Material removal state prediction method based on multi-scale attention mechanism
Through the neural network model based on the multi-scale attention mechanism processing displacement data, the problems of cumbersome detection and poor prediction accuracy in material removal state detection are solved, and efficient and accurate prediction of material removal state is achieved.
Patent Information
- Application Number
- CN202210327392.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-30
- Publication Date
- 2025-03-25
- Estimated Expiration
- 2042-03-30
AI Technical Summary
The prior art has problems such as cumbersome detection, high cost, poor prediction accuracy and low processing efficiency in material removal state detection, especially in the influence of processing errors and wear factors, it is difficult to accurately predict the material removal state.
Using a neural network model based on a multi-scale attention mechanism, the displacement data is processed through first-order difference, standardization and data amplification, combined with a multi-scale convolutional neural network and attention mechanism, the characteristics of the material removal state are extracted and classified predictions are performed.
It improves the accuracy of data feature extraction and the prediction accuracy of network model, realizes effective prediction of material removal state, reduces data volume requirements, and improves processing efficiency.
Smart Images
Figure CN114897045B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of material removal status, and in particular to a material removal status prediction method based on a multi-scale attention mechanism. Background Art
[0002] In the actual processing process, it is difficult to determine the material removal status due to factors such as abrasive belt wear and processing errors. The accurate removal of workpiece material in grinding production determines the final processing quality of the workpiece, so material removal detection of the workpiece is particularly important.
[0003] At present, the detection of material removal status is divided into direct detection and indirect detection. The main method of direct detection is to measure the removal amount through special instruments. The measurement procedure is often very cumbersome, which greatly increases the time and economic cost and affects the processing efficiency. Indirect detection can be divided into mechanism model prediction and learning model prediction. Among them, the mechanism model prediction has limited research on the material removal status of the processed workpiece. There are many factors that affect the removal rate, making it very difficult to predict the material removal status by establishing a data model or a simulation model. Learning model prediction uses artificial intelligence algorithms. Thanks to the rapid development of intelligent algorithms, the complex nonlinear mapping relationship between sensor signals and targets can be effectively established by performing model analysis on the collected signal data. However, this type of model is only suitable for target prediction under complex conditions in industry, and there are problems of low processing efficiency and poor prediction accuracy. Summary of the invention
[0004] In view of this, the purpose of the present invention is to overcome the defects in the prior art and provide a material removal state prediction method based on a multi-scale attention mechanism, which can improve the accuracy of data feature extraction and the prediction accuracy of the network model, and realize effective prediction of material removal state.
[0005] The material removal state prediction method based on the multi-scale attention mechanism of the present invention comprises the following steps:
[0006] S1. Collect displacement data of processed materials and determine the removal status of processed materials;
[0007] S2. Preprocessing the displacement data of the processed material to obtain processed displacement data; the preprocessing includes first-order difference processing, standardization processing and data amplification processing in sequence;
[0008] S3. Construct a neural network model based on multi-scale attention mechanism;
[0009] S4. Setting network training parameters and using the processed displacement data to perform network training on the neural network model to obtain a trained network model;
[0010] S5. Input the displacement data of the processing material to be tested into the trained network model, and output the removal status of the processing material to be tested.
[0011] Furthermore, the displacement data of the processed material is processed by first-order difference according to the following formula:
[0012]
[0013] Among them, y i ′ is the i-th displacement data point in the time series, y i ' +1 is the i+1th displacement data point in the time series, y i is the difference between displacement data points in adjacent time series, n=len(y′)-1, wherein len(y′) is the total number of displacement data points of displacement data y′ in the time series, and y is the displacement data after first-order difference processing.
[0014] Furthermore, the displacement data of the processed material is standardized according to the following formula:
[0015]
[0016] Wherein, μ is the data y i The mean of the data y i The standard deviation of x i For the data y i The standardized value of , x is the set of standardized values.
[0017] Furthermore, a sliding window method is used to perform data amplification processing on the normalized displacement data to obtain amplified displacement data.
[0018] Furthermore, the neural network model includes a denoising layer, a multi-scale feature extraction layer, an attention layer, and a classification layer connected in sequence;
[0019] The denoising layer is used to perform denoising and feature pre-extraction on the processed displacement data to obtain denoised displacement data;
[0020] The multi-scale feature extraction layer is used to perform multi-scale feature extraction on the denoised displacement data to obtain displacement data after feature extraction;
[0021] The attention layer is used to perform feature weighting processing on the displacement data after feature extraction to obtain feature weighted displacement data;
[0022] The classification layer is used to classify the displacement data after feature weighting to obtain the classification result of the displacement data.
[0023] Furthermore, the denoising layer includes a one-dimensional CNN network; the CNN network includes a convolutional layer and an average pooling layer.
[0024] Furthermore, the multi-scale feature extraction layer includes a one-dimensional convolutional layer and three parallel one-dimensional convolutional layers.
[0025] Furthermore, the attention layer includes a one-dimensional CNN network and a SENet network.
[0026] Furthermore, the classification layer includes a fully connected neural network layer.
[0027] Furthermore, the processed displacement data is divided into a training set and a test set by a k-fold cross validation method, and the training set and the test set are input into a neural network model for network training.
[0028] The beneficial effects of the present invention are as follows: the present invention discloses a material removal state prediction method based on a multi-scale attention mechanism, which eliminates the random trend contained in the data by using differential preprocessing displacement data, reduces the irregular fluctuations between data, and makes the data tend to be stable, which is convenient for further analysis. Data enhancement is performed on the differentiated data to expand the amount of displacement data, which reduces the difficulty of collecting large amounts of data and solves the problem of large model data requirements. The amount of data required for this method is relatively small. Combining a multi-scale convolutional neural network with an attention mechanism avoids the defects of local feature information extraction, captures information at different scales, focuses on feature information with high weights, solves the problem of information overload, improves processing efficiency and accuracy, and has a wide range of applications, and improves the accuracy of data feature extraction and the prediction accuracy of the model, thereby achieving effective prediction of material removal status. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:
[0030] Figure 1 It is a schematic diagram of the method flow of the present invention;
[0031] Figure 2 It is a schematic diagram of the data amplification principle of the present invention;
[0032] Figure 3 It is a schematic diagram of the structure of the denoising layer of the present invention;
[0033] Figure 4 It is a schematic diagram of the structure of the multi-scale feature extraction layer of the present invention;
[0034] Figure 5 Schematic diagram of the attention layer structure of the present invention;
[0035] Figure 6 Schematic diagram of the classification layer structure of the present invention. DETAILED DESCRIPTION
[0036] The present invention is further described below in conjunction with the accompanying drawings, as shown in the drawings:
[0037] The material removal state prediction method based on the multi-scale attention mechanism of the present invention comprises the following steps:
[0038] S1. Collect displacement data of processed materials and determine the removal status of processed materials;
[0039] S2. Preprocessing the displacement data of the processed material to obtain processed displacement data; the preprocessing includes first-order difference processing, standardization processing and data amplification processing in sequence;
[0040] S3. Construct a neural network model based on multi-scale attention mechanism;
[0041] S4. Setting network training parameters and using the processed displacement data to perform network training on the neural network model to obtain a trained network model;
[0042] S5. Input the displacement data of the processing material to be tested into the trained network model, and output the removal status of the processing material to be tested.
[0043] In this embodiment, in step S1, an eddy current displacement sensor and a matching NI-CRIO data acquisition system are used to collect displacement signals generated by the processed material during the processing process. The displacement signal comes from the displacement data of the processed material or workpiece during the processing process. The displacement data is actually displacement change data, and the displacement change data is the displacement change value between the actual processing state of the processed material and the original state before the formal processing. The NI-CRIO data acquisition system adopts the existing technology and will not be repeated here. And the workpiece with a displacement data △d less than 0.5mm is regarded as an unqualified workpiece, otherwise it is regarded as a qualified workpiece, that is, by defining the critical value of the collected displacement data, it is determined whether the removal state of the processed material is qualified or unqualified. Generally, the displacement data of the processed material is collected in chronological order, thereby forming displacement data points in the time series.
[0044] In this embodiment, in step S2, the displacement data collected by the displacement sensor usually contains interference factors such as noise and error. Direct analysis of the displacement data is bound to fail to obtain ideal results. Therefore, it is necessary to preprocess the data first. Performing first-order difference, standardization and data amplification processing on the data is a prerequisite for the data to be used for deep learning prediction.
[0045] In order to eliminate the influence of errors, the present invention takes displacement data as the analysis object. In a processing cycle, although the displacement data at different positions are different, the displacement change trend is similar. The displacement data is processed by first-order difference. The first-order difference is to subtract the previous data from the next data of the time series displacement data to form new displacement change data. The calculation formula is as follows:
[0046]
[0047] Among them, y′ i is the i-th displacement data point in the time series, y′ i+1 is the i+1th displacement data point in the time series, y i is the difference between displacement data points in adjacent time series, n=len(y′)-1, wherein len(y′) is the total number of displacement data points of displacement data y′ in the time series, and y is the displacement data after first-order difference processing.
[0048] Differential preprocessing has certain limitations in noise control. If the differential data is directly analyzed, the network model will consider the local noise features contained in the data, resulting in interference. In order to avoid the local features of the extracted sample data being too obvious, the data is standardized before data amplification or enhancement. Standardization can not only speed up the convergence of the model, but also improve the accuracy of the model.
[0049] The calculation formula for standardization is as follows:
[0050]
[0051] Wherein, μ is the data y i The mean of the data y i The standard deviation of x i For the data y i The standardized value of , x is the set of standardized values.
[0052] After completing the differential preprocessing and standardization of the displacement data, data augmentation is performed. The data set is enlarged using sampling and sliding window enhancement. Data augmentation methods such as Figure 2 As shown. By using this data augmentation method, the source data set with a small sample can be expanded to a capacity sufficient for training the neural network model.
[0053] In this embodiment, based on the displacement data processing, a multi-scale attention convolutional neural network (MACNN) is proposed for predicting the state of material removal. The attention mechanism in the neural network is a resource allocation scheme that allocates computing resources to more important tasks and solves the problem of information overload when computing power is limited.
[0054] In neural network learning, the more parameters a model has, the stronger its expressive power is, and the more information the model can store. However, this can lead to the problem of information overload. By introducing an attention mechanism, we can focus on the information that is more critical to the current task among a large amount of input information, reduce attention to other information, and even filter out irrelevant information. This can solve the problem of information overload and improve the efficiency and accuracy of task processing. Multi-scale feature extraction effectively avoids the defect that only local feature information can be extracted due to the use of fixed-size convolution kernels. It can capture information at different scales, and has a significant improvement in classification tasks. The neural network model includes a denoising layer, a multi-scale feature extraction layer, an attention layer, and a classification layer connected in sequence;
[0055] The denoising layer is used to perform denoising and feature pre-extraction on the preprocessed displacement data. The denoising layer is a one-dimensional CNN network, which consists of a convolutional layer and an average pooling layer. The structure is as follows: Figure 3 shown.
[0056] CNN has good feature extraction capability and can extract key features from signal data. CNN networks can be freely combined to build models with better performance. In the present invention, sliding window convolution operations are performed on displacement data through convolution checks of different lengths to extract key features from the data.
[0057] The calculation formula is as follows:
[0058] y l =k l *y l-1 +b l (1)
[0059]
[0060] Among them, y l-1 is the output of the l-1th convolutional layer, and the l-1th layer is the input of the lth convolutional layer. l is the result of convolution calculation, k l is the size of the convolution kernel, b l is the bias, f is the activation function of the parameterized rectified linear unit (PReLU), a is a fixed value; y l ′ is the value calculated by the activation function.
[0061] In order to reduce the learning difficulty of the model, improve the training speed of the network and prevent overfitting, BatchNorm (BN) processing is required after the convolution operation to keep the input of each layer of the network (i.e., y l) distribution remains unchanged. By sampling the extracted features, the computational complexity is greatly reduced without losing features. The pooling layer is divided into average pooling and maximum pooling, and the calculation formula is as follows:
[0062]
[0063]
[0064] Among them, l+1 represents the l+1th layer. and are the outputs of average pooling and maximum pooling respectively. k is the core of the pooling layer, y′ l,i The input for pooling calculation.
[0065] The use of wide-core convolution kernels can effectively eliminate interfering data, and the average pooling process can effectively smooth and denoise the data and reduce the impact of noise. The preprocessed data is processed by the denoising layer, which can not only reduce the interference factors of the data, but also map the data to a multi-dimensional feature space, making it easier for the subsequent network to further extract features.
[0066] The multi-scale feature extraction (MSFE) layer consists of a single one-dimensional convolutional layer and three parallel one-dimensional convolutional layers.
[0067] The present invention adopts a sliding window data enhancement processing method to reduce the complexity of the collected data signal. Since the simple fixed single-scale convolution processing cannot extract multi-complexity features, the present invention adopts three multi-scale parallel convolution networks to extract multiple effective feature layers. Each convolution layer is followed by BN, PReLU and max pooling. Its network structure is as follows Figure 4 shown.
[0068] In the design of the multi-scale feature extraction layer, a separate one-dimensional convolution layer is used to reduce the number of channels in the denoising layer, and the features are further filtered to remove the common features that are not initially important. Then three parallel convolution layers with different convolution kernels are used to filter the features of different scales and extract features of different frequency bands. The BN layer and PReLU are connected sequentially after each convolution layer. Finally, three parallel pooling layers are used to highlight the extracted multi-scale features.
[0069] Multi-scale networks can extract features at different levels, but also bring about the diversity and complexity of features. The importance of these features is different, so the present invention constructs an attention network layer to identify these important features.
[0070] The attention layer consists of a one-dimensional CNN with a single-layer output and a SENet. The structure of the attention layer is as follows Figure 5As shown in the figure. Since the extracted features are relatively complex, a layer of convolutional neural network is first used to extract the initial attention, and then the SENet network is used to extract the final attention. SENet is a functional sub-network that can be combined with other networks to achieve better results. The main purpose of SENet is to learn the weights of different features and automatically assign different importance according to the loss value.
[0071] SENet is combined with CNN and applied to the multi-scale feature extraction layer, followed by the fully connected layer (FC) for classification. SENet calculates the weighted results of the features extracted by the multi-scale layer integrated by the convolutional neural network and the importance of the features, and finally realizes the extraction of feature values.
[0072] The classification layer is actually a FC neural network layer. The structure of the classification layer is as follows Figure 6 The main function of each layer mentioned above is to extract the corresponding features of the target, and the function of the classification layer is to classify these features using a single-layer neural network, that is, to classify the material removal status into two categories: qualified and unqualified.
[0073] Through the denoising layer, multi-scale feature extraction layer and attention layer, a certain amount of key features are extracted from the single input batch data points. Similarly, in order to ensure that the output features of each batch of data follow the same distribution, the feature data also needs to be processed by BN before connecting to the FC neural network. The processed feature data is used as the input of the FC neural network to finally obtain the predicted classification result.
[0074] In this embodiment, the processed displacement data is divided into a training set and a test set by a k-fold cross validation method, and the training set and the test set are input into a neural network model for network training. The k-fold cross validation method of dividing the data set is different from the traditional data division method, which can avoid the limitations of fixed division data sets and is widely used in model building.
[0075] The present invention constructs a multi-scale attention convolutional neural network model with first-order difference and data enhancement processed data as input. By performing first-order difference and normalization processing on the displacement data, combined with the periodic characteristics of the processing, and enhancing the data through overlapping sliding windows, the data volume is expanded, the difficulty of collecting a large amount of data is reduced, and the problem of large demand for model data is solved, laying a technical foundation for model prediction.
[0076] The actual combination of multi-scale convolutional neural networks and attention mechanisms can use parallel convolutional layers of different sizes to extract feature data of different frequency bands, and combine the attention network to assign different weights to the features to achieve accurate extraction of important features, greatly improving the prediction accuracy of the model and achieving accurate prediction of the material removal status.
[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.
Claims
1. A material removal state prediction method based on a multi-scale attention mechanism, characterized by: The steps include: S1. Collect displacement data of processed materials and determine the removal status of processed materials; S2. Preprocessing the displacement data of the processed material to obtain processed displacement data; the preprocessing includes first-order difference processing, standardization processing and data amplification processing in sequence; S3. Construct a neural network model based on multi-scale attention mechanism; S4. Setting network training parameters and using the processed displacement data to perform network training on the neural network model to obtain a trained network model; S5. Input the displacement data of the processing material to be tested into the trained network model, and output the removal status of the processing material to be tested.
2. The material removal state prediction method based on the multi-scale attention mechanism according to claim 1 is characterized in that: The displacement data of the processed material is processed by first-order difference according to the following formula: Among them, y′ i is the i-th displacement data point in the time series, y′ i+1 is the i+1th displacement data point in the time series, y i is the difference between displacement data points in adjacent time series, n=len(y′)-1, wherein len(y′) is the total number of displacement data points of displacement data y′ in the time series, and y is the displacement data after first-order difference processing.
3. The material removal state prediction method based on the multi-scale attention mechanism according to claim 2 is characterized in that: The displacement data of the processed material is standardized according to the following formula: Wherein, μ is the data y i The mean of the data y i The standard deviation of x i For the data y i The standardized value of , x is the set of standardized values.
4. The material removal state prediction method based on multi-scale attention mechanism according to claim 3 is characterized in that: The standardized displacement data is amplified by using a sliding window method to obtain amplified displacement data.
5. The material removal state prediction method based on multi-scale attention mechanism according to claim 1, characterized in that: The neural network model includes a denoising layer, a multi-scale feature extraction layer, an attention layer and a classification layer connected in sequence; The denoising layer is used to perform denoising and feature pre-extraction on the processed displacement data to obtain denoised displacement data; The multi-scale feature extraction layer is used to perform multi-scale feature extraction on the denoised displacement data to obtain displacement data after feature extraction; The attention layer is used to perform feature weighting processing on the displacement data after feature extraction to obtain feature weighted displacement data; The classification layer is used to classify the displacement data after feature weighting to obtain the classification result of the displacement data.
6. The material removal state prediction method based on multi-scale attention mechanism according to claim 5 is characterized in that: The denoising layer includes a one-dimensional CNN network; the CNN network includes a convolutional layer and an average pooling layer.
7. The material removal state prediction method based on multi-scale attention mechanism according to claim 5, characterized in that: The multi-scale feature extraction layer includes a one-dimensional convolution layer and three parallel one-dimensional convolution layers.
8. The material removal state prediction method based on multi-scale attention mechanism according to claim 5, characterized in that: The attention layer includes a one-dimensional CNN network and a SENet network.
9. The material removal state prediction method based on multi-scale attention mechanism according to claim 5, characterized in that: The classification layer includes a fully connected neural network layer.
10. The material removal state prediction method based on multi-scale attention mechanism according to claim 1, characterized in that: The processed displacement data is divided into a training set and a test set by a k-fold cross validation method, and the training set and the test set are input into a neural network model for network training.