters at the corresponding historical moment is comprehensively reflected through the influence degree. It can be understood that the temperature data at each historical moment corresponds to an influence degree, and the vibration frequency data corresponds to an influence degree.

[0080] S5: Perform weighted summation on the influence degrees of multiple sub-parameters at each historical moment to obtain the influence degree at the corresponding historical moment.

[0081] The expression for the influence degree at the historical moment is:

[0082]

[0083] In the formula, represents the influence degree at the t-th historical moment, represents the influence degree of the first sub-parameter at the t-th historical moment, represents the influence degree of the second sub-parameter at the t-th historical moment. It can also be understood that, represents the influence degree of the temperature at the t-th historical moment, represents the influence degree of the vibration frequency at the t-th historical moment.

[0084] S6: Construct a long short-term memory network model, calculate the learning rate of the long short-term memory network model, and use the historical parameters to train the long short-term memory network model to obtain a prediction model for predicting the bearing life.

[0085] The expression of the learning rate at each historical moment is as follows: ; where represents the learning rate corresponding to the t-th historical moment, represents the preset initial learning rate, represents the influence degree of the t-th historical moment. Exemplarily, the initial learning rate is 0.1.

[0086] Combined with Figure 3 As shown, when the influence degree of the t-th historical moment is greater, the corresponding learning rate is smaller. The smaller learning rate can slow down the speed of parameter update, making the long short-term memory network model more gentle in response to abnormal data, avoiding excessive update, and thus improving the prediction accuracy for the main shaft bearing. On the contrary, when the influence degree of the t-th historical moment is smaller, the corresponding learning rate is larger, and the larger learning rate can accelerate the model convergence.

[0087] Obtain the sub-parameters of the bearing in real time, and input the sub-parameters into the trained long short-term memory network model to obtain the service life of the bearing.

[0088] The above are all the preferred embodiments of the present invention. Without limiting the protection scope of the present invention accordingly, therefore: All equivalent changes made according to the structure, shape, and principle of the present invention shall be covered within the protection scope of the present invention.

Claims

1. A method for predicting the life of a spindle bearing, characterized in that: Includes steps: Obtain the parameters of the bearing at different historical moments, the parameters include multiple sub-parameters; Constructing a long short-term memory network model, calculating the learning rate of the long short-term memory network model, and using historical parameters to train the long short-term memory network model to obtain a prediction model for predicting the bearing life; The learning rate is calculated as follows: divide the historical moments into the time period of each historical moment; calculate the first influencing factor and the second influencing factor of the sub-parameter of each historical moment, where the first influencing factor represents the average deviation of the sub-parameter in the time period of the corresponding historical moment, and the second influencing factor represents the fluctuation range of the sub-parameter in the time period of the corresponding historical moment; perform weighted summation of the first influencing factor and the second influencing factor to obtain the influence degree of the sub-parameter at each historical moment; perform weighted summation of the influence degrees of multiple sub-parameters at each historical moment to obtain the influence degree of the corresponding historical moment, and calculate the learning rate of each historical moment. The expression is: ; In the formula, represents the learning rate corresponding to the t-th historical moment, represents the preset initial learning rate, Indicates the degree of influence at the t-th historical moment; The expression of the first impact factor is: In the formula, represents the first influencing factor of the sub-parameter at the t-th historical moment, norm represents the normalization function, represents the attention weight of the sub-parameter at the t-th historical moment, , Represent the minimum and maximum values ​​of the trend component, respectively. represents the trend component at the t-th historical moment, represents the sub-parameter of the t+ith historical moment, represents the mean of all sub-parameters before the t-th historical moment, Indicates the tth The length of time between the sub-parameter of the i-th historical moment and the sub-parameter of the t-th historical moment, and n represents half of the number of remaining moments in the time period of the t-th historical moment; The expression of the second impact factor is: In the formula, represents the second impact factor corresponding to the t-th historical moment, represents the overall variance corresponding to the t-th historical moment, represents the local variance corresponding to the t-th historical moment, and β represents the preset weight; The expression of the influence degree of sub-parameters is: In the formula, represents the influence of the sub-parameter at the t-th historical moment, represents the first influencing factor of the sub-parameter at the t-th historical moment, Represents the second impact factor corresponding to the t-th historical moment.

2. A method for predicting the life of a main shaft bearing according to claim 1, characterized in that: The method of dividing the historical moments to obtain the time period of each historical moment is: taking any historical moment as the center, extending multiple moments forward and backward along the time direction to obtain the time period of the corresponding historical moment.

3. A method for predicting the life of a main shaft bearing according to claim 1, characterized in that: The method also includes: calculating the attention weight of each sub-parameter at each historical moment, the expression is: In the formula, represents the attention weight of the sub-parameter at the t-th historical moment, represents the sub-parameter at the t-th historical moment, It represents the mean of all sub-parameters before the t-th historical moment, and norm represents the normalization function.

4. A method for predicting the life of a main shaft bearing according to claim 3, characterized in that: The method also includes: using the STL algorithm to decompose the sub-parameters of the historical moments to obtain the trend components of each moment.

5. A method for predicting the life of a main shaft bearing according to claim 1, characterized in that: The method also includes: taking the variance of the sub-parameters of all moments before each historical moment as the overall variance of the corresponding historical moment, and taking the variance of the sub-parameters of each moment in the time period of each historical moment as the local variance of the corresponding historical moment.

6. A method for predicting the life of a main shaft bearing according to claim 1, characterized in that: The expression of the influence of historical moments is: In the formula, represents the degree of influence at the t-th historical moment, Indicates the influence of the first sub-parameter at the t-th historical moment, Indicates the influence of the second sub-parameter at the t-th historical moment.

7. A method for predicting the life of a main shaft bearing according to claim 1, characterized in that: The multiple sub-parameters are vibration frequency and temperature.

Images