Tantalum-niobium alloy molten pool state monitoring method and system based on machine vision

Abstract

The present application relates to the technical field of image processing, and more particularly to a tantalum-niobium alloy molten pool state monitoring method and system based on machine vision. The method comprises the following steps: acquiring a molten pool time sequence image and performing preprocessing to obtain an input feature tensor; extracting brightness features, edge gradient features and texture features to obtain a multi-component feature map; grouping and rearranging according to a radial distance to construct a ring belt index to obtain a ring belt ordered feature tensor; determining a threshold based on local energy, channel correlation and ring belt gradient parameters, dividing the feature channels into an enhancement group and an inhibition group and rearranging them respectively to obtain a double-path rearranged feature tensor; performing inter-frame compensation to obtain a time sequence alignment feature tensor; dividing and layering the time sequence alignment feature tensor to obtain a monitoring result output by a state discrimination model. Through the technical scheme of the present application, the four states of tantalum-niobium alloy molten pool stability, collapse, offset and overburning can be accurately identified, and the monitoring accuracy and anti-interference ability are improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology. More specifically, this invention relates to a machine vision-based method and system for monitoring the condition of a tantalum-niobium alloy molten pool. Background Technology

[0002] Tantalum-niobium alloys are widely used in aerospace, nuclear industry, and electronics and information technology fields due to their high melting point, corrosion resistance, and excellent high-temperature mechanical properties. During the electron beam melting process of tantalum-niobium alloys, changes in the molten pool morphology directly reflect the stability of the melting process and have a decisive impact on the uniformity of the grain structure and mechanical properties of the finished product. Traditional molten pool monitoring methods rely on manual experience or simple photoelectric sensors. Operators judge the process status by visually observing changes in the gloss of the molten pool surface. This method not only provides limited information but also has a delayed response, resulting in a high misjudgment rate when facing extremely high temperature gradients and transient thermodynamic changes.

[0003] With the development of industrial vision technology, intelligent visual monitoring technology based on temporal image feature analysis has been gradually introduced into the field of molten pool condition monitoring. This type of technology acquires molten pool image sequences through high-speed industrial cameras, extracts the brightness, edge, and texture features of the images using convolutional neural networks, and combines feature mapping rearrangement strategies such as PixelShuffle to enhance the representation of local details through channel dimension transformation and spatial resolution enhancement, thereby constructing a state classification model.

[0004] In the prior art, Chinese patent document CN112967259B discloses a method and system for predicting the piercing state of plasma arc welding based on molten pool images. This method extracts molten pool image features by defining multiple improved backbone networks, trains a prediction model by combining multiple optimizers, and selects the model with the highest prediction accuracy for piercing state prediction. This technical approach has, to some extent, promoted the progress of molten pool image analysis.

[0005] However, the above methods still have the following limitations when dealing with the complex processing conditions of tantalum-niobium alloys. Existing feature mapping strategies typically process global features uniformly, neglecting the radial gradient variation of the molten pool from the central high-temperature zone to the boundary low-temperature zone. This fails to provide a refined analysis of the physical property differences in each annular region of the molten pool, making it difficult for the model to capture subtle temperature field and morphological changes in the radial direction. At the channel processing level, the PixelShuffle method performs indiscriminate rearrangement of all feature channels, failing to distinguish between high-contribution and low-contribution channels based on local energy and correlation. This not only exacerbates the consumption of computational resources but also easily leads to key features being overwhelmed by redundant information during recombination. Furthermore, existing methods lack a spatiotemporal joint compensation mechanism for the molten pool in terms of temporal processing. They fail to fully utilize the minute deformation information of the molten pool boundary at adjacent time points for inter-frame compensation and temporal alignment. This results in feature misalignment when the model handles rapid fluctuations such as collapse and shift, limiting the accuracy and anti-interference capability of the monitoring results. Summary of the Invention

[0006] To address the technical problems of insufficient accuracy and feature misalignment in existing molten pool visual inspection methods in terms of feature mapping, channel processing, and inter-frame alignment, this invention provides solutions in the following aspects.

[0007] In a first aspect, the present invention provides a machine vision-based method for monitoring the state of a tantalum-niobium alloy molten pool, comprising:

[0008] The time series image of the molten pool is acquired and preprocessed to obtain the input feature tensor; the brightness features, edge gradient features and texture features of the input feature tensor are extracted and obtained by convolutional encoding to obtain a multi-component feature map. Construct a ring index based on the radial distance from the center of the molten pool to the boundary, and rearrange the multi-component feature maps to obtain the ring-ordered feature tensor; Based on the annular ordered feature tensor, local energy representation parameters, channel correlation parameters and annular gradient parameters are extracted, and a threshold is determined by weighted fusion. According to the threshold, the feature channels are divided into enhancement channel group and suppression channel group and rearranged respectively. Compression adjustment parameters are extracted to obtain dual-path rearranged feature tensor. The time compensation parameter at the first sampling moment is initialized to zero. Subsequently, the time compensation parameter is determined based on the change information of the melt pool boundary at adjacent moments. The time-aligned feature tensor is obtained by performing inter-frame compensation on the dual-path rearrangement feature tensor in combination with the annular gradient parameter. The temporal alignment feature tensor is divided into regions and hierarchically aggregated to extract feature parameters. The feature parameters are combined with threshold, compression adjustment parameters, time compensation parameters and channel-related parameters and input into the state discrimination model to output the monitoring results of the melt pool.

[0009] This invention constructs a ring index based on the radial distance from the center of the molten pool to the boundary, transforming the scattered two-dimensional features in the multi-component feature map into a radially ordered ring sequence, ensuring that the feature arrangement aligns with the physical property gradient distribution of the molten pool from the high-temperature zone at the center to the low-temperature zone at the boundary. At the channel processing level, enhancement and suppression channel groups are divided based on the weighted fusion results of three statistical parameters: local energy, channel correlation, and ring gradient. Spatial dimension amplification and rearrangement are performed on the enhancement channel group to highlight detailed information, while channel compression is performed on the suppression channel group to eliminate redundant noise and reduce computational cost. The inter-frame compensation stage generates a time compensation parameter matrix by extracting minute deformation information of the molten pool boundary at adjacent time points. Sub-pixel-level temporal alignment is achieved using one-dimensional deformable sampling operations. The feature parameters extracted through region division and hierarchical convergence encode the feature distribution information of each region of the molten pool in the form of multi-scale spatial statistics, which, together with thresholds, compression adjustment parameters, time compensation parameters, and channel correlation parameters, constitute a comprehensive discriminative input.

[0010] Preferably, the preprocessing to obtain the input feature tensor includes: extracting the maximum and minimum gray values ​​of all pixels in the molten pool time series image, and linearly mapping the original gray values ​​of each pixel to a closed interval of 0 to 1 through extreme value normalization calculation; calculating the centroid two-dimensional coordinate point of the molten pool region in the normalized image based on the gray distribution, and truncating a square region with a side length of 64 pixels centered on the centroid two-dimensional coordinate point as the region of interest, and performing edge zero-padding alignment processing when the truncated range exceeds the boundary of the original image; for the non-uniformly sampled molten pool time series image, setting the time alignment reference axis according to a fixed time span step value, regenerating the image data corresponding to each time node through cubic spline interpolation method, and then stitching them together in the time dimension to construct a continuous three-dimensional input feature tensor.

[0011] This invention eliminates the difference in grayscale baseline between different frames of images by extreme value normalization, and uses centroid positioning to extract the region of interest to ensure that the molten pool is always located in the center of the field of view. Cubic spline interpolation reconstructs the uniform time axis to solve the problem of temporal discontinuity caused by camera frame loss or non-uniform sampling, so that the input feature tensor has consistency in three dimensions: grayscale scale, spatial position and time interval.

[0012] Preferably, obtaining the multi-component feature map includes: retaining the input feature tensor according to the arrangement of pixel rows and columns as a single-layer original brightness feature two-dimensional matrix; applying a 3×3 Sobel filter kernel with a stride of 1 to perform convolution operations on the pixels to be processed along the X and Y directions respectively, and combining the bidirectional first-order difference results to generate a single-layer edge gradient feature two-dimensional matrix; mapping the pixel values ​​in the 0 to 1 interval proportionally to 16 discrete integer gray levels and establishing a 16th-order gray-level co-occurrence matrix in a local sliding window, calculating the angular second moment distribution value of each local window based on the gray-level co-occurrence matrix to generate a single-layer texture feature two-dimensional matrix; using three 3×3 convolutional layers with 16 channels to process the above three two-dimensional matrices in parallel, and stitching them along the channel depth direction to construct a 48-channel multi-component feature map.

[0013] The original brightness, Sobel edge gradient, and second-order moment of the gray-level co-occurrence matrix are three complementary physical information types that are encoded by independent convolutional channels and then concatenated along the channel dimension. This enables the multi-component feature map to have representation capabilities in three aspects: brightness intensity, edge sharpness, and texture roughness, providing a rich underlying feature foundation for subsequent annular grouping and rearrangement.

[0014] Preferably, obtaining the ordered feature tensor of the ring zone includes: acquiring spatial grid data of the multi-component feature map, calculating the Euclidean distance from each coordinate sampling point to the center of the feature map; establishing a discrete radius sequence with a preset number of pixels as the radius step size, and constructing a set of concentric ring zones that do not intersect from the inside to the outside; traversing the Euclidean distance and assigning a unique ring zone number to each coordinate sampling point according to its radius interval, generating a two-dimensional mapping template as the ring zone index; traversing the two-dimensional mapping template in ascending order of the ring zone number, aggregating the features within the same ring zone into a one-dimensional sequence, aligning the sequence length of each ring zone to a preset fixed number of pixels through truncation or zero-padding operations, and retaining the channel dimension to reshape it into an ordered feature tensor of the ring zone with ring zone number dimension, spatial sequence dimension and channel dimension, so that the feature arrangement is consistent with the temperature gradient distribution law of the melt pool from the central high temperature zone to the boundary low temperature zone.

[0015] Preferably, the step of dividing the feature channels into enhancement channel groups and suppression channel groups according to the threshold includes: parsing the local energy representation parameters and extracting the energy quantization value corresponding to each feature channel; comparing the energy quantization value with the threshold; assigning feature channels with energy quantization values ​​greater than the threshold to the enhancement channel group, and rounding up the number of channels in the enhancement channel group to a multiple of 2 by adding all-zero channels; assigning the remaining feature channels to the suppression channel group, and adding all-zero channels to align the number of channels in the suppression channel group to an even number when the number of channels in the suppression channel group is odd.

[0016] Preferably, the step of extracting compression adjustment parameters to obtain a dual-path rearrangement feature tensor includes: performing a one-dimensional PixelShuffle rearrangement on the features within the enhanced channel group, setting the amplification factor of the annular space dimension to 2, and expanding the length of the annular space dimension to twice its original length by transferring the data from the channel dimension to the annular space dimension; applying a 1×1 convolution to the features within the suppressed channel group to perform channel dimensionality reduction, reducing the number of channels to 50% of the original, and extracting the mean of the 1×1 convolution output features in the channel dimension as a compression adjustment parameter; and using a one-dimensional linear interpolation algorithm to align the lengths of the amplified and reduced features to the same size, and concatenating them in the channel dimension to generate a dual-path rearrangement feature tensor.

[0017] The enhanced channel group is subjected to PixelShuffle misalignment rearrangement to expand the resolution of the annular space dimension. The suppressed channel group is subjected to dimensionality reduction by 1×1 convolution to compress redundant channels and extract compression adjustment parameters for subsequent comprehensive discrimination. After alignment by linear interpolation, the two-path rearrangement feature tensors are spliced ​​in the channel dimension to generate a dual-path rearrangement feature tensor.

[0018] Preferably, obtaining the temporally aligned feature tensor includes: comparing the edge gradient features of the two-frame images obtained at the previous sampling time and the current sampling time in adjacent time steps, applying the least squares method to calculate optical flow and outputting a first-order displacement vector matrix of all discrete control base points; converting the first-order displacement vector matrix into a one-dimensional offset of the corresponding annular sequence based on the annular index, aligning the one-dimensional offset with the dimension of the annular gradient parameters through a spatial expansion operation, concatenating it with the current frame features of the dual-path rearrangement feature tensor in the channel dimension, and inputting it into a multi-layer fully connected network to calculate weights and generate a time compensation parameter matrix; and based on a one-dimensional deformable sampling operation, determining the offset of the feature sampling position according to the time compensation parameter matrix, and interpolating and aligning the current frame features to obtain the temporally aligned feature tensor.

[0019] The displacement vector matrix between adjacent frames is obtained by optical flow calculation and mapped to a one-dimensional offset by the annular index. After being aligned with the dimension of the annular gradient parameter, a time compensation parameter matrix is ​​generated by a fully connected network. The current frame features are then interpolated and aligned using a one-dimensional deformable sampling operation. The resulting temporally aligned feature tensor is then passed to the next step for region division and state discrimination.

[0020] Preferably, the output of the melt pool monitoring result includes: dividing the temporally aligned feature tensor into at least two different scales of non-overlapping grid regions in the spatial dimension; performing max pooling and average pooling on each grid region to extract local features; flattening and concatenating the pooled features into a high-dimensional aggregate vector as feature parameters; flattening the time compensation parameter into a one-dimensional vector; arranging the channel-related parameters in channel order into a one-dimensional vector; using the threshold and compression adjustment parameters as single-element values, and concatenating them with the feature parameters along the same vector dimension to form a state representation vector; after temporal feature extraction via a long short-term memory network, taking the hidden state output vector of the last time step and inputting it into a multilayer perceptron classifier to output the melt pool monitoring result.

[0021] Local statistical features are extracted at different grid granularities using multi-scale spatial pyramid pooling. Multiple intermediate parameters are concatenated with the pooling features to form a state representation vector. After temporal feature extraction via a long short-term memory network, the vector is input into a multilayer perceptron classifier, which outputs the probability distribution of four state categories.

[0022] Preferably, the weighted fusion threshold determination includes: the local energy representation parameter is obtained by normalizing the sum of squares of all elements in each feature channel; the channel correlation parameter is obtained by the mean of the Pearson correlation coefficients between the feature vectors of any two channels; the annular gradient parameter is obtained by the mean of the absolute values ​​of the differences between the feature mean values ​​of adjacent annular zones; the comprehensive contribution score of each channel is obtained by weighting and summing the three types of parameters according to preset weights, and the median of the comprehensive contribution score is taken as the threshold.

[0023] Secondly, the present invention provides a machine vision-based tantalum-niobium alloy molten pool condition monitoring system, including a processor and a memory. The memory stores computer program instructions, and the processor reads and executes the computer program instructions in the memory to perform the above-mentioned machine vision-based tantalum-niobium alloy molten pool condition monitoring method.

[0024] The beneficial effects of this invention are as follows: This invention reorganizes the scattered two-dimensional features of the molten pool image into a radially ordered sequence of rings based on radial ring indexing, so that the feature arrangement matches the physical property gradient distribution of the molten pool from the central high-temperature region to the boundary low-temperature region. According to the weighted fusion result of local energy, channel correlation and ring gradient, the feature channels are divided into enhancement channel group and suppression channel group. Spatial dimension magnification and rearrangement are performed on the enhancement channel group to retain detailed information, and channel compression is performed on the suppression channel group to suppress redundant noise and reduce the amount of computation. Differential processing is applied to feature channels with different contributions.

[0025] Furthermore, this invention introduces an inter-frame joint compensation mechanism based on optical flow tracking. By extracting minute deformation information of the molten pool boundary at adjacent time points, a time compensation parameter matrix is ​​generated. Sub-pixel-level temporal alignment is achieved using one-dimensional deformable sampling operations, which solves the problem of spatiotemporal misalignment caused by frame loss and oscillation under rapid fluctuation conditions. Combined with multi-dimensional feature parameters and path adjustment parameters, a comprehensive discrimination input is formed to accurately identify the four states of tantalum-niobium alloy molten pool: stability, collapse, displacement, and overheating. Attached Figure Description

[0026] Figure 1 This is a flowchart of the machine vision-based tantalum-niobium alloy molten pool condition monitoring method in this invention; Figure 2 This is a diagram comparing the predictive performance of different hierarchical structural models. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0028] Reference Figure 1 This invention discloses a machine vision-based method for monitoring the state of a tantalum-niobium alloy molten pool, comprising steps S1-S5: S1. Image preprocessing and multi-component feature coding.

[0029] A continuous frame sequence of original images of the electron beam melting process of tantalum-niobium alloy is acquired, and the global maximum and minimum gray values ​​of all pixels in the current image frame are extracted. An extreme value normalization algorithm is used to linearly map the original gray values ​​of each pixel to a closed interval between 0 and 1. When the maximum and minimum gray values ​​are equal, the normalization results of all pixels are uniformly set to 0.

[0030] Assuming that under stable operating conditions at room temperature, the global maximum gray value in a certain frame of the original 8-bit depth molten pool image is 238 and the minimum gray value is 12, then the floating-point number obtained by the extreme value linear mapping of the pixel with a gray value of 100 is 0.389, and the data range of the entire image converges to 0 to 1.

[0031] Based on global grayscale distribution probability statistics, the centroid two-dimensional coordinates of the high-temperature luminescent molten pool region in the normalized image are calculated. Using the centroid two-dimensional coordinates as the origin, a square image block with a width and height of 64 pixels is cropped as the region of interest. When the cropped area exceeds the original image boundary, zero-padding is performed on the excess portion to maintain dimensional consistency. Taking a normalized image frame as an example, the horizontal and vertical coordinates of the floating-point centroid position are calculated to be 245.3 and 178.4 respectively. After rounding, a square cropping window with a side length of 64×64 pixels is constructed with the coordinates (245, 178) as the geometric center. At this point, the horizontal coordinate of the left boundary is 213, and the horizontal coordinate of the right boundary is 276, ensuring the window falls completely within the image area. When the melt pool position is close to the edge of the image, for example, if the x-coordinate of the centroid of a certain frame is 28 after rounding, and the x-coordinate of the left boundary of the cropping window is 28-32=-4, which exceeds the original image boundary, then 4 columns of pixels with values ​​of 0 are added to the leftmost side of the extraction matrix to keep the output matrix dimension unchanged at 64×64.

[0032] For time-series images with uneven time intervals due to camera frame drops or non-uniform sampling in actual production, a time-aligned reference axis is configured according to a set fixed time span step value. A cubic spline interpolation method is called to fit a smoothing function along the time axis coordinate sequence of each pixel, regenerating and filling in the complete image data corresponding to each standard time node. Assuming the absolute times corresponding to the detected frames are 10ms, 23ms, 41ms, and 55ms, after configuring a fixed time span step value of 15ms, an absolute synchronization alignment reference axis is re-established, with reference nodes at 15ms, 30ms, 45ms, etc. A channel-by-channel cubic spline curve interpolation algorithm fits a cubic polynomial continuous smoothing function along the time dimension coordinate sequence of each pixel to solve for the complete reconstructed frame image data at standard times such as 15ms and 30ms. In this embodiment, the fixed time span step value is 15ms; in other embodiments, it can be adjusted within the range of [5ms, 50ms] according to the actual frame rate of the sampling device.

[0033] Within a fixed, continuous observation window, a sequence of time-uniformly aligned two-dimensional matrices is stacked and stitched together in ascending order of time. In this embodiment, the observation window length is 16 reference node time frames, forming a three-dimensional continuous input feature tensor with a time dimension depth of 16, a spatial width of 64, and a spatial height of 64. If the observation window length is too small, it will not be able to capture enough temporal change information; if it is too large, it will introduce too many redundant frames and increase the computational load. 16 reference nodes are chosen as an engineering compromise. In other embodiments, the observation window length can be adjusted within the range of [8, 32] according to the rate of change of the smelting conditions.

[0034] After obtaining the input feature tensor, it is directly retained according to the pixel row and column arrangement, serving as a single-layer original two-dimensional brightness feature matrix. A 3×3 Sobel filter kernel with a stride of 1 is invoked to perform convolution operations along the X and Y directions on the pixels to be processed, and the bidirectional first-order difference results are combined to generate a single-layer edge gradient feature two-dimensional matrix. The convolution kernel parameters and operation rules of the Sobel difference filter operator are well-defined in the field of industrial vision, and the coefficient configuration of the kernel matrix in each direction is set according to a standard form.

[0035] The normalized pixel values ​​in the 0-1 range are proportionally mapped to 16 discrete integer gray levels. Zero-padding with a width of 2 is applied around the image to maintain the same output size as the input. A 16th-order gray-level co-occurrence matrix is ​​established within a defined local sliding window. The angular second moment distribution value of each pixel within the local window is calculated using the gray-level co-occurrence matrix, generating a corresponding single-layer texture feature two-dimensional matrix. The gray-level co-occurrence matrix characterizes texture roughness and directionality by statistically analyzing the joint occurrence frequency of pixel gray-level pairs under specific spatial displacements. In this embodiment, the sliding window size is 5×5, and the calculation step is 1. In other embodiments, this can be adjusted within the range of 3×3 to 7×7 depending on the image resolution and texture scale.

[0036] Three parallel and independently operating convolutional neural network modules are constructed within a deep learning framework. Each module contains a 3×3 convolutional layer with 16 channels. The original luminance feature 2D matrix, edge gradient feature 2D matrix, and texture feature 2D matrix are input in parallel into their respective convolutional layers for processing, generating three parallel low-level convolutional feature tensors. The three sets of feature tensors are concatenated and fused along the channel dimension to construct a multi-component feature map with a total of 48 channels.

[0037] Taking a cross-sectional frame matrix in a three-dimensional continuous input feature tensor as an example, in the first branch, the 64×64 pixel matrix of the cross-section is copied as a single-layer original two-dimensional brightness feature matrix; in the second branch, the Sobel operators in the X and Y directions are called to perform convolution operations respectively, and a mirror padding with a width of 1 is configured around the image. The bidirectional first-order difference results are combined to generate a single-layer edge gradient feature two-dimensional matrix of the same size, which is also 64×64; in the third branch, the normalized gray level of the current cross-section is mapped to 16 discrete integer gray levels from 0 to 15. After zero padding with a width of 2 is configured around the image, the frequency probability of the co-occurrence of pixel pairs in each sliding neighborhood is statistically analyzed to generate a 16th-order local gray-level co-occurrence matrix. By solving the second moment of the angle in each sliding window, that is, the sum of the squares of all elements in the matrix, the single-layer texture feature two-dimensional matrix is ​​obtained. Three 64×64 two-dimensional matrices are input into three parallel convolutional layer modules for processing. Each module performs zero-padding and no downsampling, mapping to generate three low-level convolutional feature tensors, each with a size of 16×64×64. A concatenation operation is performed along the outermost channel dimension to obtain a 48-channel multi-component feature map that integrates brightness, edge, and texture information.

[0038] S2, Construction of ring index and rearrangement of feature groups.

[0039] The tantalum-niobium alloy molten pool exhibits a radial gradient distribution of temperature and physical properties from the center to the boundary. To incorporate this physical law into the feature representation, this invention constructs a concentric ring index to spatially group and rearrange the multi-component feature map.

[0040] Spatial grid data of the multi-component feature map is acquired, and the Euclidean distance from each coordinate sampling point to the geometric center of the feature map is calculated one by one. Taking the sampling point at coordinates (40, 20) as an example, the coordinates of the geometric center of the feature map are (31.5, 31.5), and the Euclidean distance from this point to the geometric center is 14.3. For the 48-channel feature tensor, a 64×64 discretized grid base map is extracted in space, containing a total of 4096 coordinate sampling points. The Euclidean distance from each sampling point to the geometric center is calculated using two-dimensional Cartesian plane ranging.

[0041] A discrete radius sequence is established using a 4-pixel step size. From the inside out, a set of non-overlapping concentric annular zones is constructed, with boundary intervals of 0 to 4, 4 to 8, 8 to 12, 12 to 16, etc. The calculated Euclidean distances are iterated through, and each sampling point is assigned a unique annular zone number based on its radius interval, generating a two-dimensional mapping template as an index for the annular zones. Sampling points with distance values ​​falling within the 12 to 16 interval are uniformly assigned annular zone number 3, thus converting all coordinates into discrete annular zone numbers.

[0042] The two-dimensional mapping template is traversed in ascending order by the ring number, and the data features belonging to the same ring are extracted and aggregated into a one-dimensional continuous sequence. Since the number of pixels covered by different rings varies, the length of the one-dimensional sequence of each ring is uniformly aligned to a preset fixed number of pixels through truncation or zero-padding operations. In this embodiment, the standard length is 128 unit nodes: for the core region numbered 0, if the statistical feature values ​​are only 50 points, the end is padded with 0s until the length reaches 128; if the outer ring has 150 elements, the extra 22 values ​​at the end are truncated, retaining only the first 128. Considering that the Euclidean distance from each sampling point to the center in the 64×64 grid base map is as far as approximately 45 pixels at the corner point, the total number of rings in this embodiment is 12 groups with a radius step of 4 pixels. In other embodiments, the radius step can be adjusted within the range of [2, 8] according to the feature map size and the melt pool scale, and the total number of rings changes accordingly.

[0043] The originally spatially scattered features are reorganized into a specific arrangement with 12 sets of spatial ring labels and 128 spatial sequences within a single ring. While retaining the 48-layer channel dimension, a ring-ordered feature tensor with a tensor size of 48 channels × 12 rings × 128 sequence lengths is generated.

[0044] S3, Channel Parameter Extraction and Dual Path Rearrangement.

[0045] The annular ordered feature tensor contains 48 channels, and each channel contributes differently to the characterization of the molten pool state. In order to distinguish between high-contribution channels and low-contribution channels and apply differential processing accordingly, three types of statistical parameters need to be extracted from the annular ordered feature tensor.

[0046] The local energy representation parameters are extracted as follows: The local energy representation parameters are obtained by normalizing the sum of squares of the elements in each feature channel of the annular ordered feature tensor. Specifically, the sum of squares of all elements within the channel is calculated along the annular dimension and the sequence dimension, and the sum is normalized to the [0, 1] interval to obtain the energy quantization value corresponding to that channel. Each of the 48 channels is assigned a one-dimensional real-number format energy quantization value; for example, the value for the first layer channel is 0.785, and the value for the second layer is 0.412. Along the channel dimension, the Pearson correlation coefficient between the feature vectors of any two channels is calculated, constructing a 48×48 channel correlation matrix. The mean of the correlation coefficients between each channel and the remaining channels is taken as the channel correlation parameter. Since the difference in the feature mean between adjacent annular zones reflects the radial gradient strength, the first-order difference of the feature mean of adjacent annular zones is calculated along the annular zone numbering direction. The absolute value of the difference result is then averaged along the sequence dimension to obtain the annular gradient parameter.

[0047] To integrate the three types of parameters into a unified channel partitioning criterion, a weighted fusion of the local energy representation parameter, channel-related parameters, and annular gradient parameters is performed to determine the threshold. The fusion weights for the three types of parameters range from [0.3, 0.6], [0.2, 0.4], and [0.1, 0.3], respectively, and their sum is 1. When the weight of the local energy representation parameter is below 0.3, the signal strength information accounts for an insufficient proportion in the fusion, and the threshold's ability to distinguish channel energy differences decreases, resulting in high-energy channels being mixed with low-energy channels. When the weight of the channel-related parameter exceeds 0.4, the degree of overlap between channel information has an excessive impact on the threshold, causing the threshold to shift excessively upward. Some independent but low-energy channels are assigned to the suppression group because they do not reach the raised threshold. In this embodiment, the fusion weights for the three types of parameters are 0.5, 0.3, and 0.2, respectively, and the median of the fusion scores of all channels is taken as the threshold. The weight allocation follows the energy priority principle: the local energy representation parameter directly reflects the signal strength of the channel and has the highest distinguishability in determining the channel contribution, so it is given the largest weight; the channel correlation parameter is used to measure the degree of information overlap between channels, so that the threshold is adaptively shifted upward when there are many redundant channels; the annular gradient parameter is used to supplement the radial structure information, and the weights of the two decrease in sequence.

[0048] For each feature channel, the local energy quantization value, the mean value of the channel's related parameters, and the annular gradient parameters are weighted and summed according to the aforementioned fusion weights to obtain the channel's comprehensive contribution score. The comprehensive contribution scores of the 48 channels are compared one by one with a threshold. Parameters under typical operating conditions are used for verification. The threshold obtained through fusion calculation is 0.635. The comprehensive contribution score of each channel is compared with this threshold. A total of 25 channels have a comprehensive contribution score greater than 0.635, and these channels are assigned to the enhancement channel group. To meet the requirement of an even multiple of the number of channels in the subsequent PixelShuffle rearrangement mechanism, all-zero channels are added at the end to round up the total number of enhancement channel groups to an even base of 26 layers.

[0049] All 23 remaining substandard feature channels were assigned to the suppression channel group. The parity of the total number of suppression channel groups was checked. When the number of channels was found to be odd, a zero-value channel was added at the end of the channel group to align it with the even-value boundary of layer 24.

[0050] For an enhanced channel group with a tensor size of 26 channels × 12 annular bands × 128 sequence length, a one-dimensional PixelShuffle rearrangement is performed. PixelShuffle improves spatial resolution without introducing additional parameters by remapping the channel dimension data to the spatial dimension according to a periodic interleaving rule. The annular band spatial dimension magnification factor is set to 2, and a portion of the channel dimension data is periodically transferred to the spatial sequence dimension through the rearrangement mechanism. This involves taking adjacent layer nodes from the original channel dimension and interleaving them into the sequence length, halving the number of data channels to 13 layers and expanding the annular band sequence length from 128 to 256, resulting in a magnified feature with a longer size.

[0051] The tensor size of the suppressed channel group is 24 channels × 12 loops × 128 sequence lengths. Channel dimensionality reduction is performed using a 1×1 convolutional kernel with learnable parameters, compressing the number of channels to 50% of the original, i.e., to 12 layers. After inference, the arithmetic mean of the 1×1 convolution output features along the channel dimension is extracted and saved as a compression adjustment parameter. Validation using actual computational data of the suppressed channel group shows that the compression adjustment parameter value is 0.231. The 1×1 convolution performs a linear weighted combination on the channel dimension while maintaining the spatial resolution, achieving channel compression.

[0052] To address the mismatch in sequence lengths between the two paths, a one-dimensional linear interpolation algorithm is used to apply a smooth weighted evaluation operation to the sequence direction of the compressed features, aligning them to a standard size of 256. A concatenation operation is then performed along the channel dimension, reconstructing and superimposing the 13-channel enlarged feature with the 12-channel interpolated compressed feature, generating a dual-path rearranged feature tensor with a total size of 25 channels × 12 loops × 256 sequence length.

[0053] S4. Inter-frame temporal compensation and feature alignment.

[0054] During the tantalum-niobium alloy smelting process, the molten pool boundary undergoes rapid and minute deformation due to thermal disturbance, which may cause the feature positions between adjacent frames to shift. To eliminate the interference of inter-frame feature misalignment on subsequent state discrimination, the time compensation parameter matrix generated in this step is the specific implementation of the time compensation parameters.

[0055] For the first sampling time in the time series, since there is no previous frame reference, the time compensation parameter matrix is ​​initialized to a zero matrix. That is, no offset compensation is applied to the features of the first frame, and the original features are directly retained as the temporal alignment feature tensor for that time. From the second sampling time onwards, the time compensation parameters are calculated as follows.

[0056] Two frames are captured from adjacent time steps, one from the previous sampling time and the other from the current sampling time. Edge gradient features generated during the preprocessing stage are extracted from each frame. The pixel differences in the edge gradient features between the two frames are compared, and high-frequency contour points are extracted as discrete control points. An analytical equation for optical flow is constructed using the discretized measurement points. A tracking model minimizing reprojection error is built using the least squares method, converging to the optimal trajectory and outputting a first-order displacement vector matrix carrying X-axis and Y-axis slippage information. Least squares optical flow estimates the pixel-level displacement field by minimizing the reprojection residual of the brightness conservation constraint equation between adjacent frames.

[0057] Based on the previously established annular index mapping template, the first-order displacement vector matrix is ​​mapped and converted into a one-dimensional offset of the corresponding annular zone. For the vectors within the outermost annular zone, a globally smoothed average offset is calculated and converted into a one-dimensional offset array. The one-dimensional offsets are then aligned with the annular gradient parameters in a unified dimension through a spatial expansion operation, and subsequently concatenated with the dual-path rearrangement feature tensor of the current frame in the channel dimension.

[0058] The concatenated hybrid tensor is flattened into a one-dimensional vector along the channel dimension and the annular sequence dimension. This vector is then input into a multilayer fully connected network to calculate weights, generating a time-compensated parameter matrix used as a reference for the node offsets in the annular sequence. The multilayer fully connected network contains a fully connected layer with 256 hidden nodes and uses a ReLU activation function for nonlinear transformation. The input is the flattened one-dimensional composite vector, and the output dimension is the same as the annular sequence dimension of the dual-path rearrangement feature tensor, i.e., a one-dimensional offset array of length 256. This is then reshaped into a 12×256 time-compensated parameter matrix, where 12 corresponds to the annular number and 256 corresponds to the annular sequence length. In this embodiment, the number of hidden nodes is 256; in other embodiments, it can be adjusted within the range [128, 512] depending on the feature dimension.

[0059] Based on the generated time-compensation parameter matrix, a one-dimensional deformable sampling operation mechanism is configured to replace the feature extraction positions originally determined according to the uniform rigidity rule with sub-pixel-level offset positions calculated based on the time-compensation parameter matrix. When the target sampling address shifts to a non-integer position due to changes, a bilinear interpolation algorithm is used for numerical padding and feature alignment, ultimately obtaining a multi-channel temporally aligned feature tensor that corrects for the effects of thermal vibration and spatiotemporal misalignment.

[0060] Taking the continuous frames acquired by the electron beam melting equipment used in this embodiment as an example, the two-frame images of the previous sampling time of 15ms and the current sampling time of 30ms are compared. After extracting discrete control base points, a first-order displacement vector matrix is ​​obtained by least-squares optical flow calculation. The vector matrix is ​​converted into a one-dimensional offset of each annulus through an annular index mapping template. After being aligned with the annular gradient parameter dimension, it is concatenated to the channel dimension of the dual-path rearrangement feature tensor. The fully connected network processes the concatenated tensor and outputs a time-compensated parameter matrix. The one-dimensional deformable sampling operation performs interpolation alignment on the current frame features based on this matrix to obtain a time-aligned feature tensor.

[0061] S5, Layered aggregation and state discrimination output.

[0062] Spatial pyramid pooling is used to spatially segment the temporally aligned feature tensor into non-overlapping grid regions of two different scales: 2×2 and 4×4. For each grid region, max pooling and average pooling functions are applied for local feature dimensionality reduction extraction. All pooled feature matrices are flattened and concatenated into a high-dimensional aggregated vector, completing hierarchical aggregation to extract global feature parameters. Spatial pyramid pooling pools the input features at multiple grid scales, and the concatenated vectors form a fixed-length aggregated representation.

[0063] From an engineering perspective, the 2×2 grid captures coarse-grained regional statistics at the annular level, while the 4×4 grid corresponds to a finer local spatial distribution. The concatenation of pooling results at these two scales allows the global feature parameters to simultaneously consider both macroscopic structure and microscopic texture.

[0064] The time-compensation parameter matrix is ​​flattened along the annular number dimension and the annular sequence dimension into a one-dimensional vector of length 3072; the channel correlation parameter is a one-dimensional vector of length 48 composed of the mean correlation coefficients of each channel arranged in channel number order; the threshold and compression adjustment parameters are all single-element values. The high-dimensional aggregated feature parameters, the flattened one-dimensional vector of time-compensation parameters, the one-dimensional vector of channel correlation parameters, the threshold, and the compression adjustment parameters are concatenated sequentially along the same vector dimension to synthesize a single state representation vector. This state representation vector is input into the hidden layer nodes of the LSTM network according to the time-step sequence to extract the long-term temporal feature variation patterns. The hidden state and cell state of the LSTM network are initialized to all zero vectors in the first time step. The LSTM selectively updates and forgets the hidden state through a gating mechanism, making it suitable for modeling long-sequence dependencies. The hidden state output vector of the LSTM network in the last time step is fed into a multilayer perceptron classifier consisting of three fully connected layers and one Softmax activation function.

[0065] The contribution mechanism of the threshold to the classification of the molten pool state lies in the fact that the threshold is obtained by weighting and fusing the local energy representation parameters, channel correlation parameters, and annular gradient parameters, and then taking the median. All three types of parameters are directly derived from the brightness, edge, and texture features of the molten pool image. The brightness distribution, edge morphology, and texture structure of the molten pool differ in four states: stable, collapsed, shifted, and overburned. In the collapsed state, the rapid reduction of the molten pool area increases the difference in the mean features of adjacent annular zones; in the shifted state, the centroid shifts, altering the feature distribution within each annular zone; and in the overburned state, the molten pool region maintains a consistently high gray level, causing an overall change in the energy distribution of each channel. Therefore, as a global statistical measure of the fusion score of the three types of parameters, the threshold's value changes with the molten pool state, carrying discriminative information representing the macroscopic state of the molten pool. Simultaneously, the threshold serves as the dividing line between the enhanced and suppressed channel groups. Inputting it into the state discrimination model allows the classifier to perceive the baseline level of the current frame's feature channel division and to interpret the features from both paths in the dual-path rearrangement feature tensor consistently. Compression adjustment parameters reflect the overall response intensity of the features after compression of the channel group, channel correlation parameters reflect the degree of information overlap between channels, and time compensation parameters reflect the amplitude and direction of the deformation of the melt pool boundary in adjacent frames. Together with the feature parameters, they constitute the comprehensive discrimination input.

[0066] The multilayer perceptron performs forward propagation to calculate and output probability distribution arrays for four categories: stable state, collapsed state, offset state, and overheated state. The `argmax` function is called to find the index of the highest value in the probability array, and this index is mapped to one of the four pre-defined category labels, which is then output as the molten pool monitoring result. In other words, the category with the highest probability is the molten pool's judgment state for the current frame, and the four states cover typical working conditions in the tantalum-niobium alloy smelting process.

[0067] The aforementioned state discrimination model requires offline training before deployment. The collected labeled molten pool time-series dataset is divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. Labels are determined into four categories based on the following quantification criteria: a stable state is indicated when the molten pool area fluctuation is within 5% of a set baseline area and the grain uniformity of the finished product is acceptable; a collapsed state is indicated when the molten pool area shrinks by more than 20% of the set baseline area within three consecutive frames; an offset state is indicated when the centroid coordinates of the molten pool relative to the previous stable frame exceed 10 pixels; and an overburned state is indicated when the highest grayscale value of the molten pool region exceeds a set upper limit for five consecutive frames and the finished product shows local grain coarsening. These thresholds are determined by process experts based on their production experience in electron beam melting of tantalum-niobium alloys. During the training phase, the cross-entropy loss function is used to measure the deviation between the model's predicted probability distribution and the true labels. The cross-entropy loss function is calculated by taking the negative logarithm of the predicted probability corresponding to the true category label. The Adam optimizer is used to perform parameter updates, with an initial learning rate of 1×10⁻⁶. -3 The learning rate is decayed to 50% of the previous stage every 50 training epochs. During training, the F1 score on the validation set is used as the early stopping criterion. Training is terminated when the F1 score on the validation set does not improve for 20 consecutive epochs, and the model weight with the highest F1 score on the validation set is retained for subsequent inference deployment.

[0068] A dataset of visual monitoring of the molten pool during electron beam melting of tantalum-niobium alloys was used as the benchmark for evaluation. This dataset includes 15,000 time series sequences of standard process good products and 5,000 time series sequences of abnormal drift accompanied by thermal vibration defects. The model evaluation metrics used were network classification prediction accuracy, F1 score, and mean absolute error. The F1 score, the harmonic mean of precision and recall, reflects the model's comprehensive ability to distinguish between different categories.

[0069] The experiment was set up with four validation groups: the basic control group was fed into the standard residual network based on the single-layer original brightness features, without using annular rearrangement and time alignment processing; variant group 1 added a three-branch encoder for brightness, edge and texture to the basic control group and combined it with annular sequence transformation processing; variant group 2 combined energy threshold partitioning and dual-path rearrangement mechanism in series with variant group 1; and the complete model group added an operation based on a one-dimensional deformable optical flow time compensation matrix to variant group 2.

[0070] Reference Figure 2 Under the same hyperparameter settings and a uniform 200 training epochs of optimization iterations, the performance of the four models on the test set is as follows: The baseline control group achieved a prediction accuracy of 82.4%, an F1 score of 80.1%, and a mean absolute error of 4.15. After introducing a three-branch encoder and annular sequence conversion, the accuracy of variant group 1 increased to 87.6%, the F1 score reached 85.3%, and the mean absolute error decreased to 3.02. Based on this, variant group 2, with cascaded energy thresholding and dual-path rearrangement, achieved an accuracy of 91.2%, an F1 score of 89.8%, and a mean absolute error of 2.11. The complete model group, including all processing steps, achieved an accuracy of 96.8% and an F1 score of 95.5%, with the mean absolute error decreasing to 1.03.

[0071] The over 5 percentage point increase in accuracy from the basic control group to variant group 1 demonstrates that the three-branch encoder (luminance, edge, and texture) combined with annular sequence transformation processing can detect melt pool boundary details more sensitively than single-layer luminance features. Variant group 2's F1 score improved by 4.5 percentage points compared to variant group 1, rising from 85.3% to 89.8%, confirming the suppressive effect of energy thresholding and dual-path rearrangement mechanisms on redundant interference from thermal radiation background. The complete model group achieved the best recognition rate of 96.8%. Optical flow tracking combined with a multi-layer feedforward network generated a compensation dictionary, solving the spatiotemporal misalignment problem caused by frame drops and oscillations under complex conditions. The sub-pixel-level alignment provided by the one-dimensional deformable resampling mechanism achieved optimal alignment accuracy for low-level features, improving the reliability of deep temporal inference.

[0072] This invention also provides a machine vision-based tantalum-niobium alloy molten pool condition monitoring system, including a processor and a memory. The memory stores computer program instructions, and the processor reads and executes the computer program instructions in the memory to perform the above-described machine vision-based tantalum-niobium alloy molten pool condition monitoring method.

[0073] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0074] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise expressly and specifically defined.

Claims

1. A machine vision-based method for monitoring the condition of a tantalum-niobium alloy molten pool, characterized in that, include: S1. Obtain the temporal image of the molten pool and preprocess it to obtain the input feature tensor; extract the brightness features, edge gradient features and texture features of the input feature tensor, and obtain a multi-component feature map through convolutional encoding; S2. Construct a ring index based on the radial distance from the center of the molten pool to the boundary, and rearrange the multi-component feature maps to obtain the ring ordered feature tensor; S3. Based on the annular ordered feature tensor, extract local energy representation parameters, channel correlation parameters, and annular gradient parameters, and determine the threshold by weighted fusion, including: the local energy representation parameters are obtained by normalizing the sum of squares of all elements in each feature channel; the channel correlation parameters are obtained by the mean of the Pearson correlation coefficients between the feature vectors of any two channels; the annular gradient parameters are obtained by the mean of the absolute values ​​of the differences between the feature mean values ​​of adjacent annular zones; the three types of parameters are weighted and summed according to preset weights to obtain the comprehensive contribution score of each channel, and the median of the comprehensive contribution score is taken as the threshold. According to the threshold, the feature channels are divided into enhancement channel group and suppression channel group and rearranged respectively. The compression adjustment parameters are extracted to obtain the dual-path rearranged feature tensor. S4. Initialize the time compensation parameter of the first sampling moment to zero. Then, determine the time compensation parameter based on the change information of the melt pool boundary at adjacent moments. Combine the ring gradient parameter to perform inter-frame compensation on the dual-path rearrangement feature tensor to obtain the temporally aligned feature tensor. S5. Perform region division and hierarchical aggregation on the time-aligned feature tensor to extract feature parameters. Combine the feature parameters with threshold, compression adjustment parameters, time compensation parameters and channel-related parameters and input them into the state discrimination model to output the monitoring results of the melt pool.

2. The machine vision-based method for monitoring the state of a tantalum-niobium alloy molten pool according to claim 1, characterized in that, The preprocessing to obtain the input feature tensor includes: extracting the maximum and minimum gray values ​​of all pixels in the molten pool time series image, and linearly mapping the original gray values ​​of each pixel to a closed interval of 0 to 1 through extreme value normalization calculation; calculating the centroid two-dimensional coordinate point of the molten pool region in the normalized image based on the gray distribution, and truncating a square region with a side length of 64 pixels centered on the centroid two-dimensional coordinate point as the region of interest, and performing edge zero-padding alignment processing when the truncated range exceeds the boundary of the original image; for the non-uniformly sampled molten pool time series image, setting the time alignment reference axis according to a fixed time span step value, regenerating the image data corresponding to each time node through cubic spline interpolation method, and then stitching them together in the time dimension to construct a continuous three-dimensional input feature tensor.

3. The machine vision-based method for monitoring the state of a tantalum-niobium alloy molten pool according to claim 1, characterized in that, The process of obtaining the multi-component feature map includes: preserving the input feature tensor according to the arrangement of pixel rows and columns, as a single-layer original two-dimensional brightness feature matrix; applying a 3×3 Sobel filter kernel with a stride of 1 to perform convolution operations on the pixels to be processed along the X and Y directions respectively, and combining the bidirectional first-order difference results to generate a single-layer edge gradient feature two-dimensional matrix; mapping the pixel values ​​in the 0 to 1 interval proportionally to 16 discrete integer gray levels and then establishing a 16th-order gray-level co-occurrence matrix within a local sliding window, calculating the angular second moment distribution value of each local window based on the gray-level co-occurrence matrix to generate a single-layer texture feature two-dimensional matrix; using three 3×3 convolutional layers with 16 channels to process the above three two-dimensional matrices in parallel, and stitching them together along the channel depth direction to construct a 48-channel multi-component feature map.

4. The machine vision-based method for monitoring the state of a tantalum-niobium alloy molten pool according to claim 1, characterized in that, The process of obtaining the ordered annular feature tensor includes: acquiring spatial grid data of multi-component feature maps, calculating the Euclidean distance from each coordinate sampling point to the center of the feature map; establishing a discrete radius sequence with a preset number of pixels as the radius step size, and constructing a set of concentric annular zones that do not intersect from the inside out; traversing the Euclidean distances and assigning a unique annular zone number to each coordinate sampling point according to its radius interval, generating a two-dimensional mapping template as an index for the annular zones; traversing the two-dimensional mapping template in ascending order of the annular zone number, aggregating the features within the same annular zone into a one-dimensional sequence, aligning the sequence length of each annular zone to a preset fixed number of pixels through truncation or zero-padding operations, and retaining the channel dimension to reshape it into an ordered annular feature tensor with an annular zone number dimension, a spatial sequence dimension, and a channel dimension.

5. The machine vision-based method for monitoring the state of a tantalum-niobium alloy molten pool according to claim 1, characterized in that, The step of dividing the feature channels into enhancement channel groups and suppression channel groups based on the threshold includes: parsing the local energy representation parameters and extracting the energy quantization value corresponding to each feature channel; comparing the energy quantization value with the threshold; assigning feature channels with energy quantization values ​​greater than the threshold to the enhancement channel group, and rounding up the number of channels in the enhancement channel group to a multiple of 2 by adding all-zero channels; assigning the remaining feature channels to the suppression channel group, and adding all-zero channels to align the number of channels in the suppression channel group to an even number when the number of channels in the suppression channel group is odd.

6. The machine vision-based method for monitoring the state of a tantalum-niobium alloy molten pool according to claim 1, characterized in that, The extraction and compression adjustment parameters to obtain the dual-path rearrangement feature tensor include: performing a one-dimensional PixelShuffle rearrangement on the features within the enhanced channel group, setting the amplification factor of the annular space dimension to 2, and expanding the length of the annular space dimension to twice by transferring the data from the channel dimension to the annular space dimension; applying a 1×1 convolution to the features within the suppressed channel group to perform channel dimensionality reduction, reducing the number of channels to 50% of the original, and extracting the mean of the 1×1 convolution output features in the channel dimension as the compression adjustment parameter; using a one-dimensional linear interpolation algorithm to align the lengths of the amplified and reduced features to the same size, and concatenating them in the channel dimension to generate the dual-path rearrangement feature tensor.

7. The machine vision-based method for monitoring the state of a tantalum-niobium alloy molten pool according to claim 1, characterized in that, The process of obtaining the temporally aligned feature tensor includes: comparing the edge gradient features of the two-frame images obtained at the previous sampling time and the current sampling time in adjacent time steps, applying the least squares method to calculate optical flow and outputting a first-order displacement vector matrix of all discrete control base points; converting the first-order displacement vector matrix into a one-dimensional offset of the corresponding annular sequence based on the annular index, aligning the one-dimensional offset with the dimension of the annular gradient parameters through a spatial expansion operation, concatenating it with the current frame features of the dual-path rearrangement feature tensor in the channel dimension, and inputting it into a multi-layer fully connected network to calculate weights and generate a time compensation parameter matrix; and interpolating and aligning the current frame features based on a one-dimensional deformable sampling operation, determining the offset of the feature sampling position according to the time compensation parameter matrix, to obtain the temporally aligned feature tensor.

8. The machine vision-based method for monitoring the state of a tantalum-niobium alloy molten pool according to claim 1, characterized in that, The output of the melt pool monitoring results includes: dividing the temporally aligned feature tensor into at least two different scales of non-overlapping grid regions in the spatial dimension; performing max pooling and average pooling on each grid region to extract local features; flattening and concatenating the pooled features into a high-dimensional aggregate vector as feature parameters; flattening the time compensation parameters into a one-dimensional vector; arranging the channel-related parameters in channel order into a one-dimensional vector; using the threshold and compression adjustment parameters as single-element values, and concatenating them with the feature parameters along the same vector dimension to form a state representation vector; after temporal feature extraction via a long short-term memory network, taking the hidden state output vector of the last time step and inputting it into a multilayer perceptron classifier to output the melt pool monitoring results.

9. A machine vision-based tantalum-niobium alloy molten pool condition monitoring system, characterized in that, It includes a processor and a memory, the memory storing computer program instructions, the processor reading and executing the computer program instructions in the memory to perform the machine vision-based tantalum-niobium alloy molten pool condition monitoring method according to any one of claims 1 to 8.

Images