A polarization binocular online measurement method for three-dimensional posture of a vacuum forging furnace forged piece

Abstract

The present application relates to the field of laser detection, and discloses a kind of vacuum forging furnace inner forging piece three-dimensional posture polarization binocular online measurement method, after being forged in vacuum high-temperature forging furnace and being jacked up, first multi-directional polarization original image and second multi-directional polarization original image are collected from two symmetrical visual angles by polarization detection system;Then based on polarization difference, forging piece region is segmented from first and second multi-directional polarization original image respectively and forging piece contour is extracted, and first forging piece contour and second forging piece contour are obtained;Finally, first forging piece contour and second forging piece contour are input to pre-established forging piece three-dimensional posture recognition model, and the model outputs forging piece pose class.The present application decomposes complex three-dimensional pose measurement problem into two two-dimensional pose recognition problems under single visual angle, reduces the engineering difficulty of data acquisition and labeling, and improves the robustness and interpretability of pose recognition.

Description

Technical Field

[0001] This invention relates to the field of laser detection technology, specifically to a polarization binocular online measurement method for the three-dimensional orientation of forgings in a vacuum forging furnace. Background Technology

[0002] Vacuum high-temperature isothermal forging (1000~1500℃) is a core forming process for critical hot-end components in aerospace and nuclear power. The three-dimensional position and spatial orientation of the forging after demolding directly determine the product quality. Defects such as misalignment, tilting, or abnormal demolding, if not detected in time, will lead to batch scrap.

[0003] In an isothermal forging environment, the forging, die, and furnace wall are at the same process temperature, and the blackbody radiation brightness and spectral distribution tend to be consistent. Conventional visual inspection and infrared inspection methods are difficult to effectively distinguish between the forging and the background because the target area lacks the physical basis for brightness and temperature differences.

[0004] To obtain the 3D pose information of a target, the commonly used method is binocular stereo vision, which simultaneously acquires images using left and right cameras, reconstructs a 3D point cloud through stereo matching and triangulation, and then registers it with a reference model to obtain pose parameters. However, this method has the following drawbacks when applied to a vacuum forging furnace: thermal expansion of the furnace body under high temperature causes deviations in parameters calibrated at room temperature, and online calibration is difficult to perform in a high-temperature environment; both the forging and the die are high-temperature metal surfaces, lacking natural texture, and under isothermal conditions, there are no differences in brightness and color, resulting in ambiguity in stereo matching over large areas of the forging surface, and the reconstructed point cloud is sparse and noisy; stereo matching has a large computational load, making it difficult to meet the requirements of online real-time detection.

[0005] Therefore, existing methods for measuring the three-dimensional posture of forgings rely on precise geometric calibration and dense stereo matching. These methods are not only difficult to adapt to high-temperature vacuum environments, but also require the acquisition of accurate three-dimensional spatial coordinates. This results in high engineering difficulty for data annotation and limited real-time performance, making it difficult to achieve stable and reliable online measurement of the three-dimensional posture of forgings under multiple constraints. Summary of the Invention

[0006] The purpose of this invention is to provide a polarization binocular online measurement method for the three-dimensional posture of forgings in a vacuum forging furnace, so as to solve the technical problem that it is difficult to achieve stable and reliable online measurement of the three-dimensional posture of forgings in the high-temperature forging environment in the prior art.

[0007] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution: A polarization binocular online measurement method for the three-dimensional orientation of forgings in a vacuum forging furnace includes the following steps: Step S1: When the forging is in the lifted state after mold opening, the polarization detection system built into the vacuum forging furnace collects the first multi-directional polarization original image and the second multi-directional polarization original image formed by the laser jointly reflected by the forging and the background area where the forging is located in different polarization directions at the two sampling points respectively. The two sampling points are in a symmetrical viewing angle relationship. Step S2: The forging region is segmented from the first multi-directional polarization original image and the second multi-directional polarization original image respectively, and the forging contour is extracted from the segmented forging region to obtain the first forging contour and the second forging contour. Step S3: Input the first forging contour and the second forging contour into the pre-established three-dimensional forging posture recognition model, and output the forging posture category from the pre-established three-dimensional forging posture recognition model; The three-dimensional pose recognition model includes a first two-dimensional pose recognition sub-network, a second two-dimensional pose recognition sub-network, and a binocular fusion sub-network; The first two-dimensional pose recognition sub-network takes the outline of the first forging as input and outputs the two-dimensional pose category of the forging from the first viewpoint. The second two-dimensional pose recognition subnetwork takes the contour of the second forging as input and outputs the two-dimensional pose category of the forging from the second perspective. The binocular fusion subnetwork takes the two-dimensional pose category of the forging in the first view and the two-dimensional pose category of the forging in the second view as input, and outputs the three-dimensional pose category of the forging.

[0008] Furthermore, both the first two-dimensional pose recognition subnetwork and the second two-dimensional pose recognition subnetwork adopt convolutional neural networks including multiple convolutional layers and fully connected layers; The convolutional layer in the first two-dimensional pose recognition sub-network is used to extract contour shape features from the input first forging contour step by step. The fully connected layer in the first two-dimensional pose recognition sub-network is followed by a Softmax activation function, which is used to output a two-dimensional pose category probability vector under the first view based on the contour shape features. The category corresponding to the maximum value in the probability vector is taken as the two-dimensional pose category under the first view. The convolutional layer in the second two-dimensional pose recognition subnetwork is used to extract contour shape features from the input second forging contour step by step. The fully connected layer in the second two-dimensional pose recognition subnetwork is followed by a Softmax activation function, which is used to output a two-dimensional pose category probability vector under the second view based on the contour shape features. The category corresponding to the maximum value in the probability vector is taken as the two-dimensional pose category under the second view. The two-dimensional pose category is used to describe the pose state of the forging in the corresponding two-dimensional image plane, including normal centering, left offset, right offset, left tilt, and right tilt.

[0009] Furthermore, the binocular fusion subnetwork adopts a multilayer perceptron structure containing multiple fully connected layers; The binocular fusion subnetwork takes as input a joint feature vector formed by concatenating the two-dimensional pose category probability vector output by the first two-dimensional pose recognition subnetwork and the two-dimensional pose category probability vector output by the second two-dimensional pose recognition subnetwork. The fully connected layer performs nonlinear fusion of the two-dimensional pose information of the first and second perspectives contained in the joint feature vector to output the three-dimensional pose category probability vector of the forging. The category corresponding to the maximum value in the three-dimensional pose category probability vector is taken as the three-dimensional pose category of the forging; The three-dimensional pose categories are used to describe the pose state of the forging in three-dimensional space within the vacuum high-temperature forging furnace, including normal, left-leaning, right-leaning, forward-leaning, backward-leaning, left-tilting, right-tilting, forward-tilting, backward-tilting, and demolding abnormality.

[0010] Furthermore, the training loss of the three-dimensional pose recognition model includes three parts: classification loss, binocular consistency loss, and two-dimensional-three-dimensional consistency loss. The classification loss includes a first two-dimensional pose classification loss, a second two-dimensional pose classification loss, and a three-dimensional pose classification loss; The first two-dimensional pose classification loss is the cross-entropy loss between the output of the first two-dimensional pose recognition sub-network and the first pose truth label of the forging. The second two-dimensional pose classification loss is the cross-entropy loss between the output of the second two-dimensional pose recognition subnetwork and the ground truth label of the forging's second pose. The 3D pose classification loss is the cross-entropy loss between the output of the binocular fusion subnetwork and the ground truth label of the 3D pose of the forging. The binocular consistency loss is used to penalize invalid 2D pose combinations that violate binocular geometric relationships in the outputs of the first 2D pose recognition subnetwork and the second 2D pose recognition subnetwork. The binocular consistency loss is calculated by summing the products of the first prediction probability and the second prediction probability corresponding to each combination in the predefined set of invalid 2D pose combinations. For any combination (i,j) in the set of invalid 2D pose combinations, the first prediction probability is the probability corresponding to the i-th category in the probability vector output by the first 2D pose recognition subnetwork, and the second prediction probability is the probability corresponding to the j-th category in the probability vector output by the second 2D pose recognition subnetwork. The formula for calculating the binocular consistency loss is as follows: ; In the formula, C is the set of invalid two-dimensional pose combinations. Each element in C is a two-dimensional pose category index combination (i,j), indicating that the first output of the i-th two-dimensional pose category and the second output of the j-th two-dimensional pose category are invalid combinations that violate the binocular geometric relationship. Let be the probability value corresponding to the i-th category in the two-dimensional pose category probability vector output by the first two-dimensional pose recognition subnetwork. The probability value corresponding to the j-th category in the two-dimensional pose category probability vector output by the second two-dimensional pose recognition subnetwork; The 2D-3D consistency loss is used to penalize invalid triplets where there is a logical contradiction between the 3D pose category output by the binocular fusion subnetwork and the two input 2D pose categories. The 2D-3D consistency loss is calculated by summing the products of the first prediction probability, the second prediction probability, and the 3D prediction probability corresponding to each combination in the predefined set of invalid triplets. For any combination (i,j,k) in the set of invalid triplets, the first prediction probability is the probability corresponding to the i-th category in the probability vector output by the first 2D pose recognition subnetwork, the second prediction probability is the probability corresponding to the j-th category in the probability vector output by the second 2D pose recognition subnetwork, and the 3D prediction probability is the probability corresponding to the k-th category in the probability vector output by the binocular fusion subnetwork. The formula for calculating the two-dimensional-three-dimensional consistency loss is as follows: ; In the formula, S is a set of invalid ternary combinations. Each element in S is a ternary combination (i,j,k) with a pose category index, indicating that the first output of the i-th 2D pose category, the second output of the j-th 2D pose category, and the binocular fusion subnetwork output of the k-th 3D pose category are invalid combinations with logical contradictions. This represents the probability value corresponding to the k-th class in the 3D pose class probability vector output by the binocular fusion subnetwork.

[0011] Furthermore, the annotation methods for the three-dimensional pose truth labels of the forging, the first pose truth label of the forging, and the second pose truth label of the forging include: The forging is placed in a vacuum high-temperature forging furnace and subjected to various positions and orientations. The first multi-directional polarization original image and the second multi-directional polarization original image are acquired as the first sample image and the second sample image. The forging region is segmented from the first sample image and the second sample image respectively, and the forging contour is extracted from the segmented forging region to obtain the first sample contour and the second sample contour. The first and second sample contours are manually annotated. The first pose truth label of the forging is assigned according to the pose state of the forging in the two-dimensional image plane under the first view. The second pose truth label of the forging is assigned according to the pose state of the forging in the two-dimensional image plane under the second view. The preset pose state of the forging in the vacuum high-temperature forging furnace when the first and second sample images of the set are collected is used as the three-dimensional pose truth value label of the forging. The first sample contour, the second sample contour, the ground truth label of the first pose of the forging, the ground truth label of the second pose of the forging, and the ground truth label of the three-dimensional pose of the forging together constitute a sample in the dataset used to train the three-dimensional pose recognition model.

[0012] Furthermore, the first multi-directional polarization original image and the second multi-directional polarization original image each include four polarization original images corresponding to polarization directions of 0°, 45°, 90°, and 135°.

[0013] Furthermore, the two sampling points are two through holes with symmetrical viewing angles on the wall of the vacuum forging furnace. A polarization detection system is set at each sampling point to form two independent detection units. The main optical axes of the two independent detection units intersect in the position area where the forging is expected to be lifted in the furnace. The two detection units include a first detection unit corresponding to the first viewing angle and a second detection unit corresponding to the second viewing angle. The polarization detection system includes a window assembly, a coaxial beam splitting and combining assembly, an active linearly polarized laser emission module, and a polarization filtering imaging receiving module.

[0014] Furthermore, the viewing window assembly is embedded in the furnace wall through hole of the vacuum high-temperature forging furnace, and the coaxial beam splitting and combining assembly is disposed outside the furnace of the viewing window assembly, for guiding the linearly polarized laser emitted by the active linearly polarized laser emitting module into the furnace, and guiding the reflected light from the furnace into the polarization filtering imaging receiving module.

[0015] Furthermore, the polarization filtering imaging receiving module includes a narrowband bandpass filter and a focal plane polarization image sensor; The center wavelength of the narrowband bandpass filter is matched with the laser wavelength of the active linearly polarized laser emission module, and the bandwidth does not exceed 3nm. The focal plane polarization image sensor integrates micro-polarization arrays in four directions (0°, 45°, 90°, and 135°) on its pixels, used for synchronously acquiring a first multi-directional polarization raw image or a second multi-directional polarization raw image in a single exposure. Compared with the prior art, the present invention has the following advantages: This invention transforms the three-dimensional pose measurement of forgings into two-dimensional pose recognition under two single-viewpoints. The first and second two-dimensional pose recognition sub-networks identify the two-dimensional pose categories from the single-viewpoint contour images, respectively. Then, the three-dimensional pose category is output after the probability vectors of the two viewpoints are nonlinearly fused by the binocular fusion sub-network. This eliminates the difficulties in camera calibration under high-temperature conditions and the difficulty in stereo matching of textureless forging surfaces in traditional binocular stereo vision. Attached Figure Description

[0016] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0017] Figure 1 A flowchart illustrating the polarization binocular online measurement method for the three-dimensional orientation of forgings in a vacuum forging furnace, as provided in this embodiment of the application.

[0018] Figure 2 This is a schematic diagram of the detection unit structure provided in an embodiment of this application.

[0019] Figure 3 This is a schematic diagram of the optical path principle of the coaxial beam splitting and combining component provided in the embodiments of this application.

[0020] Figure 4 This is a schematic diagram of the coaxial optical through-hole structure of the double-layer furnace body provided in the embodiments of this application.

[0021] Figure 5 This is a schematic diagram showing the installation positions of the coaxial beam splitter / combiner, the active linearly polarized laser emitting module, and the polarization-filtered imaging receiving module provided in the embodiments of this application.

[0022] Figure 6 This is a schematic diagram of the window component structure provided in an embodiment of this application.

[0023] Figure 7 This is a schematic diagram of the polarization state conversion between the incident laser light path and the reflected light return light path provided in the embodiments of this application.

[0024] The reference numerals in the figure are as follows: 1. Vacuum high-temperature forging furnace; 11. Outer furnace shell; 12. Inner heating furnace liner; 2. Furnace wall through holes; 21. Furnace shell through holes; 22. Furnace liner through holes; 3. Window assembly; 31. Sapphire transparent substrate; 32. Anti-metal evaporation protective coating layer; 33. Vacuum sealing structure; 4 Active linearly polarized laser emission module, 41 Linearly polarized laser, 42 Collimating and shaping lens group; 5. Coaxial beam splitter and combiner assembly, 51. Polarization beam splitter, 52. Quarter-wave plate, 5a. First port, 5b. Second port, 5c. Third port; 6. Polarization filter imaging receiver module, 61. Narrowband bandpass filter, 62. Focal plane polarization image sensor, 63. Imaging optical lens; 7. Constant temperature damping base; 8. Operational control host; 9-person human-computer interaction display terminal. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] like Figure 1 As shown, the present invention provides a polarization binocular online measurement method for the three-dimensional attitude of forgings in a vacuum forging furnace, which includes the following steps: Step S1: When the forging is in the lifted state after mold opening, the polarization detection system built into the vacuum forging furnace collects the first multi-directional polarization original image and the second multi-directional polarization original image formed by the laser jointly reflected by the forging and the background area where the forging is located in different polarization directions at the two acquisition points respectively. The two acquisition points are in a symmetrical viewing angle relationship.

[0027] Step S2: The forging region is segmented from the first multi-polarized original image and the second multi-polarized original image, and the forging contour is extracted from the segmented forging region to obtain the first forging contour and the second forging contour.

[0028] Step S3: Input the first forging profile and the second forging profile into the pre-established three-dimensional forging posture recognition model, and output the forging posture category from the pre-established three-dimensional forging posture recognition model.

[0029] The 3D pose recognition model includes a first 2D pose recognition subnetwork, a second 2D pose recognition subnetwork, and a binocular fusion subnetwork; The first two-dimensional pose recognition subnetwork takes the outline of the first forging as input and outputs the two-dimensional pose category of the forging from the first viewpoint. The second two-dimensional pose recognition subnetwork takes the contour of the second forging as input and outputs the two-dimensional pose category of the forging from the second perspective. The binocular fusion subnetwork takes the two-dimensional pose category of the forging from the first viewpoint and the two-dimensional pose category of the forging from the second viewpoint as input, and outputs the three-dimensional pose category of the forging.

[0030] This embodiment acquires polarization images of forgings simultaneously from two symmetrical perspectives. Unlike traditional binocular stereo vision, which requires precise calibration of the relative pose of the left and right cameras and pixel-level stereo matching, this embodiment only needs to ensure sufficient parallax so that different three-dimensional poses present distinguishable two-dimensional contour differences from the left and right perspectives. It does not require obtaining precise camera extrinsic parameters, thus reducing the dependence on calibration accuracy at the system architecture level.

[0031] This embodiment employs active linearly polarized laser illumination instead of relying on the forging's own thermal radiation because, under isothermal conditions, the blackbody radiation characteristics of the forging and the background are identical. The active laser introduces a narrowband signal source at a specific wavelength, significantly higher than the background signal. It utilizes the physical difference between the specular reflection on the forging surface retaining high polarization and the diffuse reflection depolarization of the mold and furnace wall surfaces. This fundamentally solves the problem of difficulty in effectively distinguishing the target from the background due to a lack of brightness and temperature differences, providing high-quality raw data for subsequent forging segmentation and contour extraction.

[0032] This embodiment independently performs forging segmentation and contour extraction from two perspectives. On one hand, the segmentation is based on the difference in polarization degree, utilizing the significant difference between the high DoP value on the forging surface and the low DoP value in the background region to achieve stable separation of the forging region under isothermal and brightness conditions. On the other hand, the segmentation result is converted into a binary contour image as input to the subsequent pose recognition network. This information compression strategy greatly simplifies the input of the pose recognition sub-network, retaining only the shape information directly related to the pose, avoiding the influence of background interference and illumination changes on pose recognition, and significantly reducing the number of parameters and computational load of the subsequent network.

[0033] It should be noted that pose recognition is based on the contour image rather than the original multi-channel polarization image because the focus in the forging process is on the spatial position and orientation of the forging, rather than its surface texture or polarization characteristics. The contour, as a concise expression of shape, already contains all the geometric information needed to determine the pose.

[0034] The three-branch continuous collaborative architecture constructed in this embodiment can decompose the complex three-dimensional pose measurement problem into two two-dimensional pose recognition problems under single viewpoints and a dual-viewpoint fusion problem, thereby avoiding the two major problems of stereo calibration and dense matching faced by traditional binocular stereo vision methods.

[0035] In this embodiment, the first two-dimensional pose recognition sub-network only needs to identify the two-dimensional pose state from a single-view forging contour image, such as whether the forging is centered, to the left or right, or tilted to the left or right in the image. This is a typical image classification task. The network structure is lightweight, and the training data only requires two-dimensional contour images and corresponding two-dimensional pose category labels. The annotation difficulty is much lower than obtaining accurate three-dimensional spatial coordinate ground truth.

[0036] In this embodiment, the second two-dimensional pose recognition subnetwork operates symmetrically to the first subnetwork. Each subnetwork independently extracts shape features from the contour image of the corresponding viewpoint and outputs a probability vector of the two-dimensional pose category. This probability vector retains the confidence information of the classification rather than a hard category label, providing a richer expression of uncertainty for subsequent fusion decisions.

[0037] In this embodiment, the binocular fusion subnetwork employs a lightweight multilayer perceptron structure. It performs nonlinear fusion within a joint feature space composed of two concatenated probability vectors, implicitly learning the mapping relationship between 2D pose combinations and 3D pose categories in a data-driven manner. This design eliminates the need for the network to explicitly know geometric parameters such as baseline distance and focal length; instead, it automatically infers the mapping rules from labeled data. This avoids the geometric calibration challenges in high-temperature environments and meets the requirements for online real-time detection with the computational efficiency of a single forward propagation.

[0038] In this embodiment, first and second 2D pose recognition sub-networks each complete 2D pose classification from single-view contours. A binocular fusion sub-network fuses and infers information from both perspectives in the joint feature space of probability vectors, outputting the 3D pose category of the forging. The binocular consistency loss and 2D-3D consistency loss introduced in the training loss encode binocular geometric rules and pose semantic logic as soft constraints into the model training process, ensuring that the model output conforms to physical laws and logical consistency even without a large amount of labeled data. This significantly reduces the engineering difficulty of data acquisition and annotation, eliminating the need for precise geometric calibration, dense stereo matching, and obtaining ground truth 3D coordinates, thereby improving the robustness and interpretability of pose recognition.

[0039] In some embodiments, the first multi-directional polarization original image and the second multi-directional polarization original image each include four polarization original images corresponding to polarization directions of 0°, 45°, 90°, and 135°.

[0040] In some embodiments, forging regions are segmented from the first multi-polarized original image and the second multi-polarized original image, and forging contours are extracted from the segmented forging regions to obtain the first forging contour and the second forging contour.

[0041] This embodiment utilizes the physical difference between the forging and the background in the polarization domain to separate the forging region from the original image with isothermal and uniform brightness. The segmentation is based on the following polarization optics principle: In a vacuum high-temperature forging furnace, the forging surface, under near-normal incident conditions, primarily exhibits specular reflection, preserving the polarization state of the reflected light and resulting in a high degree of polarization (DoP), typically ranging from 0.6 to 0.9. The mold and furnace wall surfaces, being rough and covered with an oxide layer, primarily exhibit diffuse reflection, randomizing the polarization state of the reflected light and resulting in a lower degree of polarization, typically ranging from 0.05 to 0.3. This significant difference in polarization degree between the two forms the physical basis for the segmentation.

[0042] In this embodiment, the Stokes parameters and degrees of polarization are calculated based on the first and second multi-directional polarization original images, respectively. Taking the first image as an example, the original intensity values ​​for the four polarization directions (0°, 45°, 90°, and 135°) are calculated as follows: (0° direction) (45° direction) (90° direction) (135° direction); Total luminous intensity: ; Linear polarization components (0° / 90° direction): ; Linear polarization components (45° / 135° directions): ; Degree of polarization: .

[0043] The range of DoP values ​​is: . This represents fully linearly polarized light. This represents completely unpolarized light. The second calculation method is the same.

[0044] After obtaining the polarization map, a threshold segmentation method can be used to segment the forging region. Since there is a significant difference in DoP values ​​between the forging region and the background region, binary segmentation can be performed by setting a threshold (e.g., DoP>0.6 for forging, DoP≤0.6 for background).

[0045] The threshold can be determined in the following ways: First, pre-calibration: under typical working conditions with known forging and background, polarization images are acquired, and the DoP distribution of forging pixels and background pixels is statistically analyzed to select the threshold that minimizes the segmentation error rate. Second, adaptive thresholding: using adaptive thresholding segmentation algorithms such as Otsu, the optimal threshold is automatically determined based on the bimodal distribution of the DoP histogram, eliminating the need for manual setting. In the polarization degree binary segmentation result, the boundary between the forging region and the background region is the two-dimensional contour of the forging. Edge detection of the binary image can then be used to extract the forging contour.

[0046] It should be noted that the above-described segmentation method based on polarization degree threshold is a simple and efficient implementation method. Its advantages include low computational cost, good real-time performance, and no need for pre-trained models, making it suitable for scenarios with relatively stable operating conditions. However, the segmentation of the forging region in this embodiment is not limited to the threshold method. A polarization segmentation network based on deep learning can also be used to achieve higher accuracy and stronger robustness. The specific choice can be made according to the actual deployment scenario's requirements for segmentation accuracy and computational resources.

[0047] In this embodiment, after obtaining the segmentation mask of the forging region, contour extraction is performed on the mask. Specifically, the boundary between the forging region and the background region in the binary mask is traced to obtain the forging contour represented by an ordered set of boundary points. This contour is directly input into the first and second two-dimensional pose recognition sub-networks. It should be noted that in this embodiment, the input to the pose recognition sub-network is only the binary contour image of the forging, rather than the original multi-channel polarization image. This makes the input information of the pose recognition sub-network highly concise, containing only shape information directly related to the pose, avoiding the influence of background interference and illumination changes on pose recognition, and significantly reducing the number of network parameters and computational load.

[0048] In the actual process of vacuum high-temperature forging, the focus at the forging site is not on the precise coordinates of the forging in space, but rather on whether the forging is in a normal ejection state. Whether the forging is deviated to the left or right, tilted forward or backward, or whether there is an abnormal demolding process, this information directly determines subsequent operations: if the forging's posture is normal, the next process begins; if the forging is deviated to the left, the robotic arm's gripping position is adjusted to the left; if the forging tilts beyond the safe range, an alarm is triggered and production is suspended. Process engineers need qualitative judgments that can guide operations, rather than numerical coordinates that require secondary interpretation. Dividing the posture into a finite number of discrete categories fully covers the decision-making needs of the forging site.

[0049] In some embodiments, the three-dimensional pose recognition model of the forging includes a first two-dimensional pose recognition subnetwork, a second two-dimensional pose recognition subnetwork, and a binocular fusion subnetwork. The three work together to decompose the complex three-dimensional pose measurement problem into two two-dimensional pose recognition problems under single-viewpoints and a dual-viewpoint fusion problem.

[0050] The three-dimensional pose measurement is transformed into a two-dimensional pose classification problem under two single-viewpoints. The first and second two-dimensional pose recognition sub-networks independently identify the two-dimensional pose state of the forging under that viewpoint from the single-viewpoint contour image. The binocular fusion sub-network only needs to learn the mapping relationship between the combination of the two-dimensional poses from the two viewpoints and the three-dimensional pose category.

[0051] In this embodiment, the training of the first and second sub-networks only requires two-dimensional contour images and corresponding two-dimensional pose category labels, without the need to obtain accurate three-dimensional spatial coordinate ground values, which greatly reduces the engineering difficulty of data acquisition and annotation. Moreover, binocular fusion is completed in the joint feature space of probability vectors, without relying on explicit geometric reconstruction, thus avoiding the two problems of stereo calibration and textureless matching under high temperature conditions. Finally, all three sub-networks are lightweight structures, and the entire inference process from contour images to three-dimensional pose categories only requires a single forward propagation, and the computational efficiency can meet the requirements of online real-time detection.

[0052] In some embodiments, if the contours extracted from the left and right views are found to be inconsistent (for example, the contour extraction on one side is incomplete due to window contamination), the first and second sub-networks can still independently complete the pose determination under that view. The binocular fusion sub-network has the ability to reason based on incomplete information, so the system will not completely fail due to the deterioration of information quality on one side.

[0053] In some embodiments, both the first two-dimensional pose recognition subnetwork and the second two-dimensional pose recognition subnetwork employ convolutional neural networks that include multiple convolutional layers and fully connected layers.

[0054] The convolutional layer in the first two-dimensional pose recognition subnetwork is used to extract contour shape features from the input first forging contour step by step. The fully connected layer in the first two-dimensional pose recognition subnetwork is followed by a Softmax activation function, which is used to output a two-dimensional pose category probability vector under the first view based on the contour shape features. The category corresponding to the maximum value in the probability vector is taken as the two-dimensional pose category under the first view.

[0055] The convolutional layer in the second two-dimensional pose recognition subnetwork is used to extract contour shape features from the input second forging contour step by step. The fully connected layer in the second two-dimensional pose recognition subnetwork is followed by a Softmax activation function, which is used to output a two-dimensional pose category probability vector in the second view based on the contour shape features. The category corresponding to the maximum value in the probability vector is taken as the two-dimensional pose category in the second view.

[0056] Two-dimensional pose categories are used to describe the pose state of the forging in the corresponding two-dimensional image plane, including normal centering, left offset, right offset, left tilt, and right tilt.

[0057] The task of the first and second 2D pose recognition sub-networks is to identify the 2D pose state from a single-view forging contour image. Convolutional neural networks are chosen as the basic structure because convolutional layers have natural translation invariance and local perception capabilities, enabling them to stably extract pose-related shape features from binary contour images, such as the centroid position, principal axis direction, and symmetry of the contour, while exhibiting a certain degree of robustness to changes in the absolute position and scale of the contour.

[0058] The two sub-networks are structurally symmetrical, but have different input sources. The first sub-network receives the profile of the first forging, and the second sub-network receives the profile of the second forging.

[0059] In this embodiment, the symmetrical structure allows the two sub-networks to adopt a Siamese network structure with shared weights. This halves the number of parameters, making the model more lightweight. Furthermore, the contours of the two viewpoints are mapped to the same feature space during feature extraction. During subsequent binocular fusion, the two probability vectors have consistent semantic meanings in this space, facilitating the fusion sub-network's learning of the correspondence between the difference patterns and 3D poses. However, if considering the systematic differences in imaging conditions between the two viewpoints (such as different forging contour extraction quality due to contamination on one side of the viewpoint), an independent weight structure can be used, allowing each sub-network to adapt to the specific imaging characteristics of its respective viewpoint. This can be used according to actual needs.

[0060] Setting the network output as a probability vector instead of hard category labels is to preserve the confidence information of the classification, providing richer input for subsequent binocular fusion sub-networks. For example, the first sub-network outputs probabilities of "shifting left" of 0.9 and 0.51, both classified as "shifting left" under hard category labels, but reflecting drastically different judgment reliability. The former indicates the sub-network is highly confident in its judgment, while the latter indicates the sub-network hesitates between "shifting left" and "normal centering." When the binocular fusion sub-network receives both a high-confidence "shifting left" and a high-confidence "shifting left" in the first and second instances, it can clearly determine that the forging is generally shifted to the left. When receiving a contradictory combination of a low-confidence first and a high-confidence second, the fusion sub-network needs to consider more contextual information for careful reasoning. The probability vector preserves this uncertainty, making the fusion decision more refined and reliable.

[0061] The 2D pose categories cover the possible pose states of the forging in the 2D image plane: normal centering indicates that the geometric center of the forging contour is located at the center of the image and the principal axis has no obvious tilt; left and right offsets indicate overall translation of the forging; clockwise and right tilts indicate rotation of the forging around the optical axis. These categories are sufficient to describe the pose changes of the forging in a single viewpoint, and there is a clear correspondence between the category division and the 3D pose categories output by the binocular fusion subnetwork. The translation and tilt observed by the first and second cameras, when combined, can infer the overall displacement and orientation of the forging in 3D space.

[0062] In some embodiments, the binocular fusion subnetwork employs a multilayer perceptron structure containing multiple fully connected layers.

[0063] The binocular fusion subnetwork takes the joint feature vector formed by concatenating the two-dimensional pose category probability vector output by the first two-dimensional pose recognition subnetwork and the two-dimensional pose category probability vector output by the second two-dimensional pose recognition subnetwork as input. The two-dimensional pose information of the first and second perspectives contained in the joint feature is nonlinearly fused through a fully connected layer to output the three-dimensional pose category probability vector of the forging. The category corresponding to the maximum value in the three-dimensional pose category probability vector is taken as the three-dimensional pose category of the forging; The three-dimensional pose category is used to describe the pose state of the forging in three-dimensional space inside the vacuum high-temperature forging furnace, including normal, left-leaning, right-leaning, forward-leaning, backward-leaning, left-tilting, right-tilting, forward-tilting, backward-tilting, and demolding abnormality.

[0064] The binocular fusion subnetwork employs a multilayer perceptron structure instead of more complex convolutional or recurrent structures. This is based on the analysis of the input information characteristics: the input is a joint feature vector formed by concatenating the first and second probability vectors. The dimensionality depends on the number of 2D pose categories, typically only a dozen or so. The input information is highly condensed, with each component representing the confidence level of a pose category at that viewpoint. The information density is extremely high, and spatial structure is no longer included. Under these conditions, the fully connected layers already possess sufficient nonlinear mapping capabilities to complete the inference from 2D pose combination to 3D pose categories, eliminating the need to introduce spatial feature extraction structures such as convolution. The lightweight nature of the multilayer perceptron also makes the inference time of the fusion subnetwork almost negligible, without affecting the real-time performance of the overall system.

[0065] The role of the binocular fusion sub-network is to learn the mapping relationship between the combination of two-dimensional poses from two perspectives and the actual 3D pose, rather than explicitly performing geometric reconstruction. Traditional binocular methods calculate spatial coordinates from disparity maps through calibration parameters and triangulation, a process that relies on an accurate geometric model. In this method, the binocular fusion sub-network implicitly learns this mapping in a data-driven manner: during training, the network learns, through a large number of samples, that "first leftward offset and second leftward offset" maps to 3D "leftward tilt," and "first leftward tilt and second rightward tilt" maps to 3D "forward tilt" or "backward tilt," and so on. These patterns are essentially determined by binocular disparity geometry, but the network does not need to explicitly know parameters such as baseline distance and focal length; instead, it automatically summarizes the statistical correlation between probability vector combinations and 3D pose categories from the labeled data.

[0066] It is important to emphasize that transforming the measurement target from continuous numerical values ​​(such as ΔX, ΔY, ΔZ, and Euler angles) into discrete categories is not a compromise on accuracy, but rather a reasonable setting based on the actual needs of the forging site. In the vacuum high-temperature forging process, the pose detection of the forging after demolding does not require obtaining six-degree-of-freedom values ​​accurate to the millimeter or angle. It only needs to determine whether the forging is in a normal ejection state, in which direction it is deviated, and whether the tilt exceeds the safe range. This information directly determines subsequent operations: if it is normal, it proceeds to the next process; if it is deviated to the left, the robotic arm's gripping position is adjusted; if the tilt exceeds the standard, an alarm is triggered. Dividing the pose into a finite number of discrete categories fully covers the decision-making needs of the forging site. At the same time, obtaining discrete category labels is far less difficult than obtaining continuous numerical labels. It only requires manually setting the forging to different pose states and collecting samples to label the corresponding categories during the debugging phase, without the need for high-precision spatial measurement equipment in a high-temperature environment. From an engineering feasibility perspective, this provides a guarantee for the implementation of the solution.

[0067] The 3D pose categories cover the possible pose states of the forging within the furnace: Normal indicates that the forging's position and orientation are within tolerance; Leftward, Rightward, Forward, and Backward indicate the forging's overall translation in the horizontal plane; Leftward, Rightward, Forward, and Backward indicate the forging's tilt around the horizontal axis; Demolding Abnormal indicates that the forging has not properly detached from the mold or has jammed. These categories can be directly mapped to on-site operation commands or alarm signals without secondary calculation.

[0068] In some embodiments, as shown in Table 1, the correspondence between the first and second two-dimensional pose category combinations and the three-dimensional pose categories is established based on the geometric laws of binocular vision. When a forging is translated or rotated in three-dimensional space, its two-dimensional contours under two different left and right viewpoints will exhibit specific offset and tilting patterns. Table 1 provides the correspondence between typical combinations and three-dimensional pose categories, which can serve as a reference for defining training labels or constructing invalid combination sets.

[0069] Table 1. Correspondence between the first and second 2D pose category combinations and the 3D pose categories.

[0070] In the correspondence between the first and second 2D pose category combinations and the 3D pose categories, there is a mapping rule between the offset and tilt directions in the 2D image and the actual pose in 3D space, which is determined by binocular vision geometry. The specific explanation is as follows: Translation Correspondence: In this embodiment, the first and second detection units observe the forging from the left and right viewpoints, respectively. The imaging coordinate systems of the two viewpoints are mirror-symmetrical in the horizontal direction. When the forging is translated in three-dimensional space, the offset direction of the two-dimensional contour observed from the left and right viewpoints is determined by the geometric relationship between the translation direction of the forging and the optical axis directions of the two cameras. When the forging is translated to the left in the horizontal plane, the forging moves to the right relative to both cameras. Therefore, in both the first and second views, the forging contour is shown to shift to the right of the image. When the forging is translated to the right, it is shown to shift to the left of the image in both views.

[0071] Tilting Correspondence: Left-Right Tilting (Rotation about the Vertical Axis): When the forging tilts to the left in three-dimensional space, its top appears to tilt to the left in the first image and to the right in the second image. Conversely, when the forging tilts to the right, it appears to tilt to the right in the first image and to the left in the second. The conclusion is that the top tilt direction is opposite in both perspectives. Forward-Backward Tilting (Rotation about the Horizontal Axis): When the forging tilts forward, its top appears to tilt to the left in both perspectives. When the forging tilts backward, its top appears to tilt to the right in both perspectives. The conclusion is that the top tilt direction is the same in both perspectives.

[0072] Demolding Anomaly Correspondence: A demolding anomaly indicates that the forging failed to detach from the mold normally, or that the forging encountered abnormalities such as jamming or falling during ejection. In such cases, the forging contour extracted from the first or second view may be incomplete, significantly different in shape from a normal forging contour, or the forging may not be detected at all. When the above anomaly occurs on either the first or second view, it can be directly determined as a demolding anomaly, without relying on information from the other side. This judgment logic reflects the robustness of this method to the degradation of information quality on one side. Even if contour extraction fails on one side due to viewport contamination or partial occlusion, the system can still make a correct judgment based on the abnormal signal on the other side.

[0073] It should be noted that the convolutional neural network structure used in the first and second two-dimensional pose recognition subnetworks, as well as the multilayer perceptron structure used in the binocular fusion subnetwork, are preferred embodiments of the present invention, and not the only limitation on the network structure type.

[0074] The core function of the first and second 2D pose recognition subnetworks is to extract pose-related shape features from the input forging contour image and output a 2D pose category probability vector from the corresponding viewpoint. Other network structures capable of performing the above contour feature extraction and pose classification functions are also applicable to the first and second 2D pose recognition subnetworks. For example, they can be replaced with a Transformer-based visual classification network, a lightweight classification network based on an attention mechanism, or other custom convolutional neural network structures.

[0075] The core function of the binocular fusion subnetwork is to nonlinearly fuse the 2D pose probability vectors from the first and second viewpoints, outputting a 3D pose category probability vector. Other network structures capable of achieving the aforementioned dual-view probability vector fusion and 3D classification functions are also applicable to the binocular fusion subnetwork. For example, it can be replaced with a feature fusion network based on an attention mechanism, a sequence modeling network based on Transformer, or other network structures capable of multi-input fusion classification.

[0076] The specific choice of network structure and parameter configuration can be adjusted according to the requirements of inference speed, model size, and recognition accuracy in the actual deployment scenario, without departing from the core idea of ​​this invention: transforming 3D pose measurement into 2D classification plus binocular fusion. Regardless of the specific network structure adopted, the first and second 2D pose recognition subnetworks must take the forging contour image from the corresponding viewpoint as input and output a 2D pose category probability vector. The binocular fusion subnetwork must take the concatenation of the two 2D pose category probability vectors as input and output a 3D pose category probability vector, so as to keep the core idea of ​​this embodiment unchanged.

[0077] In some embodiments, the training loss of the 3D pose recognition model includes three parts: classification loss, binocular consistency loss, and 2D-3D consistency loss.

[0078] The classification loss includes the first two-dimensional pose classification loss, the second two-dimensional pose classification loss, and the three-dimensional pose classification loss; The first two-dimensional pose classification loss is the cross-entropy loss between the output of the first two-dimensional pose recognition subnetwork and the first pose truth label of the forging. The second two-dimensional pose classification loss is the cross-entropy loss between the output of the second two-dimensional pose recognition subnetwork and the ground truth label of the forging's second pose. The 3D pose classification loss is the cross-entropy loss between the output of the binocular fusion subnetwork and the ground truth label of the 3D pose of the forging.

[0079] In this embodiment, the formula for calculating the first two-dimensional pose classification loss is: L left = - Σ i yL i· log(pL i ); In the formula, yL i For the i-th component of the first pose truth value label of the forging (with a value of 0 or 1, using one-hot encoding), pL i This loss represents the probability value corresponding to the i-th category in the two-dimensional pose category probability vector output by the first two-dimensional pose recognition subnetwork. This loss is used to optimize the parameters of the first two-dimensional pose recognition subnetwork, driving its output probability distribution towards the first ground truth label set.

[0080] In this embodiment, the formula for calculating the second two-dimensional pose classification loss is: L right = - Σ i yR i · log(pR i ); In the formula, yR i For the i-th component of the second pose truth label of the forging, pR i This is the probability value corresponding to the i-th category in the two-dimensional pose category probability vector output by the second two-dimensional pose recognition subnetwork. This loss is used to optimize the parameters of the second two-dimensional pose recognition subnetwork.

[0081] In this embodiment, the formula for calculating the 3D pose classification loss is: L 3d = - Σ k y3d k · log(p3d k ); In the formula, y3d k p3d is the k-th component of the three-dimensional pose truth label of the forging. k This is the probability value corresponding to the k-th class in the 3D pose class probability vector output by the binocular fusion subnetwork. This loss is used to optimize the parameters of the binocular fusion subnetwork, and its gradient is backpropagated through the binocular fusion subnetwork to the outputs of the first and second 2D pose recognition subnetworks, indirectly driving the outputs of the two subnetworks to have a probability distribution that is more conducive to 3D fusion.

[0082] The classification loss is a fundamental component of the training loss of a 3D pose recognition model. It consists of three sub-items, each corresponding to the independent supervision of a sub-network.

[0083] The first 2D pose classification loss is used to optimize the parameters of the first 2D pose recognition sub-network. This loss is calculated using the cross-entropy between the 2D pose category probability vector output by the first sub-network and the ground truth label of the forging's first pose. Cross-entropy loss is characterized by clear gradients and stable convergence in multi-classification tasks, driving the probability distribution output by the network towards the ground truth label. Constrained by this loss, the first sub-network progressively learns to extract pose-related shape features from the single-view contour image and establishes a mapping relationship between features and 2D pose categories.

[0084] The mechanism of the second 2D pose classification loss is symmetrical to that of the first, and it is used to optimize the parameters of the second 2D pose recognition sub-network. The first and second sub-networks independently calculate their classification losses because, although the input images from the two viewpoints are geometrically related, they are independent observations at the pixel level, requiring each to establish its own ability to recognize contours and pose categories from its own viewpoint. When the two sub-networks adopt a Siamese structure with shared weights, the gradients of the first and second classification losses are accumulated and jointly updated to update the shared parameters, allowing the same set of convolutional kernels to simultaneously serve contour feature extraction from both viewpoints.

[0085] The 3D pose classification loss is used to optimize the parameters of the binocular fusion sub-network. Simultaneously, the gradient of this loss is propagated through the binocular fusion sub-network to the outputs of the first and second sub-networks via backpropagation, indirectly driving the outputs of the two 2D pose recognition sub-networks to have a probability distribution more favorable for 3D fusion. This indirect supervision mechanism ensures that the first and second sub-networks not only pursue classification accuracy from their own perspectives but also adaptively adjust their output probability distributions during training, providing more discriminative information for binocular fusion.

[0086] The weighting of the three classification losses in the total training loss determines the convergence priority of each sub-network. Typically, the first and second classification losses have equal weights, allowing the two sub-networks to learn at the same speed. The weights of the 3D classification loss can be adjusted appropriately. In the early stages of training, a smaller weight is given to avoid excessive gradient interference from the fusion sub-network before the 2D sub-network has converged. In the later stages of training, the weights are gradually increased to shift the optimization focus towards the final 3D recognition accuracy.

[0087] It should be noted that the training ground truth labels for the three sub-networks are all discrete category labels, not continuous values. The ground truth labels for the first and second poses are two-dimensional pose categories, and the ground truth labels for the three-dimensional pose categories. Obtaining these discrete category labels only requires manually setting the pose state of the forging and collecting sample images for classification and labeling during the debugging phase. This eliminates the need for high-precision spatial measurement equipment in high-temperature environments, ensuring the engineering feasibility of the solution from the data acquisition perspective.

[0088] In some embodiments, the binocular consistency loss is used to penalize invalid two-dimensional pose combinations that violate binocular geometric relationships when the outputs of the first two-dimensional pose recognition subnetwork and the second two-dimensional pose recognition subnetwork violate binocular geometric relationships. The binocular consistency loss is calculated by summing the products of the first predicted probability and the second predicted probability corresponding to each combination in the predefined set of invalid two-dimensional pose combinations. For any combination (i,j) in the set of invalid two-dimensional pose combinations, the first predicted probability is the probability corresponding to the i-th category in the probability vector output by the first two-dimensional pose recognition subnetwork, and the second predicted probability is the probability corresponding to the j-th category in the probability vector output by the second two-dimensional pose recognition subnetwork. In this embodiment, the formula for calculating the binocular consistency loss is: ; In the formula, C is the set of invalid two-dimensional pose combinations. Each element in C is a two-dimensional pose category index combination (i,j), indicating that the first output of the i-th two-dimensional pose category and the second output of the j-th two-dimensional pose category are invalid combinations that violate the binocular geometric relationship. Let be the probability value corresponding to the i-th category in the two-dimensional pose category probability vector output by the first two-dimensional pose recognition subnetwork. This is the probability value corresponding to the j-th category in the two-dimensional pose category probability vector output by the second two-dimensional pose recognition subnetwork.

[0089] Binocular consistency loss is a unique constraint term in the training loss of this method. Its function is not to pursue complete consistency between the outputs of the first and second sub-networks, but to encode the geometric laws of binocular vision into soft constraints during the training process, thereby suppressing pose combinations of the two sub-networks that violate physical laws.

[0090] After the first and second detection units independently acquire images and extract the forging contours respectively, the two sub-networks independently perform 2D pose classification. Since the networks haven't yet learned binocular geometric constraints in the early training phase, contradictory outputs may occur, such as "the first judgment is that the forging is shifted to the right, and the second judgment is that the forging is shifted to the left." According to the geometric laws of binocular parallax, when the forging translates left and right in the horizontal plane, the contour shift directions observed by the two viewpoints should be the same; if shifts occur in opposite directions, this combination violates the basic binocular geometric relationship. The binocular consistency loss penalizes such invalid combinations, guiding the two sub-networks to output pose judgments that conform to binocular geometric logic while pursuing their own classification accuracy.

[0091] The method for calculating this loss has the following characteristics. First, it is the sum of the products of probabilities of invalid combinations, rather than directly imposing constraints on the probabilities of a single class.

[0092] When the probability of the first sub-network outputting a certain invalid category is close to zero, regardless of the output of the second sub-network, the product of these terms will approach zero, resulting in minimal contribution to the loss. This means that the loss only incurs a significant penalty when both sub-networks simultaneously assign non-negligible probabilities to invalid combinations, avoiding excessive restrictions on normal exploration. Secondly, this loss does not introduce additional truth labels. The invalid combination set C is predefined based on the arrangement of the first and second detection units and the degrees of freedom of the forging motion, as shown in Table 2. This is a one-time task and does not require labeling each sample as production conditions change. The definition of set C is based on the geometric law of binocular parallax. Specifically: when the forging translates left and right in the horizontal plane, the first and second offset directions should be the same. Therefore, "first offset to the left, second offset to the right" and "first offset to the right, second offset to the left" are invalid combinations. When the forging rotates around the vertical axis, the first and second tilt directions should be opposite. Therefore, "first tilt to the left, second tilt to the left" and "first tilt to the right, second tilt to the right" are invalid combinations. By solidifying these geometric patterns into a set of invalid combinations, binocular consistency loss achieves the injection of physical priors without increasing annotation costs.

[0093] From a training strategy perspective, the binocular consistency loss can be set with a smaller weight in the early stages of training, allowing the two sub-networks to prioritize learning contour features and basic classification capabilities from their respective perspectives. In the later stages of training, the weight is gradually increased, allowing physical constraints to take effect based on the model's existing discriminative ability, avoiding premature constraints that limit the network's learning freedom. This phased weight adjustment strategy helps achieve a balance between classification accuracy and physical consistency.

[0094] Table 2 Invalid 2D Pose Combination Set C

[0095] The invalid 2D pose combination set C listed in Table 2 represents the combinations penalized by the binocular consistency loss. The definition of set C is based on the actual arrangement of the first and second detection units on the vacuum high-temperature forging furnace and the binocular disparity geometry, and can be adjusted according to the specific arrangement. When both the first and second sub-networks simultaneously output the pose category from the above combination, it indicates that the judgments of the two sub-networks violate the binocular geometric constraints, and the binocular consistency loss will penalize the probability product corresponding to this combination.

[0096] In some embodiments, the 2D-3D consistency loss is used to penalize invalid triplets where there is a logical contradiction between the 3D pose category output by the binocular fusion subnetwork and the two input 2D pose categories. The 2D-3D consistency loss is calculated by summing the products of the first prediction probability, the second prediction probability, and the 3D prediction probability corresponding to each combination in the predefined set of invalid triplets. For any combination (i,j,k) in the set of invalid triplets, the first prediction probability is the probability corresponding to the i-th category in the probability vector output by the first 2D pose recognition subnetwork, the second prediction probability is the probability corresponding to the j-th category in the probability vector output by the second 2D pose recognition subnetwork, and the 3D prediction probability is the probability corresponding to the k-th category in the probability vector output by the binocular fusion subnetwork. In this embodiment, the formula for calculating the 2D-3D consistency loss is: ; In the formula, S is a set of invalid ternary combinations. Each element in S is a ternary combination (i,j,k) with a pose category index, indicating that the first output of the i-th 2D pose category, the second output of the j-th 2D pose category, and the binocular fusion subnetwork output of the k-th 3D pose category are invalid combinations with logical contradictions. This represents the probability value corresponding to the k-th class in the 3D pose class probability vector output by the binocular fusion subnetwork.

[0097] The 2D-3D consistency loss is the second unique constraint in the training loss of this method. Its role is to establish logical self-consistency between the input and output of the binocular fusion sub-network, and to prevent semantic contradictions between the output of the fusion network and the 2D input in the 3D pose judgment.

[0098] In this embodiment, the binocular fusion sub-network receives the two-dimensional pose category information output by the first and second sub-networks and infers the three-dimensional pose accordingly. If the first and second sub-networks both determine that the forging is "normally centered," then the fusion network outputting "leftward" is clearly unreasonable, as no offset is observed in either viewpoint, and an offset is impossible in three-dimensional space. Similarly, if both the first and second sub-networks determine that the forging is "offset to the right," then the fusion network outputting "rightward" or "normal" violates basic correspondence logic. The two-dimensional-three-dimensional consistency loss ensures that the output of the fusion network remains semantically consistent with its input information by penalizing such contradictory ternary combinations.

[0099] Unlike the relationship between the first and second constraints in binocular consistency loss, the 2D-3D consistency loss constrains the relationship between the input (2D pose combination) and the output (3D pose). Together, they constitute a complete logical guarantee from input to output: binocular consistency loss ensures that the 2D judgments of the two perspectives conform to binocular geometry, and 2D-3D consistency loss ensures that the fused network does not make logical errors based on correct input. This dual constraint approach, combining front-end geometric constraints with back-end logical constraints, allows the model to output physically reasonable and logically consistent results even without a large amount of labeled data.

[0100] The calculation of the 2D-3D consistency loss also exhibits soft-constraint characteristics. The loss value is only significant when the first probability, second probability, and 3D probability simultaneously assign non-negligible probabilities to invalid combinations; any probability approaching zero will cause the contribution of this loss term to approach zero. The set S of invalid ternary combinations is predefined by those skilled in the art based on the semantic and logical relationship between the 2D pose category and the 3D pose category. Specifically, combinations where the 3D pose category is "normal" and the first or second output is offset or tilted, combinations where the 3D pose category is "leftward" and neither the first nor second output is "rightward offset," and any other combinations other than those listed in Table 1 that are valid correspondences can be classified into the set S of invalid ternary combinations, as shown in Table 3 (a portion is shown, not the entire set). The definition of this set is completed once and does not require sample-by-sample labeling as production conditions change, improving the model's generalization ability and interpretability without increasing data acquisition costs.

[0101] Table 3. Set of Invalid Triple Combinations S (Excerpt)

[0102] The invalid triplet combination set S listed in Table 3 represents the combinations penalized in the 2D-3D consistency loss. "Any" means that the position can take any valid 2D or 3D pose category. The definition of set S is based on the inverse set of the valid correspondences in Table 1; that is, any triplet combination not in the valid correspondences in Table 1 is an invalid combination. When the combination between the 3D pose category output by the binocular fusion subnetwork and the two input 2D pose categories belongs to set S, it indicates a logical contradiction between the output and input information of the fusion network. The 2D-3D consistency loss will penalize the product of the three probabilities corresponding to this combination.

[0103] The specific settings of the two-dimensional pose category and the three-dimensional pose category in this embodiment can be flexibly adjusted according to the actual process requirements of the vacuum high-temperature forging furnace and the degree of freedom of movement of the forging.

[0104] In some embodiments, the annotation method for the three-dimensional pose truth value label of the forging, the first pose truth value label of the forging, and the second pose truth value label of the forging includes: The forging is placed in a vacuum high-temperature forging furnace and subjected to various positions. The first and second detection units acquire the first and second multi-directional polarization original images as the first and second sample images. The forging region is segmented from the first sample image and the second sample image respectively, and the forging contour is extracted from the segmented forging region to obtain the first sample contour and the second sample contour. The first and second sample contours are manually annotated. The first pose truth label of the forging is assigned according to the pose state of the forging in the two-dimensional image plane under the first view. The second pose truth label of the forging is assigned according to the pose state of the forging in the two-dimensional image plane under the second view. The preset pose state of the forging in the vacuum high-temperature forging furnace when the first and second sample images of the set are collected is used as the three-dimensional pose truth value label of the forging. In this embodiment, the first sample contour, the second sample contour, the ground truth label of the first pose of the forging, the ground truth label of the second pose of the forging, and the ground truth label of the three-dimensional pose of the forging together constitute a sample in the dataset for training the three-dimensional pose recognition model.

[0105] The preset pose states cover various pose conditions that may occur in actual forging production, including normal states and abnormal states such as leftward, rightward, forward, backward, leftward tilt, rightward tilt, forward tilt, and backward tilt. By manually setting the pose state of the forging during the debugging phase, the actual pose of the forging in three-dimensional space can be determined simultaneously with data acquisition, and this pose can be used as a three-dimensional pose truth label. This approach avoids the difficulty of using external measuring equipment to record the spatial coordinates of the forging in high-temperature environments. The three-dimensional truth label comes directly from the known preset state of the forging at the time of acquisition, rather than from post-measurement or calculation.

[0106] After acquiring the first and second sample images, forging region segmentation and contour extraction were performed on them respectively to obtain the first sample contour and the second sample contour. Segmentation was based on the difference in polarization degree to ensure stable separation of the forging from the background under isothermal and brightness conditions. The first and second segmentation processes were performed independently, each generating a forging contour image from the corresponding viewpoint.

[0107] The annotation of the first and second pose truth labels is directly based on the contour images of the corresponding viewpoints: the annotator observes the pose state of the forging in the first sample contour and assigns the first pose truth label according to the offset and tilt direction of the forging in the two-dimensional image plane under the first viewpoint; the annotator observes the pose state of the forging in the second sample contour and assigns the second pose truth label according to the offset and tilt direction of the forging in the two-dimensional image plane under the second viewpoint. This annotation method only requires the annotator to judge the offset and tilt direction of the contour, such as whether the forging contour is biased to the left or right in the image, and whether the principal axis is tilted to the left. It is an intuitive judgment at the visual level, with low annotation difficulty and good consistency.

[0108] The first sample contour, the second sample contour, the first pose ground truth label, the second pose ground truth label, and the 3D pose ground truth label are combined into a single training sample. This sample provides complete supervision signals for the three sub-networks: the first sample contour and the first pose ground truth label are used to train the first 2D pose recognition sub-network, the second sample contour and the second pose ground truth label are used to train the second 2D pose recognition sub-network, and the probability vectors extracted from the first and second sample contours by their respective sub-networks are input into the binocular fusion sub-network, which is then fused and output under the supervision of the 3D pose ground truth label.

[0109] It should be noted that all truth labels in this annotation method are discrete category labels, not continuous numerical values. This annotation method ensures good consistency in category label annotation, with different annotators showing a high degree of consistency in their judgments of "leaning left" and "leaning right." In contrast, the annotation of continuous numerical values ​​can lead to deviations between annotators and within the annotator themselves at different times. Moreover, obtaining category labels does not require precision measuring equipment; it only requires manually setting the forging pose and recording the corresponding category during the debugging phase, simplifying engineering implementation. Finally, the category labels directly correspond to the final application requirements. What is needed on-site in forging is a qualitative judgment of "which way it's leaning," rather than a precise numerical offset. The training objective and inference objective are completely consistent with the category labeling objective.

[0110] In some embodiments, the two sampling points are two through holes with symmetrical viewing angles on the wall of the vacuum forging furnace. A polarization detection system is provided at each sampling point to form two independent detection units. The main optical axes of the two independent detection units intersect in the position area where the forging is expected to be lifted in the furnace. The two detection units include a first detection unit corresponding to the first viewing angle and a second detection unit corresponding to the second viewing angle. The polarization detection system includes a window assembly, a coaxial beam splitting and combining assembly, an active linearly polarized laser emission module, and a polarization filtering imaging receiver module.

[0111] In some embodiments, the window assembly is embedded in a through hole in the furnace wall of a vacuum high-temperature forging furnace, and the coaxial beam splitter is disposed outside the furnace of the window assembly to guide the linearly polarized laser emitted by the active linearly polarized laser emission module into the furnace and to guide the reflected light from the furnace into the polarization filter imaging receiver module.

[0112] In some embodiments, the polarization filter imaging receiver module includes a narrowband bandpass filter and a focal plane polarization image sensor. The center wavelength of the narrowband bandpass filter is matched with the laser wavelength of the active linearly polarized laser emission module, and the bandwidth does not exceed 3nm. The pixel of the split-focus plane polarization image sensor integrates a micro-polarization array in four directions: 0°, 45°, 90°, and 135°, for synchronous acquisition of a first multi-directional polarization original image or a second multi-directional polarization original image in a single exposure.

[0113] The first and second multi-polarized original images are obtained by irradiating the forging and background area with linearly polarized laser light emitted by the active linearly polarized laser emitting modules of the first and second detection units, and by collecting the reflected light through their respective polarization filtering imaging receiving modules.

[0114] This step employs active linearly polarized laser illumination instead of relying on the forging's own thermal radiation because, under isothermal conditions of 1000–1500°C, the blackbody radiation emitted by the forging, mold, and furnace wall is identical in brightness and spectrum, making it impossible to extract distinguishable information from it. Active laser illumination introduces a narrowband signal source at a specific wavelength, significantly higher than the blackbody radiation background, providing a signal basis for subsequent polarization analysis. The laser's polarization extinction ratio is typically no less than 100:1 to ensure the determinism of the incident polarization state.

[0115] The first and second detection units simultaneously acquire data from two different viewpoints. The purpose is to obtain the two-dimensional contour information of the forging from these two viewpoints, providing input for the subsequent binocular fusion sub-network to infer the three-dimensional pose using the pose differences between the two viewpoints. Unlike traditional binocular stereo vision, which requires precise calibration of the relative pose relationship between the left and right cameras and pixel-level stereo matching, in this embodiment, the relative positional relationship between the first and second detection units only needs to ensure sufficient parallax between the two viewpoints so that different three-dimensional poses present distinguishable two-dimensional contour differences from the left and right viewpoints, without needing to obtain precise camera extrinsic parameters. This reduces the requirements for calibration accuracy.

[0116] In step S1 of this embodiment, the polarization detection system is a system capable of acquiring multi-directional polarized original images formed by lasers reflected from the forging and background areas in different polarization directions. Its implementation mainly relies on active linear polarization laser illumination technology, which utilizes the physical difference between the specular reflection of the forging surface retaining high polarization and the diffuse reflection of the mold and furnace wall surface depolarizing, to provide the original data basis for subsequently distinguishing the forging from the background from the polarization dimension.

[0117] A preferred embodiment of a polarization detection system is further provided: such as Figure 2 As shown, the polarization detection system in this embodiment mainly consists of four core functional units: window assembly 3, coaxial beam splitting and combining assembly 5, active linear polarization laser emission module 4, and polarization filtering imaging receiving module 6. The four are connected in series through optical paths to form a complete polarization detection channel from outside the furnace to inside the furnace and back to outside the furnace.

[0118] like Figure 4 , Figure 6 As shown, the viewing window assembly 3 is embedded in the furnace wall through-hole 2 of the vacuum high-temperature forging furnace 1, serving as the only optical interface between the high-temperature environment inside the furnace and the ambient temperature environment outside. The viewing window assembly 3 is made of high-temperature resistant and light-transmitting sapphire material. Its mechanical strength and optical performance remain stable within a working temperature range of 1000~1500℃. It exhibits high transmittance and minimal attenuation for 808nm and 1064nm wavelength lasers, has a Mohs hardness of 9, and good wear resistance, enabling it to withstand the erosion of metal vapors and high-temperature dust inside the furnace.

[0119] The normal direction of the window assembly 3 defines the direction of the main optical axis of the device. After passing through the window assembly 3, the main optical axis points to the area where the forging is located in the furnace through the coaxial optical through hole.

[0120] The inner surface of the window assembly 3 is coated with an anti-metal evaporation protective coating layer 32. This coating layer is a SiO2 / TiO2 multilayer high-temperature resistant dielectric film with a transmittance of not less than 96% for the working laser wavelength (1064nm). At the same time, it reduces the adhesion of metal vapor to the sapphire surface and extends the effective service life of the window.

[0121] A vacuum sealing structure 33 is provided between the viewing window assembly 3 and the furnace wall through hole 2, including an oxygen-free copper vacuum sealing gasket and a floating elastic compression flange. The oxygen-free copper sealing gasket undergoes controllable deformation under the pre-tightening force of the disc spring group of the elastic compression flange, forming a high-vacuum sealing contact surface with the outer peripheral surface of the sapphire transparent substrate and the inner wall surface of the furnace wall through hole. The adaptive deformation capability of the elastic element can compensate for the axial and radial thermal expansion displacement generated by the furnace body during the thermal cycle of 1000-1500℃, maintaining the sealing reliability across the entire temperature range.

[0122] The coaxial beam splitter / combiner 5 is located outside the furnace of the viewing window assembly 3 and is the core optical coupling element of this device. For example... Figures 3-5As shown, the coaxial beam splitter / combiner 5 has three optical ports: First port 5a: facing the active linearly polarized laser emission module 4, serving as the laser incident port; Second port 5b: Oriented toward polarization filter imaging receiver module 6, it is the reflected light output port; The third port 5c: facing the window assembly 3, along the main optical axis, is the optical port leading into the furnace.

[0123] The first port 5a and the second port 5b are arranged in mutually orthogonal directions, and the third port 5c is arranged along the main optical axis. The function of the coaxial beam splitting and combining component 5 is to guide the laser light input from the first port 5a to the output of the third port 5c, and then through the viewing window into the furnace; and to guide the reflected light returning from the third port 5c to the output of the second port 5b, and then into the polarization filtering imaging receiving module 6.

[0124] The laser emitter and the imaging receiver need to share the same optical path (coaxial) passing through the furnace wall to simplify the structure, reduce the window size, reduce the opening area of ​​the furnace wall, reduce the impact on the structural strength of the furnace body and the vacuum seal, and at the same time ensure that the incident light and the reflected light are always accurately aligned.

[0125] However, the emitted and received light must be effectively separated to prevent high-power laser light from directly entering the imaging sensor and causing saturation or permanent damage. Therefore, an optical beam splitter / combiner is needed to selectively split the incident and reflected light by utilizing the difference in polarization states, guiding them to their respective ports.

[0126] An active linearly polarized laser emitting module 4 is installed outside the first port 5a of the coaxial beam splitter / combiner 5, with its emitting end facing the first port 5a. The active linearly polarized laser emitting module 4 emits linearly polarized laser light with a wavelength of 808 nm or 1064 nm. The emitted linearly polarized laser light enters the coaxial beam splitter / combiner 5 from the first port 5a, is guided to the third port 5c for output, and then passes through the window assembly 3 along the main optical axis into the furnace.

[0127] This device employs "active" linearly polarized laser illumination instead of relying on the thermal radiation of the forging itself. In an isothermal environment of 1000–1500°C, the blackbody radiation emitted by the forging, mold, and furnace wall is identical in brightness and spectrum, making them difficult to distinguish. Active laser illumination introduces a narrowband signal source at a specific wavelength (e.g., 1064 nm) that is significantly higher than the blackbody radiation background, providing a signal basis for subsequent narrowband filtering and polarization analysis. Simultaneously, the laser is required to have highly linearly polarized characteristics (polarization extinction ratio ≥ 100:1) because only when the polarization state of the incident light is determined and known can the forging and background be distinguished by analyzing the changes in the polarization state of the reflected light (spectral reflection maintains polarization contrast, diffuse reflection depolarizes).

[0128] The polarization-filtered imaging receiver module 6 is mounted outside the second port 5b of the coaxial beam splitter / combiner assembly 5, with its receiving end facing the second port 5b. The polarization-filtered imaging receiver module 6 receives and analyzes the polarization characteristics of the light output from the second port 5b, and outputs polarization image data.

[0129] The light output from port 5b has been polarization-selectively split by the coaxial beam splitter, containing a highly polarized signal reflected from the forging specular surface and a low-polarized signal diffusely reflected from the mold / furnace wall. The polarization filtering imaging receiver module further performs narrowband filtering (to suppress blackbody radiation noise) and polarization imaging (to quantize the polarization degree of each pixel) on this light signal, providing the image data basis for subsequent polarization degree thresholding.

[0130] The linearly polarized laser emitted by the active linearly polarized laser emission module 4 passes sequentially through the first port 5a and the third port 5c, through the window assembly 3, and enters the furnace along the main optical axis to irradiate the surface of the forging. The reflected light from the surface of the forging and the background inside the furnace (mold, furnace wall) returns along the original path along the main optical axis (this is a common geometric characteristic of specular reflection and diffuse reflection; under near-normal incident conditions, the reflected light returns near the incident direction), and then sequentially enters the polarization filtering imaging receiving module 6 through the third port 5c and the second port 5b.

[0131] The incident and reflected light paths share the same principal optical axis segment passing through the viewing window, achieving "coaxiality". Utilizing the geometric property that reflected light returns near the incident direction, only one viewing window is needed to simultaneously complete laser incident and reflected light reception.

[0132] Therefore, the device of the present invention, by embedding a high-temperature resistant light-transmitting material window assembly in the through hole of the furnace wall and setting a coaxial beam splitting and combining assembly on the outside of the window furnace, couples the active linear polarization laser emitting module and the polarization filtering imaging receiving module to the same window, sharing the same main optical path passing through the furnace wall, thus constructing a complete polarization detection channel from outside the furnace to inside the furnace and back. Furthermore, by utilizing the physical difference between the polarization characteristics preserved by the specular reflection of the forging surface and the depolarization by the diffuse reflection of the mold / furnace wall, it achieves accurate differentiation between the forging and the background under isothermal, same brightness and same spectrum conditions.

[0133] like Figures 3-5 As shown, the coaxial beam splitter and combiner 5 includes a polarization beam splitter 51 and a quarter-wave plate 52.

[0134] The polarization beam splitter 51 is located at the intersection of the first port 5a, the second port 5b, and the third port 5c, i.e., the geometric intersection center of the three-port optical path. The quarter-wave plate 52 is located on the principal optical axis between the polarization beam splitter 51 and the window assembly 3, downstream of the polarization beam splitter 51 and upstream of the window assembly 3.

[0135] In this embodiment, the combination of a polarization beam splitter and a quarter-wave plate is a classic optical architecture for achieving coaxial transmit and receive separation, which is widely used in optical systems such as LiDAR, confocal microscopes, and optical coherence tomography (OCT).

[0136] The isolation of incident and reflected light is achieved by two orthogonal transformations of polarization states: linearly polarized light → quarter-wave plate → circularly polarized light → target reflection → circularly polarized light → quarter-wave plate → orthogonally linearly polarized light. The orthogonally linearly polarized light is split into different directions by a polarization beam splitter.

[0137] In this embodiment, the architecture not only achieves separation of transmission and reception, but also combines the specular reflection characteristics of the forging surface and the diffuse reflection characteristics of the mold / furnace wall surface to provide an inherent contrast enhancement effect between the forging and the background before it enters the imaging sensor.

[0138] In some embodiments, such as Figure 5 As shown, the polarization beam splitter 51 is a polarization beam-splitting cube prism (PBS Cube), which is formed by bonding two right-angle prisms along their diagonal surfaces. The bonding diagonal surfaces are coated with a polarization-selective multilayer dielectric film (beam-splitting surface). The polarization beam-splitting cube prism is placed at the intersection center of the three ports.

[0139] The beam splitter 51 has high transmittance (T) for linearly polarized light incident along the principal optical axis in the first polarization direction (denoted as P polarization direction - the electric field vector is located in the incident plane). p ≥95%), causing it to pass through the beam-splitting surface and exit from the third port 5c along the principal optical axis. It has high reflectivity (R) for linearly polarized light incident from the third port 5c along the principal optical axis and with a second polarization direction orthogonal to the first polarization direction (denoted as the S polarization direction – the electric field vector is perpendicular to the incident plane). s (≥95%), causing it to be reflected by the beam splitter and rotated 90°, outputting from the second port 5b.

[0140] In this embodiment, the beam-splitting surface of the polarization beam-splitting cubic prism is formed by alternating layers of high and low refractive index dielectric films, and its optical thickness satisfies the Brewster angle condition, so that P-polarized light is transmitted with extremely low loss and S-polarized light is reflected with high reflectivity.

[0141] The cubic prism is chosen over the flat polarization beam splitter because both transmitted and reflected light from the cubic prism propagate in a straight line along the inside of the prism, and the exit surface is parallel to the incident surface, which does not introduce beam deflection or astigmatism, making it suitable for polarization imaging systems that require high beam quality.

[0142] The P / S polarization direction is defined to be consistent with the incident surface of the beam splitter, ensuring that the P-polarized light output by the laser emission module is efficiently transmitted and the reflected S-polarized light is efficiently reflected to the imaging module.

[0143] In some embodiments, the fast axis direction of the quarter-wave plate 52 forms a 45° angle with the polarization direction (i.e., the P-polarization direction) of the linearly polarized laser emitted by the active linearly polarized laser emission module 4 (in actual engineering, an installation tolerance of ±3° is allowed, i.e., 42°~48°).

[0144] The condition for a quarter-wave plate to convert linearly polarized light into circularly polarized light is that the polarization direction of the linearly polarized light forms a 45° angle with the fast axis of the wave plate. At this time, the amplitudes of the components of the linearly polarized light along the fast and slow axes are equal, and the fast-axis component passes through the wave plate before the slow-axis component, introducing a 90° phase difference (i.e., a quarter-wavelength optical path difference) between them. The resulting light is circularly polarized.

[0145] The specific derivation is as follows: Let the electric field of the incident ray-polarized light be... The polarization direction makes a 45° angle with the fast axis. On the fast axis... and slow axis The directional components are: ; ; After passing through a quarter-wave plate, the fast-axis component produces a relative output to the slow-axis component. (90°) Phase advance: ; ; The two beams have a 90° phase difference and equal amplitude, and when combined, they produce circularly polarized light. The direction of rotation of circularly polarized light depends on the relative orientation of the fast axis and the incident polarization direction.

[0146] When circularly polarized light is reflected by the mirror on the surface of the forging and then passes through the quarter-wave plate again, the component of the circularly polarized light in the fast and slow axis directions introduces a 90° phase difference again, and the cumulative phase difference becomes 180°. After synthesis, it becomes linearly polarized light, and its polarization direction is orthogonal to the incident direction, that is, it changes from P polarization to S polarization.

[0147] If the angle between the fast axis and the polarization direction deviates by 45°, the amplitudes of the fast and slow axis components will be unequal, resulting in elliptically polarized light rather than ideally circularly polarized light. After reflection and a second passage through the waveplate, although the polarization direction of the outgoing light remains approximately orthogonal to the incident light, the polarization purity decreases. This causes some light energy to be transmitted to the first port (returning to the laser direction) by the polarization beam splitter instead of being reflected to the second port (imaging direction), reducing system efficiency. An installation tolerance of ±3° is acceptable in engineering, and the decrease in polarization purity is within 5%.

[0148] Example 2 describes the complete optical path process and polarization state change from laser emission to the reflected light entering the imaging module, which is the physical core of polarization detection in this invention. The following, in conjunction with... Figure 7 Please provide a detailed explanation.

[0149] The light path (path A) of specular reflection from the surface of the forging: The linearly polarized laser emitted by the active linearly polarized laser emission module 4 (the first linearly polarized laser) enters the polarization beam splitter 51 with the first polarization direction (P-polarization direction). The polarization beam splitter 51 has high transmittance for P-polarized light, which is transmitted and output from the third port 5c along the principal optical axis and directed towards the quarter-wave plate 52.

[0150] When P-polarized light passes through a quarter-wave plate 52, it is converted into circularly polarized light (assuming right-handed circular polarization) because the fast axis is at a 45° angle to the polarization direction.

[0151] Right-hand circularly polarized light passes through the viewing window assembly 3, enters the furnace cavity through the coaxial optical through hole, and irradiates the surface of the forging.

[0152] Specular reflection occurs on the surface of a forging under near-normal incident conditions. The physical essence of specular reflection is coherent scattering of light on a smooth surface. The polarization state of the reflected light is preserved (not randomized), but the rotation direction of circularly polarized light is reversed during reflection; right-handed circular polarization becomes left-handed circular polarization. This is because reflection introduces an additional π-phase change, altering the rotation direction of circular polarization but not changing its essential characteristics.

[0153] The left-hand circularly polarized light returns along the original optical path (principal optical axis), passes through the window assembly 3, and then passes through the quarter-wave plate 52 again. The quarter-wave plate 52 introduces another 90° phase difference into the left-hand circularly polarized light. After superimposing with the 180° phase change introduced by reflection, the total phase difference between the fast and slow axis components is 270° (equivalent to -90°). After synthesis, it becomes linearly polarized light (the second linearly polarized laser), and its polarization direction is orthogonal to the incident P direction—that is, it becomes the S polarization direction.

[0154] S-polarized light enters polarization beamsplitter 51, which has high reflectivity for S-polarized light, and reflects it 90°, outputting it from the second port 5b and entering polarization filtering imaging receiver module 6.

[0155] Energy efficiency analysis: Assuming the polarization beam splitter has a 95% transmittance for P-light and a 95% reflectance for S-light, a 97% transmittance for the quarter-wave plate, and a 90% transmittance for the viewing window, the total transmission efficiency of the light reflected from the forged mirror from the laser emission point to the imaging sensor is approximately: η 锻件 ≈0.95 (PBS transmission) × 0.97 (QWP) × 0.90 (viewing window) × ρ 锻件 (Forging reflectivity) × 0.90 (Viewport return) × 0.97 (QWP return) × 0.95 (PBS reflectivity) ≈ 0.69 × ρ 锻件 .

[0156] Where, ρ 锻件 The specular reflectivity of the forging surface is typically 0.3-0.6 under 1064nm wavelength and high-temperature metal surface conditions.

[0157] The light path (path B) of diffuse reflection from the mold / furnace wall: The process from laser emission to irradiating the inner surface of the furnace is the same as path A: the P-linearly polarized laser is transmitted through a polarization beam splitter, converted into right-hand circularly polarized light by a quarter-wave plate, passes through the viewing window and enters the furnace to irradiate the mold and furnace wall surface.

[0158] The rough surfaces of the mold and furnace walls, along with the presence of an oxide layer, produce diffuse reflection. The physical essence of diffuse reflection is the incoherent scattering of light on a rough surface. The random orientation of numerous microscopic scattering units randomizes (depolarizes) the polarization state of the reflected light. Specifically, diffusely reflected light contains both S-polarization and P-polarization components, as well as various elliptic polarization states, and the phase relationships of these polarization states are random.

[0159] The reflected light, in these randomly mixed polarization states, returns along the original optical path (diffuse reflection component near the normal incident direction), passing through window assembly 3 and quarter-wave plate 52. Since the incident light is already in a randomly polarized state, the 90° phase difference introduced by the quarter-wave plate to the linearly polarized component of the randomly polarized state only changes the polarization state form of each component (linear polarization → elliptic polarization, elliptic polarization → another type of elliptic polarization), but does not change the randomness of the polarization state. Therefore, the light after passing through the quarter-wave plate is still in a randomly mixed polarization state.

[0160] When light with randomly mixed polarization states enters the polarization beam splitter 51, the polarization beam splitter splits the light according to the polarization direction: The S-polarization component is reflected to the second port 5b and enters the polarization filtering imaging receiver module 6. The P-polarization component is transmitted back to the first port 5a, that is, back to the laser direction, and is attenuated by the isolator or optical path inside the laser. The remaining elliptical polarized light is distributed to the two ports according to the S1 component of the Stokes parameter.

[0161] For completely depolarized diffuse reflection light ( Its S-polarization component and P-polarization component have equal energy. The polarization beam splitter reflects about 50% of the energy to the second port 5b and transmits about 50% of the energy back to the first port (5a).

[0162] Energy efficiency analysis: Only about 50% of the diffuse reflected light reaches the imaging sensor after passing through the polarization beam splitter, while nearly 100% of the specular reflected light reaches the imaging sensor. Therefore, before entering the imaging sensor, the coaxial beam splitter and combiner already provides an inherent contrast enhancement effect of approximately 2:1 (100%:50%), making the signal intensity in the forging region approximately twice that of the background region, thus laying a high signal-to-noise ratio foundation for subsequent polarization degree thresholding.

[0163] This enhanced contrast is the key to the innovative application of the classic transceiver separation optical path architecture to the polarization detection of high-temperature forgings in this invention. In traditional LiDAR or OCT systems, polarization beamsplitters and quarter-wave plates are only used for transceiver isolation, without distinguishing between target and background signals. However, in this embodiment, the physical characteristics of the forging's specular reflection preserving polarization and the mold / furnace wall's diffuse reflection depolarizing perfectly couple with the polarization-selective beam splitter's characteristics. This allows the same optical structure to achieve selective enhancement of the forging and background signals while simultaneously completing transceiver separation.

[0164] Polarization degree imaging analysis: The light output from port 5b enters the polarization-filtered imaging receiver module 6. This light contains two signals: 1. Highly polarized light reflected from the mirror surface of the forging: After a complete P→circular polarization→S polarization conversion path, it reaches the sensor with nearly 100% efficiency, and the polarization direction is highly consistent (S direction), with a high degree of polarization.

[0165] 2. Low-polarized light diffusely reflected from the mold / furnace wall: After a polarization state randomization process, only about 50% of the energy reaches the sensor, and the polarization direction is disordered with low polarization degree.

[0166] The two signals are superimposed in the same imaging field of view. By calculating the degree of polarization (DoP) of each pixel, the forging pixels (high DoP) and the background pixels (low DoP) can be distinguished.

[0167] like Figure 6 As shown, the viewing window assembly 3 includes a sapphire transparent substrate 31. The inner surface of the sapphire transparent substrate 31 is coated with an anti-metal evaporation protective coating layer 32. A vacuum sealing structure 33 is provided between the sapphire transparent substrate 31 and the furnace wall through hole 2.

[0168] The sapphire transparent substrate 31 is a sapphire disc with a diameter of 80 mm and a thickness of 8 mm, made of high-purity optical-grade α-Al2O3 single crystal. The initial transmittance of sapphire to 808 nm and 1064 nm wavelength lasers is not less than 85%, and the transmittance decay does not exceed 10% of the initial value after continuous operation at 1500℃ for 500 hours.

[0169] The anti-metal evaporation protective coating layer 32 is a SiO2 / TiO2 multilayer dielectric film, deposited on the furnace-side surface of the sapphire transparent substrate 31 using electron beam evaporation physical vapor deposition (EBSD). The total thickness of the film layer is approximately 2 μm. This coating layer has two functions: firstly, the contact angle of metal vapor on the coating layer surface is larger than that on the bare sapphire surface, reducing the adhesion strength of metal vapor and facilitating subsequent cleaning and maintenance; secondly, even if there is slight metal deposition on the coating layer surface, the impact on the transmittance at a specific laser wavelength (1064 nm) is controlled within a small range.

[0170] The vacuum sealing structure 33 includes an oxygen-free copper C-section vacuum sealing gasket and a floating elastic compression flange. The oxygen-free copper sealing gasket is arranged around the outer peripheral surface of the sapphire transparent substrate 31 and the inner wall surface of the furnace wall through hole 2.

[0171] Oxygen-free copper (C10200) exhibits excellent plastic deformation capabilities under high-temperature vacuum conditions, without releasing gases or polluting the vacuum environment. The floating elastic clamping flange applies an axial preload of approximately 8 kN to the sapphire transparent substrate via a disc spring assembly, causing the oxygen-free copper sealing gasket to undergo controllable plastic deformation to form a high-vacuum level sealing contact surface.

[0172] The elastic deformation capability of the disc spring assembly can automatically compensate for the axial and radial thermal expansion displacement generated by the furnace body during the thermal cycle at 1000~1500℃, and maintain the stability of the sealing surface pressure throughout the entire temperature range.

[0173] The viewing window assembly 3 is the only component in the entire device that directly withstands the high temperature and high vacuum inside the furnace, and its long-term reliability directly determines the availability of the entire detection system. The high-temperature stability of sapphire ensures that its light transmittance does not decrease over long periods at 1000-1500℃; the anti-metal evaporation protective coating ensures an extended effective service life of the viewing window in a metal vapor environment; and the oxygen-free copper seal combined with the elastic floating clamping flange ensures that the vacuum seal does not fail during frequent temperature cycles. These three elements work together to enable the viewing window assembly to operate stably for extended periods in extreme environments.

[0174] In some embodiments, such as Figure 6 As shown, the sapphire light-transmitting substrate 31 is installed obliquely in the furnace wall through hole 2. The normal direction of the sapphire light-transmitting substrate 31 forms a preset oblique angle with the main optical axis (typical value is 5°~15°, preferably 10°), so that the reflected light from the front and rear surfaces of the sapphire light-transmitting substrate 31 deviates from the main optical axis, reducing backlight interference.

[0175] The front and back surfaces of any optical window / plate will produce Fresnel reflections (reflectivity of approximately 4% per surface, and approximately 7% per surface for sapphire with a refractive index of 1.76). If the sapphire is mounted perpendicular to the principal optical axis (with the normal coinciding with the principal optical axis), the reflected light from the front and back surfaces will return along the principal optical axis, enter the coaxial beam splitter 5, be reflected by the polarization beam splitter to the second port 5b, and form strong background stray light / ghost image on the imaging sensor, severely interfering with polarization imaging.

[0176] When the sapphire is mounted at an angle, the reflected light from its front and rear surfaces deviates from the principal optical axis. Taking a 10° angle as an example, the incident angle is 10°, and according to Fresnel's law of reflection, the reflection angle is also 10°, resulting in a 20° deviation of the reflected light from the incident direction. The reflected light from the rear surface of the sapphire is refracted within the sapphire, and its exit direction also deviates from the principal optical axis. These reflected lights, deviating from the principal optical axis, do not enter the third port 5c of the coaxial beam splitter / combiner component 5, and therefore do not reach the imaging sensor, effectively eliminating backlight interference.

[0177] The effect of angled installation on transmitted light (laser incident light and reflected signal light) is minimal: the additional optical path introduced by a 10° angle is approximately (θ' is the angle of refraction), its impact on image quality is negligible. The lateral beam offset introduced by the oblique placement is approximately This can be compensated by adjusting the position of the coaxial beam splitting and combining component 5.

[0178] like Figure 2 , Figure 4 As shown, the vacuum high-temperature forging furnace 1 adopts a double-layer structure of an outer furnace shell 11 and an inner heating furnace liner 12. The furnace wall through-holes 2 include a furnace shell through-hole 21 opened on the furnace shell 11 and a furnace liner through-hole 22 opened at the corresponding position on the furnace liner (12), both of which are coaxially arranged. The viewing window assembly 3 is embedded in the furnace shell through-hole (21), and its main optical axis is consistent with the common axis of the furnace shell through-hole 21 and the furnace liner through-hole 22.

[0179] In this embodiment, the double-layer structure of the vacuum high-temperature forging furnace is its basic structural feature: the outer furnace shell 11 is a vacuum container, maintaining a high vacuum environment inside the furnace; the inner heating furnace liner 12 is a heating element, heating the space inside the furnace to the forging process temperature (1000~1500℃). A gap is left between the two layers for placing heating elements and heat insulation materials.

[0180] In the double-layer structure, the laser incident beam must pass through both the furnace shell and the furnace liner simultaneously to reach the mold cavity inside the furnace. If only a hole is made in the furnace shell and not in the furnace liner, the laser will be blocked by the opaque furnace liner wall and will not be able to reach the forging. Therefore, through holes must be made at corresponding positions in the furnace shell and furnace liner, and both must be coaxially arranged so that the laser and reflected light can pass through without obstruction along a common axis. The viewing window assembly is installed in the through hole in the furnace shell (the furnace shell is the outer wall of the vacuum container, and the viewing window achieves vacuum sealing here), while the through hole in the furnace liner is a free through hole (the interior of the furnace liner is the high-temperature forging space and does not require additional sealing).

[0181] In some embodiments, such as Figure 2 , Figure 4 As shown, the furnace liner through hole 22 is an oblong hole or a waist-shaped hole. The length direction of the oblong hole is consistent with the thermal expansion direction of the inner heating furnace liner 12, so as to compensate for the thermal expansion displacement of the heating furnace liner 12 during the heating process and ensure that the main optical axis is not blocked by the heating furnace liner 12 in the entire temperature range.

[0182] The furnace chamber 12 will undergo thermal expansion during the heating process. Taking typical parameters as an example: the furnace chamber material is a nickel-based high-temperature alloy, with a linear expansion coefficient of... Temperature rise from room temperature (25°C) to 1500°C The characteristic length of the furnace lining from a reference point (e.g., one end of the furnace lining) to the location of the through hole. Then the thermal expansion displacement is: ; If the furnace liner through-hole is circular and its diameter matches the beam diameter (approximately 15-20 mm), during thermal expansion, the furnace liner through-hole may shift by more than 10 mm relative to the furnace shell through-hole, causing the main optical axis to fall into the solid part of the furnace liner and the optical path to be blocked.

[0183] The elongated hole design extends the furnace liner through-hole along the direction of thermal expansion, with a length of: ; in, To meet the minimum diameter required for the beam to pass through (typically 1.5-2 times the beam diameter). This represents the maximum thermal expansion displacement. Taking the above parameters as an example, The width of the oblong hole Not less than It is usually 15-20mm.

[0184] The inner wall of the oblong hole is chamfered (to remove sharp edges and prevent laser diffraction) and polished (to reduce surface roughness). (To reduce multiple reflections of the laser on the hole wall) and avoid stray light interference.

[0185] Within the entire temperature range (room temperature to 1500℃), after the furnace liner expands thermally, the furnace liner through hole shifts along the length of the elongated hole. However, due to the length margin design of the elongated hole, no matter how much the furnace liner expands, the opening range of the furnace liner through hole always covers the main optical axis position of the furnace shell through hole, ensuring that the optical path is unobstructed.

[0186] like Figure 5 As shown, the polarization filtering imaging receiving module 6 includes a narrow-band bandpass filter 61 and a focal plane polarization image sensor 62 arranged sequentially along the optical path. The center wavelength of the narrow-band bandpass filter 61 matches the laser wavelength emitted by the active linear polarization laser emission module 4, and is used to filter out blackbody thermal radiation noise in the furnace. The focal plane polarization image sensor 62 integrates micro-polarization arrays in multiple directions on its pixels for real-time acquisition of polarization images in a single exposure.

[0187] In this embodiment, within the furnace environment of 1000~1500℃, all high-temperature surfaces (forgings, molds, furnace walls) emit strong blackbody thermal radiation. According to Planck's blackbody radiation law, at a wavelength of 1064nm, the spectral radiance of a blackbody at 1500℃ is approximately: ; Although 1064 nm is located on the short-wavelength side of the blackbody radiation curve (with a peak at approximately 1.6 μm), and its radiance is much lower than the peak, it is still not negligible. Especially in imaging sensors, the integral effect of blackbody thermal radiation will generate a uniform background signal (dark current + thermal radiation signal) across the entire target surface, reducing image contrast.

[0188] The narrowband bandpass filter 61 restricts the spectral range received by the sensor to an extremely narrow bandwidth of ±1.5 nm of the laser wavelength. Outside this bandwidth, blackbody radiation is attenuated by more than 100,000 times by the out-of-band optical density (OD≥5) of the filter; within this bandwidth, the integrated energy of blackbody radiation is extremely small (because the bandwidth is only 3 nm), while the laser reflection signal is concentrated within this bandwidth and its intensity is much higher than that of blackbody radiation. Therefore, after narrowband filtering, the sensor mainly receives the laser reflection signal, and blackbody thermal radiation noise is effectively suppressed.

[0189] In addition, after obtaining the optical signal after narrowband filtering, it is necessary to analyze the degree of polarization (DoP) of each pixel to distinguish between high-polarization forging pixels and low-polarization background pixels. There are various methods to measure the degree of polarization, among which the division-of-focal-plane (DoFP) method best meets the real-time requirements of this application.

[0190] The split-focus plane polarization image sensor integrates a micro-polarizer (with linewidth in the micrometer range) above each pixel. Every four adjacent pixels form a "polarization superpixel," corresponding to the four polarization directions of 0°, 45°, 90°, and 135°, respectively. A single exposure can simultaneously acquire light intensity information in all four polarization directions, eliminating the need for rotating the polarizer or multiple frame acquisitions. This makes it suitable for real-time detection in dynamic scenarios such as forging ejection during forging (frame rate ≥ 30fps).

[0191] In some embodiments, such as Figure 5 As shown, the narrowband bandpass filter 61 has a bandwidth ≤ 3nm (full width at half maximum, FWHM). The focal plane polarization image sensor 62 integrates micro-polarization arrays in four directions: 0°, 45°, 90°, and 135° on its pixels.

[0192] The choice of bandwidth needs to strike a balance between two constraints: The narrower the bandwidth, the better the blackbody thermal radiation noise suppression effect and the higher the signal-to-noise ratio. The narrower the bandwidth, the higher the requirement for laser wavelength stability, and the peak transmittance of the filter may decrease due to manufacturing precision limitations.

[0193] The 3nm bandwidth corresponds to a relative bandwidth of approximately 0.28% at a wavelength of 1064nm, which fully meets the wavelength stability requirements (temperature drift ≤0.1nm / ℃, and wavelength drift ≤0.3nm after temperature control) for linearly polarized laser diodes (LDs) or solid-state lasers. Furthermore, narrowband filters (center wavelength 1064nm, bandwidth 3nm, peak transmittance ≥85%) can be mass-produced using current optical thin-film manufacturing technology at a controllable cost.

[0194] Based on a bandwidth of 1064nm and 3nm, the transmission band of the filter is 1062.5-1065.5nm. Within this band, the radiation energy density of a 1500℃ blackbody is extremely low. After filtering, the blackbody radiation noise is suppressed by more than 100 times (compared to broadband reception without a filter), ensuring high signal-to-noise ratio imaging.

[0195] The micro-polarization arrays at 0°, 45°, 90°, and 135° were chosen because polarization degree calculation based on Stokes parameters requires light intensity measurements in at least three different polarization directions. This four-directional configuration provides information redundancy, improving the accuracy and robustness of polarization degree calculations while meeting the requirements of real-time computation.

[0196] The extinction ratio (defined as the ratio of maximum to minimum transmittance of light in the same polarization direction) of micropolarizers in all directions must be ≥100:1. A higher extinction ratio results in a wider dynamic range for polarization degree measurement and a stronger ability to distinguish between high polarization (forgings) and low polarization (background). Currently, micropolarizers based on nanowire grating technology can achieve an extinction ratio of over 200:1 at a wavelength of 1064 nm, meeting the requirements.

[0197] like Figure 5 As shown, the active linearly polarized laser emission module 4 includes a linearly polarized laser 41 and a collimating and shaping lens group 42. The laser emitted by the linearly polarized laser 41 is collimated and shaped by the collimating and shaping lens group 42 and then enters the first port 5a as a parallel beam.

[0198] In this embodiment, the linearly polarized laser 41 is an active illumination light source, requiring the output of highly linearly polarized laser light with a polarization extinction ratio ≥100:1. Optional solutions include: Linearly polarized laser diodes (LDs) directly output linearly polarized light. They are compact, highly efficient, and suitable for wavelengths of 1064nm or 808nm.

[0199] Solid-state lasers (such as Nd:YAG) with external polarizers offer high output power and good beam quality, but their structure is relatively complex. The actual choice depends on the detection distance, spot size, and power requirements.

[0200] The collimating and shaping lens group 42 collimates the diverging beam output from the laser into a parallel beam and shapes the beam diameter and cross-sectional energy distribution (e.g., Gaussian beam → flat-top beam). The purpose of collimation is to ensure that the beam maintains a small divergence angle after passing through the coaxial beam splitter 5 and the window assembly 3, forming a suitably sized and uniformly energetic illumination spot on the forging surface. The beam diameter is typically designed to be 5-10 mm, slightly smaller than the effective aperture of the polarizing beam splitter 51 and the window assembly 3 (typically ≥25 mm), leaving sufficient centering margin.

[0201] like Figure 5 As shown, the polarization filtering imaging receiver module 6 also includes an imaging optical lens 63, which is disposed on the second port 5b side of the narrow bandpass filter 61 and is used to image the light filtered by the narrow bandpass filter 61 onto the target surface of the focal plane polarization image sensor 62.

[0202] In this embodiment, the narrowband filter only performs spectral filtering (suppressing blackbody thermal radiation noise) and does not have imaging function. The light output from the second port 5b is divergent reflected light from the forging and background inside the furnace. It needs to be focused by the imaging optical lens 63 and formed into a clear spatial image on the sensor target surface, so that the spatial position information of the forging and background is encoded into the pixel coordinates of the image.

[0203] The field of view and focal length of the imaging optical lens 63 are designed based on the size of the forging in the furnace and the detection distance. For example, if the detection distance (distance from the viewing window to the forging) is 500mm and the imaging area of ​​the forging is 150mm × 150mm, then the required field of view is approximately [missing information]. The corresponding lens focal length (calculated based on a sensor target size of 11.3mm) is approximately Using a standard industrial lens with a focal length of 50mm can achieve a field of view of approximately 12°, which meets the imaging coverage requirements.

[0204] The imaging optical lens 63 must operate in a band that covers the selected laser wavelength (1064nm) and have good transmittance and image quality at that wavelength. An achromatic lens or a dedicated near-infrared lens can be selected.

[0205] The following sources of interference exist during the operation of the vacuum high-temperature forging furnace: 1. Vibration: Strong mechanical vibrations are generated during the forging process and transmitted to the external mounting structure through the furnace body. Vibration can cause changes in the relative positions of optical elements, resulting in optical path misalignment and blurred imaging.

[0206] 2. Heat conduction: Although the temperature of the outer wall of the furnace is much lower than that inside the furnace, it is still higher than room temperature (possibly 100-300℃), affecting nearby optical modules through solid-state conduction and thermal radiation. Temperature fluctuations can lead to laser output power and wavelength drift, changes in the splitting ratio of the polarization beam splitter, and increased dark current noise in the sensor.

[0207] Based on this, such as Figure 2 , Figure 4 , Figure 5 As shown, the active linearly polarized laser emitting module 4, the coaxial beam splitting and combining component 5, and the polarization filtering imaging receiving module 6 are all mounted on the constant temperature and vibration damping base 7. The constant temperature and vibration damping base 7 is set independently of the vacuum high-temperature forging furnace 1 and is not rigidly connected to the furnace body.

[0208] In this embodiment, the constant temperature damping base 7 is not rigidly connected to the furnace body; instead, a flexible connection or air gap isolates the vibration transmission path. Furthermore, an air-floating damper or spring damping system is used to attenuate the furnace body vibration to within the amplitude range allowed by the optical elements. In addition, the constant temperature damping base 7 has a built-in temperature control system (such as constant temperature water circulation or Peltier temperature control) to control the working temperature of the optical module within a constant range (such as 25±0.5℃) and eliminate the impact of temperature fluctuations on optical performance.

[0209] like Figure 2As shown, the computing control host 8 is electrically connected to the polarization filter imaging receiving module 6. It is used to receive the polarization image output by the polarization filter imaging receiving module 6, calculate the polarization degree of each pixel, divide the pixels into forging area and background area according to the polarization degree threshold, extract the forging contour and solve the forging position and orientation.

[0210] In this embodiment, in the original image output by the focal plane polarization image sensor 62, each polarization superpixel (2×2 pixel group) contains light intensity values ​​in four polarization directions: (0° direction) (45° direction) (90° direction) (135° direction). The computing and control host calculates the Stokes parameters based on these four light intensity values: Total luminous intensity: ; Linear polarization components (0° / 90° direction): ; Linear polarization components (45° / 135° directions): ; Degree of polarization: .

[0211] The range of DoP values ​​is: . This represents fully linearly polarized light. This indicates completely unpolarized light.

[0212] The DoP value of pixels in the forging area is high (theoretically close to 1.0, but in practice, due to system extinction ratio limitations and surface scattering, the typical value is 0.6-0.9). The DoP value of pixels in the mold / furnace wall area is low (theoretically close to 0, but in practice, due to residual polarization selectivity of the coaxial beam splitter and combiner and sensor noise, the typical value is 0.05-0.3). There is a significant difference in DoP between the two, which can be used for binary segmentation by setting a threshold (e.g., DoP>0.6 for forging, DoP≤0.6 for background).

[0213] The threshold can be determined in the following ways: Pre-calibration is performed, and polarization images are acquired under typical working conditions with known forging / background conditions. The DoP distribution of forging pixels and background pixels is statistically analyzed, and the threshold that minimizes the segmentation error rate is selected.

[0214] Adaptive thresholding uses adaptive thresholding algorithms such as Otsu to automatically determine the optimal threshold based on the bimodal distribution of the DoP histogram, eliminating the need for manual setting.

[0215] In the polarization-degree binary segmentation result, the boundary between the forging region and the background region is the two-dimensional contour of the forging. The computing control host 8 performs edge detection on the binary image to extract the forging contour. When this device adopts a binocular configuration, it sets up two independent sets of viewing window components, coaxial beam splitting and combining components, active linearly polarized laser emitting modules, and polarization filtering imaging receiving modules, which serve as the first detection unit and the second detection unit, respectively, to simultaneously acquire first and second multi-directional polarized raw images from two different perspectives. The computing control host 8 runs a segmentation algorithm on the first and second images respectively, segmenting the first forging region and the second forging region and extracting the contours to obtain the first forging contour and the second forging contour.

[0216] The computing control host 8 inputs the first and second forging contours into a pre-established three-dimensional pose recognition model for the forging. This model includes a first two-dimensional pose recognition sub-network, a second two-dimensional pose recognition sub-network, and a binocular fusion sub-network: the first sub-network identifies the two-dimensional pose category from the first forging contour under a first viewpoint; the second sub-network identifies the two-dimensional pose category from the second forging contour under a second viewpoint; and the binocular fusion sub-network nonlinearly fuses the probability vectors of the two two-dimensional pose categories from the two viewpoints to output the three-dimensional pose category of the forging. The three-dimensional pose category describes the spatial state of the forging within the furnace, including normal, left-leaning, right-leaning, forward-leaning, backward-leaning, left-tilting, right-tilting, forward-tilting, backward-tilting, and demolding abnormality.

[0217] When using a monocular configuration, only a single-sided detection unit and polarization segmentation network are run to obtain the two-dimensional contour of the forging, but it can be used to determine the integrity of the forging contour (such as whether there are defects such as missing material or underpressure).

[0218] The calculation results and judgment outputs of the computing control host 8 are transmitted to the human-machine interface display terminal 9 via a wired communication interface (such as Gigabit Ethernet, CameraLink, or USB 3.0). The human-machine interface display terminal 9 is an industrial-grade touch screen or an external monitor for an industrial control computer, located at the operating table in the forging workshop, allowing operators to monitor the detection status in real time. The display interface of the human-machine interface display terminal 9 includes the following information areas: Real-time polarization image area: Displays the original polarization image acquired by the focal plane polarization image sensor 62 and the pseudo-color polarization image after polarization degree calculation. In the pseudo-color image, high polarization degree areas (forgings) are displayed in warm colors, and low polarization degree areas (background) are displayed in cool colors, allowing operators to intuitively distinguish the forgings from the background.

[0219] Forging contour overlay area: The forging contour lines extracted after polarization degree threshold segmentation are overlaid on the real-time polarization image. The contour lines are drawn with a highlight color, allowing operators to intuitively judge the completeness and accuracy of the contour extraction.

[0220] The forging outline and posture overlay area displays the extracted forging outline on the real-time polarized image, and labels the currently identified 3D pose category or 2D outline information with text.

[0221] Judgment Result and Alarm Area: Displays the judgment result of the current inspection ("Pass" or "Fail"). When judged as unqualified, it will alert the operator in a conspicuous manner (such as flashing red text and sound alarm) to the deviation or demolding abnormality, and automatically record the alarm time, deviation value and corresponding polarization image for post-process review and defect analysis.

[0222] This invention transforms the problem of three-dimensional pose measurement of forgings into a two-dimensional pose recognition problem under two single-viewpoints. The first and second two-dimensional pose recognition sub-networks identify the two-dimensional pose categories from the single-viewpoint contour images, respectively. Then, the probability vectors of the two viewpoints are nonlinearly fused by the binocular fusion sub-network to output the three-dimensional pose category. This eliminates the difficulties of camera calibration under high-temperature conditions and stereo matching of textureless forging surfaces in traditional binocular stereo vision.

[0223] This invention transforms the complex six-DOF pose numerical regression into a discrete category classification problem. Network training only requires two-dimensional contour images and corresponding two-dimensional pose category labels, eliminating the need for precise three-dimensional spatial coordinate ground truth, thus reducing the engineering difficulty of data acquisition and annotation. Binocular fusion is completed in a single forward propagation within the joint feature space of probability vectors, meeting the requirements for online real-time detection. The binocular consistency loss and two-dimensional-three-dimensional logical consistency loss introduced into the training loss encode binocular geometric laws and pose semantic logic as soft constraints into the model training process, ensuring that the model output conforms to physical laws and improving the robustness and interpretability of inference.

[0224] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and not to limit them. Although this application 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. All such modifications or substitutions should be covered within the protection scope of this application, and should not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A polarization binocular online measurement method for the three-dimensional attitude of forgings in a vacuum forging furnace, characterized in that, Includes the following steps: Step S1: When the forging is in the lifted state after mold opening, the polarization detection system built into the vacuum forging furnace collects the first multi-directional polarization original image and the second multi-directional polarization original image formed by the laser jointly reflected by the forging and the background area where the forging is located in different polarization directions at the two sampling points respectively. The two sampling points are in a symmetrical viewing angle relationship. Step S2: The forging region is segmented from the first multi-directional polarization original image and the second multi-directional polarization original image respectively, and the forging contour is extracted from the segmented forging region to obtain the first forging contour and the second forging contour. Step S3: Input the first forging contour and the second forging contour into the pre-established three-dimensional forging posture recognition model, and output the forging posture category from the pre-established three-dimensional forging posture recognition model; The three-dimensional pose recognition model includes a first two-dimensional pose recognition sub-network, a second two-dimensional pose recognition sub-network, and a binocular fusion sub-network; The first two-dimensional pose recognition sub-network takes the outline of the first forging as input and outputs the two-dimensional pose category of the forging from the first viewpoint. The second two-dimensional pose recognition subnetwork takes the contour of the second forging as input and outputs the two-dimensional pose category of the forging from the second perspective. The binocular fusion subnetwork takes the two-dimensional pose category of the forging in the first view and the two-dimensional pose category of the forging in the second view as input, and outputs the three-dimensional pose category of the forging.

2. The polarization binocular online measurement method for the three-dimensional attitude of forgings in a vacuum forging furnace according to claim 1, characterized in that, Both the first two-dimensional pose recognition subnetwork and the second two-dimensional pose recognition subnetwork adopt convolutional neural networks including multiple convolutional layers and fully connected layers; The convolutional layer in the first two-dimensional pose recognition sub-network is used to extract contour shape features from the input first forging contour step by step. The fully connected layer in the first two-dimensional pose recognition sub-network is followed by a Softmax activation function, which is used to output a two-dimensional pose category probability vector under the first view based on the contour shape features. The category corresponding to the maximum value in the probability vector is taken as the two-dimensional pose category under the first view. The convolutional layer in the second two-dimensional pose recognition subnetwork is used to extract contour shape features from the input second forging contour step by step. The fully connected layer in the second two-dimensional pose recognition subnetwork is followed by a Softmax activation function, which is used to output a two-dimensional pose category probability vector under the second view based on the contour shape features. The category corresponding to the maximum value in the probability vector is taken as the two-dimensional pose category under the second view. The two-dimensional pose category is used to describe the pose state of the forging in the corresponding two-dimensional image plane.

3. The polarization binocular online measurement method for the three-dimensional attitude of forgings in a vacuum forging furnace according to claim 2, characterized in that, The binocular fusion subnetwork adopts a multilayer perceptron structure containing multiple fully connected layers; The binocular fusion subnetwork takes as input a joint feature vector formed by concatenating the two-dimensional pose category probability vector output by the first two-dimensional pose recognition subnetwork and the two-dimensional pose category probability vector output by the second two-dimensional pose recognition subnetwork. The fully connected layer performs nonlinear fusion of the two-dimensional pose information of the first and second perspectives contained in the joint feature vector to output the three-dimensional pose category probability vector of the forging. The category corresponding to the maximum value in the three-dimensional pose category probability vector is taken as the three-dimensional pose category of the forging; The three-dimensional pose category is used to describe the pose state of the forging in three-dimensional space inside the vacuum high-temperature forging furnace.

4. The polarization binocular online measurement method for the three-dimensional attitude of forgings in a vacuum forging furnace according to claim 3, characterized in that, The training loss of the three-dimensional pose recognition model includes three parts: classification loss, binocular consistency loss, and two-dimensional-three-dimensional consistency loss. The classification loss includes a first two-dimensional pose classification loss, a second two-dimensional pose classification loss, and a three-dimensional pose classification loss; The first two-dimensional pose classification loss is the cross-entropy loss between the output of the first two-dimensional pose recognition sub-network and the first pose truth label of the forging. The second two-dimensional pose classification loss is the cross-entropy loss between the output of the second two-dimensional pose recognition subnetwork and the ground truth label of the forging's second pose. The 3D pose classification loss is the cross-entropy loss between the output of the binocular fusion subnetwork and the ground truth label of the 3D pose of the forging. The binocular consistency loss is used to penalize invalid 2D pose combinations that violate binocular geometric relationships in the outputs of the first 2D pose recognition subnetwork and the second 2D pose recognition subnetwork. The binocular consistency loss is calculated by summing the products of the first prediction probability and the second prediction probability corresponding to each combination in the predefined set of invalid 2D pose combinations. For any combination (i,j) in the set of invalid 2D pose combinations, the first prediction probability is the probability corresponding to the i-th category in the probability vector output by the first 2D pose recognition subnetwork, and the second prediction probability is the probability corresponding to the j-th category in the probability vector output by the second 2D pose recognition subnetwork. The formula for calculating the binocular consistency loss is as follows: ； In the formula, C is the set of invalid two-dimensional pose combinations. Each element in C is a two-dimensional pose category index combination (i,j), indicating that the first output of the i-th two-dimensional pose category and the second output of the j-th two-dimensional pose category are invalid combinations that violate the binocular geometric relationship. Let be the probability value corresponding to the i-th category in the two-dimensional pose category probability vector output by the first two-dimensional pose recognition subnetwork. The probability value corresponding to the j-th category in the two-dimensional pose category probability vector output by the second two-dimensional pose recognition subnetwork; The 2D-3D consistency loss is used to penalize invalid triplets where there is a logical contradiction between the 3D pose category output by the binocular fusion subnetwork and the two input 2D pose categories. The 2D-3D consistency loss is calculated by summing the products of the first prediction probability, the second prediction probability, and the 3D prediction probability corresponding to each combination in the predefined set of invalid triplets. For any combination (i,j,k) in the set of invalid triplets, the first prediction probability is the probability corresponding to the i-th category in the probability vector output by the first 2D pose recognition subnetwork, the second prediction probability is the probability corresponding to the j-th category in the probability vector output by the second 2D pose recognition subnetwork, and the 3D prediction probability is the probability corresponding to the k-th category in the probability vector output by the binocular fusion subnetwork. The formula for calculating the two-dimensional-three-dimensional consistency loss is as follows: ; In the formula, S is a set of invalid ternary combinations. Each element in S is a ternary combination (i,j,k) with a pose category index, indicating that the first output of the i-th 2D pose category, the second output of the j-th 2D pose category, and the binocular fusion subnetwork output of the k-th 3D pose category are invalid combinations with logical contradictions. This represents the probability value corresponding to the k-th class in the 3D pose class probability vector output by the binocular fusion subnetwork.

5. The polarization binocular online measurement method for the three-dimensional attitude of forgings in a vacuum forging furnace according to claim 4, characterized in that, The annotation methods for the three-dimensional pose truth value label of the forging, the first pose truth value label of the forging, and the second pose truth value label of the forging include: The forging is placed in a vacuum high-temperature forging furnace and subjected to various positions and orientations. The first multi-directional polarization original image and the second multi-directional polarization original image are acquired as the first sample image and the second sample image. The forging region is segmented from the first sample image and the second sample image respectively, and the forging contour is extracted from the segmented forging region to obtain the first sample contour and the second sample contour. The first and second sample contours are manually annotated. The first pose truth label of the forging is assigned according to the pose state of the forging in the two-dimensional image plane under the first view. The second pose truth label of the forging is assigned according to the pose state of the forging in the two-dimensional image plane under the second view. The preset pose state of the forging in the vacuum high-temperature forging furnace when the first and second sample images of the set are collected is used as the three-dimensional pose truth value label of the forging. The first sample contour, the second sample contour, the ground truth label of the first pose of the forging, the ground truth label of the second pose of the forging, and the ground truth label of the three-dimensional pose of the forging together constitute a sample in the dataset used to train the three-dimensional pose recognition model.

6. The polarization binocular online measurement method for the three-dimensional attitude of forgings in a vacuum forging furnace according to claim 5, characterized in that, The first and second multi-directional polarization original images each include four polarization original images corresponding to polarization directions of 0°, 45°, 90°, and 135°, respectively.

7. The polarization binocular online measurement method for the three-dimensional attitude of forgings in a vacuum forging furnace according to claim 6, characterized in that, The two sampling points are two through holes with symmetrical viewing angles on the wall of the vacuum forging furnace. A polarization detection system is set at each sampling point to form two independent detection units. The main optical axes of the two independent detection units intersect in the position area where the forging is expected to be lifted in the furnace. The two detection units include a first detection unit corresponding to the first viewing angle and a second detection unit corresponding to the second viewing angle. The polarization detection system includes a window assembly, a coaxial beam splitting and combining assembly, an active linearly polarized laser emission module, and a polarization filtering imaging receiving module.

8. The polarization binocular online measurement method for the three-dimensional attitude of forgings in a vacuum forging furnace according to claim 7, characterized in that, The viewing window assembly is embedded in the through hole of the furnace wall of the vacuum forging furnace, and the coaxial beam splitting and combining assembly is disposed outside the furnace of the viewing window assembly. It is used to guide the linearly polarized laser emitted by the active linearly polarized laser emitting module into the furnace and guide the reflected light from the furnace into the polarization filtering imaging receiving module.

9. The polarization binocular online measurement method for the three-dimensional attitude of forgings in a vacuum forging furnace according to claim 8, characterized in that, The polarization filtering imaging receiver module includes a narrowband bandpass filter and a focal plane polarization image sensor. The center wavelength of the narrowband bandpass filter is matched with the laser wavelength of the active linearly polarized laser emission module, and the bandwidth does not exceed 3nm. The focal plane polarization image sensor integrates micro-polarization arrays in four directions (0°, 45°, 90°, and 135°) on its pixels, which are used to synchronously acquire a first multi-directional polarization original image or a second multi-directional polarization original image in a single exposure.

Images