Environment-friendly pipeline mold preparation method

Abstract

The present application relates to the technical field of mold preparation, and particularly relates to a green and environment-friendly pipeline mold preparation method, which comprises the following steps: extracting machining traces to generate machining local image features by means of an image processing mapping algorithm to extract a visual image of a machining surface of a pipeline mold workpiece, and extracting reflectivity differences of multiple wave bands to generate machining micro-spectrum features by means of a calibration model to extract a spectral image; constructing a mapping learning model based on an improved multi-modal CNN model architecture and having a visual input branch and a spectral input branch, inputting the machining local image features into the visual input branch, inputting the machining micro-spectrum features into the spectral input branch, and generating a machining surface roughness; and adjusting machining parameters of the machining surface according to a comparison result of the machining surface roughness and a threshold value. The present application realizes the optimized prediction of the surface roughness, thereby reducing invalid machining actions, and realizes the green and environment-friendly preparation process of power saving, pollution reduction and improvement of the pipeline mold forming qualification rate.

Description

Technical Field

[0001] This invention relates to the field of mold preparation technology, and in particular to a green and environmentally friendly method for preparing pipe molds. Background Technology

[0002] In the green machining process of high-speed milling of pipe molds, the surface roughness of the pipe mold workpiece is usually monitored by multi-source data. Then, the cutting parameters, such as feed rate and depth of cut, are dynamically adjusted according to the surface roughness to improve the first pass rate, reduce the scrap rate, and thus reduce energy consumption and waste generation.

[0003] For example, the patent with application number 202311199634.X discloses a water-guided laser drilling device and method based on spectral detection and visual detection, including a water-guided laser system, a visual analysis and control system, and a spectral signal analysis and control system. However, it does not use a deep learning-based multi-source data fusion algorithm to perform comprehensive prediction and analysis of the surface roughness of the processed surface.

[0004] Therefore, in the green manufacturing process of pipe molds, how to use deep learning models to predict the surface roughness of the machined surface and control the machining parameters is a technical problem that needs to be solved. Summary of the Invention

[0005] Therefore, this invention provides a green and environmentally friendly method for preparing pipe molds. By setting an improved CNN architecture with a dual-input branch network structure and a cross-modal feature coupling model, the method achieves optimized prediction of surface roughness, thereby reducing ineffective processing actions and realizing a green and environmentally friendly preparation process that saves electricity, reduces pollution, and improves the qualification rate of pipe mold forming.

[0006] To achieve the above objectives, this invention proposes a green and environmentally friendly method for preparing pipe molds, comprising:

[0007] Acquire visual and spectral images of the machined surface of the pipe mold workpiece;

[0008] The visual image is processed by an image processing mapping algorithm to extract processing traces and generate local processing image features. The spectral image is then processed by a calibration model to extract reflectance differences across multiple bands and generate microspectral processing features.

[0009] A mapping learning model based on an improved multimodal CNN model architecture with visual input and spectral input branches is constructed. The local image features of the processing are input into the visual input branch, and the microspectral features of the processing are input into the spectral input branch. The mapping learning model generates the surface roughness of the processing.

[0010] The tool's machining parameters on the machined surface are adjusted based on the comparison between the surface roughness and the threshold.

[0011] Furthermore, the mapping learning model has a feature fusion regression module connected to the visual input branch and the spectral input branch, and the process of generating the surface roughness includes:

[0012] The processed local image features are preprocessed visually and then input into the visual input branch to determine the critical defect area.

[0013] The processed microspectral features are used to determine a multi-band reflectance trend sequence in the spectral input branch through an improved TCN-GRU unit;

[0014] The surface roughness of the processed surface is determined by performing cross-modal fusion feature extraction and regression prediction on the key defect region and the temporal trend of the band sequence reflectance through the feature fusion regression module.

[0015] Furthermore, the spectral input branch has a mapping extraction unit, and the improved TCN-GRU unit includes a TCN subunit, a GRU subunit, and an attention subunit. The process of determining the multi-band reflectance trend sequence through the improved TCN-GRU unit includes:

[0016] After the processed microspectral features are preprocessed, the local correlation features of the band spectrum are determined by the mapping extraction unit.

[0017] The local correlation features of the spectrum in the aforementioned band are used to generate a long-range parameter sequence through a TCN subunit;

[0018] The long-range parameter sequence is used to generate the multi-band reflectance trend sequence as its hidden state through the GRU subunit, and the attention subunit controls the updating of the hidden state by the GRU subunit.

[0019] Furthermore, the hidden state includes the previous hidden state, the current hidden state, and the transitive hidden state. The process of controlling the GRU subunit to update the hidden state through the attention subunit includes:

[0020] The current hidden state and the target vector are used to generate attention weights through an attention subunit;

[0021] The transitive hidden state is generated based on the attention weight, the previous hidden state, and the current hidden state, and then the transitive hidden state is input to the GRU subunit for iterative calculation.

[0022] Furthermore, the mapping extraction unit has a fully connected layer and a one-dimensional convolutional layer, and the process of determining the local correlation features of the band spectrum includes:

[0023] After the processed microspectral features are preprocessed, the nonlinear relationships between bands are extracted through the fully connected layer to generate correlated band spectra.

[0024] The associated band spectra are processed through the one-dimensional convolutional layer to extract local correlation features between adjacent bands, thereby generating local correlation features of the band spectra.

[0025] Furthermore, the feature fusion regression module includes a feature fusion unit and a regression output unit, and the process of determining the surface roughness of the processed surface includes:

[0026] The key defect region and the temporal trend of reflectance of the band sequence are combined across modal features and multi-scale convolutional fusion through a feature fusion unit to generate visual spectral fusion features.

[0027] The visual spectral fusion features are upsampled and activated function feature mapping is performed on the regression output unit to determine the surface roughness of the processed surface.

[0028] Furthermore, the visual input branch is sequentially configured with a first two-dimensional convolutional layer, a max pooling layer, a residual connection layer, and a second two-dimensional convolutional layer. The first two-dimensional convolutional layer is used to extract macroscopic texture features, the second two-dimensional convolutional layer is used to extract microscopic roughness peak features, and the channel spatial attention mechanism is used to determine key regions of roughness defects.

[0029] Furthermore, the surface roughness is the Ra value distribution of the processed surface, and the process of constructing the mapping learning model includes:

[0030] The luminance regression term is calculated based on the ratio of the product of the true Ra value distribution and the predicted Ra value distribution to the sum of the mean squares.

[0031] Based on the local window covariance, true local variance, and predicted local variance of the true Ra value distribution and the predicted Ra value distribution, calculate the contrast regression term;

[0032] A similarity loss function is constructed based on the brightness regression term and the contrast regression term, and the mapping learning model is optimized and trained using the similarity loss function.

[0033] The above solution achieves breakthroughs in core indicators such as surface quality inspection accuracy, processing stability, and resource utilization efficiency through multimodal data fusion, temporal-spatial joint modeling, and dynamic closed-loop control, providing an intelligent solution for green and environmentally friendly mold manufacturing.

[0034] Furthermore, the visual preprocessing includes histogram equalization to highlight surface defects;

[0035] The calibration model includes differential calculations for key bands.

[0036] Furthermore, the threshold includes a first threshold, a second threshold, and a duration threshold that increase sequentially, and the processing parameters include the cutting speed, feed rate, and depth of cut of the tool on the pipe mold workpiece;

[0037] If the duration of the surface roughness being greater than the first threshold is greater than the duration threshold, then reduce the feed rate and depth of cut.

[0038] If the duration for which the surface roughness is greater than the first threshold is less than the duration threshold, then the cutting speed is reduced.

[0039] If the surface roughness of the machined surface is less than a first threshold and greater than a second threshold, then the cutting speed and feed rate are reduced.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] 1. By setting an improved CNN architecture with a dual-input branch network structure, a cross-modal feature coupling model is used to optimize the prediction of surface roughness, thereby reducing invalid processing actions and realizing a green and environmentally friendly manufacturing process that saves electricity, reduces pollution, and improves the qualification rate of pipe mold forming.

[0042] 2. Through multimodal data fusion, temporal-spatial joint modeling, and dynamic closed-loop control, breakthroughs have been achieved in core indicators such as surface quality inspection accuracy, processing stability, and resource utilization efficiency, providing an intelligent solution for green and environmentally friendly mold manufacturing. Attached Figure Description

[0043] Figure 1 This is a schematic flowchart of the green and environmentally friendly pipe mold preparation method according to an embodiment of the present invention;

[0044] Figure 2 This is a flowchart illustrating the mapping learning model of the green and environmentally friendly pipe mold preparation method according to an embodiment of the present invention.

[0045] Figure 3 This is a schematic diagram of the spectral input branch of the green and environmentally friendly pipe mold preparation method according to an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram of the process for adjusting the processing parameters in the green and environmentally friendly pipe mold preparation method according to an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0048] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0049] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0050] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0051] like Figures 1 to 4 As shown, this invention provides a green and environmentally friendly method for preparing pipe molds. By setting an improved CNN architecture with a dual-input branch network structure and a cross-modal feature coupling model, it achieves optimized prediction of surface roughness, thereby reducing ineffective processing actions and realizing a green and environmentally friendly preparation process that saves electricity, reduces pollution, and improves the qualification rate of pipe mold forming.

[0052] like Figures 1 to 4 As shown, this embodiment proposes a green and environmentally friendly method for preparing pipe molds, including:

[0053] Acquire visual and spectral images of the machined surface of the pipe mold workpiece;

[0054] The visual image is processed by an image processing mapping algorithm to extract processing traces and generate local processing image features. The spectral image is then processed by a calibration model to extract reflectance differences across multiple bands and generate microspectral processing features.

[0055] A mapping learning model based on an improved multimodal CNN model architecture with visual input and spectral input branches is constructed. The local image features of the processing are input into the visual input branch, and the microspectral features of the processing are input into the spectral input branch. The mapping learning model generates the surface roughness of the processing.

[0056] The tool's machining parameters on the machined surface are adjusted based on the comparison between the surface roughness and the threshold.

[0057] It should be noted that the above-mentioned pipe mold preparation method is preferably applied to high-speed milling green machining processes. By integrating a vision system and a multispectral sensor, it captures visual images characterizing the machined surface and spectral images of the micro-morphology of the machined surface in real time. Then, by using a mapping learning model with an improved multimodal CNN model architecture to fuse macroscopic images and micro-morphology, it can predict and monitor the roughness of the machined surface to a certain extent. This allows for dynamic adjustment of machining parameters, including cutting speed, feed rate, and depth of cut, thereby reducing energy consumption, resource waste, waste generation, and environmental pollution, and achieving a green and environmentally friendly pipe mold preparation process.

[0058] It should be noted that the surface roughness is preferably the Ra (arithmetic mean deviation of profile) value distribution of the processed surface. The Ra value distribution reflects the microscopic geometric characteristics by quantifying the height difference of the surface profile deviating from the average line. The larger the Ra value distribution, the greater the overall undulation of the processed surface of the pipe mold. Consequently, the inner and outer surfaces of the pipe formed by injection molding through the pipe mold are more undulating and rougher, leading to unqualified pipe performance. Specifically, the Ra value distribution is closely related to several pipe performance characteristics. For example, an Ra value distribution of less than 0.8 μm can reduce the friction coefficient of the inner surface of the pipe and extend the pipe life. An appropriate Ra value distribution can avoid stress concentration that reduces the fatigue resistance of the pipe material. Therefore, this embodiment achieves accurate prediction and monitoring of the Ra value distribution through a mapping learning model, realizing the green and environmentally friendly performance of the pipe mold itself.

[0059] like Figure 2 As shown, further, the mapping learning model has a feature fusion regression module connected to the visual input branch and the spectral input branch, and the process of generating the surface roughness includes:

[0060] The processed local image features are preprocessed visually and then input into the visual input branch to determine the critical defect area.

[0061] The processed microspectral features are used to determine a multi-band reflectance trend sequence in the spectral input branch through an improved TCN-GRU unit;

[0062] The surface roughness of the processed surface is determined by performing cross-modal fusion feature extraction and regression prediction on the key defect region and the temporal trend of the band sequence reflectance through the feature fusion regression module.

[0063] It is understandable that the visual input branch and the spectral input branch can be used to extract, in parallel, the continuity of the processing texture within the macroscopic critical defect area and the microscopic defects characterized by the multi-band reflectance trend sequence. It is also understandable that the spectral images acquired by the multispectral sensor include multiple reflectance data generated by irradiating the workpiece surface with multiple different bands. The reflectance of each band reflects the absorption and scattering characteristics of the workpiece surface, and thus reflects the microscopic morphology of the workpiece surface. By extracting the reflectance corresponding to multiple sub-bands in the near-infrared band (700-850nm), if the reflectance increases with the increase of the sub-band, and all... If the value is greater than the normal level, it indicates the presence of micro-roughness peaks on the workpiece surface. These micro-roughness peaks lead to an increase in the corresponding surface roughness (Ra value distribution). However, judging the surface roughness using only one of the multispectral or visual sensors can result in conclusions that are too macroscopic or too microscopic. Therefore, this embodiment uses a visual input branch, a spectral input branch, and a feature fusion regression module to achieve a joint prediction and judgment that takes into account both the global surface roughness and the local detailed surface roughness. By fusing the continuity of the processing texture that the visual input branch focuses on and the micro-defects that the spectral input branch focuses on, the false detection rate of surface roughness defects in the pipe mold is reduced.

[0064] like Figure 3 As shown, further, the spectral input branch has a mapping extraction unit, and the improved TCN-GRU unit includes a TCN subunit, a GRU subunit, and an attention subunit. The process of determining the multi-band reflectance trend sequence through the improved TCN-GRU unit includes:

[0065] After the processed microspectral features are preprocessed, the local correlation features of the band spectrum are determined by the mapping extraction unit.

[0066] The local correlation features of the spectrum in the aforementioned band are used to generate a long-range parameter sequence through a TCN subunit;

[0067] The long-range parameter sequence is used to generate the multi-band reflectance trend sequence as its hidden state through the GRU subunit, and the attention subunit controls the updating of the hidden state by the GRU subunit.

[0068] like Figure 3 As shown, further, the hidden state includes the previous hidden state, the current hidden state, and the transitive hidden state. The process of controlling the GRU subunit to update the hidden state through the attention subunit includes:

[0069] Attention weights are generated by the attention subunit based on the current hidden state and the target vector; the transitive hidden state is generated based on the attention weights, the previous hidden state and the current hidden state, and the transitive hidden state is input to the GRU subunit for iterative calculation.

[0070] Specifically, the process of generating attention weights is as follows:

[0071]

[0072] In the formula, Attention t The attention weights are represented by softmax(e). b ) represents the target vector e at time t controlled by the softmax activation function. b Whether to update, h t Let represent the current hidden state, and W represent the weight vector.

[0073] Specifically, the process of generating the transmitted hidden state is as follows:

[0074]

[0075] r t =Swish(W r e t +U r h t-1 +b r )

[0076] In the formula, h t h t-1 h t ′ represent the current hidden state, the previous hidden state, and the passed hidden state, respectively, e t Let W represent the input embedding at time t, where PReLU is the PReLU activation function, Swish is the Swish activation function, and W... h U h W r U r b is a learnable weight vector. h b r All of these are learnable bias terms.

[0077] Specifically, the process by which the GRU subunit generates the multi-band reflectance trend sequence as its hidden state is as follows:

[0078]

[0079] z t =Swish(W z e t +U z ht-1 +b z )

[0080] In the formula, h t h t-1 h t '' represents the current hidden state, the previous hidden state, and the passed hidden state, respectively. t e t r t Let W represent the update gate, the input embedding, and the reset gate at time t, respectively. Let Swish be the Swish activation function. z U z Both are learnable weight vectors, b z All are learnable bias terms. Specifically, the previous hidden state and the transitive hidden state are updated according to the update gate, and then summed to obtain the current hidden state h. t The value of the hidden state is passed through the passing sub-unit for updating, and the reset gate is used to determine whether the input embedding and the previous hidden state need to be updated.

[0081] Furthermore, the mapping extraction unit has a fully connected layer and a one-dimensional convolutional layer, and the process of determining the local correlation features of the band spectrum includes:

[0082] After the processed microspectral features are preprocessed, the nonlinear relationships between bands are extracted through the fully connected layer to generate correlated band spectra.

[0083] The associated band spectra are processed through the one-dimensional convolutional layer to extract local correlation features between adjacent bands, thereby generating local correlation features of the band spectra.

[0084] Specifically, the processed microspectral features are a processed one-dimensional multi-band reflectance data sequence. After mapping the nonlinear relationship of the band reflectance data through a fully connected layer with 256 neurons in the spectral input branch, a 1×256-sized associated band spectrum is generated. The associated band spectrum is then passed through a one-dimensional convolutional layer composed of 32 1×3-sized convolutional kernels to generate a 1×32-sized local band spectrum correlation feature. Finally, the local band spectrum correlation feature is passed through an improved TCN-GRU unit with 64 hidden units to generate a 1×64-sized multi-band reflectance trend sequence.

[0085] Furthermore, the visual input branch is sequentially configured with a first two-dimensional convolutional layer, a max pooling layer, a residual connection layer, and a second two-dimensional convolutional layer. The first two-dimensional convolutional layer is used to extract macroscopic texture features, the second two-dimensional convolutional layer is used to extract microscopic roughness peak features, and the channel spatial attention mechanism is used to determine key regions of roughness defects.

[0086] Specifically, the processed local image features, after visual preprocessing, generate 512×512×3 visual features, which are then sequentially input into a first two-dimensional convolutional layer with 64 7x7 kernels and a stride of 2, a max pooling layer with a 3x3 window size and a stride of 2, a residual connection layer with 3x3 kernels and 64 channels, and a second two-dimensional convolutional layer with 128 3x3 kernels. The output is a 64x64x128 defect key region. The max pooling layer is used for dimensionality reduction and to preserve key spatial information, while the residual connection layer is used to avoid gradient vanishing and enhance detailed features.

[0087] Furthermore, the feature fusion regression module includes a feature fusion unit and a regression output unit, and the process of determining the surface roughness of the processed surface includes:

[0088] The key defect region and the temporal trend of reflectance of the band sequence are combined across modal features and multi-scale convolutional fusion through a feature fusion unit to generate visual spectral fusion features.

[0089] The visual spectral fusion features are upsampled and activated function feature mapping is performed on the regression output unit to determine the surface roughness of the processed surface.

[0090] Specifically, the cross-modal feature stitching is a concatenate operation, used to stitch together the key defect region and the temporal trend of the band sequence reflectance. The multi-scale convolutional fusion is set with a two-dimensional convolutional layer and a fusion layer. The two-dimensional convolutional layer is set with 256 1x1 size convolutional kernels for compression. The fusion layer is set with parallel 3x3 convolutional blocks, 5x5 convolutional blocks and 7x7 convolutional blocks, used to correlate and fuse textures and spectra at different scales.

[0091] Specifically, the regression output unit is configured with 64 two-dimensional convolutional layers of size 3x3 for normalization, the upsampling operation is bilinear interpolation with a magnification factor of 2, and the activation function feature mapping operation is ReLU activation function operation.

[0092] Furthermore, the surface roughness is the Ra value distribution of the processed surface, and the process of constructing the mapping learning model includes:

[0093] The luminance regression term is calculated based on the ratio of the product of the true Ra value distribution and the predicted Ra value distribution to the sum of the mean squares.

[0094] Based on the local window covariance, true local variance, and predicted local variance of the true Ra value distribution and the predicted Ra value distribution, calculate the contrast regression term;

[0095] A similarity loss function is constructed based on the brightness regression term and the contrast regression term, and the mapping learning model is optimized and trained using the similarity loss function.

[0096] Specifically, the similarity loss function is:

[0097]

[0098] In the formula, Loss represents the similarity loss function, α and β represent the first and second weights respectively, and k represents the k-th sliding window. These represent the luminance regression term and the contrast regression term, respectively, μ ky μ ky′ Let σ represent the true Ra value distribution and the predicted Ra value distribution of the k-th sliding window, respectively. ky σ ky′ σ kyy′ Let represent the predicted local variance, the true local variance, and the local in-window covariance of the k-th sliding window, respectively. Preferably, both the true Ra value distribution and the predicted Ra value distribution are local means within the sliding window. The brightness regression term is used to evaluate the brightness consistency of the generated Ra value distribution, and the contrast regression term is used to measure the grayscale level consistency of the generated Ra value distribution. Therefore, compared to the traditional MSE loss function, the similarity loss function can guarantee the spatial continuity of the predicted Ra value distribution, avoid local abrupt changes or ambiguity, and effectively balance global accuracy and local details for the boundaries of surface defects such as processing marks, pores, and surface roughness.

[0099] The above solution achieves breakthroughs in core indicators such as surface quality inspection accuracy, processing stability, and resource utilization efficiency through multimodal data fusion, temporal-spatial joint modeling, and dynamic closed-loop control, providing an intelligent solution for green and environmentally friendly mold manufacturing.

[0100] Furthermore, the visual preprocessing includes histogram equalization to highlight surface defects; the calibration model includes key band differential calculation.

[0101] Specifically, the histogram equalization is global histogram equalization, which is used to count the number of pixels at each gray level in the image, calculate the cumulative probability distribution function (CDF) for each gray level, map the original gray level to the new gray level, enhance global contrast, and highlight processing marks and surface texture.

[0102] Specifically, within the roughness-sensitive wavelength range of 550nm to 1050nm, multiple key wavelength ranges are selected based on the optical properties of the pipe mold material. For example, key wavelength ranges for aluminum alloy materials include blue light band 450-500nm, green light band 550-570nm, near-infrared band 800-900nm, and short-wave infrared band 2000-2200nm. Differential calculation of the reflectivity detected by multiple key wavelength ranges can suppress the influence of light intensity.

[0103] like Figure 4 As shown, further, the threshold includes a first threshold, a second threshold, and a duration threshold that increase sequentially, and the processing parameters include the cutting speed, feed rate, and depth of cut of the tool on the pipe mold workpiece;

[0104] If the duration of the surface roughness being greater than the first threshold is greater than the duration threshold, then reduce the feed rate and depth of cut.

[0105] If the duration for which the surface roughness is greater than the first threshold is less than the duration threshold, then the cutting speed is reduced.

[0106] If the surface roughness of the machined surface is less than a first threshold and greater than a second threshold, then the cutting speed and feed rate are reduced.

[0107] Specifically, the first threshold is 0.8 μm, the duration threshold is 0.7 s, and the second threshold is 0.4 μm. If the duration of the surface roughness being greater than the first threshold is longer than the duration threshold, the feed rate is reduced to 80% of its original value, and the depth of cut is reduced to 90% of its original value. If the duration of the surface roughness being greater than the first threshold is shorter than the duration threshold, the cutting speed is reduced to 85% of its original value. If the surface roughness is less than the first threshold but greater than the second threshold, the cutting speed is reduced to 90% of its original value, and the feed rate is reduced to 95% of its original value. It is understood that when the Ra value distribution exceeds the first threshold and lasts for a long time, reducing the feed rate and depth of cut can increase the rigidity of the pipe mold, reduce the vibration amplitude, and eliminate surface damage caused by vibration marks. When the Ra value distribution exceeds the first threshold and lasts for a short time, reducing the cutting speed lowers the average temperature of the cutting zone, preventing thermal softening of the pipe mold material surface, thereby reducing surface damage caused by adhesive wear between the tool rake face and the workpiece. When the Ra value distribution is less than the first threshold and greater than the second threshold, the cutting speed and feed rate are reduced simultaneously to reduce the cutting thickness of a single tooth and suppress surface damage caused by the propagation of microcracks due to cyclic thermal stress.

[0108] Therefore, in the above scheme, by adjusting the cutting speed, feed rate and depth of cut, damage to the surface of the pipe mold can be suppressed while taking into account processing efficiency and stability.

[0109] As a specific implementation process in this embodiment, the CNC programming software is CATIA, version V5. Its programming program carries the original machining parameters of the pipe mold machining surface of the high-speed cutting machine tool, and connects to the external platform of Python environment through the open API interface provided by CATIA. The external platform integrates the deep learning framework TensorFlow. In TensorFlow, the mapping learning model of the improved multimodal CNN model architecture described in this embodiment is set. The visual image and spectral image of the machining surface of the pipe mold workpiece collected by the high-speed cutting machine tool are input into the mapping learning model. The mapping learning model generates its predicted machining surface roughness and transmits the machining surface roughness back to the CATIA programming program through the API interface. The programming program adjusts the cutting speed, feed rate and depth of cut of the high-speed cutting machine tool on the pipe mold workpiece according to the machining surface roughness.

[0110] As a specific implementation process of this embodiment, the dataset of the mapping learning model, including the sample set and the test set, is constructed by the trial method. Specifically, visual images of the dataset of the machined surface of the pipe mold workpiece are acquired by a Basler ace 2 high-resolution industrial camera with a ring light source, spectral images of the dataset of the machined surface of the pipe mold workpiece are acquired by a HySpex SWIR-384 hyperspectral camera, and Ra value distribution is obtained by a Mitutoyo Surftest SJ-410 contact profilometer as the label of the dataset.

[0111] As a specific implementation of this embodiment, the main parameters of the mapping learning model include: gradient descent algorithm in Adam through training optimization of the dataset; Mini Batch Size for each training iteration, set to 4; Maximum Epochs for training, set to 300; Initial LearnRate, set to 0.7e-5; Learn Rate Drop Factor, set to 0.6; and LearnRate Drop Period, set to 100. These parameter settings are adapted to the high-speed cutting process of pipe molds on high-speed cutting machine tools. The mapping learning model is preferably integrated into hardware of an NVIDIA RTX 4090 processor.

[0112] In this embodiment, an improved CNN architecture with a dual-input branch network structure and cross-modal feature coupling model is used to optimize the prediction of surface roughness, thereby reducing ineffective machining operations and achieving a green and environmentally friendly manufacturing process that saves energy, reduces pollution, and improves the yield of pipe mold forming. Through multimodal data fusion, temporal-spatial joint modeling, and dynamic closed-loop control, breakthroughs are achieved in core indicators such as surface quality detection accuracy, machining process stability, and resource utilization efficiency, providing an intelligent solution for green and environmentally friendly mold manufacturing. By adjusting the cutting speed, feed rate, and depth of cut, damage to the pipe mold surface can be suppressed while maintaining machining efficiency and stability.

[0113] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0114] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A green and environmentally friendly method for preparing pipe molds, characterized in that, include: Acquire visual and spectral images of the machined surface of the pipe mold workpiece; The visual image is processed by an image processing mapping algorithm to extract processing traces and generate local processing image features. The spectral image is then processed by a calibration model to extract reflectance differences across multiple bands and generate microspectral processing features. A mapping learning model based on an improved multimodal CNN model architecture with visual input and spectral input branches is constructed. The local image features of the processing are input into the visual input branch, and the microspectral features of the processing are input into the spectral input branch. The mapping learning model generates the surface roughness of the processing. Adjust the tool's machining parameters on the machined surface based on the comparison result between the machined surface roughness and the threshold. The mapping learning model has a feature fusion regression module connected to the visual input branch and the spectral input branch. The process of generating the surface roughness of the processed surface includes: determining a multi-band reflectance trend sequence by using an improved TCN-GRU unit to extract the microspectral features of the processed surface in the spectral input branch; inputting the local image features of the processed surface into the visual input branch after visual preprocessing to determine the key defect region; and performing cross-modal fusion feature extraction and regression prediction on the key defect region and the multi-band reflectance trend sequence through the feature fusion regression module to determine the surface roughness of the processed surface. The spectral input branch has a mapping extraction unit, and the improved TCN-GRU unit includes a TCN subunit, a GRU subunit, and an attention subunit. The process of determining the multi-band reflectance trend sequence includes: after spectral preprocessing of the processed microspectral features, determining the local correlation features of the band spectra through the mapping extraction unit; generating a long-range parameter sequence from the local correlation features of the band spectra through the TCN subunit; generating the multi-band reflectance trend sequence as its hidden state through the GRU subunit using the long-range parameter sequence, and controlling the updating of the hidden state by the GRU subunit through the attention subunit.

2. The method for preparing a green and environmentally friendly pipe mold according to claim 1, characterized in that, The hidden state includes the previous hidden state, the current hidden state, and the transitive hidden state. The process of controlling the GRU subunit to update the hidden state through the attention subunit includes: The current hidden state and the target vector are used to generate attention weights through an attention subunit; The transitive hidden state is generated based on the attention weight, the previous hidden state, and the current hidden state, and then the transitive hidden state is input to the GRU subunit for iterative calculation.

3. The method for preparing a green and environmentally friendly pipe mold according to claim 1, characterized in that, The mapping extraction unit has a fully connected layer and a one-dimensional convolutional layer. The process of determining the local correlation features of the band spectrum includes: After the processed microspectral features are preprocessed, the nonlinear relationships between bands are extracted through the fully connected layer to generate correlated band spectra. The associated band spectra are processed through the one-dimensional convolutional layer to extract local correlation features between adjacent bands, thereby generating local correlation features of the band spectra.

4. The method for preparing a green and environmentally friendly pipe mold according to claim 1, characterized in that, The feature fusion and regression module includes a feature fusion unit and a regression output unit. The process of determining the roughness of the processed surface includes: The key defect region and the multi-band reflectance trend sequence are combined across modal features and multi-scale convolutional fusion through a feature fusion unit to generate visual spectral fusion features. The visual spectral fusion features are upsampled and activated function feature mapping is performed on the regression output unit to determine the surface roughness of the processed surface.

5. The method for preparing a green and environmentally friendly pipe mold according to claim 1, characterized in that, The visual input branch is sequentially configured with a first two-dimensional convolutional layer, a max pooling layer, a residual connection layer, and a second two-dimensional convolutional layer. The first two-dimensional convolutional layer is used to extract macroscopic texture features, the second two-dimensional convolutional layer is used to extract microscopic roughness peak features, and the channel spatial attention mechanism is used to determine the key regions of roughness defects.

6. The method for preparing a green and environmentally friendly pipe mold according to claim 1, characterized in that, The surface roughness is the Ra value distribution of the processed surface, and the process of constructing the mapping learning model includes: The luminance regression term is calculated based on the ratio of the product of the true Ra value distribution and the predicted Ra value distribution to the sum of the mean squares. Based on the local window covariance, true local variance, and predicted local variance of the true Ra value distribution and the predicted Ra value distribution, calculate the contrast regression term; A similarity loss function is constructed based on the brightness regression term and the contrast regression term, and the mapping learning model is optimized and trained using the similarity loss function.

7. The method for preparing a green and environmentally friendly pipe mold according to claim 1, characterized in that, The visual preprocessing includes histogram equalization to highlight surface defects; The calibration model includes differential calculations for key bands.

8. The method for preparing a green and environmentally friendly pipe mold according to any one of claims 1 to 7, characterized in that, The thresholds include a first threshold, a second threshold, and a duration threshold, and the processing parameters include the cutting speed, feed rate, and depth of cut of the tool on the pipe mold workpiece. If the duration of the surface roughness being greater than the first threshold is greater than the duration threshold, then reduce the feed rate and depth of cut. If the duration for which the surface roughness is greater than the first threshold is less than the duration threshold, then the cutting speed is reduced. If the surface roughness of the machined surface is less than a first threshold and greater than a second threshold, then the cutting speed and feed rate are reduced.

Images