Multi-source information fusion driven end-to-end surface roughness prediction method

Abstract

The invention discloses a multi-source information fusion driven end-to-end surface roughness prediction method, which comprises the following steps of: firstly, preprocessing original dynamic signal data, intercepting an effective part of a signal, and segmenting the signal by adopting a data enhancement method and a window method to fulfill the aim of expanding a data set; then, according to the method of the invention, a parallel convolutional neural network is utilized to directly carry out multi-scale feature processing on an input signal, and more important features can be reserved by utilizing an automatic feature extraction process. A mixed attention mechanism is introduced, the feature screening capability of a network model and the fusion capability of multi-sensor signal features are enhanced, finally, dynamic factors and static factors are combined to jointly predict the surface roughness, and the prediction precision of the surface roughness is further improved. According to the surface roughness prediction method, features can be automatically extracted by using multi-source data, complementation of processing information is realized, and the accuracy and reliability of roughness prediction are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of industrial manufacturing, and particularly relates to an end-to-end surface roughness prediction method driven by multi-source information fusion. BACKGROUND

[0002] In the field of industrial manufacturing, surface roughness can effectively reflect the micro-geometric morphology of a workpiece and has certain influence on the wear resistance and electrical conductivity of a product, so that the surface roughness is usually regarded as a key evaluation index of product quality. However, in the actual production process, the surface roughness is usually manually measured by a contact type roughness measuring instrument after machining, and the cutting parameters are adjusted according to the measurement results, so that the production efficiency is reduced, the cost is increased, and the machining quality of the workpiece cannot be timely and accurately improved. Therefore, accurately predicting the surface roughness of the workpiece has important significance for improving the production efficiency, reducing the cost and improving the surface quality of the workpiece.

[0003] At present, the surface roughness prediction method based on data driving has become a research hotspot. This method collects and analyzes the cutting process parameters and sensor signal data, and uses an artificial intelligence algorithm to construct a prediction model to realize the prediction of the surface roughness. However, the existing method has the following problems:

[0004] 1. Feature loss caused by relying on manual feature extraction: in the feature processing process, most of the current methods use Pearson correlation coefficients to select features from the time domain, frequency domain and time-frequency domain of the signal, and do not fully utilize all the information of the original signal sequence, which easily causes feature loss and reduces the prediction accuracy of the model.

[0005] 2. Single traditional model for predicting the effect of cutting process parameters: the current method usually uses a single traditional machine learning model or a simple neural network to predict the surface roughness, and the data used is limited to static factors such as feed speed and spindle speed, and dynamic factors such as tool wear and machining vibration are ignored, which limits the prediction effect. SUMMARY

[0006] To solve the above problems and difficulties, the application provides an end-to-end surface roughness prediction method driven by multi-source information fusion.

[0007] The technical scheme adopted by the application to achieve the above purpose is:

[0008] An end-to-end surface roughness prediction method driven by multi-source information fusion, comprising the following steps:

[0009] 1) Obtain dynamic sensor data and static machining process parameters of multiple sources, and pre-process them;

[0010] 2) using a multi-scale convolutional neural network to extract the pre-processed dynamic data features to obtain a feature vector;

[0011] 3) performing feature enhancement on the feature vector through a mixed space and channel attention mechanism;

[0012] 4) performing feature representation on the static parameters through a static information encoding network;

[0013] 5) fitting the dynamic features and static features through a fully connected layer to obtain a surface roughness prediction result.

[0014] The multi-scale convolutional neural network includes four parallel branches, wherein the first three branches have the same structure, each including a one-dimensional convolution layer, a batch normalization layer and a Relu activation function layer connected in sequence, and the fourth branch includes a maximum pooling layer, a one-dimensional convolution layer, a batch normalization layer and a Relu activation function layer connected in sequence. The outputs of the four branches are added to serve as the output of the multi-scale convolutional neural network.

[0015] The mixed space and channel attention mechanism includes a channel attention module and a space attention module, wherein:

[0016] The channel attention module obtains the average features and prominent features of each channel by sequentially passing the feature map through an average pooling layer and a maximum pooling layer, and inputs the average features and prominent features into a shared multilayer perceptron to obtain a channel attention map.

[0017] The space attention module performs two pooling operations on the feature map along the channel axis to obtain average pooling features and maximum pooling features of the entire channel, and extracts features through a convolution layer to obtain a space attention map.

[0018] The step 3) is specifically:

[0019] F CAM =MLP(AvgPool(F))+MLP(MaxPool(F))

[0020] F SAM =Conv(Concat[AvgPool(F),MaxPool(F)])

[0021] [M C (F),M S (F)]=MLP CS (Concat[F CAM ,F SAM ])

[0022]

[0023] wherein F is an input feature, MLP is a multi-layer perceptron, AvgPool(·) represents an average pooling operation, MaxPool(·) represents a maximum pooling operation, Conv(·) represents a one-dimensional convolution operation, Concat represents a feature concatenation operation, sigma represents a Sigmoid function, F CAM is a channel attention map, F SAM is a spatial attention map, M C 9F) is a channel attention weight map, M S (F) is a spatial attention weight map.

[0024] The static information encoding network is a three-layer convolutional layer, and a batch normalization layer is added after each convolutional layer.

[0025] A multi-source information fusion driven end-to-end surface roughness prediction system comprises the following steps:

[0026] A data acquisition module is configured to acquire multi-source heterogeneous dynamic sensor data and static machining process parameters, and to preprocess the same.

[0027] A multi-scale convolutional neural network module is configured to extract features of the preprocessed dynamic data by using a multi-scale convolutional neural network to obtain a feature vector.

[0028] A hybrid attention mechanism module is configured to enhance the feature vector by using a hybrid spatial and channel attention mechanism.

[0029] A static information encoding network module is configured to represent static parameters by using a static information encoding network.

[0030] A fully connected layer module is configured to fit the dynamic features and the static features by using a fully connected layer to obtain a surface roughness prediction result.

[0031] A multi-source information fusion driven end-to-end surface roughness prediction device comprises a memory and a processor; the memory is configured to store a computer program; the processor is configured to implement the multi-source information fusion driven end-to-end surface roughness prediction method when the computer program is executed.

[0032] A computer readable storage medium stores a computer program thereon, and the computer program, when executed by a processor, implements the multi-source information fusion driven end-to-end surface roughness prediction method.

[0033] The present application has the following advantages and benefits:

[0034] 1. High accuracy prediction: the method of the present application uses convolutional neural network for feature extraction, introduces hybrid attention mechanism for feature enhancement, and further improves the prediction accuracy. It is superior to the method based on single sensor signal or traditional machine learning model method.

[0035] 2. Multi-source information fusion: the method of the present application uses feed speed, spindle speed, cutting depth, etc. as static factors, vibration signal, cutting force signal and current signal as dynamic factors, and comprehensively considers the influencing factors of surface roughness. Through the fusion of dynamic information and static machining process parameters, the problem that the prediction accuracy is limited due to considering only single influencing factor in part of the modeling method is solved.

[0036] 3. Comprehensive feature extraction: the method of the present application uses parallel multi-scale convolutional neural network to directly mine feature information from original signal, reduces the dependence on professional knowledge, and does not need manual feature extraction and selection. It can better capture the change rule of tool wear and improve the prediction accuracy.

[0037] 4. Strong practicability: the method of the present application is based on deep learning algorithm, and has high practicability and operability. For surface roughness prediction in industrial manufacturing field, it provides a simple, efficient and reliable solution, and provides an effective way for fast and accurate prediction of product surface quality in intelligent manufacturing environment, further improving production efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is the module flow chart of the method of the present application;

[0039] Figure 2 is the schematic diagram of data processing part;

[0040] Figure 3 is the MSCNN-CSAM model structure diagram;

[0041] Figure 4 is the hybrid attention module structure diagram. DETAILED DESCRIPTION

[0042] The present application will be further described in detail below in combination with the drawings and examples.

[0043] As shown in Figure 1 , the method of the present application comprises the following steps:

[0044] 1. First, data acquisition and preprocessing, collect multi-source heterogeneous dynamic sensor data including vibration signal, cutting force signal and motor current signal, and static machining process parameters such as spindle speed, feed speed and cutting depth through designing reasonable machining experiment, then perform effective partial interception, expansion, normalization and other preprocessing on these original data to obtain available one-dimensional sequence signal.

[0045] For signal sequence data, since the collected signal sequence includes the part where the cutter does not contact the workpiece at the beginning of the milling process, the useful data of each group is intercepted by calculating the root mean square value of the window signal using the threshold method, and since the signals are collected at the same time, the effective data position of the x-axis direction vibration data can be determined, and the corresponding part of other signal data is intercepted. Due to the limitations of processing environment and processing cost, in the case of limited amount of available training data, in order to carry out high-precision model training, the method uses window method to intercept signal sequence for data set expansion. By setting a reasonable small size time window, multiple local subsequences can be cut from the same signal sequence, these subsequences share the same surface roughness target value and processing parameters, thereby expanding the sample size of training, the data processing part is shown in Figure 2 .

[0046] 2、As shown in Figure 3 , the preprocessed effective dynamic signal data and static process data are integrated into a data set available for the model, 80% is randomly selected as the training set, and 20% is randomly selected as the test set. The dynamic signal data is input into the parallel multi-scale convolutional neural network MSCNN to extract signal features. The method includes three parallel multi-scale convolutional neural networks, each of which is composed of four convolution branches, three of which use convolution kernels of different sizes, [4, 6, 10], [15, 30, 45] and [50, 70, 90], respectively. The same step size is set in each network branch, the number of convolution kernels is 6, and the same length of each branch output feature is ensured by setting different padding. The number of output feature maps of each parallel network is 18. The first convolution branch first performs maximum pooling to match the feature dimension with the other three branches, uses a 1×1 convolution kernel and an activation function, and extracts global features and introduces nonlinear transformation. The different scale features obtained by the first three branches are connected and added to the global features obtained by the fourth branch. Batch normalization operation is used after the convolution layer to speed up the convergence speed of the training process, and maximum pooling operation is added after the convolution network to reduce the dimension and reduce the memory. Then the pooling layer improves the expression ability of the features and reduces the dimension.

[0047] 3、As shown in Figure 4As shown, the feature vector output by the MSCNN is passed through a hybrid attention module CSAM, which is composed of two sub-modules: a channel attention module (CAM) and a spatial attention module (SAM). The channel attention module passes the feature map through an average pooling layer and a maximum pooling layer to obtain the average feature and prominent feature of each channel, inputs them into a shared multilayer perceptron (MLP), and obtains a channel attention map. The spatial attention performs two pooling operations on the feature information along the channel axis to obtain the average pooling feature and the maximum pooling feature of the entire channel, and then extracts the features through a convolution layer. This method mixes spatial attention and channel attention. First, the channel attention map and the spatial attention map are calculated in parallel, then the correlation between the spatial features and the channel features is established through an MLP, and finally the spatial attention and the channel attention are divided into two parts again, the attention weight is calculated using a Sigmoid function, and an attention weight map is obtained. The original input feature map is multiplied element by element with the generated attention weight map to realize the weighting of the original feature map, comprehensively capture the global information and local information of the input feature, fuse the features from different sensors from multiple angles, and realize feature enhancement.

[0048] 4. The static cutting parameters affecting the surface roughness are input into a convolutional coding network for feature representation. First, the discrete static cutting parameters are input into a multi-layer convolutional neural network, and a Relu activation function is connected after each convolutional layer. Then, the correlation between the parameters is captured by using a fully connected layer to provide auxiliary information for subsequent roughness prediction.

[0049] 5. Finally, the fully connected regression prediction layer, the fully connected layer is the last part of the model, the output of the hybrid attention mechanism is dimensionally reduced through the pooling layer, and a Dropout layer is added to prevent overfitting, the signal features and the static features are spliced and input into the fully connected layer for surface roughness regression prediction.

[0050] The surface roughness prediction method of the application realizes an end-to-end prediction model, without manually extracting features based on prior knowledge, improves the comprehensive acquisition ability of signal sequence features, and considers the synergistic complementarity of multiple information sources, comprehensively considers the influencing factors of surface roughness, and further improves the prediction performance and accuracy of surface roughness.

[0051] Through the comprehensive application of the above processing techniques, the surface roughness prediction method of the application can effectively utilize dynamic signal data and static process parameters for prediction, and by virtue of the attention fusion mechanism and effective feature extraction strategy, good prediction results are achieved in the end-to-end prediction task of surface roughness, laying a solid foundation for realizing intelligent detection of machining quality.

Claims

1. A multi-source information fusion driven end-to-end surface roughness prediction method, characterized in that, The method comprises the following steps: 1) acquiring multi-source heterogeneous dynamic sensor data and static process parameters, and preprocessing the same; 2) extracting features of the preprocessed dynamic data by using a multi-scale convolutional neural network to obtain a feature vector; 3) performing feature enhancement on the feature vector by using a mixed spatial and channel attention mechanism; 4) performing feature representation on the static parameters by using a static information encoding network; 5) fitting the dynamic features and the static features in a full connection layer to obtain a surface roughness prediction result.

2. The multi-source information fusion driven end-to-end surface roughness prediction method according to claim 1, characterized in that, The multi-scale convolutional neural network comprises four parallel branches, wherein the first three branches have the same structure, each comprising a one-dimensional convolutional layer, a batch normalization layer and a Relu activation function layer connected in sequence, and the fourth branch comprises a max-pooling layer, a one-dimensional convolutional layer, a batch normalization layer and a Relu activation function layer connected in sequence, and the outputs of the four branches are added together as the output of the multi-scale convolutional neural network.

3. The multi-source information fusion driven end-to-end surface roughness prediction method according to claim 1, characterized in that, The mixed spatial and channel attention mechanism comprises a channel attention module and a spatial attention module, wherein: The channel attention module obtains average features and prominent features of each channel by sequentially passing the feature map through an average pooling layer and a max-pooling layer, and inputs the average features and the prominent features into a shared multilayer perceptron to obtain a channel attention map; The spatial attention module performs two pooling operations on the feature map along the channel axis to obtain average-pooling features and max-pooling features of the entire channel, and extracts features through a convolutional layer to obtain a spatial attention map.

4. The multi-source information fusion driven end-to-end surface roughness prediction method according to claim 1, characterized in that, The step 3) is specifically: F CAM = MLP(AvgPool(F)) + MLP(MaxPool(F)) F SAM = Conv(Concat[AvgPool(F), MaxPool(F)]) [M C (F),M S (F)]=MLP CS (Concat[F CAM ,F SAM ]) where F is the input feature, MLP is a multi-layer perceptron, AvgPool(·) denotes an average pooling operation, MaxPool(·) denotes a max pooling operation, Conv(·) denotes a one-dimensional convolution operation, Concat denotes a feature concatenation operation, σ represents a Sigmoid function, F CAM is a channel attention map, F SAM is a spatial attention map, M C (F) is a channel attention weight map, M S (F) is a spatial attention weight map.

5. The multi-source information fusion driven end-to-end surface roughness prediction method according to claim 1, characterized in that, The static information encoding network is a three-layer convolutional layer, and a batch normalization layer is added after each convolutional layer.

6. A multi-source information fusion driven end-to-end surface roughness prediction system, characterized in that, The method comprises the following steps: A data acquisition module is configured to acquire multi-source heterogeneous dynamic sensor data and static process parameters, and to preprocess the same; A multi-scale convolutional neural network module is configured to extract features of the preprocessed dynamic data by using a multi-scale convolutional neural network to obtain a feature vector; A mixed attention mechanism module is configured to perform feature enhancement on the feature vector by using a mixed spatial and channel attention mechanism; A static information encoding network module is configured to perform feature representation on the static parameters by using a static information encoding network; A full connection layer module is configured to fit the dynamic features and the static features in a full connection layer to obtain a surface roughness prediction result.

7. A multi-source information fusion driven end-to-end surface roughness prediction apparatus, characterized in that, The memory is configured to store a computer program, and the processor is configured to implement the method according to any one of claims 1-5 when executing the computer program.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, and the computer program is configured to implement the method according to any one of claims 1-5 when executed by a processor.

Images