Visual observation system for high temperature and high pressure test of silicone sealant

Abstract

The present application belongs to the technical field of material testing and visual detection, and particularly relates to a visual observation system for high-temperature and high-pressure testing of organic silicon sealant, which comprises a high-pressure sealing cavity, a transparent observation window, a high-frame-rate industrial camera, a multi-spectrum coaxial light source system, an image acquisition and control module, and an image processing and analysis module; multi-scale space-time features are extracted through construction of an image pyramid and a three-dimensional convolutional neural network, a constitutive model of the organic silicon sealant is embedded into a light flow calculation framework to reconstruct a deformation field and a stress field in accordance with physical laws, a space-time graph convolutional network is used to identify rupture, flow and debonding failure modes, a physical information neural network is used to predict failure time and an extension trajectory, and finally, a safety working condition window is output by fusing temperature and pressure data. The present application realizes intelligent, quantitative and accurate observation of the dynamic failure process of the organic silicon sealant under high-temperature and high-pressure working conditions, and provides reliable data support for sealant selection and injection molding process optimization.

Description

Technical Field

[0001] This invention relates to the field of materials testing and visual inspection technology, and in particular to a visual observation system for high-temperature and high-pressure testing of silicone sealant. Background Technology

[0002] In the manufacturing process of automotive transmission parts, coils often experience insulation layer damage due to instantaneous high-temperature and high-pressure impacts during injection molding, leading to short circuits. To address this issue, silicone sealant has been introduced as a protective layer material due to its superior high-temperature resistance, high-pressure resistance, and insulation properties. However, in actual process development, accurately evaluating the dynamic behavior of silicone sealant under injection molding conditions, especially its deformation, flow, and cracking characteristics under high-temperature and high-pressure environments, has become a technical bottleneck.

[0003] Currently, high-temperature and high-pressure testing of silicone sealants mainly relies on offline testing methods, such as: Static thermal aging test: The sealant sample is placed in a high-temperature chamber and heated, and then taken out to test its performance changes. It cannot simulate the dynamic pressure impact process.

[0004] Pressure vessel testing: The sealant is placed in a sealed container and pressure is applied, but there is a lack of real-time observation methods, making it difficult to capture transient failure processes.

[0005] Human observation or video playback: Manual observation through an observation window outside the high-temperature and high-pressure equipment poses safety hazards and cannot achieve quantitative analysis.

[0006] The above methods generally have the following problems: they cannot observe the dynamic behavior of sealants in real time, non-contact, and at high resolution under actual injection molding conditions. Therefore, it is urgent to develop a system that can perform real-time visual observation of silicone sealants under high temperature and high pressure environments to support sealant material selection, process parameter optimization, and protection effect verification. Summary of the Invention

[0007] Based on the technical problems existing in the prior art, this invention proposes a visual observation system for high-temperature and high-pressure testing of silicone sealants.

[0008] This invention proposes a visual observation system for high-temperature and high-pressure testing of silicone sealants, comprising: The high-pressure sealed cavity has a sample stage inside for placing silicone sealant samples; A transparent observation window is located at the top of the high-pressure sealed cavity to provide an optical observation channel; A high frame rate industrial camera, positioned above a transparent observation window, is used to acquire dynamic image sequences of sealant samples under high temperature and high pressure conditions. Multispectral coaxial light source system is used to provide multi-band, uniform, and non-reflective illumination for high frame rate industrial cameras; The image acquisition and control module is used to synchronously control the image acquisition and operating parameter acquisition of a high frame rate industrial camera. The image processing and analysis module is used to extract multi-scale spatiotemporal features, reconstruct deformation fields of physical constraints, identify multimodal failure modes and predict dynamic behavior from the acquired dynamic image sequences, and output quantitative performance parameters of the sealant.

[0009] Preferably, the image processing and analysis module includes: a multi-scale spatiotemporal feature extraction unit, used to construct an image pyramid and use a three-dimensional convolutional neural network to extract multi-level features of the image sequence in spatial scale and temporal dimension; a physical constraint deformation field reconstruction unit, used to embed the constitutive model of the silicone sealant as a physical constraint into the optical flow calculation framework to reconstruct the deformation field and stress field that conform to the material's physical properties; a multi-modal failure mode recognition unit, used to classify and locate the sealant's cracking, flow, and debonding failure modes based on a graph neural network; a dynamic behavior prediction unit, used to predict the sealant's failure time, failure location, and failure propagation trajectory based on a long short-term memory network and a physical information neural network; and a data fusion and process window modeling unit, used to fuse the image analysis results with temperature and pressure data, establish a correlation model between the sealant's performance and operating parameters, and output a safe operating condition window.

[0010] Preferably, the multi-scale spatiotemporal feature extraction unit adopts a feature pyramid network structure to extract image features at at least three spatial scales. The length of the time window corresponding to each scale is inversely proportional to the spatial resolution, wherein the highest spatial resolution corresponds to the shortest time window to capture rapid transient deformation, and the lowest spatial resolution corresponds to the longest time window to capture slow flow trends.

[0011] Preferably, the physical constraint deformation field reconstruction unit embeds the Mooney-Rivlin hyperelastic constitutive model or Maxwell viscoelastic constitutive model of the silicone sealant into the optical flow energy functional in the form of partial differential equations, solves the displacement field that simultaneously satisfies the image brightness constant constraint and the physical constitutive constraint through variational method, and outputs the strain tensor field and stress tensor field that conform to the material physical properties.

[0012] Preferably, the multimodal failure mode recognition unit uses a spatiotemporal graph convolutional network, taking each pixel in the image sequence as a graph node, and the spatial adjacency and temporal correspondence between pixels as graph edges. The spatiotemporal neighborhood information is aggregated through graph convolution operations, and the failure mode label and corresponding confidence level of each pixel are output.

[0013] Preferably, the dynamic behavior prediction unit adopts a physical information neural network with an encoder-decoder structure. The encoder part uses a convolutional long short-term memory network to extract spatiotemporal features, and the decoder part embeds the viscoelastic constitutive equation of the silicone sealant as a physical constraint loss term, so as to ensure that the prediction results conform to physical laws while predicting the deformation field and failure area of ​​the sealant in the future.

[0014] Preferably, the multispectral coaxial light source system includes a combination structure of at least three LED light source modules with different wavelengths, a wavelength-tunable filter, a ring light source and a coaxial optical adapter, wherein the brightness and flicker timing of the LED light source modules with different wavelengths can be independently controlled to obtain multispectral images of the sealant in different spectral bands, so as to distinguish the sealant body, crack area, flow front and impurities.

[0015] Preferably, the device also includes an ultrasonic vibration-assisted observation device, which is located below the sample stage and is used to apply high-frequency, low-amplitude vibrations to the sealant sample to accelerate the microstructure evolution process of the sealant under high temperature and high pressure, shorten the test time, and capture the vibration-induced micro-deformation characteristics through a high frame rate industrial camera.

[0016] Preferably, it also includes a terahertz time-domain spectrometer, which emits terahertz waves to the sealant sample through a transparent observation window and receives the reflected signals. It is used for non-destructive testing of the internal crosslinking density change and micro-defect evolution of the sealant during high temperature and high pressure processes, and forms a multi-modal fusion observation with the visible light images of a high frame rate industrial camera.

[0017] Compared with the prior art, the present invention provides a visual observation system for high-temperature and high-pressure testing of silicone sealants, which has the following beneficial effects: 1. By constructing an image pyramid and a three-dimensional convolutional neural network, the rapid transient deformation and slow flow trend of sealant under high temperature and high pressure are captured simultaneously, solving the technical problem that it is difficult to take into account the feature extraction of multiple time scales in a single scale.

[0018] 2. By embedding the constitutive model of silicone sealant into the optical flow calculation framework, the reconstructed deformation and stress fields conform to the physical properties of the material, avoiding non-physical results that may be generated by pure data-driven methods, and improving the accuracy and reliability of the analysis.

[0019] 3. Spatiotemporal graph convolutional networks are used to classify and locate failure modes such as rupture, flow, and debonding, achieving refined and intelligent identification of failure processes. Compared with the traditional threshold method, it has higher accuracy and generalization ability.

[0020] 4. Based on physical information neural networks, the failure time, failure location, and failure propagation trajectory of sealants are predicted, providing a forward-looking tool for process prediction and quality control.

[0021] 5. By integrating visual analysis results with temperature and pressure data, a quantitative correlation model between sealant performance and operating parameters is established, and a safe operating condition window is output, which directly serves the selection of sealant and the optimization of injection molding process, and has significant engineering application value. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2 This is a block diagram of the image processing and analysis module of the present invention.

[0023] In the diagram: 1. High-pressure sealed cavity; 2. Transparent observation window; 3. High frame rate industrial camera; 4. Multispectral coaxial light source system; 5. Image acquisition and control module; 6. Image processing and analysis module; 61. Multi-scale spatiotemporal feature extraction unit; 62. Physical constraint deformation field reconstruction unit; 63. Multimodal failure mode recognition unit; 64. Dynamic behavior prediction unit; 65. Data fusion and process window modeling unit; 7. Temperature sensor; 8. Pressure sensor; 9. Terahertz time-domain spectrometer; 10. Display terminal. Detailed Implementation

[0024] 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 only some embodiments of the present invention, and not all embodiments.

[0025] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0026] Example 1 like Figure 1 As shown, this embodiment provides a visual observation system for high-temperature and high-pressure testing of silicone sealant, including a high-pressure sealed cavity 1, a transparent observation window 2, a high frame rate industrial camera 3, a multispectral coaxial light source system 4, an image acquisition and control module 5, an image processing and analysis module 6, a temperature sensor 7, a pressure sensor 8, an ultrasonic vibration-assisted observation device, and a terahertz time-domain spectrometer 9.

[0027] The high-pressure sealed chamber 1 is a cylindrical stainless steel structure with an inner diameter of 100mm, a depth of 80mm, and a wall thickness of 15mm, designed to withstand a pressure of 30MPa. A flange interface is located at the top of the chamber, and a transparent observation window 2 is fixed to the flange interface via a metal pressure ring and a fluororubber sealing ring, providing an optical observation channel. The transparent observation window 2 is made of sapphire glass, with a diameter of 60mm and a thickness of 15mm, coated with anti-reflective films on both sides, achieving a transmittance greater than 95% in the 400nm-1000nm wavelength range.

[0028] The chamber has a high-pressure gas inlet and a pressure relief valve on its side. The high-pressure gas inlet connects to a nitrogen cylinder and a pressure regulating valve to adjust the internal pressure. A heating plate is located at the bottom of the chamber, and a PID controller is used for closed-loop temperature control with an accuracy of ±1℃. A sample stage is located in the center of the chamber, with a 25mm diameter circular groove on its surface for positioning the silicone sealant sample. Below the sample stage is an ultrasonic vibration-assisted observation device, consisting of a piezoelectric ceramic transducer and an ultrasonic power supply. Operating at a frequency of 40kHz and an amplitude of 5μm, this device applies high-frequency, low-amplitude vibrations to the sealant sample to accelerate the microstructure evolution process.

[0029] The terahertz time-domain spectrometer 9 is located outside the high-pressure sealed cavity 1. Its transmitter and receiver are coupled to the top of the transparent observation window 2 via optical fiber. The terahertz wave is incident vertically on the surface of the sealant sample, and the receiver collects the reflected signal. This is used for non-destructive testing of the changes in the internal crosslinking density and the evolution of micro-defects in the sealant during the high-temperature and high-pressure process.

[0030] The high frame rate industrial camera 3 is mounted directly above the transparent observation window 2. It uses a CMOS image sensor with a resolution of 2048×2048 pixels, a pixel size of 5.5μm×5.5μm, a maximum frame rate of 1000 frames / second, a 2x telecentric lens, a working distance of 110mm, a field of view of 25mm×25mm, and a depth of field of 2mm.

[0031] A multispectral coaxial light source system 4 provides multi-band, uniform, and reflection-free illumination for a high-frame-rate industrial camera 3. The system includes three LED light source modules with wavelengths of 470nm (blue light), 525nm (green light), and 635nm (red light), as well as a wavelength-tunable filter, a ring light source, and a coaxial optical adapter. The ring light source, located at the front of the telecentric lens, consists of 48 evenly arranged LEDs, each with independently controllable brightness. The coaxial optical adapter, integrated between the telecentric lens and the high-frame-rate industrial camera 3, projects light perpendicularly onto the sample surface via a beam splitter, achieving coaxial illumination. The wavelength-tunable filter, located at the front of the ring light source, selects a specific wavelength of illumination. The three LED light source modules can flash sequentially to acquire images at different spectral bands.

[0032] The image acquisition and control module 5 includes an FPGA controller and a synchronization trigger, used for synchronously controlling the image acquisition and operating parameter acquisition of the high frame rate industrial camera 3. The FPGA controller uses a Xilinx Zynq UltraScale+ series chip, integrating an ARM Cortex-A53 processor and FPGA logic units. The synchronization trigger generates TTL level pulse signals, simultaneously triggering the exposure of the high frame rate industrial camera 3, the strobe of the LED light source, and the sampling of the temperature sensor 7 and pressure sensor 8, ensuring that all data are strictly aligned on the time axis, with a time synchronization accuracy better than 1ms.

[0033] The image processing and analysis module 6 is used to perform multi-scale spatiotemporal feature extraction, deformation field reconstruction based on physical constraints, multi-modal failure mode recognition, and dynamic behavior prediction on the acquired dynamic image sequences, outputting quantitative performance parameters of the sealant. In this embodiment, the image processing and analysis module 6 is an embedded industrial control computer equipped with the NVIDIA Jetson AGX Orin computing platform and has built-in image processing algorithms. Its specific structure is as follows: Figure 2 As shown, it includes a multi-scale spatiotemporal feature extraction unit 61, a physical constraint deformation field reconstruction unit 62, a multi-modal failure mode recognition unit 63, a dynamic behavior prediction unit 64, and a data fusion and process window modeling unit 65.

[0034] Multi-scale spatiotemporal feature extraction unit: used to construct an image pyramid and use a three-dimensional convolutional neural network (3D-CNN) to extract multi-level features of the image sequence in spatial and temporal dimensions.

[0035] Specifically, for an input high frame rate image sequence I(x, y, t), this unit first constructs an L-layer Gaussian pyramid, with each layer having an image size half that of the previous layer, forming a total of L spatial scales. At each spatial scale, a 3D convolutional kernel with a time window length of T_k is used for feature extraction, where T_k is inversely proportional to the spatial scale, i.e.: T_k = T_0 × 2^(k-1), k = 1, 2, ..., L; where k = 1 corresponds to the highest spatial resolution (original size), T_1 = 5 frames (shortest time window), used to capture rapid transient deformations; k = L corresponds to the lowest spatial resolution, T_L = 40 frames (longest time window), used to capture slow flow trends.

[0036] The 3D convolutional kernel has a size of 3×3×3 and a stride of 1×1×1. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. After three 3D convolutional layers, the feature maps at each scale are restored to their original resolution through upsampling, and a feature pyramid fusion strategy is used to weight and fuse the features at each scale, outputting a multi-scale spatiotemporal feature map F_multiscale(x, y, t).

[0037] Physically Constrained Deformation Field Reconstruction Unit: This unit is used to embed the constitutive model of silicone sealant as a physical constraint into the optical flow calculation framework to reconstruct the deformation field and stress field that conform to the physical properties of the material.

[0038] This unit first establishes a constitutive model for silicone sealant. For silicone sealant, the Mooney-Rivlin hyperelastic constitutive model is used to describe its mechanical behavior under large deformation. The strain energy density function is: W = C_10(I_1-3) + C_01(I_2-3) + (1 / d)(J-1)^2; where I_1 and I_2 are the first and second invariants of the Cauchy-Green deformation tensor, J is the volume change ratio, and C_10, C_01, and d are material constants, which were calibrated through previous uniaxial tensile tests.

[0039] This unit embeds the aforementioned constitutive model into the variational optical flow energy functional in the form of partial differential equations. The traditional variational optical flow method minimizes the following energy functional: E(u, v) = ∫∫[(I_xu + I_yv + I_t)^2 + α(||∇u||^2 + ||∇v||^2)] dxdy; Based on this, the present invention adds a physical constraint term E_phys(u, v) to form the total energy functional: E_total(u, v) = E_data(u, v) + αE_smooth(u, v) + βE_phys (u, v); where the physical constraint term E_phys(u, v) is defined as: E_phys(u, v) = ∫∫||F(u, v) - σ(ε(u, v))||^2dxdy; where ε(u, v) is the strain tensor calculated from the displacement fields u, v, σ(ε) is the nominal stress tensor calculated from the Mooney-Rivlin constitutive model, and F(u, v) is the stress divergence calculated from the displacement fields through finite differences.

[0040] The energy functional is solved using a variational method, and then iteratively optimized using gradient descent to obtain a displacement field that simultaneously satisfies the constraints of constant image brightness and physical constitutive properties. Based on this, the Green strain tensor E = 1 / 2(∇u + ∇u^T + ∇u^T∇u) and the second Piola-Kirchhoff stress tensor S = ∂W / ∂E are further calculated, outputting the strain and stress tensors of the sealant region.

[0041] Multimodal failure mode identification unit: used to classify and locate the cracking, flow, and debonding failure modes of sealant based on spatiotemporal graph convolutional networks.

[0042] This unit first treats each pixel in the image sequence as a graph node to construct a spatiotemporal graph G=(V, E_s, E_t), where V is the node set (all pixels), E_s is the spatial edge set (the spatial adjacency relationship between adjacent pixels), and E_t is the temporal edge set (the temporal correspondence of the same pixel between adjacent frames).

[0043] The node feature vector v_i is extracted from the multi-scale spatiotemporal feature map F_multiscale(x_i, y_i, t), which has 128 dimensions. The spatial edge weight w_s(i, j) is defined as the inverse of the Euclidean distance between pixels i and j multiplied by the feature similarity, and the temporal edge weight w_t(i, j) is defined as the optical flow consistency at the same pixel position in consecutive frames.

[0044] This unit employs a spatiotemporal graph convolutional network (ST-GCN) for failure mode identification. The graph convolution operation is defined as: h_i^(l+1) = σ(∑_{j∈N(i)}(1 / √(d_id_j))W^(l)h_j^(l)); where N(i) is the spatiotemporal neighborhood of node i (including spatially adjacent nodes and time-corresponding nodes), d_i is the degree of node i, W^(l) is the learnable weight matrix of the l-th layer, and σ is the ReLU activation function.

[0045] The network structure consists of three spatiotemporal graph convolutional layers, each followed by a spatiotemporal pooling layer. Finally, a fully connected layer outputs the failure mode probability distribution (rupture, flow, deadhesion, normal) for each node, and the cross-entropy loss function is used for training.

[0046] Dynamic Behavior Prediction Unit: Used to predict the failure time, failure location, and failure propagation trajectory of sealant based on long short-term memory network and physical information neural network.

[0047] This unit employs a Physical Information Neural Network (PINN) with an encoder-decoder structure. The encoder part uses a Convolutional Long Short-Term Memory (ConvLSTM) network, and its gating structure is defined as follows: i_t=σ(W_xi*X_t+W_hi*H_{t-1}+b_i); f_t=σ(W_xf*X_t+W_hf*H_{t-1}+b_f); o_t=σ(W_xo*X_t+W_ho*H_{t-1}+b_o); g_t=tanh(W_xg*X_t+W_hg*H_{t-1}+b_g); C_t=f_t⊙C_{t-1}+i_t⊙g_t; H_t = o_t⊙tanh(C_t; Where X_t is the multi-scale spatiotemporal feature map input at time t, H_t is the hidden state, C_t is the cell state, * represents convolution operation, and ⊙ represents Hadamard product.

[0048] The encoder encodes historical image sequences into spatiotemporal feature sequences using ConvLSTM, and the decoder uses deconvolution layers to gradually restore spatial resolution, outputting the predicted deformation field and failure probability map for future moments.

[0049] The decoder embeds a physical constraint loss term, L_physics, which uses the Maxwell viscoelastic constitutive equation of the silicone sealant as a partial differential equation constraint: ∂σ / ∂t + (σ / η) = E∂ε / ∂t; where E is the elastic modulus and η is the viscosity coefficient. This constitutive equation is discretized in finite difference form and added as a soft constraint to the loss function: L_physics = ∑_{t}||σ_pred(t) - σ_phys(t)||^2; the total loss function is the weighted sum of the prediction error and the physical constraint: L_total = L_mse + λL_physics; the network is trained through backpropagation to predict the deformation field and failure region of the sealant at future moments, and outputs the failure time, failure location, and failure propagation trajectory.

[0050] Data fusion and process window modeling unit: used to fuse image analysis results with temperature and pressure data, establish a correlation model between sealant performance and operating parameters, and output a safe operating condition window.

[0051] This unit first aligns the parameters output by the image analysis unit, such as sealant area, maximum strain, rupture time, rupture propagation rate, average flow rate, and failure mode, with the operating condition data collected by the temperature and pressure sensors according to the timestamps to construct a multi-dimensional time-series dataset.

[0052] This unit employs Gaussian process regression (GPR) to establish a correlation model between the sealant failure time t_fail and operating parameters (temperature T, pressure P): t_fail(T, P) ~ GP(m(T, P), k((T, P), (T', P'))); where the mean function m(T, P) is a quadratic polynomial, and the kernel function k is a combination of a squared exponential kernel and a white noise kernel. The hyperparameters are solved using Bayesian optimization to obtain the probability distribution of the failure time.

[0053] This unit further employs a multi-objective optimization algorithm, aiming to maximize failure time, using temperature and pressure as decision variables, to solve for the Pareto optimal frontier and output a safe operating condition window (i.e., the temperature-pressure combination region that satisfies the condition that the failure time is greater than the threshold). This window is output to the display terminal 10 in the form of a two-dimensional contour map.

[0054] The image processing and analysis method of this embodiment will be described in detail below with reference to the specific test process.

[0055] (I) Multi-scale spatiotemporal feature extraction: After the test began, the high frame rate industrial camera 3 continuously acquired images at a frame rate of 500 frames per second, for a total of 1200 frames, corresponding to a test duration of 2.4 seconds. The multi-scale spatiotemporal feature extraction unit 61 received the image sequence I (x, y, t), where x and y are spatial coordinates and t is the time index.

[0056] This unit first constructs a four-layer Gaussian pyramid (L=4). The first layer is the original image size of 2048×2048 pixels, the second layer is downsampled to 1024×1024 pixels, the third layer is downsampled to 512×512 pixels, and the fourth layer is downsampled to 256×256 pixels. At each spatial scale, a three-dimensional convolutional kernel is used to extract spatiotemporal features. The kernel size is 3×3×3, and the number of channels is 64.

[0057] The temporal window length T_k is inversely correlated with the spatial scale: T_1 = 5 frames (corresponding to 0.01 seconds), T_2 = 10 frames (0.02 seconds), T_3 = 20 frames (0.04 seconds), and T_4 = 40 frames (0.08 seconds). After each scale passes through three 3D convolutional layers, it is upsampled to the original resolution through bilinear interpolation. The feature maps of each scale are weighted and fused according to the weight w_k = 1 / 2^(k-1) to output a multi-scale spatiotemporal feature map F_multiscale(x, y, t) with dimensions of 2048×2048×128.

[0058] (II) Physically constrained deformation field reconstruction: The physical constraint deformation field reconstruction unit 62 receives multi-scale spatiotemporal feature maps and reconstructs the deformation field using the variational optical flow method. First, the Mooney-Rivlin constitutive model parameters of the sealant are calibrated through a previous uniaxial tensile test: C_10=0.45MPa, C_01=0.12MPa, d=0.02MPa^(-1).

[0059] The constitutive model is embedded into the optical flow energy functional, and the total energy functional is: E_total(u,v)=∫∫[(I_xu+I_yv+I_t)^2+α(||∇u||^2+||∇v||^2)+β||F(u,v)-σ(ε(u,v))||^2]dxdy; where α=10, β=100. The gradient descent method is used for iterative solution, with 500 iterations and a learning rate of 0.01.

[0060] After obtaining the displacement fields u(x, y) and v(x, y), the Green strain tensor E and the second Piola-Kirchhoff stress tensor S are calculated, and the strain tensor field and stress tensor field are output. Taking the strain field as an example, the output is a six-component (ε_xx, ε_yy, ε_xy, ε_zz, ε_xz, ε_yz) with a spatial resolution of 50 μm / pixel.

[0061] (III) Multimodal Failure Mode Identification: The multimodal failure mode identification unit 63 constructs a spatiotemporal graph G. Approximately 16,000 nodes are generated (calculated at 2048×2048 / 256) with each 16×16 pixel region as a superpixel node. Spatial edges connect adjacent superpixel nodes, and temporal edges connect corresponding superpixel nodes in consecutive frames.

[0062] The node feature vector is composed of the average pooling result of the multi-scale spatiotemporal feature map in the corresponding region, with a dimension of 128. A three-layer spatiotemporal graph convolutional network is adopted, with the first layer outputting a dimension of 64, the second layer outputting a dimension of 32, and the third layer outputting a dimension of 4 (corresponding to four modes: rupture, flow, deadhesion, and normal). Global average pooling is followed by graph convolution operations, and the failure mode probability distribution of each node is output through a fully connected layer.

[0063] The training dataset used a labeled dataset containing 2000 high-temperature and high-pressure test videos. The pixel-level failure mode annotations for each frame of each video were completed by materials experts. The cross-entropy loss function, Adam optimizer, and learning rate of 0.001 were employed. After 50 training epochs, the test accuracy reached 94.2%.

[0064] In the test of this embodiment, at the 8.3-second mark, the multimodal failure mode recognition unit 63 detected a cracking mode in the edge area of ​​the sealant with a confidence level of 0.92, and output the cracking location coordinates (x=12.3mm, y=8.7mm) and cracking length of 0.6mm.

[0065] (iv) Dynamic behavior prediction: The dynamic behavior prediction unit 64 employs a physical information neural network with an encoder-decoder structure. The encoder consists of three ConvLSTM layers, each with a 3×3 kernel size and 64, 128, and 256 channels, respectively. The input sequence length is 40 frames (corresponding to 0.08 seconds), and the output hidden state dimension is 256×32×32.

[0066] The decoder contains three deconvolutional layers that gradually restore the feature map to its original size of 2048×2048 and output the predicted deformation field and failure probability map for the next 20 frames (corresponding to 0.04 seconds).

[0067] The physical constraints are based on Maxwell's viscoelastic constitutive equation: ∂σ / ∂t + (σ / η) = E∂ε / ∂t, where the elastic modulus E = 1.2 MPa and the viscosity coefficient η = 0.8 MPa·s, calibrated through prior rheological experiments. The physical constraint term L_physics calculates the mean square error between the predicted stress field and the stress field calculated by the constitutive equation in finite difference form.

[0068] The total loss function is L_total = L_mse + 0.1L_physics. The Adam optimizer is used with a learning rate of 0.0005 and the training time is 100 epochs.

[0069] In this embodiment, based on the first 40 frames (0.08 seconds), the network predicts the deformation field and failure probability map for the next 20 frames (0.04 seconds). The prediction results show that the fracture length expands to 1.2 mm at 8.5 seconds, with an error of less than 0.1 mm compared to the actual observed value of 1.1 mm. The prediction accuracy meets engineering requirements.

[0070] (v) Data fusion and process window modeling: The data fusion and process window modeling unit 65 fuses the above analysis results with data from temperature sensor 7 and pressure sensor 8. The temperature sampling frequency is 100Hz, the pressure sampling frequency is 100Hz, and after alignment with timestamps, the data is interpolated to an image sampling frequency of 500Hz.

[0071] Construct a multidimensional time-series dataset containing the following variables: time (ms), temperature (°C), pressure (MPa), sealant area (mm²), maximum strain (%), rupture time (ms), rupture length (mm), rupture propagation rate (mm / s), average flow rate (mm / s), and failure mode coding.

[0072] A Gaussian process regression model was used to establish the correlation between failure time t_fail and temperature T and pressure P. A combination of a squared exponential kernel and a white noise kernel was used as the kernel function, and the hyperparameters were solved through maximum likelihood estimation. A grid search was performed within a temperature range of 150℃-300℃ and a pressure range of 5MPa-20MPa to calculate the predicted failure time.

[0073] Multi-objective optimization aims to maximize failure time and solves for the Pareto optimal front. A failure time threshold of 30 seconds is set, and the temperature-pressure combination region satisfying the conditions is output. The safe operating condition window is the region with temperature ≤220℃ and pressure ≤12MPa, where the failure time is greater than 30 seconds. The results are output to the display terminal 10 in the form of a two-dimensional contour map.

[0074] (vi) Multimodal fusion observation: This embodiment integrates both an ultrasonic vibration-assisted observation device and a terahertz time-domain spectrometer 9, forming a multi-modal fusion observation capability.

[0075] The ultrasonic vibration-assisted observation device applied vibration to the sample stage at a frequency of 40 kHz and an amplitude of 5 μm during the test, accelerating the evolution of the microstructure inside the sealant. Comparative tests showed that the sealant's rupture time was shortened from 8.3 seconds to 4.2 seconds after vibration was applied, improving the test efficiency by approximately 50%. Simultaneously, the vibration-induced micro-deformation characteristics were clearly captured by a high-frame-rate industrial camera.

[0076] The terahertz time-domain spectrometer 9 acquires terahertz time-domain waveforms every 5 seconds during testing. Frequency-domain spectra are obtained through fast Fourier transform, and the average refractive index and absorption coefficient of the sealant in the 0.5-2.0 THz frequency range are calculated. The terahertz spectral parameters are linearly correlated with the sealant's crosslinking density, allowing real-time monitoring of the sealant's chemical structure evolution during high-temperature and high-pressure processes. A sudden change in the absorption coefficient of 0.05-0.1 before cracking can serve as an early warning signal of failure.

[0077] Terahertz data and visible light image data are synchronized through the image acquisition and control module 5. After multimodal data fusion, the data is input into the image processing and analysis module 6 to further improve the accuracy and reliability of failure identification.

[0078] Example 2 This embodiment is basically the same as Embodiment 1, except that: the silicone sealant uses a viscoelastic constitutive model instead of a hyperelastic constitutive model, which is suitable for describing the viscoelastic behavior of the sealant at lower temperatures or shorter time scales.

[0079] The one-dimensional form of Maxwell's viscoelastic constitutive model is: σ + τdσ / dt = ηdε / dt, where τ is the relaxation time and η is the viscosity. This model is embedded in the optical flow energy functional as a partial differential equation and solved by finite difference discretization.

[0080] This embodiment uses the Maxwell model for physical constraint deformation field reconstruction, suitable for test conditions where the test temperature is below 180℃ or the test time is less than 5 seconds. Under these conditions, the Maxwell model has higher fitting accuracy than the Mooney-Rivlin model, and the deformation field reconstruction error is reduced by approximately 15%.

[0081] Example 3 This embodiment is basically the same as embodiment 1, except that the multimodal failure mode recognition unit 63 adopts a spatiotemporal attention mechanism to enhance the representation ability of key areas.

[0082] Specifically, a self-attention module is introduced after the spatiotemporal graph convolutional layer to calculate the attention weight for each node: a_ij = softmax((W_qh_i)^T(W_kh_j) / √d_k); where h_i and h_j are the node feature vectors, W_q and W_k are the learnable projection matrices, and d_k is the feature dimension. The attention weights are used to weighted aggregate neighborhood information, enhancing attention to key regions such as the rupture front and flow boundaries.

[0083] In the test of this embodiment, the introduction of the spatiotemporal attention mechanism improved the failure mode recognition accuracy from 94.2% to 96.8%, especially in the classification of broken boundary pixels, the accuracy was significantly improved.

[0084] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A visual observation system for high-temperature and high-pressure testing of silicone sealant, characterized in that, include: A high-pressure sealed cavity (1) is provided inside, which is used to place silicone sealant samples; A transparent observation window (2) is set on the top of the high-pressure sealed cavity (1) to provide an optical observation channel; A high frame rate industrial camera (3) is set above the transparent observation window (2) to acquire dynamic image sequences of sealant samples under high temperature and high pressure. A multispectral coaxial light source system (4) is used to provide multi-band, uniform, and non-reflective illumination for high frame rate industrial cameras (3); Image acquisition and control module (5) is used to synchronously control the image acquisition and working condition parameter acquisition of high frame rate industrial camera (3); The image processing and analysis module (6) is used to perform multi-scale spatiotemporal feature extraction, physical constraint deformation field reconstruction, multi-modal failure mode recognition and dynamic behavior prediction on the acquired dynamic image sequence, and output the quantitative performance parameters of the sealant.

2. The visual observation system for high-temperature and high-pressure testing of silicone sealant according to claim 1, characterized in that, The image processing and analysis module (6) includes: A multi-scale spatiotemporal feature extraction unit (61) is used to construct an image pyramid and use a three-dimensional convolutional neural network to extract multi-level features of the image sequence in spatial scale and temporal dimension; The physical constraint deformation field reconstruction unit (62) is used to embed the constitutive model of the silicone sealant as a physical constraint into the optical flow calculation framework to reconstruct the deformation field and stress field that conform to the physical properties of the material. A multimodal failure mode identification unit (63) is used to classify and locate the cracking, flow and debonding failure modes of sealant based on graph neural network; The dynamic behavior prediction unit (64) is used to predict the failure time, failure location and failure propagation trajectory of the sealant based on the long short-term memory network and the physical information neural network. The data fusion and process window modeling unit (65) is used to fuse image analysis results with temperature and pressure data, establish a correlation model between sealant performance and operating parameters, and output a safe operating window.

3. The visual observation system for high-temperature and high-pressure testing of silicone sealant according to claim 2, characterized in that, The multi-scale spatiotemporal feature extraction unit (61) adopts a feature pyramid network structure to extract image features at at least three spatial scales. The length of the time window corresponding to each scale is inversely proportional to the spatial resolution. The highest spatial resolution corresponds to the shortest time window to capture rapid transient deformation, and the lowest spatial resolution corresponds to the longest time window to capture slow flow trends.

4. The visual observation system for high-temperature and high-pressure testing of silicone sealant according to claim 2, characterized in that, The physical constraint deformation field reconstruction unit (62) embeds the Mooney-Rivlin hyperelastic constitutive model or Maxwell viscoelastic constitutive model of the silicone sealant into the optical flow energy functional in the form of partial differential equations. It solves the displacement field that simultaneously satisfies the image brightness constant constraint and the physical constitutive constraint by variational method, and outputs the strain tensor field and stress tensor field that conform to the physical properties of the material.

5. The visual observation system for high-temperature and high-pressure testing of silicone sealant according to claim 2, characterized in that, The multimodal failure mode recognition unit (63) adopts a spatiotemporal graph convolutional network, taking each pixel in the image sequence as a graph node, and the spatial adjacency relationship and temporal correspondence between pixels as graph edges. It aggregates spatiotemporal neighborhood information through graph convolution operations and outputs the failure mode label and corresponding confidence level of each pixel.

6. The visual observation system for high-temperature and high-pressure testing of silicone sealant according to claim 2, characterized in that, The dynamic behavior prediction unit (64) adopts a physical information neural network with an encoder-decoder structure. The encoder part uses a convolutional long short-term memory network to extract spatiotemporal features, and the decoder part embeds the viscoelastic constitutive equation of the silicone sealant as a physical constraint loss term. While predicting the deformation field and failure area of ​​the sealant in the future, it ensures that the prediction results conform to physical laws.

7. The visual observation system for high-temperature and high-pressure testing of silicone sealant according to claim 1, characterized in that, The multispectral coaxial light source system (4) includes a combination structure of at least three LED light source modules with different wavelengths, a wavelength adjustable filter, a ring light source and a coaxial optical adapter. The LED light source modules with different wavelengths can independently control the brightness and flicker timing to obtain multispectral images of the sealant in different spectral bands, so as to distinguish the sealant body, crack area, flow front and impurities.

8. The visual observation system for high-temperature and high-pressure testing of silicone sealant according to claim 1, characterized in that, It also includes an ultrasonic vibration-assisted observation device, which is set below the sample stage to apply high-frequency, low-amplitude vibrations to the sealant sample to accelerate the microstructure evolution process of the sealant under high temperature and high pressure, shorten the test time, and capture the vibration-induced micro-deformation characteristics through a high frame rate industrial camera (3).

9. The visual observation system for high-temperature and high-pressure testing of silicone sealant according to claim 1, characterized in that, It also includes a terahertz time-domain spectrometer (9), which emits terahertz waves to the sealant sample through a transparent observation window (2) and receives the reflected signals. It is used for non-destructive testing of the internal crosslinking density change and micro-defect evolution of the sealant during the high temperature and high pressure process, and forms a multi-modal fusion observation with the visible light image of the high frame rate industrial camera (3).

Images