Milling surface roughness prediction method based on heterogeneous temperature data fusion

Abstract

The application belongs to the technical field of numerical control machining, and particularly relates to a milling surface roughness prediction method based on heterogeneous temperature data fusion. The roughness prediction model used by the method comprises a 1D multi-scale residual convolutional neural network, a bidirectional long short-term memory neural network, a 2D multi-scale residual convolutional neural network, a normalization layer, a projection layer and a KAN network; the pretreated maximum temperature sequence and average temperature sequence are subjected to the 1D multi-scale residual convolutional neural network to obtain maximum temperature sequence features and average temperature sequence features; the maximum temperature sequence features and average temperature sequence features are subjected to the bidirectional long short-term memory neural network to obtain maximum temperature long short-term dependence features and average temperature long short-term dependence features; the pretreated thermal imaging diagram is subjected to the 2D multi-scale residual convolutional neural network to obtain a thermal imaging feature map; the maximum temperature long short-term dependence features, the average temperature long short-term dependence features, the thermal imaging feature map, process parameters and tool profile degree are subjected to normalization, projection and splicing, and then are mapped by the KAN network to obtain surface roughness. The method solves the problems of insufficient spatial information of the temperature sequence and time continuity of the thermal imaging diagram, provides more comprehensive temperature information, and improves the prediction accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of numerical control machining, and particularly relates to a milling surface roughness prediction method based on heterogeneous temperature data fusion. BACKGROUND

[0002] Milling is an important numerical control machining method, which is widely used in various product machining due to its high efficiency and productivity. In milling, surface roughness is one of the most common and important indicators to characterize the surface quality of a workpiece. Therefore, accurate prediction of surface roughness is of great significance for optimizing machining parameters, improving production efficiency and ensuring product quality. Milling temperature refers to the temperature generated by friction between the tool and the workpiece during milling. It has an impact on the machining process and results, and is directly related to the surface quality of the workpiece, machining accuracy, tool life and machining efficiency. In particular, the temperature change in the milling area directly affects the surface roughness of the workpiece, so the temperature can be used to predict the surface roughness.

[0003] Traditional surface roughness prediction mainly relies on empirical formulas and regression analysis, which cannot accurately reflect the dynamic changes and nonlinear relationships in the actual machining process. When the working conditions change, it may not be able to make predictions, so there are problems such as poor robustness and poor universality. With the improvement of computing power and the progress of machine learning technology, researchers have begun to explore the use of data-driven methods to improve surface roughness prediction, especially in recent years, deep learning technology has shown excellent ability in handling complex pattern recognition tasks, providing a new solution for surface roughness prediction. However, existing deep learning-based surface roughness prediction methods often rely on single-source temperature data, which cannot fully reflect the complex effects of temperature on the machining process. Therefore, the present application proposes a milling surface roughness prediction method based on heterogeneous temperature data fusion, which realizes accurate prediction of surface roughness by fusing heterogeneous temperature data and combining process parameters. SUMMARY

[0004] To solve the technical problems of the prior art, the present application provides a milling surface roughness prediction method based on heterogeneous temperature data fusion.

[0005] The technical problem solved by the present application adopts the following technical solution:

[0006] A milling surface roughness prediction method based on heterogeneous temperature data fusion, characterized in that the method comprises the following steps:

[0007] Step 1: Milling the workpiece under different process parameters, collecting thermal imaging, maximum temperature sequence and average temperature sequence during the milling process; after milling, measure the surface roughness of the workpiece;

[0008] Step two: preprocess the thermal image, maximum temperature sequence and average temperature sequence, and construct a data set;

[0009] Step three: construct a roughness prediction model, including a 1D multi-scale residual convolutional neural network, a bidirectional long short-term memory neural network, a 2D multi-scale residual convolutional neural network, a normalization layer, a projection layer and a KAN network;

[0010] The preprocessed maximum temperature sequence and average temperature sequence are input into the 1D multi-scale residual convolutional neural network to obtain maximum temperature sequence features and average temperature sequence features; the maximum temperature sequence features and the average temperature sequence features are input into the bidirectional long short-term memory neural network to obtain maximum temperature long short-term dependence features and average temperature long short-term dependence features; the preprocessed thermal image is input into the 2D multi-scale residual convolutional neural network to obtain a thermal image feature map; the maximum temperature long short-term dependence features, the average temperature long short-term dependence features, the thermal image feature map, the process parameters and the tool profile are normalized and then projected to obtain maximum temperature projection features, average temperature projection features, thermal image projection features, process projection features and tool profile projection features; all the projection features are spliced and then input into the KAN network to obtain the surface roughness;

[0011] Step four: use the data set to train the roughness prediction model, and use the trained roughness prediction model for surface roughness prediction in milling.

[0012] Further, the 1D multi-scale residual convolutional neural network comprises a multi-scale one-dimensional convolution module, the output features of the current one-dimensional convolution module are connected with the input features to obtain one-dimensional residual features, and the one-dimensional residual features are input into the next one-dimensional convolution module after a pooling operation; the input features of the one-dimensional convolution module are sequentially input into a one-dimensional convolution, batch normalization and activation operation to obtain the output features of the module;

[0013] The 2D multi-scale residual convolutional neural network comprises a multi-scale two-dimensional convolution module, the output features of the current two-dimensional convolution module are connected with the input features to obtain two-dimensional residual features, and the two-dimensional residual features are input into the next two-dimensional convolution module after a pooling operation; the input features of the two-dimensional convolution module are sequentially input into a two-dimensional convolution, batch normalization and activation operation to obtain the output features of the module.

[0014] Further, the maximum temperature sequence refers to a time sequence formed by the maximum temperature in the machining area at each sampling time, and the average temperature sequence refers to a time sequence formed by the average temperature in the machining area at each sampling time.

[0015] Further, the preprocessing of the thermal imaging image includes denoising, alignment, cropping, graying and normalization processing, and the preprocessing of the maximum temperature sequence and the average temperature sequence includes padding and normalization.

[0016] Compared with the prior art, the present application has the following beneficial effects:

[0017] (1) The temperature in the milling process is an important factor affecting the surface roughness of the product, and the traditional surface roughness prediction method often relies on single source temperature data, only focusing on the time domain characteristics or spatial characteristics of the temperature data. The temperature sequence reflects the dynamic change of the temperature in the time domain, but lacks spatial information, and the thermal imaging image reflects the spatial distribution of the temperature, but lacks time continuity, therefore the present application fuses the two heterogeneous temperature sequences and thermal imaging images, which complement each other to solve the problems of insufficient spatial information of the temperature sequence and time continuity of the thermal imaging image, and provide more comprehensive temperature information, thereby improving the prediction accuracy. The fusion of heterogeneous temperature data improves the robustness and stability of the model, reduces the noise interference of a single data source, and can alleviate the decline in prediction accuracy caused by sensor errors or data quality problems, and can also enhance the generalization ability of the model, so that it can adapt to different working conditions and environments, and improve the universality of the application.

[0018] (2) The roughness prediction model uses a 1D multi-scale residual convolutional neural network to capture temperature sequence features at different scales, and the temperature sequence features are passed through a bidirectional long short-term memory neural network to use past and future information of the temperature sequence to better understand the relationship between the current time information and the past and future information, and to capture the long and short term dependencies of the temperature sequence, thereby obtaining more comprehensive temperature sequence features; a 2D multi-scale residual convolutional neural network is used to extract features of the thermal imaging image, and finally a KAN network is used to map the feature dimension to the predicted surface roughness, the whole process realizes the nonlinear mapping between the heterogeneous temperature data and the surface roughness, and realizes the accurate prediction of the surface roughness. The 1D multi-scale residual convolutional neural network and the 2D multi-scale residual convolutional neural network avoid the gradient vanishing problem through residual connection, improve the training efficiency and performance of the model. The KAN network fuses and predicts different modal data, and the KAN network can handle complex nonlinear relationships and provide a smooth activation function, thereby more accurately capturing the subtle and complex changes of multi-modal data in the processing process, and improving the prediction accuracy. The adaptive network and regularization mechanism of KAN enhance the generalization ability of the model, avoid overfitting, and improve the robustness, and have unique advantages in handling continuous data and complex feature interactions.

[0019] (3) In order to obtain a rich data set, an optimal Latin hypercube design milling experiment is adopted to ensure that the process parameters are uniformly distributed, so that more rich thermal images, maximum temperature sequences and average temperature sequences under different process parameters can be obtained, and the model can learn more information, which helps to improve the prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0020] Fig. 1 is the overall flowchart of the present application;

[0021] Fig. 2 is a structural schematic diagram of the roughness prediction model of the present application;

[0022] Fig. 3 is a roughness prediction result graph of the milling surface of the present application. DETAILED DESCRIPTION

[0023] The specific embodiments are described below in conjunction with the drawings, which are only used to specifically introduce the technical solutions of the present application, and do not limit the protection scope of the present application.

[0024] The present application provides a kind of milling surface roughness prediction method based on heterogeneous temperature data fusion (referred to as method, see Figs. 1-3 ), comprising the following steps:

[0025] Step one: under different process parameters, workpiece is milled, and thermal imaging, maximum temperature sequence and average temperature sequence in the milling process are collected by using thermal imager;After milling, the surface roughness of the workpiece is measured by using 3D topography scanner as label;

[0026] In order to ensure that the parameters are uniformly distributed, the main process parameters affecting the surface roughness, including feed speed, cutting depth, spindle speed, etc. are used as variables, and the optimal Latin hypercube (Optimal Latin Hypercube) method is used to design milling experiment. The size of the workpiece in this embodiment is 20x70x70mm, the workpiece surface is initially milled to ensure that the actual cutting depth when collecting data is the cutting depth designed in the experiment, and the surface of the workpiece 70x70mm is used as the processing surface. The processing surface is evenly divided into multiple processing areas, and each processing area is milled with different process parameters and consistent milling path.

[0027] Multiple sampling points are uniformly arranged in the processing area, the maximum temperature sequence refers to the time sequence formed by the maximum temperature in the processing area at each sampling time, and the average temperature sequence refers to the time sequence formed by the average temperature in the processing area at each sampling time.

[0028] Step two: pre-process thermal imaging, maximum temperature sequence and average temperature sequence, and construct data set;

[0029] Preprocessing of thermal imaging images: first, denoising the thermal imaging images to remove random noise in the thermal imaging images and improve image quality; then, aligning and cropping the denoised thermal imaging images according to the position and size of the processing area to ensure that all thermal imaging images maintain consistent positional relationships, facilitating the capture of temperature change trends during processing; performing grayscale processing on the cropped thermal imaging images to reduce the consumption of computing resources while improving the stability and generalization ability of the model; and finally, performing normalization processing on the grayscale processed thermal imaging images to unify the pixels to the same range.

[0030] The preprocessing of the maximum temperature sequence and the average temperature sequence includes padding and normalization; since the feed rates of different process parameters are different, resulting in different lengths of temperature sequences, Newton's cooling law is used to pad the temperature sequences to make the lengths of all temperature sequences consistent; Newton's cooling law describes the heat transfer rate between an object and its surroundings, and the law followed by an object that is gradually cooling by transferring heat to the surrounding medium when the temperature of the object is higher than that of the surroundings; when there is a temperature difference between the surface of the object and the surrounding environment, the heat dissipated per unit area per unit time is proportional to the temperature difference, and the mathematical expression is:

[0031]

[0032] In the formula, T and T env are the temperatures of the object and the surrounding environment at time t, respectively; k is the cooling coefficient, which depends on the physical properties of the object and its contact conditions with the environment.

[0033] Step three: constructing a roughness prediction model, including a 1D multi-scale residual convolutional neural network, a bidirectional long short-term memory neural network, a 2D multi-scale residual convolutional neural network, a normalization layer, a projection layer, and a KAN network;

[0034] The preprocessed maximum temperature sequence and the average temperature sequence are extracted by a 1D multi-scale residual convolutional neural network to obtain maximum temperature sequence features and average temperature sequence features; the maximum temperature sequence features and the average temperature sequence features are captured by a bidirectional long short-term memory neural network to capture bidirectional long-distance dependency relationships, obtaining maximum temperature long short-term dependency features and average temperature long short-term dependency features; the preprocessed thermal imaging images are extracted by a 2D multi-scale residual convolutional neural network to obtain thermal imaging feature maps; after the maximum temperature long short-term dependency features, the average temperature long short-term dependency features, the thermal imaging feature maps, the process parameters, and the tool profile are processed by the normalization layer, they are projected by the projection layer to map all features into the same space, obtaining maximum temperature projection features, average temperature projection features, thermal imaging projection features, process projection features, and tool profile projection features; after these projection features are spliced, the surface roughness is obtained by the KAN network.

[0035] The 1D multi-scale residual convolutional neural network comprises a multi-scale one-dimensional convolutional module, the output feature of the current one-dimensional convolutional module is connected with the input feature in residual to obtain a one-dimensional residual feature, and the one-dimensional residual feature is subjected to a pooling operation to obtain the input feature of the next one-dimensional convolutional module; the one-dimensional convolutional module adopts one-dimensional convolution operation to extract features of the input, the extracted features are subjected to batch normalization operation to ensure the stability of the training process, and the features subjected to batch normalization are subjected to ReLU activation operation to introduce nonlinearity to enhance the expression ability of the model; the multi-scale convolutional module uses different sizes of convolution kernels to extract multi-scale features of the temperature sequence, the size of the convolution kernel determines the size of the receptive field of the convolution operation, and a smaller convolution kernel can capture local details, and a larger convolution kernel has a larger receptive field and can capture dependent information in a larger range. The expression of the one-dimensional convolutional module is:

[0036] yi = ReLU(BatchNorm(Conv1d(xi))) (2)

[0037] In the formula, x i , y i represent the input and output features of the i-th convolutional module, ReLU(·) represents the ReLU activation operation, BatchNorm(·) represents the batch normalization operation, and Conv1d(·) represents the one-dimensional convolution operation.

[0038] The residual connection and the pooling operation are represented as:

[0039] xi +1 = Pool(Conv1d(xi)+yi) (3)

[0040] In the formula, x i+1 represents the input feature of the i+1-th convolutional module, and Pool(·) represents the pooling operation.

[0041] The bidirectional long short-term memory neural network is obtained by extending the long short-term memory neural network (LSTM), and simultaneously captures information from the forward and backward directions, so that the model can fully utilize the bidirectional time dependence in the sequence. The core of the LSTM is the memory cell state and the gating mechanism, including the forgetting gate, the input gate and the output gate, and the gating mechanism uses the Sigmoid activation function to control the degree of information flow; the processing process of the LSTM is represented as:

[0042] The forgetting gate is used to control the information to be forgotten, and is represented as:

[0043] f t = σ(W f ·[h t-1 ,x t ]+bf ) (4)

[0044] where f t denotes the output of the forget gate at time step t, σ denotes the Sigmoid activation function, W f and b f denote the weight matrix and bias term of the forget gate at time step t, h t-1 denotes the hidden state at time step t-1, x t denotes the input at time step t.

[0045] The input gate is used to control how much information of the current input will be written into the memory cell, denoted as:

[0046] i t = σ(W i · [h t-1 , x t ] + b i ) (5)

[0047] where i t denotes the output of the input gate at time step t, W i and b i denote the weight matrix and bias term of the input gate at time step t.

[0048] The candidate memory cell state is generated as:

[0049]

[0050] where C denotes the candidate memory state at time step t, tanh(·) denotes the tanh activation function, W C and b C denote the weight matrix and bias term of the candidate memory.

[0051] The memory cell state is updated according to the candidate memory cell state and the memory cell state at the previous time step as:

[0052]

[0053] where C t and C t-1 denote the memory cell states at time steps t and t-1, * denotes the convolution operation.

[0054] The output gate controls the output at the current time step as:

[0055] o t = σ(W o · [h t-1 , x t ] + b o ) (8)

[0056] where o t represents the output of the t time step output gate, W o , b o represents the weight matrix and bias term of the output gate.

[0057] Hidden state:

[0058] h t = o t ·tanh(C t ) (9)

[0059] where h t represents the hidden state of the t time step, which contains the long and short term memory information of the current time step, and is subjected to tanh activation to compress the value to the range of (-1, 1) to control the flow of information.

[0060] The output feature of the bidirectional LSTM is obtained by concatenating the hidden states of the two LSTMs, that is:

[0061] h t bi = Concate(h t forward , h t backward ) (10)

[0062] where htbi represents the output feature of the bidirectional LSTM, Concate(·) represents the concatenation operation, h t forward is the hidden state of the LSTM processed from front to back, and h t backward is the hidden state of the LSTM processed from back to front.

[0063] The 2D multi-scale residual convolutional neural network comprises a multi-scale two-dimensional convolution module, the output feature of the current two-dimensional convolution module is connected in residual with the input feature to obtain a two-dimensional residual feature, and the two-dimensional residual feature is subjected to a pooling operation to obtain the input feature of the next two-dimensional convolution module; the two-dimensional convolution module adopts a two-dimensional convolution operation to extract features from the input, and the extracted features are subjected to batch normalization and ReLU activation operation to obtain the output feature of the two-dimensional convolution module.

[0064] Since the traditional multi-layer perceptron often has difficulty in dealing with high-dimensional data, the KAN network changes the fixed activation function to a learnable activation function, so that each weight parameter in the KAN network can be replaced by a univariate function, and these univariate functions are parameterized in the form of a spline function, thereby providing extremely high flexibility and using fewer parameters to simulate complex functions, thereby enhancing the interpretability of the model.

[0065] Step four: using the data set to train the roughness prediction model, and using the trained roughness prediction model for milling surface roughness prediction.

[0066] Fig. 3 The prediction results of the method of the application are compared in the graph, wherein the mean square error MSE is 0.026, the mean absolute error MAE is 0.0498, and the determination coefficient R 2 is 0.9636. The mean square error MSE of the prediction result using only the temperature sequence (including the maximum and average temperature sequence) is 0.0785, the mean absolute error MAE is 0.1430, and the determination coefficient R 2 is 0.8366. The mean square error MSE of the prediction result using only the thermal imaging graph is 0.1014, the mean absolute error MAE is 0.2259, and the determination coefficient R 2 is 0.8074. It can be seen that the mean square error, the mean absolute error and the determination coefficient of the application are significantly better than the prediction results using only the temperature sequence and the thermal imaging graph, because the heterogeneous data fusion of the temperature sequence and the thermal imaging graph solves the problem of insufficient spatial information of the temperature sequence and the time continuity of the thermal imaging graph, and can provide more comprehensive temperature information, thereby improving the prediction accuracy, and enhancing the robustness and generalization ability of the model.

[0067] The unmentioned parts of the application are applicable to the prior art.

Claims

1. A method for predicting surface roughness of milling based on heterogeneous temperature data fusion, characterized in that, The method comprises the following steps: Step one: milling the workpiece under different process parameters, collecting thermal images, maximum temperature sequences and average temperature sequences during the milling process; after milling, measuring the surface roughness of the workpiece; Step two: preprocessing the thermal images, maximum temperature sequences and average temperature sequences, and constructing a data set; Step three: constructing a roughness prediction model, including a 1D multi-scale residual convolutional neural network, a bidirectional long short-term memory neural network, a 2D multi-scale residual convolutional neural network, a normalization layer, a projection layer and a KAN network; The preprocessed maximum temperature sequences and average temperature sequences are input into the 1D multi-scale residual convolutional neural network to obtain maximum temperature sequence features and average temperature sequence features; the maximum temperature sequence features and the average temperature sequence features are input into the bidirectional long short-term memory neural network to obtain maximum temperature long short-term dependence features and average temperature long short-term dependence features; the preprocessed thermal images are input into the 2D multi-scale residual convolutional neural network to obtain thermal image feature maps; the maximum temperature long short-term dependence features, the average temperature long short-term dependence features, the thermal image feature maps, the process parameters and the tool profile are normalized and then projected to obtain maximum temperature projection features, average temperature projection features, thermal image projection features, process projection features and tool profile projection features; all the projection features are spliced and then input into the KAN network to obtain the surface roughness; Step four: using the data set to train the roughness prediction model, and using the trained roughness prediction model for milling surface roughness prediction.

2. The method of claim 1, wherein the method is based on a heterogeneous temperature data fusion. The 1D multi-scale residual convolutional neural network comprises a multi-scale one-dimensional convolution module, the output features of the current one-dimensional convolution module are connected with the input features to obtain one-dimensional residual features, and the one-dimensional residual features are input into the next one-dimensional convolution module after a pooling operation; The input features of the one-dimensional convolution module are sequentially input into a one-dimensional convolution, batch normalization and activation operation to obtain the output features of the module; The 2D multi-scale residual convolutional neural network comprises a multi-scale two-dimensional convolution module, the output features of the current two-dimensional convolution module are connected with the input features to obtain two-dimensional residual features, and the two-dimensional residual features are input into the next two-dimensional convolution module after a pooling operation; the input features of the two-dimensional convolution module are sequentially input into a two-dimensional convolution, batch normalization and activation operation to obtain the output features of the module.

3. The method of claim 1 or 2, wherein the method is based on a heterogeneous temperature data fusion. The maximum temperature sequence refers to a time sequence formed by the maximum temperature in the machining area at each sampling time, and the average temperature sequence refers to a time sequence formed by the average temperature in the machining area at each sampling time.

4. The method of claim 1, wherein the method is based on a heterogeneous temperature data fusion. The preprocessing of the thermal images includes denoising, alignment, cropping, grayscale and normalization processing, and the preprocessing of the maximum temperature sequences and the average temperature sequences includes padding and normalization.

Images