Bearing size prediction method and system based on CNN-LSTM-Attention neural network

Abstract

The invention relates to the technical field of bearings, and provides a bearing size prediction method based on a CNN-LSTM-Attention neural network, and the method comprises the steps: S1 to S4, collecting the geometric size data and environmental data of a bearing through bearing measurement equipment, and carrying out the initialization processing of the data, and generating a first data set; and then maximum-minimum normalization processing is carried out on the data set. And constructing a prediction model composed of an input layer, a CNN module, an Attention module, an LSTM module and an output layer, and training the model until the prediction precision reaches a preset value. And collecting the geometric dimension and environmental data of the bearing in real time, and inputting the data into the trained model to obtain a prediction result of the geometric dimension of the bearing. According to the method and the system, complex nonlinear errors such as temperature change, mechanical structure deformation and sensor nonlinear characteristics can be effectively compensated, and the bearing size with higher accuracy can be obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of bearings, and particularly to a bearing size prediction method and system based on a CNN-LSTM-Attention neural network. Background Art

[0002] In the field of modern mechanical manufacturing, as one of the core components, the dimensional accuracy of bearings plays a crucial role in the performance and service life of equipment. In particular, key dimensions such as the inner diameter, roundness, and taper of bearings often require extremely high precision requirements. Therefore, equipped with high-precision dimensional inspection equipment is a key link to ensure the production quality of bearings.

[0003] CN113538378A discloses an on-line bearing size detection system based on deep learning. The on-line detection steps are as follows: Step 1: Overall architecture of the on-line visual detection system for automotive wheel hub bearings: Establish the hardware architecture and software architecture of the system; The system hardware includes four parts: a light source system, a camera system, an image acquisition card, and a computer; Among them, the light source system includes a light source and a mounting bracket, and the camera system includes an industrial camera, a lens, and a camera bracket. The image signal is transmitted to the computer through the image acquisition card, and the computer completes the corresponding image processing work; Select models for each module according to the system function requirements, and build the hardware platform of the on-line visual detection system for automotive wheel hub bearings: Hardware is the basis of the visual detection system, and software is the key to realizing the detection function of the system; According to the research purpose and system requirements analysis, the system software module is mainly divided into five parts: an image acquisition module, a distortion correction module, an interference area detection and processing module, an image edge detection module, and an automotive wheel hub bearing size measurement module; Step 2: Detection and processing of interference areas on the surface of automotive wheel hub bearings: A convolutional neural network is a hierarchical structure, consisting of an input layer, multiple hidden layers such as convolutional layers, pooling layers, activation layers, fully connected layers, and an output layer; The input image learns the original features from the image through multiple operations such as convolution and pooling, and finally converts tasks such as classification into objective functions through the fully connected layer and the output layer; Using the backpropagation algorithm, the loss value between the network prediction value and the true value of the training set is fed back from the last layer forward to update the weights. After multiple forward feedbacks and backpropagations until the model converges, a trained model is obtained. This model consists of seven convolutional layers and is therefore called a fully convolutional neural network; And use the trained model as a pre-trained model; By analyzing the composition logic of the FCN, the idea of fine-tuning is adopted to train the interference area detection model of the automotive wheel hub bearing image; Step 3: Image edge detection: First, perform rough edge detection on the automotive wheel hub bearing image, and then perform fine positioning on the edges of the automotive wheel hub bearing image; Adopt the Canny adaptive edge detection method to divide the entire image into several sub-images. In order to make the contours continuous, there is a certain overlapping area between the sub-images. The proportion parameter of the overlapping area in the sub-image is d, and then the high and low thresholds of each sub-image are adaptively set according to the result after non-maximum suppression; Use the interpolation method to perform sub-pixel positioning on the automotive wheel hub bearing image. The cubic spline interpolation method is based on the edge coordinates located by the rough edge image of the automotive wheel hub bearing, and the discrete pixel gray values are interpolated using the cubic spline function on the image after the interference area is processed to obtain a continuous edge gray distribution curve, improving the accuracy of edge detection to the sub-pixel level; Step 4: Platform construction and application verification: 4.1. Hardware platform construction: The hardware part mainly includes a light source system, an industrial camera, a lens, an image acquisition card, and a computer; a. Selection of industrial camera: Select a CMOS camera; b.Lens: According to the working environment in the machining of automotive wheel bearings, the lens should be selected with high clarity, a compact shape, excellent seismic resistance, and characteristics of high and low temperatures; c. Light source: Select an LED ring light and directly install it coaxially with the industrial lens; 4.2. Application verification: Convert and measure the dimensions of automotive wheel bearings according to the pixel equivalent, and the experimental measurement results will be displayed on the software interface; After detecting the images of automotive wheel bearings, the interference areas can be effectively detected and processed, and then an edge image with good integrity and no noise can be obtained for dimension detection and error analysis.

[0004] CN115060497B discloses a bearing fault diagnosis method based on CEEMD energy entropy and optimized PNN, including: S1. Collect bearing vibration signals in different fault states; S2. Use the complementary ensemble empirical mode decomposition algorithm to decompose the bearing vibration signals to obtain intrinsic mode components; S3. Calculate the correlation coefficients of the intrinsic mode components and the bearing vibration signal data and perform screening processing to obtain effective intrinsic mode components; S4. Extract the energy entropy of the effective intrinsic mode components to form a feature vector matrix;

[0005] S5. Build a probabilistic neural network model, input the feature vector matrix into the probabilistic neural network model, train and optimize the probabilistic neural network model to obtain a probabilistic neural network bearing fault diagnosis model; S6. Input the feature vector matrix into the probabilistic neural network bearing fault diagnosis model, build a probabilistic neural network fault diagnosis model optimized by the improved sparrow search algorithm, and complete fault identification and classification.

[0006] CN119374902A discloses a sliding bearing fault detection method and system based on IGWO-VMD-CNN, including the following steps: Obtain the acoustic emission signals of the sliding bearing, and the acoustic emission signals of the sliding bearing include the acoustic emission signals of the sliding bearing without faults and the acoustic emission signals of the sliding bearing with faults; Introduce an improved tent mapping, a nonlinear convergence operator, and an improved wolf position update mechanism to construct an improved grey wolf optimization algorithm (IGWO); Based on the improved grey wolf optimization algorithm, combine the variational mode decomposition algorithm (VMD) to reconstruct the acoustic emission signals of the sliding bearing to obtain the reconstructed acoustic emission signals of the sliding bearing; Build a sliding bearing fault detection neural network (CNN) to perform fault detection and identification processing on the reconstructed acoustic emission signals of the sliding bearing to obtain the sliding bearing fault detection results.

[0007] When facing complex non - linear errors, the existing comprehensive detection equipment for bearing inner dimensions is difficult to meet the actual needs of high - precision production scenarios due to the limitations of its linear correction and simple error correction methods. In the actual production environment, the operating conditions of the equipment are complex and variable, and the traditional calibration and correction methods are difficult to adapt to these changes and cannot adjust the measurement errors in real - time and accurately. Summary of the Invention

[0008] Through long - term practice, it is found that most detection equipment is based on linear correction methods. This method assumes a linear relationship between errors and measured values. However, in actual measurements, errors often have complex non - linear characteristics. Usually, it can only handle some basic and highly regular errors and cannot effectively compensate for complex non - linear errors. For example, during the bearing size detection process, the actual measurement errors may be affected by a combination of factors such as temperature changes, mechanical structure deformation, and the non - linear characteristics of sensors. These factors cause the errors to exhibit complex non - linear characteristics. In high - precision production scenarios, the accuracy requirements for bearing sizes are extremely high. The traditional linear correction and simple error correction methods cannot meet this high - precision requirement, resulting in technical problems such as a large deviation between the measurement result and the true size.

[0009] In view of this, the present invention aims to propose a bearing size prediction method based on a CNN - LSTM - Attention neural network, including:

[0010] Step S1: Measure the geometric dimension data of the bearing through a bearing measurement device. The geometric dimension data at least includes the inner diameter of the bearing, the roundness of the inner ring, and the taper of the inner ring; and collect environmental data. The environmental data at least includes the temperature value; respectively perform initialization processing on the geometric dimension data and the environmental data to obtain a first data set.

[0011] Step S2: Perform maximum - minimum normalization processing on the first data set and divide it into a training set and a test set according to 8:2. The training set is used as the training data for the prediction model, and the test set is the test data for the prediction model.

[0012] Step S3: Construct a prediction model. The prediction model includes an input layer, a CNN module, an Attention module, an LSTM module, and an output layer connected in series; first convert the training set data into two - dimensional feature variables as the input of the input layer. The convolutional layer uses a two - dimensional convolutional layer with a convolutional kernel size of 1*1 and the number of channels being 32 and 64 respectively, and uses the Relu activation function; the Attention module includes two fully - connected layers, consisting of 32 neurons and 64 neurons respectively; the LSTM module includes 6 hidden neurons; input the training set, and use the true geometric dimension data of the bearing as the output. After training the prediction model, when the prediction accuracy reaches a preset value, a trained prediction model is obtained.

[0013] Step S4: Use a sensor to collect the geometric dimension data and the environmental data of the bearing in real time, input them into the trained prediction model, and obtain a prediction result.

[0014] In one embodiment, in step S2, the geometric dimension data and the environmental data in the first data set are respectively normalized and spliced to form an array.

[0015] In one embodiment, the array is converted into a 4D array.

[0016] In one embodiment, in step S3, after constructing the prediction model, the hyperparameters of the neural network model are optimized and set by an intelligent algorithm, and the intelligent algorithm includes a particle swarm optimization algorithm.

[0017] In one embodiment, the Attention module includes at least two self-attention mechanism layers.

[0018] The present invention also discloses a system for executing the bearing size prediction method based on the CNN-LSTM-Attention neural network as described above. The system includes:

[0019] A data acquisition unit for measuring the geometric dimension data of the bearing through a bearing measuring device. The geometric dimension data at least includes the inner diameter of the bearing, the roundness of the inner ring, and the taper of the inner ring; and collecting environmental data, the environmental data at least includes a temperature value; respectively performing initialization processing on the geometric dimension data and the environmental data to obtain a first data set;

[0020] A data preprocessing unit for performing maximum-minimum normalization processing on the first data set and dividing it into a training set and a test set according to 8:2. The training set is used as the training data of the prediction model, and the test set is the test data of the prediction model;

[0021] A model construction unit for constructing a prediction model. The prediction model includes an input layer, a CNN module, an Attention module, an LSTM module, and an output layer connected in series; converting the training set data into two-dimensional feature variables as the input of the input layer. The convolutional layer uses a two-dimensional convolutional layer, the convolutional kernel size is 1*1, and the number of channels is 32 and 64 respectively, and the Relu activation function is used; the Attention module includes two fully connected layers, which are respectively composed of 32 neurons and 64 neurons; the LSTM module includes 6 hidden neurons; inputting the training set, and using the real geometric dimension data of the bearing as the output, after training the prediction model, when the prediction accuracy reaches a preset value, a trained prediction model is obtained;

[0022] A prediction unit, configured to collect the geometric dimension data and the environmental data of the bearing in real time by using a sensor, input the collected data into a trained prediction model, and obtain a prediction result.

[0023] In one embodiment, the system further includes a tuning unit, configured to optimize the hyperparameters of the neural network model by using an intelligent algorithm, where the intelligent algorithm includes a particle swarm optimization algorithm.

[0024] In one embodiment, the system further includes an evaluation unit, configured to evaluate the prediction performance of the prediction model.

[0025] The present invention also discloses an electronic device, including at least one processor; and

[0026] a memory communicatively connected to the at least one processor; wherein,

[0027] the memory stores instructions executable by the at least one processor, and when the instructions are executed by the at least one processor, the at least one processor is enabled to execute the bearing dimension prediction method based on the CNN-LSTM-Attention neural network as described above.

[0028] The present invention also discloses a machine-readable storage medium, on which instructions are stored, and the instructions are used to cause a machine to execute the bearing dimension prediction method based on the CNN-LSTM-Attention neural network as described above in this application.

[0029] The bearing size prediction method based on the CNN-LSTM-Attention neural network disclosed in the present invention, through steps S1-S4, first collects the geometric size data (including inner diameter, inner ring roundness, and inner ring taper) and environmental data (such as temperature value) of the bearing through a bearing measurement device, and performs initialization processing on these data to generate a first data set; then performs maximum-minimum normalization processing on the data set and divides it into a training set and a test set according to a ratio of 8:2. Subsequently, a prediction model composed of an input layer, a CNN module, an Attention module, an LSTM module, and an output layer is constructed. Among them, the CNN module uses a 1×1 convolution kernel and a ReLU activation function, the Attention module contains two fully connected layers, and the LSTM module contains 6 hidden neurons. The model is trained with the training set until the prediction accuracy reaches a preset value. Finally, the geometric size and environmental data of the bearing are collected in real time by sensors and input into the trained model to obtain the prediction result of the bearing geometric size. This method combines multi-source data fusion and deep learning technology, uses CNN to extract spatial features, LSTM to capture temporal features, and the Attention mechanism to enhance key information, and can effectively predict the geometric size of the bearing to perform high-precision compensation for the measurement error of the bearing. The present invention also discloses a system for executing the above method. This method and system can effectively compensate for complex non-linear errors, such as temperature changes, mechanical structure deformations, sensor non-linear characteristics, etc., to obtain more accurate bearing sizes.

[0030] Other features and advantages of the present invention will be described in detail in the subsequent specific implementation part. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0032] In the accompanying drawings:

[0033] Figure 1 Schematic diagram of the bearing size prediction method based on the CNN-LSTM-Attention neural network according to an embodiment of the present invention;

[0034] Figure 2 Graph of the bearing size prediction result based on the CNN-LSTM-Attention neural network according to an embodiment of the present invention;

[0035] Figure 3 Inner size comprehensive detection device of the bearing for collecting data in the bearing size prediction method based on the CNN-LSTM-Attention neural network according to an embodiment of the present invention;

[0036] Figure 4 Prediction result graph of the bearing size prediction method based on support vector regression according to an embodiment of the present invention;

[0037] Figure 5 Prediction result graph of the bearing size prediction method of the PSO - BP neural network according to an embodiment of the present invention;

[0038] Figure 6 Prediction result graph of the bearing size prediction method of the LSTM neural network according to an embodiment of the present invention.

[0039] Explanation of reference numerals:

[0040] 1, lifting mechanism; 2, rotating mechanism; 3, inner diameter measurement system; 4, placement plane; 5, connection plane; 6, base. Detailed implementation manners

[0041] The following will describe in detail the specific implementation manners of the present invention with reference to the accompanying drawings. It should be understood that the specific implementation manners described herein are only for explaining and illustrating the present invention, and are not used to limit the present invention.

[0042] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a 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 of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0043] It should be noted that the terms "first", "second", "third", etc. in the specification and claims of the present invention and the above - mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances for the embodiments of the present invention described herein. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non - exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these process, method, product or device.

[0044] To solve the problems in the prior art, most detection devices are based on linear correction methods. This method assumes a linear relationship between errors and measured values. However, in actual measurements, errors often have complex non-linear characteristics. Usually, only some basic and highly regular errors can be processed, and complex non-linear errors cannot be effectively compensated. For example, during the bearing size detection process, the actual measurement errors may be affected by a combination of factors such as temperature changes, mechanical structure deformations, and non-linear characteristics of sensors. These factors result in complex non-linear error characteristics. In high-precision production scenarios, the precision requirements for bearing sizes are extremely high. Traditional linear correction and simple error correction methods cannot meet such high-precision requirements, leading to large deviations between measurement results and true sizes and other technical problems. The present invention provides a bearing size prediction method based on a CNN-LSTM-Attention neural network, as Figure 1-2 shown in the schematic diagram of a bearing size prediction method based on a CNN-LSTM-Attention neural network according to an embodiment of the present invention. The bearing size prediction method based on a CNN-LSTM-Attention neural network includes

[0045] Step S1: Measure the geometric size data of the bearing through a bearing measurement device. The geometric size data at least includes the inner diameter of the bearing, the roundness of the inner ring, and the taper of the inner ring; and collect environmental data, which at least includes the temperature value; respectively perform initialization processing on the geometric size data and the environmental data to obtain a first data set;

[0046] Step S2: Perform maximum-minimum normalization processing on the first data set and divide it into a training set and a test set according to a ratio of 8:2. The training set is used as the training data for the prediction model, and the test set is the test data for the prediction model;

[0047] Step S3: Build a prediction model. The prediction model includes an input layer, a CNN module, an Attention module, an LSTM module, and an output layer connected in series; first convert the training set data into two-dimensional feature variables as the input of the input layer. The convolutional layer uses a two-dimensional convolutional layer with a kernel size of 1*1 and the number of channels being 32 and 64 respectively, and uses the Relu activation function; the Attention module includes two fully connected layers, which are composed of 32 neurons and 64 neurons respectively; the LSTM module includes 6 hidden neurons; input the training set, and use the true geometric size data of the bearing as the output. After training the prediction model, when the prediction accuracy reaches a preset value, a trained prediction model is obtained;

[0048] Step S4: Use a sensor to real-time collect the geometric size data and the environmental data of the bearing, input them into the trained prediction model, and obtain a prediction result.

[0049] The bearing size prediction method based on the CNN-LSTM-Attention neural network disclosed by the present invention, through steps S1-S4, first collects the geometric size data (including inner diameter, inner ring roundness and inner ring taper) and environmental data (such as temperature value) of the bearing through a bearing measurement device, and performs initialization processing on these data to generate a first data set; then performs maximum-minimum normalization processing on the data set and divides it into a training set and a test set according to a ratio of 8:2. Subsequently, a prediction model composed of an input layer, a CNN module, an Attention module, an LSTM module and an output layer is constructed. Among them, the CNN module uses a 1×1 convolutional kernel and a ReLU activation function, the Attention module includes two fully connected layers, and the LSTM module includes 6 hidden neurons. The model is trained with the training set until the prediction accuracy reaches a preset value. Finally, the geometric size and environmental data of the bearing are collected in real time by a sensor and input into the trained model to obtain the prediction result of the bearing geometric size. This method combines multi-source data fusion and deep learning technology, uses CNN to extract spatial features, LSTM to capture temporal features, and the Attention mechanism to enhance key information, and can effectively predict the geometric size of the bearing and perform high-precision compensation for the measurement error of the bearing. This method can effectively compensate for complex non-linear errors, such as temperature changes, mechanical structure deformations, sensor non-linear characteristics, etc., and obtain more accurate bearing sizes.

[0050] Taking Figure 3 the bearing inner diameter comprehensive measurement device shown as an example, the bearing inner diameter comprehensive measurement device includes a connecting plane 5 and a base 6, and the connecting plane 5 is vertically fixed on the base 6. Place the bearing to be measured on the placing plane 4, and under the mutual cooperation of the lifting mechanism 1, the rotating mechanism 2 and the inner diameter measurement system 3, collect the inner size data of the bearing to be measured, that is, the bearing measurement device measures the geometric size data of the bearing. In step S2, perform maximum-minimum normalization processing on the first data set, that is where, X norm is the data after maximum-minimum normalization, X is the original data, and X max and X min are the maximum and minimum values in the data set respectively.

[0051] In order to be directly input into a machine learning or deep learning model for training and prediction, the data is flattened to adapt to the input format of the neural network. In a more preferred embodiment, in step S2, normalization processing is respectively performed on the geometric dimension data and the environmental data in the first dataset, and they are concatenated to form an array. Normalizing the geometric dimension data and the environmental data respectively avoids the influence of different dimensions. Wherein, when the environmental data includes the temperature measurement value of the environment. For example, when the environmental data only has the temperature measurement value, and the geometric dimension data is the inner diameter, inner ring roundness, and inner ring taper, they are concatenated according to the sample dimension to form a unified input array. Then, before input, the array needs to be converted into a 4D array. In this way, it is better applicable to the scenario of multi-source data fusion.

[0052] In order to find the optimal combination of hyperparameters more efficiently to improve the performance and prediction accuracy of the model. In a more preferred embodiment, in step S3, after constructing the prediction model, the hyperparameters of the neural network model are optimized and set through an intelligent algorithm, and the intelligent algorithm includes the particle swarm optimization algorithm. For example, the hyperparameters include the learning rate (LearningRate), batch size (Batch Size), number of network layers (Number of Layers), number of neurons in each layer (Numberof Neurons), activation function type (Activation Function), regularization parameters (RegularizationParameters). Manually adjusting these hyperparameters is not only time-consuming but also difficult to find the global optimal solution. Therefore, using an intelligent algorithm for automated optimization can significantly improve efficiency and model performance. The hyperparameters of the neural network model are optimized and set through the particle swarm optimization algorithm to find the optimal combination of hyperparameters and improve the prediction accuracy and generalization ability. As Figure 1 shown, for example, the input layer accepts 2D feature variables as input, including geometric dimension data and environmental data; the convolutional layer uses a two-dimensional convolutional layer, the convolutional kernel size is 1*1, and the number of channels is 32 and 64 respectively. The ReLU activation function is used to extract the features of the input data and further enhance the feature expression ability. The pooling layer uses a global average pooling layer, which retains the average information of each channel, reduces the complexity of the model, and avoids overfitting. The Attention module is used to make the model more focused on the parts of the input sequence that are more important for the current task; it contains two fully connected layers, consisting of 32 neurons and 64 neurons respectively, for further expanding and refining the features. The LSTM module contains 6 hidden neural units, which are used to capture the long-term dependencies in the sequence data and process the time series features; it contains 1 neuron, which maps the processed features to a single predicted value, that is, the predicted result of the inner dimension of the bearing.

[0053] The self-attention mechanism is a mechanism that can capture the internal dependencies of input data and is widely used in natural language processing (NLP) and computer vision (CV) tasks. Its core idea is to dynamically assign weights to each element in the input sequence by calculating the correlation between each element and other elements, thereby highlighting important information. In time series or spatial data, there may be important relationships between elements at a long distance, and the self-attention mechanism can directly model these relationships. In a more preferred embodiment, the Attention module includes at least two layers of self-attention mechanism layers. In order to gradually extract higher-level features through multi-layer stacking. The purpose of designing at least two layers of self-attention mechanism layers in the Attention module is to further optimize the ability of feature extraction and dependency modeling. The first layer of the self-attention mechanism is used to capture the local dependencies in the input data and extract preliminary important features. The second layer of the self-attention mechanism is used to further capture higher-level dependencies on the basis of the features extracted by the first layer and extract more abstract features. Through gradually extracting features, the multi-layer self-attention mechanism can avoid the model from overfitting to noise or local features prematurely, thereby improving the generalization ability.

[0054] In order to verify the prediction ability of the neural network model for bearing inner dimension prediction provided by the present invention, when establishing the CNN-LSTM-Attention neural network, three different machine learning models were also established, trained synchronously using the same data packet, and three other network models were established. The three different machine learning models specifically used include SVR (Support Vector Regression), PSO-BP neural network, and LSTM, and the prediction results are as Figures 4 to 6 shown.

[0055] Among them, in order to comprehensively verify the prediction ability of the neural network model for bearing inner dimension prediction, the present invention evaluates the provided neural network model and the three machine learning models established simultaneously. The evaluation indicators used include the mean square error MSE, root mean square error RMSE, and coefficient of determination R 2 . The prediction results of the bearing dimension prediction method based on the CNN-LSTM-Attention neural network proposed by the present invention are closely fitted to the true values, with small errors, and can meet the measurement requirements of the current bearing inner dimension comprehensive detection equipment. According to the prediction results and the true values, MAE, RMSE, and R 2 are calculated and the results are output.

[0056] Table 1: R 2 , MAE, and RMSE based on SVR, PSO-BP, LSTM, and CNN-LSTM-Attention models

[0057]

[0058] The present invention also discloses a system for executing the bearing size prediction method based on the CNN-LSTM-Attention neural network as described above. The system includes,

[0059] A data acquisition unit, configured to measure the geometric dimension data of the bearing through a bearing measuring device. The geometric dimension data at least includes the bearing inner diameter, inner ring roundness, and inner ring taper; and collect environmental data, where the environmental data at least includes a temperature value; respectively perform initialization processing on the geometric dimension data and the environmental data to obtain a first data set;

[0060] A data preprocessing unit, configured to perform max-min normalization processing on the first data set and divide it into a training set and a test set according to 8:2. The training set is used as the training data of the prediction model, and the test set is the test data of the prediction model;

[0061] A model construction unit, configured to construct a prediction model. The prediction model includes an input layer, a CNN module, an Attention module, an LSTM module, and an output layer connected in series; convert the training set data into two-dimensional feature variables as the input of the input layer. The convolutional layer uses a two-dimensional convolutional layer, the convolutional kernel size is 1*1, and the number of channels is 32 and 64 respectively. The Relu activation function is used; the Attention module includes two fully connected layers, which are composed of 32 neurons and 64 neurons respectively; the LSTM module includes 6 hidden neurons; input the training set, and use the true geometric dimension data of the bearing as the output. After training the prediction model, when the prediction accuracy reaches a preset value, a trained prediction model is obtained;

[0062] A prediction unit, configured to use a sensor to collect the geometric dimension data and the environmental data of the bearing in real time, input the trained prediction model, and obtain a prediction result.

[0063] The system for implementing the bearing size prediction method based on the CNN-LSTM-Attention neural network as described above collects the geometric size data (including inner diameter, inner ring roundness, and inner ring taper) and environmental data (such as temperature values) of the bearing through a data acquisition unit, and initializes these data to generate a first data set. The data preprocessing unit performs maximum-minimum normalization on the data set and divides it into a training set and a test set according to a ratio of 8:2. The model construction unit constructs a prediction model composed of an input layer, a CNN module, an Attention module, an LSTM module, and an output layer. The CNN module uses a 1×1 convolutional kernel and a ReLU activation function. The Attention module contains two fully connected layers. The LSTM module contains 6 hidden neurons. The model is trained with the training set until the prediction accuracy reaches a preset value. Finally, the prediction unit uses sensors to collect the geometric size and environmental data of the bearing in real time, inputs them into the trained model, and obtains the prediction results of the bearing geometric size. This system combines multi-source data fusion and deep learning technologies, uses CNN to extract spatial features, LSTM to capture temporal features, and the Attention mechanism to enhance key information, and can effectively predict the geometric size of the bearing, and perform high-precision compensation for the measurement error of the bearing. This system can effectively compensate for complex non-linear errors, such as temperature changes, mechanical structure deformations, sensor non-linear characteristics, etc., and obtain more accurate bearing sizes.

[0064] In order to automatically find the optimal combination of hyperparameters to improve the performance, prediction accuracy, and generalization ability of the model. In a more preferred embodiment, the system further includes a tuning unit, and the tuning unit is used to optimize the hyperparameters of the neural network model through an intelligent algorithm, and the intelligent algorithm includes a particle swarm optimization algorithm. Manually adjusting these hyperparameters is not only time-consuming but also difficult to find the global optimal solution. Therefore, using an intelligent algorithm for automatic optimization can significantly improve efficiency and model performance. Therefore, the tuning unit optimizes the hyperparameters of the neural network model through an intelligent algorithm (such as the particle swarm optimization algorithm, PSO).

[0065] In order to quantitatively evaluate the performance of the model and ensure its reliability, accuracy, and robustness in practical applications. In a more preferred embodiment, the system further includes an evaluation unit, and the evaluation unit is used to evaluate the prediction performance of the prediction model. As Figures 4 - 6 shown in Table 1. The evaluation unit is one of the core components of the system and is responsible for comprehensively evaluating the performance of the trained prediction model. For example, the prediction ability of the model is quantitatively evaluated and optimized through evaluation metrics (such as accuracy, precision, recall, F1 score, etc.).

[0066] The present invention also discloses an electronic device, at least one processor; and

[0067] a memory communicatively connected to the at least one processor; wherein,

[0068] the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the bearing size prediction method based on the CNN-LSTM-Attention neural network as described above.

[0069] The present invention also discloses a machine-readable storage medium, on which instructions are stored, and the instructions are used to cause a machine to execute the bearing size prediction method based on the CNN-LSTM-Attention neural network as described above in the present application.

[0070] It should be noted that, for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0071] In the above embodiments, the descriptions of the respective embodiments have their own emphases. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0072] In addition, in each embodiment of the present invention, the functional units can be integrated in one processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above integrated units can be implemented in the form of hardware or in the form of software functional units.

[0073] The foregoing is only the preferred embodiments of the present invention and is not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A bearing size prediction method based on CNN-LSTM-Attention neural network, characterized in that: The bearing size prediction method based on the CNN-LSTM-Attention neural network includes: Step S1, measuring the geometric dimension data of the bearing by means of a bearing measuring device, wherein the geometric dimension data at least includes the inner diameter of the bearing, the inner ring roundness and the inner ring taper; and collecting environmental data, wherein the environmental data at least includes a temperature value; initializing the geometric dimension data and the environmental data to obtain a first data set; Step S2, performing maximum-minimum normalization processing on the first data set, and dividing it into a training set and a test set according to an 8:2 ratio, the training set is used as training data of the prediction model, and the test set is used as test data of the prediction model; Step S3, constructing a prediction model, the prediction model includes a serially connected input layer, a CNN module, an Attention module, an LSTM module and an output layer; the training set data is first converted into a two-dimensional feature variable as the input of the input layer, the convolution layer adopts a two-dimensional convolution layer, the convolution kernel size is 1*1, the number of channels is 32 and 64 respectively, and the ReLU activation function is adopted; the Attention module includes two fully connected layers, which are composed of 32 neurons and 64 neurons respectively; the LSTM module includes 6 hidden neurons; the training set is input, and the actual geometric size data of the bearing is used as the output. After the prediction model is trained, when the prediction accuracy reaches the preset value, the trained prediction model is obtained; Step S4, using sensors to collect the geometric dimension data and environmental data of the bearing in real time, inputting the data into the trained prediction model to obtain a prediction result.

2. The bearing size prediction method based on CNN-LSTM-Attention neural network according to claim 1 is characterized in that: In step S2, the geometric dimension data and the environmental data in the first data set are normalized respectively and concatenated to form an array.

3. The bearing size prediction method based on CNN-LSTM-Attention neural network according to claim 2 is characterized in that: Convert the array to a 4D array.

4. The bearing size prediction method based on CNN-LSTM-Attention neural network according to claim 1 is characterized in that: In step S3, after the prediction model is constructed, the hyperparameters of the neural network model are optimized and set by an intelligent algorithm, wherein the intelligent algorithm includes a particle swarm optimization algorithm.

5. The bearing size prediction method based on CNN-LSTM-Attention neural network according to any one of claims 1 to 4, characterized in that: The Attention module includes at least 2 layers of self-attention mechanism.

6. A system for executing the bearing size prediction method based on CNN-LSTM-Attention neural network as described in any one of claims 1 to 5, characterized in that: The system comprises, A data acquisition unit is used to measure the geometric dimension data of the bearing by means of a bearing measuring device, wherein the geometric dimension data at least includes the inner diameter of the bearing, the inner ring roundness and the inner ring taper; and to acquire environmental data, wherein the environmental data at least includes a temperature value; and to initialize the geometric dimension data and the environmental data to obtain a first data set; A data preprocessing unit, used for performing maximum-minimum normalization processing on the first data set, and dividing it into a training set and a test set according to an 8:2 ratio, the training set is used as training data of the prediction model, and the test set is used as test data of the prediction model; A model building unit is used to build a prediction model, which includes a serially connected input layer, a CNN module, an Attention module, an LSTM module and an output layer; the training set data is first converted into a two-dimensional feature variable as the input of the input layer, the convolution layer adopts a two-dimensional convolution layer, the convolution kernel size is 1*1, the number of channels is 32 and 64 respectively, and the ReLU activation function is adopted; the Attention module includes two fully connected layers, which are respectively composed of 32 neurons and 64 neurons; the LSTM module includes 6 hidden neurons; the training set is input, and the actual geometric size data of the bearing is used as the output. After the prediction model is trained, when the prediction accuracy reaches a preset value, a trained prediction model is obtained; The prediction unit is used to collect the geometric dimension data and the environmental data of the bearing in real time using a sensor, input the data into a trained prediction model, and obtain a prediction result.

7. The system according to claim 6, characterized in that The system also includes a tuning unit, which is used to optimize the hyperparameters of the neural network model through an intelligent algorithm, and the intelligent algorithm includes a particle swarm optimization algorithm.

8. The bearing size prediction method based on CNN-LSTM-Attention neural network according to claim 7 is characterized in that: The system further comprises an evaluation unit, which is used for evaluating the prediction performance of the prediction model.

9. An electronic device, characterized in that: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the bearing size prediction method based on the CNN-LSTM-Attention neural network described in any one of claims 1-5.

10. A machine-readable storage medium, characterized in that: The machine-readable storage medium stores instructions for enabling a machine to execute a bearing size prediction method based on a CNN-LSTM-Attention neural network as described in any one of claims 1 to 5 of the present application.

Images