Novel ionization vacuum gauge hot cathode filament damage state monitoring device and method

Abstract

The invention is suitable for the field of hot cathode ionization vacuum gauge filament life prediction and monitoring, and provides a novel ionization vacuum gauge hot cathode filament damage state monitoring device and method, and the device comprises a damage device and a state monitoring device. The damage device comprises a hand-cranking digital display sliding table, a first flange nut, a second flange nut, a third flange nut, a fourth flange nut, a short pressing plate, a first screw rod, a second screw rod, an iridium-yttrium oxide lamp filament, a cross-shaped sliding table, a long pressing plate, a bread board, a nut and a cutter fixing clamp. The state monitoring device comprises a molecular pump set, a vacuum valve, a cold cathode ionization vacuum gauge, a vacuum chamber, a hot cathode ionization vacuum gauge, a controller, an upper computer and a high-speed camera. The operation process of the damage system provided by the invention completely meets the requirement for damage to the coating width and the depth of the lamp filament, and the precision is relatively accurate. Moreover, the change of the surface of the lamp filament in the working process of the lamp filament can be observed, and real-time damage detection can be carried out on the lamp filament on the basis of processing image information in the experiment process.

Description

Technical Field

[0001] The present invention relates to the field of prediction and monitoring of the filament life of a hot cathode ionization vacuum gauge, and particularly to a monitoring device and method for the damage state of the hot cathode filament of a new type of ionization vacuum gauge. Background Art

[0002] As the core sensor for high / ultra-high vacuum measurement, the performance of the hot cathode ionization vacuum gauge directly affects the stability of the equipment. In cutting-edge fields such as semiconductor manufacturing, particle accelerators, and space simulation chambers, during long-term operation, due to factors such as high-temperature evaporation, local hot spot formation, or mechanical vibration of the filament, the surface morphology gradually deteriorates, ultimately leading to a decrease in emission efficiency or fracture failure.

[0003] With the development of cutting-edge fields and the increasing requirements for vacuum measurement, the requirements for hot cathode ionization vacuum gauges have led to new research and application values for the research of hot cathode ionization vacuum gauges. Summary of the Invention

[0004] The present invention provides a monitoring device and method for the damage state of the hot cathode filament of a new type of ionization vacuum gauge. Through the quantitative damage of the filament coating and the internal filament, and a damage factor prediction method that integrates image processing technology and Graph Neural Network (GraphSAGE) in the experiment, through real-time image analysis of the filament morphology and combined with dynamic feature modeling, accurate prediction of the damage degree is achieved.

[0005] A monitoring device for the damage state of the hot cathode filament of a new type of ionization vacuum gauge includes a damage device and a state monitoring device. Among them, the damage device includes:

[0006] A hand-crank digital display slide table, first to fourth flange nuts, a short pressure plate, a first screw, a second screw, an iridium-yttrium oxide filament, a cross slide table, a long pressure plate, a breadboard, nuts, and a tool fixing clamp;

[0007] The breadboard is threadedly connected with a hand-crank digital display slide table, a first screw, a second screw, and a cross slide table; a tool fixing clamp is fixedly connected to the hand-crank digital display slide table by threading; the first screw is threadedly connected with a first flange nut and a second flange nut; the second screw is threadedly connected with a third flange nut and a fourth flange nut, and a short pressure plate is clamped between the first flange nut and the second flange nut; a long pressure plate is clamped between the third flange nut and the fourth flange nut; the iridium-yttrium oxide filament is fixed on the cross slide table by the short pressure plate and the long pressure plate. The cross slide table is threadedly connected to the breadboard; the cross slide table is vertically assembled with both the short pressure plate and the long pressure plate, where: both the short pressure plate and the long pressure plate are located above the cross slide table, and the short pressure plate and the long pressure plate are parallel to the cross slide table.

[0008] Preferably, the tool fixing clamp consists of two parts, the upper part is directly connected to the hand-operated digital display slide table, and the lower part is connected to the upper part by threads and fixed to the hand-operated digital display slide table with a nut.

[0009] Preferably, the cross slide table is assembled from two one-way slide tables and connected by threads.

[0010] Preferably, the condition monitoring device includes:

[0011] A molecular pump group (composed of a first molecular pump and a second molecular pump), a vacuum valve, a cold cathode ionization vacuum gauge, a vacuum chamber, a hot cathode ionization vacuum, a controller, a host computer, and a high-speed camera; the molecular pump group is connected to the vacuum valve, the vacuum valve is connected to the vacuum chamber, one end of the vacuum chamber is connected to the cold cathode ionization vacuum gauge, and the other end is connected to the hot cathode ionization vacuum gauge. The hot cathode ionization vacuum gauge is connected to its controller and the host computer to monitor the filament, and the hot cathode ionization vacuum gauge is connected to the high-speed camera to photograph the surface of the filament in the hot cathode lamp ionization vacuum gauge during operation.

[0012] Preferably, the vacuum chamber is a KF25 three-way joint.

[0013] Another object of the present invention is to provide a method for monitoring the damage state of the hot cathode filament of a new type of ionization vacuum gauge, which is characterized in that the method is implemented based on the above-mentioned device for monitoring the damage state of the hot cathode filament of a new type of ionization vacuum gauge, and the method includes the following steps:

[0014] Step 1: First, install the blade on the tool fixing clamp, place the iridium-yttrium oxide filament on the cross slide table, and use the first to fourth flange nuts to control the short pressing plate and the long pressing plate to fix the filament. Stick the blade to the filament and record the number A on the hand-operated digital display slide table.

[0015] Step 2: Determine the damage position, width, and depth of the iridium-yttrium oxide filament. Use the cross slide table to determine the position, and use the forward and backward movement of the cross slide table to damage the filament in terms of width; realize the depth damage of the filament by moving the cross slide table left and right, and the depth is determined by the hand-operated digital display slide table.

[0016] Step 3: After damaging the filament, loosen the first to fourth flange nuts to release the short pressing plate and the long pressing plate, and remove the filament above to conduct a filament life experiment.

[0017] Step 4: Weld the removed filament into the hot cathode ionization vacuum gauge, start the molecular pump group, open the vacuum valve, start the controller and the host computer, and observe the cathode current, cathode voltage, filament temperature, bias voltage, grid electron current, and vacuum chamber pressure of the filament during the change process. Wait for the pressure to reach 10 -4 order of magnitude;

[0018] Step 5: After reaching the negative fourth order of magnitude, use the host computer to record data, observe the characteristic changes on the surface of the filament with a high-speed camera, and denoise and enhance the contrast of the original image to improve the recognizability of the damaged area;

[0019] Step 6: Based on the filament hot spot model, define a filament damage factor to quantitatively characterize the damage degree of evaporation or morphological changes on the filament surface; establish a mapping relationship between the image information and the damage factor to form a sample set, and divide the training set and the test set proportionally; perform image processing on the training set: use the watershed algorithm to segment the damaged area and the normal area in the filament image; extract the filament contour features through edge detection technology; fuse the temporal features of the filament at different time points to construct a dynamic damage evolution data set; construct a GraphSAGE graph neural network prediction model, with image features and temporal features as inputs and the damage factor as the output, and train the model;

[0020] Step 7: Real-time prediction - Input the real-time collected image into the trained prediction model to output the damage factor D.

[0021] Preferably, in Step 5, BM3D is used for denoising the original image.

[0022] Preferably, in Step 5, the contrast enhancement is a Transformer-based method.

[0023] Preferably, based on the filament hot spot model, defining the filament damage factor includes the following steps:

[0024] Filament hot spot model: Based on the heat conduction equation and evaporation kinetics, define the damage factor D as:

[0025] ;

[0026] Where: A hotspot / A total : The area ratio of the hot spot region (calibrated by an infrared thermal imager); : The local temperature change rate; ΔS: The contour distortion index (quantified by edge detection); α, β, γ are weight coefficients, calibrated through experiments.

[0027] Preferably, EdgeNeXt is used for the edge detection, and the watershed algorithm needs to preset the prior morphological parameters of the filament structure to improve the segmentation accuracy.

[0028] Preferably, the graph structure of the GraphSAGE graph neural network prediction model uses the local region of the filament as nodes and the spatial or temporal association as edges, and uses a mean aggregator to update the node features.

[0029] The beneficial effects achieved by the present invention:

[0030] (1) In the present invention, quantitative coating of the filament and damage to the filament surface can be carried out by a mechanical device. Since the hand-cranked digital display slide table enables the accuracy to reach the micron level, the width damage to the filament can be controlled by the cross slide table, and at the same time, the hand-cranked digital display slide table is used to control the depth damage.

[0031] (2) In the present invention, the blade fixing clamp adopts a triangular form, and the blade is fixed by a nut and a screw rod. Therefore, during the cutting process, the cutting can be kept stable.

[0032] (3) In the present invention, high-precision dynamic monitoring: through the watershed algorithm and temporal sequence fusion, the spatio-temporal evolution of damage is accurately captured. Graph structure modeling: GraphSAGE can effectively characterize the local and global associations of the filament, which is superior to the traditional pixel-level processing method.

[0033] (4) In the present invention, through a special device, the evaluation has high accuracy, few influencing factors, high efficiency, and is convenient and simple to operate. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic structural diagram of the damage device in the present invention.

[0035] Figure 2 It is a schematic structural diagram of the state monitoring device in the present invention.

[0036] Figure 3 It is a flow chart for evaluating the damage of the hot cathode filament in the present invention.

[0037] Figure 4 It is a comparison diagram of filament processing images in the present invention.

[0038] Figure 5 It is a flow chart of the previous pretreatment in the present invention.

[0039] Figure 6 It is a flow chart for constructing the GraphSAGE graph neural network prediction model in the present invention.

[0040] ANNEX MARKING NOTES:

[0041] 1 - Hand-cranked digital display slide table, 2 - First flange nut, 3 - Short pressure plate, 4 - First flange nut, 5 - First screw rod, 6 - Iridium-yttrium oxide filament, 7 - Cross slide table, 8 - Third flange nut, 9 - Second screw rod, 10 - Third flange nut, 11 - Long pressure plate, 12 - Breadboard, 13 - Nut, 14 - Tool fixing clamp, 15 - First molecular pump, 16 - Second molecular pump; 17 - Vacuum valve; 18 - Cold cathode ionization vacuum gauge; 19 - Vacuum chamber; 20 - Hot cathode ionization vacuum gauge; 21 - Controller; 22 - Host computer; 23 - High-speed camera. DETAILED DESCRIPTION OF THE INVENTION

[0042] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Therefore, the detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents the selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0043] In the description of the present invention, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality of" means two or more, unless otherwise specifically defined.

[0044] Please refer to Figure 1 and Figure 2 , an embodiment of the present invention provides a monitoring device for the damage state of the hot cathode filament of a new type of ionization gauge, including a damage device and a state monitoring device. Among them, the damage device includes:

[0045] A hand-cranked digital display slide table 1, first to fourth flange nuts 2 / 4 / 8 / 10, a short pressure plate 3, a first screw 5, a second screw 9, an iridium-yttrium oxide filament 6, a cross slide table 7, a long pressure plate 11, a breadboard 12, a nut 13, and a tool fixing clip 14;

[0046] The breadboard 12 is threadedly connected with a hand-cranked digital display slide table 1, a first screw 5, a second screw 9, and a cross slide table 7; the tool fixing clip 14 is fixedly connected to the hand-cranked digital display slide table 1 by threading; the first screw 5 is threadedly connected with a first flange nut 2 and a second flange nut 4; the second screw 9 is threadedly connected with a third flange nut 8 and a fourth flange nut 10, and a short pressure plate 3 is clamped between the first flange nut 2 and the second flange nut 4; a long pressure plate 11 is clamped between the third flange nut 8 and the fourth flange nut 10; the iridium-yttrium oxide filament 6 is fixed on the cross slide table 7 by the short pressure plate 3 and the long pressure plate 11. The cross slide table 7 is threadedly connected to the breadboard 12; the cross slide table 7 is vertically assembled with both the short pressure plate 3 and the long pressure plate 11, where: both the short pressure plate 3 and the long pressure plate 11 are located above the cross slide table 7, and the short pressure plate 3 and the long pressure plate 11 are parallel to the cross slide table 7.

[0047] In this embodiment, the tool fixing clamp 14 consists of two parts, the upper part is directly connected to the hand-operated digital display slide table 1, and the lower part is connected to the upper part by threads and fixed to the hand-operated digital display slide table 1 with a nut 13.

[0048] In this embodiment, the cross slide table 7 is assembled by two one-way slide tables and connected by threads.

[0049] Please refer to Figure 2 , in this embodiment, the state monitoring device includes:

[0050] A molecular pump group (composed of a first molecular pump 15 and a second molecular pump 16), a vacuum valve 17, a cold cathode ionization vacuum gauge 18, a vacuum chamber 19, a hot cathode ionization vacuum gauge 20, a controller 21, a host computer 22, and a high-speed camera 23; the molecular pump group is connected to the vacuum valve 17, the vacuum valve 17 is connected to the vacuum chamber 19, one end of the vacuum chamber 19 is connected to the cold cathode ionization vacuum gauge 18, the other end is connected to the hot cathode ionization vacuum gauge 20, the hot cathode ionization vacuum gauge 20 is connected to its controller 21 and the host computer 22 to monitor the filament, and the hot cathode ionization vacuum gauge 20 is connected to the high-speed camera 23 to capture the surface of the filament in the hot cathode ionization vacuum gauge 20 during operation.

[0051] In order to better record the surface morphology of the filament, a high-speed camera 23 is used to record the morphology during the process.

[0052] In order to facilitate data recording and comprehensive data recording, the host computer 22 is used for data acquisition.

[0053] In order to ensure the capture of fine damage on the filament surface, a high-speed camera 23 is used and equipped with a resolution microscopic lens.

[0054] In order to ensure that the image sequence can completely capture the dynamic process of damage, the sampling frequency of the high-speed camera 23 needs to cover the working cycle of the filament.

[0055] In order to describe the damage of the filament, the damage factor is a combined parameter of the evaporation rate of the filament surface, the area ratio of the hot spot region, or the contour distortion index.

[0056] In order to improve the segmentation accuracy, EdgeNeXt is used for edge detection, and the prior morphological parameters of the filament structure need to be preset for the watershed algorithm.

[0057] In order to update the node features, the graph structure of the GraphSAGE model uses the local area of the filament as nodes and spatial or temporal associations as edges, and a mean aggregator is used.

[0058] In this embodiment, the vacuum chamber 19 is a KF25 three-way joint.

[0059] Please refer to Figures 3 to 6 , another object of the present invention is to provide a method for monitoring the damage state of the hot cathode filament of a new type of ionization vacuum gauge, characterized in that the method is implemented based on the above-mentioned device for monitoring the damage state of the hot cathode filament of a new type of ionization vacuum gauge, and the method includes the following steps:

[0060] Step 1: First, install the blade on the tool fixing clamp 14, place the iridium-yttrium oxide filament 6 on the cross slide 7, and use the first to fourth flange nuts 2 / 4 / 8 / 10 to control the short pressing plate 3 and the long pressing plate 11 to fix the filament. Stick the blade to the filament and record the number A on the hand-cranked digital display slide 1;

[0061] Step 2: Determine the damage position, width, and depth of the iridium-yttrium oxide filament. Use the cross slide 7 to determine the position, and use the forward and backward movement of the cross slide 7 to damage the filament in terms of width; realize the depth damage of the filament by moving the cross slide 7 left and right, and the depth is determined by the hand-cranked digital display slide 1;

[0062] Step 3: After damaging the filament, loosen the first to fourth flange nuts 2 / 4 / 8 / 10 to release the short pressing plate 3 and the long pressing plate 11, and remove the filament above to conduct the filament life experiment;

[0063] Step 4: Weld the removed filament into the hot cathode ionization vacuum gauge 20 and connect it to the device in Figure 2 . Start the molecular pump set, open the vacuum valve 17, start the controller 21 and the upper computer 22, and observe the cathode current, cathode voltage, filament temperature, bias voltage, grid electron current, and vacuum chamber pressure of the filament during the change process. Wait until the pressure reaches 10 -4 order of magnitude;

[0064] Step 5: After reaching the 10 -4 order of magnitude, use the upper computer 22 to record data, and use the high-speed camera 23 to observe the characteristic changes on the filament surface, and perform denoising and contrast enhancement on the original image to improve the recognizability of the damaged area;

[0065] Step 6: Image feature extraction and dataset construction, graph neural network model construction and training;

[0066] Based on the filament hot spot model, a filament damage factor is defined to quantitatively characterize the damage degree of evaporation or morphological changes on the filament surface. A mapping relationship is established between the image information and the damage factor to form a sample set, and the training set and test set are divided according to a certain proportion. Image processing is performed on the training set: the watershed algorithm is used to segment the damaged area and the normal area in the filament image; the filament contour features are extracted through edge detection technology; the temporal features of the filament at different time points are fused to construct a dynamic damage evolution data set; a GraphSAGE graph neural network prediction model is constructed, with image features and temporal features as inputs and the damage factor as the output, and the model is trained.

[0067] In the watershed algorithm segmentation of this embodiment, first, preprocessing is performed. The image is grayscale and binarized, and morphological operations (erosion - dilation) are used to eliminate small noises. Then, label generation is carried out. Based on the prior morphology of the filament (such as a straight or spiral structure), initial seed points are manually marked to guide the watershed algorithm to accurately segment the damaged area. Then, edge detection and contour extraction are performed. The EdgeNeXt is used to extract the filament contour, and parameters such as contour curvature and length change rate are calculated. Then, temporal feature fusion is carried out. The image features at consecutive time points (such as: hot spot area, contour distortion) are stitched according to a time window (such as: 10 frames) to construct a spatio - temporal feature matrix. For details, see Figure 4 , the first to fourth columns respectively represent: the image sequence of the filament with undamaged surface under actual working conditions, the image sequence of the filament with undamaged surface in the laboratory environment, the image sequence of the filament with a 2 - mm surface damage, and the image sequence of the filament with a 4 - mm surface damage; the first to fifth rows respectively represent: the original image, grayscale and binarization, unsharp masking, feature extraction, and segmentation of the damaged area based on the watershed algorithm.

[0068] Step 7: Real - time prediction - The real - time collected images are input into the trained damage factor prediction model, and the damage factor D (range 0 - 1, 0 for no damage, 1 for critical failure) is output; Verification method - Compare the actual observed filament break time with the moment when D = 1 predicted by the model, with an error ≤ 3%. The importance of temporal features is verified through ablation experiments (the prediction error increases by 15% after removal).

[0069] In this embodiment, in step 5, the steps for denoising the original image include:

[0070] Noise estimation: First, the background area is extracted, and the area in the image without the filament is selected; then, the standard deviation is calculated, and the pixel standard deviation σ of the background area is calculated as the noise intensity estimation value;

[0071] Block matching: Input image - noisy image I noisy; First, perform block selection with reference. Divide the image into overlapping reference blocks (e.g., 8×8 pixels) and process them block by block. Then, conduct similar block search. Within the search window (e.g., 32×32 pixels), calculate the Euclidean distance between other blocks and the reference block, and select the top N blocks with the smallest distance (e.g., N = 16) to form a 3D block group;

[0072] 3D transformation and threshold processing: First, perform 3D transformation, that is, perform the following transformations on each 3D block group: 2D transformation, that is, perform 2D discrete cosine transformation on each block; 1D transformation, that is, perform 1D Haar wavelet transformation along the third dimension (the stacking direction of similar blocks), and then apply hard threshold filtering: Apply a hard threshold (e.g., λ = 2.7σ) to the coefficients after 3D transformation to suppress noise; then perform inverse 3D transformation: Inverse-transform the filtered coefficients back to the image domain to obtain the denoised 3D block group;

[0073] Aggregation: Weighted average, that is, aggregate the denoised block group to the original position according to the weight (based on block matching quality) to generate the basic estimated image I basic ;

[0074] Quadratic block matching: Input the basic estimated image I basic , similar block search, that is, the same as the basic estimation step, but use I basic for more accurate block matching;

[0075] Wiener filtering: 3D transformation, that is, perform 3D transformation on the block groups of the noisy image and the basic estimated image respectively, and calculate the Wiener coefficient W:

[0076] ;

[0077] where, T basic is the transformation coefficient of the basic estimated block, and σ is the standard deviation of the noise. Apply the Wiener coefficient to the transformation coefficients of the noisy image, perform frequency-domain filtering, and conduct inverse 3D transformation to restore the filtered block group.

[0078] Quadratic aggregation and post-processing: Weighted average: Aggregate the finally estimated block group to the original position to generate the denoised image I final ; Post-processing: Unsharp masking, with parameters Gaussian kernel radius = 2, intensity = 0.3, to restore the high-frequency details lost due to threshold processing; Local gamma correction, adaptively adjust the gamma value (γ = 0.9) for the damaged area to avoid overexposure, and the output is the denoised image I denoised , PSNR > 40 dB, SSIM > 0.95;

[0079] Among them, the first four steps are basic estimation, and the last three steps are final estimation.

[0080] In the present invention, in step 5, the contrast enhancement is a Transformer-based method, including the following steps:

[0081] Data preprocessing: Block processing, that is, the image is segmented into blocks of 256×256 pixels with an overlapping area of 64 pixels (25% overlap rate), and normalization, that is, the pixel values are linearly mapped from [0, 255] to [-1, 1];

[0082] Model architecture: The core architecture is Uformer - Encoder: Transformer layers, using the SwinTransformer module to capture local and global features through window self-attention; Downsampling, using stride convolution to gradually reduce the resolution, Decoder: Upsampling, restoring the resolution through transposed convolution or pixel shuffling; Skip connection, that is, fusing the multi-scale features of the encoder with the corresponding layers of the decoder to retain detailed information, Contrast enhancement head: Dynamic curve prediction, that is, outputting a pixel-level contrast adjustment curve; Illumination estimation module, that is, predicting the global illumination distribution map to guide local contrast stretching;

[0083] Training strategy:

[0084] (1) Loss function design: Pixel-level loss - L1 loss, forcing the pixel alignment between the enhanced image and the target image:

[0085] ;

[0086] where E represents the expected value, characterizing the average of the batch data; I output and I target represent the model output and the true reference image respectively. The SSIM loss is introduced to maintain structural similarity, that is:

[0087] ;

[0088] where SSIM represents the structural similarity metric function.

[0089] (2) Perceptual loss: Extract features through a pre-trained VGG network to constrain the high-frequency details of the enhanced image:

[0090] ;

[0091] where : The feature extraction function of the VGG network.

[0092] (3) Adversarial loss: Introduce a discriminator network to enhance the visual authenticity of the enhancement result:

[0093] ;

[0094] where D: Discriminator, distinguishing real and generated images; G: Generator, generating enhanced images.

[0095] (4) Training strategy: Two-stage training - pre-training, that is, optimizing the L1+SSIM loss on synthetic data for fast convergence; fine-tuning, that is, jointly optimizing all losses (L1+SSIM+perceptual loss+adversarial loss) on real hot cathode filament data; learning rate scheduling, that is, dynamically adjusting the learning rate using cosine annealing, with the initial value set to 1e-4, and early stopping method, that is, monitoring the validation set loss and terminating the training if it does not decrease for 10 consecutive rounds;

[0096] (5) Model inference and post-processing: Inference process: Input chunking, that is, dividing the high-resolution image into chunks and inputting them into the model one by one; inter-block fusion, that is, performing weighted averaging on the overlapping regions to eliminate boundary artifacts; Post-processing optimization: Unsharp masking, that is, sharpening the damaged edges (radius = 2, intensity = 0.3); Dynamic gamma correction, that is, adaptively adjusting the gamma value according to the local contrast (γ = 0.8 - 1.2); Output: Enhanced image I enhanced , SSIM>0.95, local contrast improvement rate ≥50%.

[0097] In the present invention, by using the heat conduction equation and the filament coating evaporation kinetics, a filament hot spot model is introduced to obtain the filament coating damage factor D:

[0098] ;

[0099] Where: A hotspot / A total : Proportion of the hot spot area in the total area (calibrated with the assistance of an infrared thermal imager); : Local temperature change rate; ΔS: Contour distortion index (quantified by edge detection); α, β, γ are weight coefficients, calibrated through experiments. The damage factor is a combined parameter of the filament surface evaporation rate, the proportion of the hot spot area, or the contour distortion index.

[0100] In the present invention, the watershed algorithm segmentation includes the following steps:

[0101] (1) Gradient calculation: Sobel operator: Horizontal gradient kernel: ; Vertical gradient kernel: ; Gradient magnitude: , where G x and G y are the horizontal and vertical gradients respectively;

[0102] (2) Marker generation: Foreground marker (damaged area): Threshold segmentation, that is: Using the Otsu algorithm to adaptively determine the threshold and segment the high-gradient area; Morphological opening operation, that is: Erosion with a 3×3 kernel followed by dilation to remove noise points; Background marker (normal area): Distance transformation, that is: Calculating the distance from each pixel in the gradient map to the nearest edge; Selecting the farthest area, that is: Taking the top 10% pixels with the largest distance values as the background marker;

[0103] (3) Watershed transformation: Input, i.e., gradient map + foreground / background marker; water filling simulation, i.e., start filling water from the marked area, and the watershed boundary is formed where the water meets; over-segmentation processing, i.e., merge regions with an area < 50 pixels and eliminate fragments; post-processing: morphological closing operation, i.e., dilation with a 3×3 kernel followed by erosion to connect broken edges; artifact filtering, i.e., remove isolated regions (area < 50 pixels, aspect ratio < 1.5); Output: binary mask M damage , Dice coefficient ≥ 0.85.

[0104] In the present invention, EdgeNeXt is used for contour extraction, including the following steps:

[0105] (1) Model configuration: Encoder-decoder architecture: multi-scale Transformer blocks, i.e., extract features at 4 resolution levels (256 → 128 → 64 → 32); noise-adaptive attention, i.e., dynamically adjust the attention weights according to the local noise level;

[0106] (2) Loss function: weighted cross-entropy loss, set the weight of damaged edge pixels = 5; contrastive learning loss, i.e., maximize the similarity of positive sample pairs (different enhanced views of the same image);

[0107] (3) Training process: Data augmentation: randomly add Gaussian noise (σ = 10 - 30), motion blur (kernel size = 5); two-stage training: pre-training, i.e., optimize the cross-entropy loss on synthetic data; fine-tuning, i.e., jointly optimize the cross-entropy and contrastive losses on real data;

[0108] (3) Post-processing and feature extraction: Non-maximum suppression, i.e., retain the local maximum response along the gradient direction, refine the edge to a single pixel, morphological closing operation, i.e., use a 3×3 kernel to connect broken edges;

[0109] (4) Feature calculation: Morphological features include area, perimeter, shape factor; texture features include contrast and energy of gray-level co-occurrence matrix (GLCM);

[0110] (5) Output: feature vector F contour ∈R 64 , including 16-dimensional morphological features and 48-dimensional texture features.

[0111] In the present invention, constructing a GraphSAGE graph neural network prediction model includes the following steps:

[0112] (1)Construct the graph structure: Define nodes, that is, divide the image into 16×16 grids. Each node feature includes the local grayscale mean, gradient intensity, damage label (0 / 1), and temporal change rate. Define edges: Spatial adjacency edges, that is, adjacent grid nodes (up, down, left, right); Temporal correlation edges, that is, across time frames at the same position, with a similarity > 0.8 (cosine similarity). (2) Configure the GraphSAGE model: Embedding dimension - 64, use the mean aggregator as the aggregation function, the number of network layers is 2-layer graph convolution, followed by the ReLU activation function and Dropout (ratio = 0.2) after each layer, and the loss function is the mean squared error - the difference between the predicted damage factor and the true value.

[0113] (3)Temporal consistency loss: Penalize the mutation of prediction results between adjacent frames:

[0114] ;

[0115] Select Adam as the optimizer (learning rate = 0.001). The training strategy is to divide the dataset into a training set and a test set at a ratio of 8:2. Introduce early stopping (patience value = 10) during training to prevent overfitting, and update online, that is: for each newly added frame of image, dynamically update the graph structure and re-predict.

[0116] (4)Multimodal extension: Integrate infrared thermogram data to enhance the node feature dimension.

[0117] (5)Output: A damage factor prediction model with an MAE < 2% on the test set and a real-time inference speed ≥ 25 FPS.

[0118] The filament damage data obtained using the damage device is relatively accurate, and the control process of the device is not complex, easy to operate, and can solve the current accuracy problem of filament damage. And it can observe the changes on the filament surface during the operation of the filament. Through the watershed algorithm and temporal fusion, accurately capture the spatio-temporal evolution of damage. GraphSAGE effectively represents the local and global associations of the filament, superior to traditional pixel-level processing methods. Only requires conventional image acquisition equipment and does not require modification of existing hardware.

[0119] It should be noted that in this article, the term "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article, or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, article, or device. Without further limitations, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, article, or device including that element.

[0120] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be similarly included in the patent protection scope of the present invention.

Claims

1. A monitoring device for the damage state of the hot cathode filament of a new type of ionization vacuum gauge, characterized in that, Comprising a damage device and a state monitoring device, wherein the damage device includes: A hand-cranked digital display slide, first to fourth flange nuts, a short pressure plate, a first screw, a second screw, an iridium-yttrium oxide filament, a cross slide, a long pressure plate, a breadboard, nuts, and a tool fixing clamp; The breadboard is threadedly connected with a hand-cranked digital display slide, a first screw, a second screw, and a cross slide; the tool fixing clamp is fixedly connected to the hand-cranked digital display slide by threading; the first screw is threadedly connected with a first flange nut and a second flange nut; the second screw is threadedly connected with a third flange nut and a fourth flange nut, and a short pressure plate is clamped between the first flange nut and the second flange nut; a long pressure plate is clamped between the third flange nut and the fourth flange nut; the iridium-yttrium oxide filament is fixed on the cross slide by the short pressure plate and the long pressure plate; the cross slide is threadedly connected to the breadboard.

2. A monitoring device for the damage state of the hot cathode filament of a novel ionization vacuum gauge according to claim 1, characterized in that The tool fixing clamp consists of upper and lower parts. The upper part is directly connected to the hand-cranked digital display slide, and the lower part is connected to the upper part by threading and fixed to the hand-cranked digital display slide with a nut.

3. A monitoring device for the damage state of the hot cathode filament of a novel ionization vacuum gauge according to claim 2, characterized in that The cross slide is assembled by two linear slides and is connected by threading.

4. A monitoring device for the damage state of the hot cathode filament of a novel ionization vacuum gauge according to claim 3, characterized in that, The state monitoring device includes: A molecular pump group, a vacuum valve, a cold cathode ionization vacuum gauge, a vacuum chamber, a hot cathode ionization vacuum gauge, a controller, a host computer, and a high-speed camera; the molecular pump group is connected to the vacuum valve, the vacuum valve is connected to the vacuum chamber, one end of the vacuum chamber is connected with a cold cathode ionization vacuum gauge, and the other end is connected with a hot cathode ionization vacuum gauge. The hot cathode ionization vacuum gauge is connected to its controller and the host computer to monitor the filament, and the hot cathode ionization vacuum gauge is connected to the high-speed camera to photograph the surface of the filament in the hot cathode lamp ionization vacuum gauge during operation.

5. A monitoring device for the damage state of the hot cathode filament of a novel ionization vacuum gauge according to claim 4, characterized in that, The vacuum chamber is a KF25 three-way joint.

6. A method for monitoring the damage state of the hot cathode filament of a new type of ionization vacuum gauge, characterized in that, The method is implemented based on a novel ionization vacuum gauge hot cathode filament damage state monitoring device according to claim 5, and the method includes the following steps: Step 1: First, install the blade on the tool fixing clamp, place the iridium-yttrium oxide filament on the cross slide, and use the first to fourth flange nuts to control the short pressure plate and the long pressure plate to fix the filament. Stick the blade to the filament and record the number A on the hand-cranked digital display slide. Step 2: Determine the damage position, width, and depth of the iridium-yttrium oxide filament. Use the cross slide to determine the position and move the cross slide back and forth to damage the filament in terms of width; achieve the depth damage of the filament by moving the cross slide left and right, and the depth is determined by the hand-cranked digital display slide. Step 3: After damaging the filament, loosen the first to fourth flange nuts to release the short pressure plate and the long pressure plate, and remove the filament above to conduct the filament life experiment. Step 4: Weld the removed filament into a hot-cathode ionization vacuum gauge, start the molecular pump set, open the vacuum valve, start the controller and the host computer, and observe the cathode current, cathode voltage, filament temperature, bias voltage, grid electron current, and vacuum chamber pressure of the filament during the change process, and wait for the pressure to reach 10 -4 order of magnitude; Step 5: After reaching the negative fourth order of magnitude, record the data with the host computer, observe the characteristic changes on the filament surface with the high-speed camera, and perform denoising and contrast enhancement on the original image to improve the distinguishability of the damaged area. Step 6: Based on the filament hot spot model, define the filament damage factor to quantitatively characterize the damage degree of evaporation or morphological changes on the filament surface; establish a mapping relationship between the image information and the damage factor to form a sample set, and divide the training set and the test set according to a certain proportion; perform image processing on the training set: use the watershed algorithm to segment the damaged area and the normal area in the filament image; extract the filament contour features through edge detection technology; fuse the temporal features of the filament at different time points to construct a dynamic damage evolution data set; construct a GraphSAGE graph neural network prediction model, with image features and temporal features as inputs and the damage factor as the output, and train the model. Step 7: Real-time prediction - Input the real-time collected image into the trained prediction model to output the damage factor D.

7. A method for monitoring the damage state of the hot cathode filament of a novel ionization vacuum gauge according to claim 6, characterized in that, In step 5, BM3D is used for denoising the original image, and the contrast enhancement is based on the Transformer method.

8. The method for monitoring the damage state of the hot cathode filament of a novel ionization vacuum gauge according to claim 6, characterized in that, Based on the filament hot spot model, the steps for defining the filament damage factor include the following: Filament hot spot model: Based on the heat conduction equation and evaporation kinetics, the damage factor D is defined as: Among them, A hotspot / A total : Proportion of the area of the hot spot region; : Local temperature change rate; ΔS: Contour distortion index; α, β, γ are weighting coefficients.

9. The method for monitoring the damage state of the hot cathode filament of a novel ionization vacuum gauge according to claim 6, characterized in that, The edge detection uses EdgeNeXt, and the watershed algorithm needs to preset the prior morphological parameters of the filament structure to improve the segmentation accuracy.

10. The method for monitoring the damage state of the hot cathode filament of a novel ionization gauge according to claim 6, wherein, The graph structure of the GraphSAGE graph neural network prediction model takes the local area of the filament as nodes and the spatial or temporal association as edges, and uses the mean aggregator to update the node features.

Images