Method for detecting key parts of machine tool based on industrial CT

Abstract

The application relates to the field of precision detection technology and discloses a machine tool key component detection method based on industrial CT, which comprises the following steps: CT scanning of a machine tool key component to be detected to obtain CT image data; generating a three-dimensional model of the key component based on the CT image data; and defect segmentation of the three-dimensional model by using a deep learning segmentation network, wherein the deep learning segmentation network adopts an encoding-decoding structure, a channel attention unit is introduced in an encoder, a spatial attention unit is introduced in a skip connection, and internal defects are identified; the application realizes nondestructive testing of machine tool key components by using the industrial CT technology, the key components do not need to be disassembled, a deep learning segmentation network is used and a double-attention mechanism is introduced, the identification capability of small-size defects is significantly improved, accurate positioning and classification of defects are realized, and the high-precision detection requirement of machine tool key components is met.

Description

Technical Field

[0001] This invention relates to the field of precision testing technology, and in particular to a method for testing key machine tool components based on industrial CT. Background Technology

[0002] The machining accuracy and internal quality of key machine tool components (such as spindles, lead screws, housings, gears, etc.) directly determine the machine tool's assembly accuracy, operational stability, and service life. Traditional inspection methods mainly include: manual visual inspection, which can only observe surface defects and cannot detect internal defects; ultrasonic testing, which has limited accuracy in identifying micro-cracks and is difficult to quantify defect dimensions; two-dimensional X-ray inspection, which lacks three-dimensional spatial information and is prone to missing hidden defects; and coordinate measuring machines, which can only measure surface dimensions and cannot achieve integrated inspection of internal quality and geometric tolerances.

[0003] Existing testing methods have significant limitations: on the one hand, it is difficult to simultaneously detect internal defects and dimensional deviations without disassembling parts, as the disassembly process may damage the parts; on the other hand, the testing accuracy is greatly affected by manual operation, resulting in low efficiency and failing to meet the testing needs of large-scale production in the machine tool manufacturing industry.

[0004] Existing technical solutions attempt to use industrial CT technology for component inspection, but the following problems still exist in the defect identification stage: Current solutions mainly use traditional image processing algorithms or standard deep learning networks for defect identification, which lack sufficient accuracy in identifying common small-sized defects in machine tool components (such as microcracks and tiny pores), easily leading to missed or false detections. Although the standard U-Net network is widely used in image segmentation tasks, when processing CT images of machine tool components, the segmentation accuracy is difficult to meet actual production requirements due to factors such as indistinct defect features and background noise interference.

[0005] Therefore, there is an urgent need for a detection method that can achieve high-precision identification of internal defects in key machine tool components, in order to solve the technical problem that traditional methods cannot detect internal defects in a non-destructive and comprehensive manner. Summary of the Invention

[0006] The purpose of this invention is to provide a method for detecting key machine tool components based on industrial CT, so as to solve the technical problem that existing technologies cannot accurately identify small-sized defects inside machine tool components.

[0007] To achieve this objective, the present invention adopts the following technical solution: A method for inspecting key machine tool components based on industrial CT, comprising the following steps: S1. Perform CT scans on the key components of the machine tool to be inspected to obtain CT image data; S2. Generate a three-dimensional model of the component based on the CT image data; S3. A deep learning segmentation network is used to segment defects in the three-dimensional model. The deep learning segmentation network adopts an encoder-decoder structure, introduces channel attention units in the encoder, and introduces spatial attention units in the skip connections to identify internal defects.

[0008] Preferably, before step S1, the key components of the machine tool to be tested are cleaned, and tooling fixtures are selected for positioning and fixing according to the characteristics of the components.

[0009] Preferably, before step S1, a scanning parameter database is established based on the component features. The scanning parameter database contains the mapping relationship between different component types and scanning parameters. The scanning parameters include tube voltage, tube current, and exposure time. For new types of parts, parameters are determined through orthogonal experiments, and the optimal scanning parameters are determined by comparing the image quality indicators under different combinations of scanning parameters. The image quality indicators include signal-to-noise ratio and contrast-to-noise ratio.

[0010] Preferably, before step S2, the CT image data is subjected to contrast enhancement processing and denoising processing, wherein the contrast enhancement processing adopts a limited contrast adaptive histogram equalization algorithm, and the denoising processing adopts a non-local mean denoising algorithm.

[0011] Preferably, after step S3, the three-dimensional model is spatially registered with the standard model, and the spatial registration adopts a point cloud registration algorithm.

[0012] Specifically, based on the registered 3D model, the key dimensions and geometric tolerances of the parts are extracted, and the deviation between the actual value and the standard value is calculated. The key dimensions include at least one of diameter, length, hole diameter, and wall thickness, and the geometric tolerances include at least one of coaxiality, parallelism, and roundness.

[0013] In some preferred embodiments, an inspection report is generated based on the defect segmentation results and geometric tolerance deviations. The inspection report includes defect distribution information, dimensional deviation information, and pass / fail determination results.

[0014] Preferably, the deep learning segmentation network is a ResU-Net network, which is based on the U-Net structure and introduces residual connections. The channel attention unit is an SE module, which is embedded in the residual module of the ResU-Net network encoder. The spatial attention unit is a CBAM module, which is embedded in the skip connection layer of the ResU-Net network.

[0015] Specifically, the deep learning segmentation network is trained using a CT image dataset containing defect types. After data augmentation processing including rotation, flipping, and noise addition, the dataset is trained using the Dice-BCE hybrid loss function. The Dice-BCE hybrid loss function L... total The calculation formula is: L total =w1×L Dice +w2×L BCE , Among them, L Dice The Dice loss function is calculated using the following formula: L Dice =1-(2×|P∩G|) / (|P|+|G|), Where P is the set of pixels in the predicted defect region, G is the set of pixels in the defect region of the true label, |P∩G| represents the intersection of the two, and |P| and |G| represent the number of pixels in each of the two. L BCE Let be the binary cross-entropy loss function, and its calculation formula is: L BCE =-(1 / N)×∑[g i ×log(p i )+(1-g i )×log(1-p i )], Where N is the total number of pixels, g i p is the true label of the i-th pixel (value 0 or 1). i Let be the probability that the i-th pixel is predicted to be a defect; w1 and w2 are weighting coefficients, and their calculation formula is as follows: w1=N pos / (N pos +N neg ), w2=N neg / (N pos +N neg ), Where, N pos N represents the number of positive sample pixels. neg This represents the number of negative sample pixels.

[0016] Preferably, the key components of the machine tool include at least one of a spindle, a lead screw, a housing, and a gear, and the defect type includes at least one of looseness, cracks, and pores.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention achieves non-destructive testing of key machine tool components using industrial CT technology, eliminating the need to disassemble components and avoiding damage during the testing process. It also comprehensively detects hidden defects such as internal porosity, cracks, and air bubbles. Furthermore, it employs a deep learning segmentation network, introducing channel attention units in the encoder and spatial attention units in skip connections. By adaptively adjusting feature channel weights and enhancing the feature representation of key spatial regions, it significantly improves the ability to identify small-sized defects. Finally, it generates 3D models of components based on CT image data and performs defect segmentation, achieving precise defect localization and classification, thus meeting the high-precision testing requirements of key machine tool components.

[0018] The present invention has other features and advantages, which will be apparent from or will be set forth in detail in the accompanying drawings and the following detailed description, which together serve to explain the particular principles of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a machine tool key component inspection method based on industrial CT according to an embodiment of the present invention. Detailed Implementation

[0021] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended to limit the scope of protection of this application.

[0022] Please see Figure 1 The machine tool key component inspection method based on industrial CT in this embodiment is illustrated with a machine tool spindle as an example. Of course, this inspection method can also be applied to the inspection of other components, and is not limited here.

[0023] The machine tool key component inspection method based on industrial CT in this embodiment includes the following steps: S1. Perform CT scans on the key components of the machine tool to be inspected to obtain CT image data.

[0024] In this embodiment, the component to be inspected is a machine tool spindle. Before CT scanning, the spindle is first cleaned by using anhydrous ethanol to clean the surface oil and remove rust and other attached impurities. Based on the spindle's dimensions and weight, a three-jaw centering fixture is selected for positioning and fixing. Soft pads are used at the contact points between the fixture and the spindle to prevent scratching the surface and to ensure that the spindle does not shift during the scanning process.

[0025] Before performing a CT scan, a scanning parameter database is established based on the spindle's material density and structural characteristics. This database contains mappings between different machine tool component types (such as spindles, lead screws, housings, gears, etc.) and optimal scanning parameters. These scanning parameters include tube voltage, tube current, and exposure time.

[0026] For new types of components, an orthogonal experimental design was employed, setting multiple combinations of tube voltage (100-200kV), tube current (50-150μA), and exposure time (0.2-0.8s). By calculating the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) of the CT images, the optimal parameter combination for image quality was selected. In this embodiment, for stainless steel spindles, the optimal parameter combination stored in the scanning parameter database was used. When inspecting the same type of spindle subsequently, the matching parameters in the database were directly called, improving inspection efficiency.

[0027] An industrial CT scanner, comprising an X-ray source, detector, and rotary stage, is used to perform a 360° rotational scan of the spindle. The spindle, fixed to a fixture, is placed on the rotary stage. The X-ray source emits X-rays that penetrate the spindle, and the detector receives the transmitted X-ray signals. The rotary stage rotates the spindle 360°, acquiring two-dimensional projection data from different angles. A filtered back-projection algorithm is then used to reconstruct the three-dimensional volume data from the two-dimensional projection data, obtaining the CT image data of the spindle.

[0028] S2. Generate a three-dimensional model of the component based on the CT image data.

[0029] Before generating the 3D model, the CT image data is preprocessed. First, a contrast-limited adaptive histogram equalization algorithm (CLAHE algorithm in this embodiment) is used for contrast enhancement. The CLAHE algorithm divides the CT slice image into several sub-regions (each sub-region is 8×8 pixels in this embodiment), performs histogram equalization on each sub-region, and limits the magnitude of contrast enhancement (the contrast limit threshold is set to 4.0 in this embodiment) to avoid excessive noise amplification. This processing can enhance local contrast and highlight the grayscale difference between defects and the substrate.

[0030] Then, a nonlocal means denoising algorithm is used to remove noise and artifacts generated during the scanning process. This algorithm utilizes a weighted average of pixel information from similar regions in the image (in this embodiment, the search window size is 21×21 pixels, and the similarity window size is 7×7 pixels), effectively suppressing noise while preserving edges and details. Finally, effective ROI regions along the main axis are extracted through image cropping to reduce background interference.

[0031] Based on preprocessed CT image data, a volume rendering algorithm is used to generate a high-precision 3D model of the principal axis. The volume rendering algorithm uses grayscale information from the CT images and techniques such as ray casting and texture mapping to visualize the 3D volume data as a 3D model, fully presenting the internal structure of the principal axis.

[0032] S3. A deep learning segmentation network is used to segment defects in the three-dimensional model. The deep learning segmentation network adopts an encoder-decoder structure, introduces channel attention units in the encoder, and introduces spatial attention units in the skip connections to identify internal defects.

[0033] This embodiment uses the ResU-Net network as the deep learning segmentation network. The ResU-Net network is based on the U-Net structure and introduces residual connections, which include an encoder, a decoder and skip connections.

[0034] The encoder uses ResNet as its backbone network, comprising multiple residual modules. These residual modules introduce cross-layer connections, enabling gradients to propagate directly back to shallower layers, effectively mitigating the gradient vanishing problem in deep networks. SE modules are embedded within the residual modules of the encoder as channel attention units.

[0035] The SE module comprises a global average pooling layer and fully connected layers. The global average pooling layer performs global average pooling on the input feature map, extracting global information for each channel to obtain channel descriptors. Two fully connected layers perform a non-linear transformation on the channel descriptors: the first fully connected layer performs dimensionality reduction (reduction rate r=16), followed by ReLU activation; the second fully connected layer performs dimensionality increase, followed by Sigmoid activation to obtain channel weights. These channel weights are multiplied channel-by-channel with the input feature map, achieving adaptive weighting of key feature channels, enhancing the response to important features while suppressing irrelevant or redundant features.

[0036] It is worth noting that the introduction of the SE module needs to consider the following: ResNet's residual module can extract deep features, while the SE module adaptively adjusts the importance of each channel through a global channel attention mechanism, strengthening key features and suppressing redundant features. Furthermore, the residual module has a certain mitigating effect on gradient vanishing; the introduction of the SE module does not disrupt this mechanism but optimizes feature selection locally, improving overall model performance. Therefore, embedding the SE module into the residual module can avoid its impact on the overall network gradient propagation. Since ResNet's residual module is the bottleneck residual module, the SE module is embedded within the bottleneck residual module. Preferably, in this embodiment, the SE module is embedded in the residual modules of the 3rd, 4th, and 5th layers of the encoder.

[0037] The decoder uses upsampling to gradually restore the spatial resolution of the feature map, and passes the shallow details of the encoder to the decoder through skip connections, effectively alleviating the problem of information loss in deep networks.

[0038] A CBAM module is embedded in the skip connections as a spatial attention unit. The CBAM module includes a channel attention submodule and a spatial attention submodule. The channel attention submodule performs global max pooling and global average pooling on the input feature map to obtain two channel descriptors, which are then summed after passing through a shared fully connected layer and activated by Sigmoid to obtain the channel weights.

[0039] The spatial attention submodule performs max pooling and average pooling on the channel-weighted feature maps to obtain two spatial feature maps. These two maps are concatenated and then subjected to a 7×7 convolution to generate a spatial attention map. The spatial weights are then obtained by sigmoid activation. Multiplying the spatial weights by the feature maps achieves adaptive weighting for key spatial regions.

[0040] Understandably, the CBAM module combines channel attention and spatial attention, adaptively enhancing features in key regions and suppressing redundant or irrelevant information by strengthening feature representation in both channel and spatial dimensions. This design not only optimizes the quality of features received by the decoder, contributing to improved segmentation accuracy, but also, due to the low computational cost of the CBAM module, improves performance while maintaining a lightweight model, making it highly practical.

[0041] For the CBAM module embedding skip connections, the purpose of skip connections in its U-Net structure is to pass shallow detail information from the encoder to the decoder, effectively mitigating the problem of information loss in deep networks. Combining the feature enhancement capabilities of the CBAM module in both channel and spatial dimensions, this embodiment introduces the CBAM module into the skip connection part of the U-Net to adaptively enhance the feature representation of key regions and suppress redundant or irrelevant information.

[0042] The training process of the ResU-Net network in this embodiment is as follows: First, a CT image dataset containing common defect types (porosity, cracks, and air holes) of machine tool parts was constructed. This dataset contains 5000 CT images of parts of different materials and sizes, along with corresponding defect annotations, covering 1000 part samples. Data augmentation was performed on the dataset, including rotation (±15°), flipping (horizontal and vertical flipping), and noise addition (Gaussian noise, mean 0, standard deviation 0.01), to increase data diversity and improve the model's generalization ability.

[0043] The model is trained using the Dice-BCE hybrid loss function. In this embodiment, the Dice-BCE hybrid loss function L... total The calculation formula is: L total =w1×L Dice +w2×L BCE , Among them, L Dice The Dice loss function is calculated using the following formula: L Dice =1-(2×|P∩G|) / (|P|+|G|), Where P is the set of pixels in the predicted defect region, G is the set of pixels in the defect region of the ground truth label, |P∩G| represents the intersection of the two, and |P| and |G| represent the number of pixels in each region, respectively. The Dice loss is evaluated directly based on the overlapping area between the predicted result and the ground truth label, and it performs well in image segmentation tasks that handle class imbalance, especially when the number of defect pixels is much smaller than the number of background pixels.

[0044] L BCE Let be the binary cross-entropy loss function, and its calculation formula is: L BCE =-(1 / N)×∑[g i ×log(p i )+(1-g i )×log(1-p i )], Where N is the total number of pixels, g i p is the true label of the i-th pixel (value 0 or 1). i Let be the probability that the i-th pixel is predicted as a defect. The binary cross-entropy loss is used to measure the difference between the probability distribution predicted by the model and the true label distribution. By minimizing the loss, the predicted probability distribution of the model gradually approaches the true label.

[0045] w1 and w2 are weighting coefficients used to balance the effects of Dice loss and binary cross-entropy loss. The weighting coefficients are adaptively calculated based on the number of positive and negative samples, and their calculation formula is as follows: w1=N pos / (N pos +N neg ), w2=N neg / (N pos +N neg ), Where, N pos N represents the number of positive sample pixels (defective pixels). neg This represents the number of negative sample pixels (background pixels). By adaptively adjusting the weight coefficients, the effects of the two loss functions can be better balanced, further optimizing the model's segmentation and classification accuracy.

[0046] Furthermore, in this embodiment, the training process uses the Adam optimizer, with an initial learning rate of 0.001, a batch size of 16, and 100 training epochs. The training termination condition is that the Dice coefficient on the validation set no longer increases for 10 consecutive epochs, or the maximum number of training epochs is reached.

[0047] After training, the 3D model of the main axis is input into the trained ResU-Net network for defect segmentation. The network analyzes the 3D model layer by layer and outputs the probability that each pixel belongs to a defect.

[0048] Understandably, by setting a threshold (in this embodiment, the threshold is set to 0.5), pixels with a probability greater than the threshold are marked as defective pixels, thereby achieving defect segmentation. Based on the volume, shape, and location characteristics of the defects, the defect type (porosity, cracks, pores) inside the spindle is identified, and they are classified according to severity into minor defects (volume <1mm³, does not affect use), moderate defects (volume 1-5mm³, requires rework), and severe defects (volume >5mm³ or located in critical parts, directly scrapped).

[0049] In this embodiment, the ResU-Net network identifies internal porosity defects and microcracks in the 3D model of the spindle. For example, a porosity defect located in the middle section of the spindle, with a volume of approximately 2.3 mm³, is classified as a moderate defect, while a microcrack located at the end of the spindle, with a length of approximately 0.8 mm, is classified as a minor defect.

[0050] After identifying internal defects, the 3D model of the main spindle is spatially registered with a standard model. Specifically, the standard model is the CAD model of the main spindle from the design phase. Spatial registration employs a point cloud registration algorithm; this embodiment uses the Iterative Closest Point (ICP) algorithm. The ICP algorithm iteratively calculates the optimal rotation matrix and translation vector between two point clouds, aligning the 3D model of the main spindle with the standard model in spatial position. In this embodiment, the iteration termination condition is that the change in distance between points is less than 0.001 mm or the number of iterations is greater than 50. After registration, the positional deviation matrix between the two is calculated to provide a benchmark for subsequent dimensional measurements and deviation analysis. In this embodiment, the registration error is less than 0.002 mm.

[0051] Based on the registered 3D model, the key dimensions and geometric tolerances of the spindle are extracted, and the deviations between the actual values ​​and the standard values ​​are calculated. The key dimensions in this embodiment include the key diameter, length, bore diameter, and wall thickness of the spindle. The correspondence between pixels and physical dimensions is established using the calibration parameters (pixel-physical size mapping relationship) of the industrial CT equipment, and the actual dimensions are measured.

[0052] The geometric tolerances in this embodiment include coaxiality, parallelism, and roundness. For roundness measurement, the least squares method is used to fit the cylindrical surface of the main shaft section and calculate the deviation between the actual roundness and the standard value. For coaxiality measurement, the central axis of the main shaft is fitted and its offset from the standard central axis is calculated.

[0053] In this embodiment, the actual value of the critical diameter of the spindle deviates from the standard value by +0.015mm, which is within the tolerance range specified in the drawing (±0.02mm). The actual value of the roundness is 0.008mm, while the standard value is 0.005mm, which exceeds the tolerance range and requires rework.

[0054] Furthermore, this embodiment can also generate an inspection report based on the defect segmentation results and geometric tolerance deviations. Preferably, the inspection report of this embodiment includes at least a three-dimensional model of the spindle, a defect distribution heat map, a critical dimension deviation table, geometric tolerance inspection results, and a pass / fail judgment conclusion. The defect distribution heat map uses different colors to mark the locations of defects of different severity.

[0055] For example, for the spindle in this embodiment, the inspection report clearly marks the location of porosity defects and microcracks, the data of out-of-tolerance roundness, and gives the qualification judgment conclusion as "unqualified, it is recommended to rework the out-of-tolerance roundness parts and repair the porosity defects", providing a basis for production rework or quality traceability.

[0056] Combination Figure 1This invention achieves non-destructive testing of key machine tool components using industrial CT technology, eliminating the need for component disassembly and avoiding damage during the testing process. In this embodiment, the spindle can be inspected for internal defects without disassembly, maintaining the integrity of the components. Furthermore, it employs a ResU-Net network, embedding an SE module in the encoder's residual module and a CBAM module in the skip connections. By adaptively adjusting feature channel weights and enhancing the feature representation of key spatial regions, it significantly improves the ability to identify small-sized defects. In this embodiment, a microcrack with a length of 0.8 mm was successfully identified, improving the accuracy by approximately 30% compared to traditional methods. Finally, it generates a 3D model of the components based on CT image data and performs defect segmentation, achieving precise defect location and classification. Simultaneously, registration with a standard model integrates dimensional measurement and geometric tolerance detection, increasing the inspection efficiency by approximately 50% compared to traditional multi-device joint inspection, meeting the high-precision inspection requirements of key machine tool components.

[0057] Experimental results show that, compared with traditional detection methods, the present invention improves the crack detection accuracy from 58.1% to 92.98%, an improvement of 34.88 percentage points; and the porosity detection accuracy from 54.36% to 91.57%, an improvement of 37.21 percentage points. Compared with the standard U-Net network, the present invention improves the crack detection accuracy from 75.94% to 92.98%, an improvement of 17.04 percentage points; and the porosity detection accuracy from 73.18% to 91.57%, an improvement of 18.39 percentage points. These data fully demonstrate the superiority of the technical solution of the present invention.

[0058] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for inspecting key machine tool components based on industrial CT, characterized in that, Includes the following steps: CT scans were performed on the key components of the machine tool to be inspected to obtain CT image data; A three-dimensional model of the component is generated based on the CT image data; A deep learning segmentation network is used to segment defects in the 3D model. The deep learning segmentation network adopts an encoder-decoder structure, introducing channel attention units in the encoder and spatial attention units in the skip connections to identify internal defects.

2. The method for inspecting key machine tool components based on industrial CT as described in claim 1, characterized in that, Before acquiring CT image data, the key components of the machine tool to be inspected are cleaned, and tooling fixtures are selected for positioning and fixing according to the characteristics of the components.

3. The method for inspecting key machine tool components based on industrial CT as described in claim 1, characterized in that, Before performing a CT scan, a scanning parameter database is established based on the characteristics of the components. The scanning parameter database contains the mapping relationship between different component types and scanning parameters, including tube voltage, tube current, and exposure time. For new types of parts, parameters are determined through orthogonal experiments, and the optimal scanning parameters are determined by comparing the image quality indicators under different combinations of scanning parameters. The image quality indicators include signal-to-noise ratio and contrast-to-noise ratio.

4. The method for inspecting key machine tool components based on industrial CT as described in claim 1, characterized in that, Before generating the 3D model, the CT image data is subjected to contrast enhancement and denoising processing. The contrast enhancement processing adopts a limited contrast adaptive histogram equalization algorithm, and the denoising processing adopts a non-local mean denoising algorithm.

5. The method for inspecting key machine tool components based on industrial CT as described in claim 1, characterized in that, After identifying internal defects, the 3D model is spatially registered with the standard model using a point cloud registration algorithm.

6. The method for inspecting key machine tool components based on industrial CT as described in claim 5, characterized in that, Based on the registered 3D model, the key dimensions and geometric tolerances of the parts are extracted, and the deviation between the actual value and the standard value is calculated. The key dimensions include at least one of diameter, length, hole diameter, and wall thickness, and the geometric tolerances include at least one of coaxiality, parallelism, and roundness.

7. The method for inspecting key machine tool components based on industrial CT as described in claim 6, characterized in that, An inspection report is generated based on the defect segmentation results and geometric tolerance deviations. The inspection report includes defect distribution information, dimensional deviation information, and pass / fail determination results.

8. The method for inspecting key machine tool components based on industrial CT as described in claim 1, characterized in that, The deep learning segmentation network is a ResU-Net network, which is based on the U-Net structure and introduces residual connections. The channel attention unit is an SE module, which is embedded in the residual module of the ResU-Net network encoder. The spatial attention unit is a CBAM module, which is embedded in the skip connection layer of the ResU-Net network.

9. The method for inspecting key machine tool components based on industrial CT as described in claim 8, characterized in that, The deep learning segmentation network is trained using a CT image dataset containing defect types. After data augmentation processing including rotation, flipping, and noise addition, the dataset is trained using the Dice-BCE hybrid loss function. The Dice-BCE hybrid loss function L... total The calculation formula is: L total =w1×L Dice +w2×L BCE , Among them, L Dice The Dice loss function is calculated using the following formula: L Dice =1-(2×|P∩G|) / (|P|+|G|), Where P is the set of pixels in the predicted defect region, G is the set of pixels in the defect region of the true label, |P∩G| represents the intersection of the two, and |P| and |G| represent the number of pixels in each of the two. L BCE Let be the binary cross-entropy loss function, and its calculation formula is: L BCE =-(1 / N)×∑[g i ×log(p i )+(1-g i )×log(1-p i )], Where N is the total number of pixels, g i p is the true label of the i-th pixel (value 0 or 1). i Let be the probability that the i-th pixel is predicted to be a defect; w1 and w2 are weighting coefficients, and their calculation formula is as follows: w1=N pos / (N pos +N neg ), w2=N neg / (N pos +N neg ), Where, N pos N represents the number of positive sample pixels. neg This represents the number of negative sample pixels.

10. The method for inspecting key machine tool components based on industrial CT as described in claim 1, characterized in that, The key components of the machine tool include at least one of the following: spindle, lead screw, housing, and gear; and the defect types include at least one of the following: looseness, cracks, and porosity.

Images