Bearing heat treatment result prediction method based on deep learning

Abstract

The invention relates to an innovative bearing heat treatment result prediction technology, and the powerful capability of a deep learning network is fully utilized. The method is characterized in that key parameters, such as quenching time, quenching temperature and other conditions possibly influencing a heat treatment result, of the bearing in the heat treatment process are integrated to serve as a feature set of input data, and the physical or chemical attribute value of the bearing after heat treatment is completed serves as a target variable of model prediction. According to the characteristic that data is high in time sensitivity, a long short-term memory (LSTM) network is selected to learn a complex mapping relation between input features and output results, time sequence dependency in the data is effectively captured, and time sequence features of the data are grasped. The trained model not only can provide an accurate prediction result for an operator, but also can help the operator to understand how different parameters affect the quality of a final product, so that product loss caused by experiments is reduced.

Description

Technical Field

[0001] The present invention belongs to the field of bearing heat treatment, and specifically is a bearing heat treatment result prediction method based on deep learning. Background Art

[0002] In the field of bearing heat treatment, precise control of the processing environment is crucial to ensure the final quality of bearing products. The optimized configuration of environmental parameters can not only significantly improve the performance of bearings, but also extend their service life. However, traditional processing environment control often relies on the experience and intuition of technicians. This method has certain instability and it is difficult to ensure continuous high-quality output. The deep learning model can learn the complex relationship between environmental parameters and bearing quality through a large amount of data training. This data-driven method not only overcomes the limitations of manual experience, but also provides technicians with stable and repeatable data support, greatly improving the efficiency of the bearing heat treatment process and product quality.

[0003] The patent "A heat treatment method for extra-large bearings made of medium-carbon alloy steel" (publication number CN117802302A) applied by Luoyang Xinqianglian Slewing Bearing Co., Ltd. proposes an innovative heat treatment solution. This method uses induction heating to perform medium-frequency heating and quenching treatment on the ring mounting surface and working surface of the extra-large bearing respectively. Precise heating is performed during the high-speed rotation of the ring by multiple induction heaters to ensure uniform surface structure and small deformation of the quenched surface. The "A heat treatment method and system for bearing steel parts" (application number CN202311736098.2) applied by the Jincheng Nanjing Electromechanical Hydraulic Engineering Research Center of China Aviation Industry Corporation effectively improves the impact toughness of GCr15 bearing steel parts by precisely controlling the heating temperature, insulation time and cooling method, and solves the problem of early failure under impact load conditions. This patented technology is of great significance for improving the service life of bearings under harsh working conditions.

[0004] In general, among the existing technical solutions, there is still a lack of relevant methods to integrate deep learning into the field of bearing heat treatment, which leads to a high dependence on manual experience and a lack of stability and reliability. Summary of the invention

[0005] The present invention proposes a bearing heat treatment result prediction method based on deep learning, the purpose of which is to use a deep learning network to predict the final attribute value of the product using the equipment parameters preset by the operator without conducting actual production experiments. The core of this prediction method is the Long Short-Term Memory (LSTM) network. The structural design of the LSTM network enables it to learn the complex mapping relationship between input features and output results, and effectively capture the temporal dependencies in the data. The method of the present invention can train the LSTM model to identify the association between the preset parameters of the equipment and the physical or chemical properties of the bearing after heat treatment. After the model training is completed, the operator can quickly obtain the expected attribute value of the product after heat treatment by inputting new equipment parameters, thereby assisting relevant personnel to make more scientific decisions before production.

[0006] The technical solution adopted by the present invention is:

[0007] A bearing heat treatment result prediction method based on deep learning comprises the following steps:

[0008] S1. Build a training data set using the equipment’s operating parameters and product test results;

[0009] S2. Preprocess the actual data of the equipment, clean the data, and fill in missing values;

[0010] S3, standardize the data;

[0011] S4, split the data into a training set for model training and a test set for model testing;

[0012] S5. Define the LSTM model and set the model hyperparameters;

[0013] S6. Using the actual bearing heat treatment data after preprocessing and standardization to train the model;

[0014] S7. Use the test dataset to test the model performance.

[0015] Furthermore, step S1 is specifically as follows:

[0016] The actual parameters during equipment operation, such as gas environment parameters, quenching area temperature, quenching time, cold treatment method, air drying temperature and other data are selected as features, and the deformation, hardness, organization and other properties of the product are taken as prediction targets.

[0017] Furthermore, step S2 is specifically as follows:

[0018] S21. In the actual operation of the heat treatment equipment, some preset data may be abnormal due to human operation. The support vector machine (SVM) method is used to classify normal and abnormal data, so as to eliminate abnormal data. The segmentation hyperplane of SVM is expressed as:

[0019] α T x+β=0 (1)

[0020] Among them, α and β are the parameters of SVM, which are used to determine different segmentation planes. Then the distance from any sample point x to the hyperplane is:

[0021]

[0022] Among them, d is the distance from the sample point to the hyperplane. When a sample point is too far from the hyperplane, that is, the d value is too large, it will be marked as a data anomaly.

[0023] S22. In the process of data transmission and retention of bearing heat treatment equipment, due to the fluctuation of front-end and back-end communication of the equipment, the attribute values ​​of some samples are missing. When a certain attribute of the preset data of the equipment is missing, the mean filling method is adopted to perform effective data recovery. Specifically, all non-missing values ​​in the attribute column are first summed up and then divided by the number of non-missing values ​​to obtain the mean of the attribute column, which is used to fill the missing values. The specific disclosure is:

[0024]

[0025] Among them, v missing is the missing value of an attribute column, is the sum of all non-missing values ​​of the attribute column, v num is the number of all non-missing values ​​in the attribute column. The purpose of this is to maintain the distribution characteristics of the data on this attribute and avoid introducing additional bias, which will cause the prediction performance of the model to deteriorate.

[0026] Furthermore, step S3 is specifically as follows:

[0027] In order to solve the problem of large numerical differences caused by different dimensions of operation data of different components during the actual operation of the equipment, all pre-processed equipment operation data are standardized using the Z-score Normalization method, that is,

[0028]

[0029] Among them, z is the standardized data value, ∈ is the mean of the attribute column data, and δ is the standard deviation of the attribute column data. This method makes the standardized data have the characteristics of mean 0 and standard deviation 1. It can make the feature data in a similar range, so that the model converges faster during training, reducing the difficulty of model training while improving stability.

[0030] Furthermore, step S4 is specifically as follows:

[0031] The processed data is randomly divided into training set and test set in a ratio of 8:2. A larger proportion of data is used for model training to ensure that there is enough data for model training, and a smaller proportion of test set data is used to test the prediction performance of the model.

[0032] Furthermore, step S5 is specifically as follows:

[0033] S51. During the heat treatment process of the bearing heat treatment equipment, the quality of the product shows obvious time series characteristics, so LSTM is selected as the prediction model for the product results. At the same time, the hyperparameters of the model are further refined to adapt to the specific working conditions of the bearing heat treatment. When defining the loss function of the LSTM model, the mean square error (MSE) is selected as the error function. It calculates the average of the squares of the difference between the model prediction value and the actual value, which can provide intuitive feedback on the prediction accuracy for the model. The specific formula is:

[0034]

[0035] Among them, v loss is the model prediction loss value, n is the total number of samples in a batch, Y i is the true value, is the predicted value of the model.

[0036] S52, use the Adam optimization algorithm as the optimizer of the model to speed up the convergence of the model. For each iteration step t of the model, first calculate the gradient:

[0037]

[0038] Among them, g t represents the gradient at iteration step t, represents the gradient of the parameter θ, f t (θ t-1 ) is the objective function to be optimized. Then update the first-order moment vector m, the formula is defined as:

[0039] m t =β 1 ·m t-1 +(1-β1 )·g t (7)

[0040] m t is the first-order moment vector at time step t, β 1 is the exponential decay rate of the first-order moment estimate, g t is the gradient at time step t. Update the second-order moment vector v, the formula is defined as:

[0041]

[0042] Among them, v t is the second-order moment vector at time step t, β 2 is the exponential decay rate of the second-order moment estimate. Afterwards, the correction m t and v t The deviation is defined as:

[0043]

[0044]

[0045] Finally, use the corrected and Update the model parameters. The specific update formula is:

[0046]

[0047] Among them, θ t represents the model parameters before update, θ t+1 represents the updated model parameters, α is the learning rate of the model, and ∈ is a small constant used to prevent the denominator from being zero and maintain numerical stability.

[0048] S53. Considering that the actual operation data of the equipment is not complicated, in order to avoid overfitting caused by over-training of the model, the number of iterations of model training is set to 200 times, and the batch size is set to 32.

[0049] S54, two different activation functions are used to activate the neurons of the model. One is the Sigmoid activation function, and the activation formula is as follows:

[0050]

[0051] Among them, z is the original neuron output value, and σ(z) is the value after activation. The other is the tanh activation function, and the activation formula is as follows:

[0052]

[0053] Among them, tanh(z) is the output value after the neuron is activated.

[0054] Further, step S6 is specifically as follows:

[0055] In order to better grasp the time series characteristics in the operation data of the heat treatment equipment, the model structure is defined as a forget gate, an input gate, and an output gate. The forget gate is used to discard values ​​that have little impact on the results, thereby reducing the data burden of the model and improving the prediction accuracy. The formula for defining the forget gate is as follows:

[0056] f t =σ(α f ·[f t-1 ,x t ]+β f ) (14)

[0057] Among them, f t is the output value of the forget gate, α f , β f The learnable parameters of the model, x t is the input value. The formula used by the input gate is:

[0058] i t =σ(α i ·[h t-1 ,x t ]+β i ) (15)

[0059]

[0060] Among them, i t It is the information that the model needs to decide to update. is the candidate cell state. Thus, the data state C at the current time t is obtained using the following formula: t :

[0061]

[0062] Then, the formula used by the output gate is:

[0063] o t =σ(α o ·[h t-1 ,x t ]+β o ) (18)

[0064] h t =o t tanh(C t ) (19)

[0065] The model uses o t As the output of the output gate, ht That is, the hidden state of the model at time step t, reflecting the information that the network should remember at the current time step.

[0066] Further, step S7 is specifically as follows:

[0067] In order to further improve the adaptability of the model under the actual working conditions of bearing heat treatment equipment, the performance of the model is tested using the training data set after the model training is completed. During the test, the attributes of some test samples are randomly manually erased to determine the robustness of the model. According to the final test results of the model, the model training iterations or learning rate and other hyperparameters are adjusted to achieve the best effect.

[0068] Compared with the prior art, the beneficial technical effects of the present invention are:

[0069] 1. The deep learning-based bearing heat treatment result prediction method proposed in this invention successfully applies the deep learning method to the field of bearing heat treatment. It effectively predicts the physical and chemical property values ​​of products without conducting actual production experiments, reducing product losses caused by product experiments and parameter tuning. At the same time, it provides data support for technicians to adjust production parameters to achieve the optimal processing effect.

[0070] 2. The deep learning prediction model selected by the present invention can learn the complex mapping relationship between input features and output results, effectively capture the time series dependency in the data, and is suitable for prediction tasks with high time dependency such as bearing heat treatment. At the same time, the model complexity is low, the number of parameters required for training is small, and it has strong portability. It also has the advantages of fast training speed, convenient adjustment, not easy to overfit, and simple implementation, which provides convenience for actual deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a flow chart of a method according to an embodiment of the present invention.

[0072] Figure 2 It is a model training flow chart of the present invention.

[0073] Figure 3 It is a schematic diagram of model construction of the present invention. DETAILED DESCRIPTION

[0074] like Figure 1 As shown, a bearing heat treatment result prediction method based on deep learning includes the following steps:

[0075] S1. Build a training data set using the equipment’s operating parameters and product test results;

[0076] The actual parameters during equipment operation, such as gas environment parameters, quenching area temperature, quenching time, cold treatment method, air drying temperature and other data are selected as features, and the deformation, hardness, organization and other properties of the product are taken as prediction targets.

[0077] S2. Preprocess the actual data of the equipment, clean the data, and fill in missing values;

[0078] S21. In the actual operation of the heat treatment equipment, some preset data may be abnormal due to human operation. The support vector machine (SVM) method is used to classify normal and abnormal data, so as to eliminate abnormal data. The segmentation hyperplane of SVM is expressed as:

[0079] α T x+β=0 (1)

[0080] Among them, α and β are the parameters of SVM, which are used to determine different segmentation planes. Then the distance from any sample point x to the hyperplane is:

[0081]

[0082] Among them, d is the distance from the sample point to the hyperplane. When a sample point is too far from the hyperplane, that is, the d value is too large, it will be marked as a data anomaly.

[0083] S22. In the process of data transmission and retention of bearing heat treatment equipment, due to the fluctuation of front-end and back-end communication of the equipment, the attribute values ​​of some samples are missing. When a certain attribute of the preset data of the equipment is missing, the mean filling method is adopted to perform effective data recovery. Specifically, all non-missing values ​​in the attribute column are first summed up and then divided by the number of non-missing values ​​to obtain the mean of the attribute column, which is used to fill the missing values. The specific disclosure is:

[0084]

[0085] Among them, v missing is the missing value of an attribute column, is the sum of all non-missing values ​​of the attribute column, v num is the number of all non-missing values ​​in the attribute column. The purpose of this is to maintain the distribution characteristics of the data on this attribute and avoid introducing additional bias, which will cause the prediction performance of the model to deteriorate.

[0086] S3, standardize the data;

[0087] In order to solve the problem of large numerical differences caused by different dimensions of operation data of different components during the actual operation of the equipment, all pre-processed equipment operation data are standardized using the Z-score Normalization method, that is,

[0088]

[0089] Among them, z is the standardized data value, ∈ is the mean of the attribute column data, and δ is the standard deviation of the attribute column data. This method makes the standardized data have the characteristics of mean 0 and standard deviation 1. It can make the feature data in a similar range, so that the model converges faster during training, reducing the difficulty of model training while improving stability.

[0090] S4, split the data into a training set for model training and a test set for model testing;

[0091] Considering that the operating data of bearing heat treatment equipment is difficult to obtain and the acquisition cost is high, the processed data is randomly divided into training set and test set at a ratio of 8:2. A larger proportion of data is used for model training to ensure that there is enough data for model training, and a smaller proportion of test set data is used to test the prediction performance of the model.

[0092] S5. Define the LSTM model and set the model hyperparameters;

[0093] S51. During the heat treatment process of the bearing heat treatment equipment, the quality of the product shows obvious time series characteristics, so LSTM is selected as the prediction model for the product results. At the same time, the hyperparameters of the model are further refined to adapt to the specific working conditions of the bearing heat treatment. When defining the loss function of the LSTM model, the mean square error (MSE) is selected as the error function. It calculates the average of the squares of the difference between the model prediction value and the actual value, which can provide intuitive feedback on the prediction accuracy for the model. The specific formula is:

[0094]

[0095] Among them, v loss is the model prediction loss value, n is the total number of samples in a batch, Y i is the true value, is the predicted value of the model.

[0096] S52, use the Adam optimization algorithm as the optimizer of the model to speed up the convergence of the model. For each iteration step t of the model, first calculate the gradient:

[0097]

[0098] Among them, g t represents the gradient at iteration step t, represents the gradient of the parameter θ, f t (θ t-1 ) is the objective function to be optimized. Then update the first-order moment vector m, the formula is defined as:

[0099] m t =β 1 ·m t-1 +(1-β 1 )·g t (7)

[0100] m t is the first-order moment vector at time step t, β 1 is the exponential decay rate of the first-order moment estimate, g t is the gradient at time step t. Update the second-order moment vector v, the formula is defined as:

[0101]

[0102] Among them, v t is the second-order moment vector at time step t, β 2 is the exponential decay rate of the second-order moment estimate. Afterwards, the correction m t and v t The deviation is defined as:

[0103]

[0104]

[0105] Finally, use the corrected and Update the model parameters. The specific update formula is:

[0106]

[0107] Among them, θ t represents the model parameters before update, θ t+1 represents the updated model parameters, α is the learning rate of the model, and ∈ is a small constant used to prevent the denominator from being zero and maintain numerical stability.

[0108] S53. Considering that the actual operation data of the equipment is not complicated, in order to avoid overfitting caused by over-training of the model, the number of iterations of model training is set to 200 times, and the batch size is set to 32.

[0109] S54, two different activation functions are used to activate the neurons of the model. One is the Sigmoid activation function, and the activation formula is as follows:

[0110]

[0111] Among them, z is the original neuron output value, and σ(z) is the value after activation. The other is the tanh activation function, and the activation formula is as follows:

[0112]

[0113] Among them, tanh(z) is the output value after the neuron is activated.

[0114] S6. Using the actual bearing heat treatment data after preprocessing and standardization to train the model;

[0115] In order to better grasp the time series characteristics in the operation data of the heat treatment equipment, the model structure is defined as a forget gate, an input gate, and an output gate. The forget gate is used to discard values ​​that have little impact on the results, thereby reducing the data burden of the model and improving the prediction accuracy. The formula for defining the forget gate is as follows:

[0116] f t =σ(α f ·[f t-1 ,x t ]+β f ) (14)

[0117] Among them, f t is the output value of the forget gate, α f , β f The learnable parameters of the model, x t is the input value. The formula used by the input gate is:

[0118] i t =σ(α i ·[h t-1 ,x t ]+β i ) (15)

[0119]

[0120] Among them, i t It is the information that the model needs to decide to update. is the candidate cell state. Thus, the data state C at the current time t is obtained using the following formula: t :

[0121]

[0122] Then, the formula used by the output gate is:

[0123] o t =σ(α o ·[h t-1 ,x t ]+β o ) (18)

[0124] h t =o t tanh(C t ) (19)

[0125] The model uses o t As the output of the output gate, h t That is, the hidden state of the model at time step t, reflecting the information that the network should remember at the current time step.

[0126] S7. Use the test dataset to test the model performance.

[0127] In order to further improve the adaptability of the model under the actual working conditions of bearing heat treatment equipment, the performance of the model is tested using the training data set after the model training is completed. During the test, the attributes of some test samples are randomly manually erased to determine the robustness of the model. According to the final test results of the model, the model training iterations or learning rate and other hyperparameters are adjusted to achieve the best effect.

[0128] The above is a detailed description of the embodiments of the present invention in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Within the knowledge scope of ordinary technicians in the field, various changes can be made without departing from the purpose of the present invention, which are all within the protection scope of this technology.

Claims

1. A bearing heat treatment result prediction method based on deep learning, characterized in that: The following steps are involved: S1. Build a training data set using the equipment’s operating parameters and product test results; S2. Preprocess the actual data of the equipment, clean the data, and fill in missing values; S3, standardize the data; S4, split the data into a training set for model training and a test set for model testing; S5. Define the LSTM model and set the model hyperparameters; S6. Using the actual bearing heat treatment data after preprocessing and standardization to train the model; S7. Use the test dataset to test the model performance.

Images