Method for plasma bombardment energy distribution detection and life prediction of silicon carbide part

Abstract

The invention provides a method for plasma bombardment energy distribution detection and life prediction of a silicon carbide part, and belongs to the technical field of semiconductor processing based on deep learning. The method comprises the following steps: firstly, arranging a magnetic probe array to obtain local electromagnetic parameters, and obtaining plasma luminescence characteristics by using optical imaging and spectrum acquisition equipment; preprocessing the acquired data to obtain multi-source fusion features; designing an energy distribution reconstruction framework based on a generative adversarial network, wherein a generator is used for deducing a three-dimensional energy distribution field of plasma; and inputting the three-dimensional energy distribution field time sequence characteristics into a bombardment dynamic characteristic modeling time sequence depth network, establishing a nonlinear mapping relation between energy distribution and material aging, further obtaining life prediction parameters, and finally obtaining the remaining service life of the silicon carbide part. According to the method, a complex degradation mechanism can be reflected more accurately, and the precision and robustness of life prediction are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor processing technology based on deep learning, and particularly relates to a method for detecting the energy distribution of plasma bombardment and predicting the lifetime of silicon carbide components. Background Technology

[0002] Silicon carbide (SiC), as a third-generation wide-bandgap semiconductor material, has been widely used in power electronic devices, semiconductor manufacturing equipment, aerospace devices, and critical components in high-temperature extreme environments due to its high thermal conductivity, high breakdown electric field strength, low power loss, and good radiation resistance. Especially during processes such as plasma etching and deposition, SiC components are often exposed to strong ion currents and high-energy particle bombardment for extended periods, leading to complex damage phenomena on their surfaces, including defect accumulation, lattice dislocations, and compositional migration. These damage processes inevitably result in component performance degradation and shortened lifespan, thus affecting the stability and reliability of the entire system. However, the plasma bombardment process has the following significant characteristics and technical challenges: Highly transient: Plasma parameters change rapidly on the nanosecond to millisecond scale, and the energy distribution exhibits strong temporal fluctuations, making it difficult for traditional testing methods to capture the complete dynamic process.

[0003] Complex spatial distribution: Energy bombardment exhibits non-uniformity and anisotropy in three-dimensional space, and the global energy distribution cannot be reconstructed by relying on a single-point probe or a limited number of observation points.

[0004] Nonlinear interaction mechanism: The interaction process between plasma and silicon carbide surface involves multiple nonlinear mechanisms such as electromagnetic field coupling, particle collision and heat conduction, which are difficult to describe accurately by traditional analytical models.

[0005] High uncertainty in lifetime prediction: Most existing lifetime prediction methods rely on empirical formulas or data fitting based on accelerated experiments, lacking an effective characterization of the relationship between dynamic energy distribution and material aging under real service conditions, resulting in large deviations in prediction results. Summary of the Invention

[0006] To address the above problems, this invention proposes a method for detecting the energy distribution of plasma bombardment and predicting the lifetime of silicon carbide components, comprising the following steps: S1, deploy a magnetic probe array to obtain local electromagnetic parameters, and simultaneously use optical imaging and spectral acquisition equipment to obtain the luminescence characteristics of the plasma; S2 uses environmental processing parameters to preprocess the acquired data, mapping data from different sources to a unified coordinate frame and time reference, and extracting key feature parameters closely related to energy input to obtain multi-source fusion features, including electromagnetic and optical features under plasma bombardment. and aligned process condition sequence ; S3, utilizing Generative Adversarial Networks (GANs), based on and By combining physical constraints and data-driven methods, the three-dimensional energy deposition distribution inside the workpiece can be reconstructed. S4, a temporal deep network based on bombardment dynamic feature modeling, first constructs a learnable frame-level vector based on the three-dimensional energy deposition distribution, generating a low-dimensional feature vector for each frame. and with process condition sequence Concatenate to form a joint feature vector ; Combine frame-level feature sequences The time-level hidden states are obtained by inputting a bidirectional LSTM, and event detection and temporal attention are built on top of them to obtain a global temporal representation vector. Train a regression mapping Energy time series characteristics Converted into material aging index prediction results ; S5, in the obtained stigma , , , Based on this, a life degradation model is established to estimate the remaining service life of silicon carbide components.

[0007] Preferably, the data acquired by S1 includes the magnetic field vector obtained from the discrete probe. Magnetic probe output voltage The original images recorded by the camera are Raw spectral data recorded by the spectrometer Process condition vector ; in, Here are the pixel coordinates, and t is the absolute timestamp. For wavelength, index Indicates the first Spectral lines; Process condition vector This includes radio frequency power, bias voltage, gas flow rate, cavity pressure, and applied magnetic field strength.

[0008] Preferably, the environmental processing parameters described in S2 include the response function of each probe. noise spectral density Camera intrinsic matrix K, camera extrinsic matrix R, probe coordinate set Save the set of distortion parameters The mapping matrix from the probe coordinate system to the workpiece coordinate system is denoted as M, and the system clock jitter is... The local current density obtained by the magnetic probe through electromagnetic inversion calculation ; The specific preprocessing steps include: First, the magnetic probe output voltage Low-frequency drift and high-frequency noise are removed using a bandpass filter; in the frequency domain, for each Perform probe response compensation and suppress noise spectrum In the frequency domain representation, the Fourier transform of the original voltage is: Using a known single probe, the number of coil turns, the effective coil area, and the sampling time step can be obtained. The voltage integral is normalized to restore the instantaneous magnetic field. Then the time derivative of the magnetic field was calculated. This quantity is used as a dynamic proxy for energy input; after obtaining the magnetic field distribution, it is combined with the obtained local current density estimated from the magnetic field. Kernel matrix With regularization parameters The current density estimate is obtained by solving a linear inverse problem with regularization. ; For optical images, the original image of the optical channel is first... Pixel dark current background obtained after dark field correction The mean value is obtained by calculating the set of flat field images. The corrected brightness distribution is obtained based on the pixel correction formula. ; Subsequently, the pixel coordinates were determined using the camera intrinsic and extrinsic parameter matrices K and R. Based on the coordinate transformation performed by K, the image is projected and back-projected onto the three-dimensional coordinates (X1, Y1, Z1) of the workpiece surface. Then, three-dimensional interpolation is used to map the image brightness onto the same workpiece surface grid as the magnetic field probe, obtaining the luminous intensity distribution in physical space. ; Raw spectral data After calibration and background subtraction at the wavelength, several spectral lines suitable for diagnosis are selected, and the electron temperature is estimated according to the Boltzmann relation. All data is jittered by the clock. Corrections were made to ensure that optical and electromagnetic data were synchronized on the same time reference. Set the sliding window length in the time dimension as... Select the number of frames within the time window Subsequent spatial resampling and feature construction are performed with a step size of [missing value]. The feature tensor is uniformly resampled to a 128*128 grid in the spatial dimension, and the final multi-source fusion feature tensor is: Where T=100 represents the number of time frames, H=W=128 represents the spatial resolution, and the number of channels C=6, and: ,in, From the current density amplitude With the rate of change of magnetic field The product of and the corrected luminous intensity distribution It is composed of weighted combinations.

[0009] Preferably, the target output of the generative adversarial network (GAN) during the training phase is a three-dimensional voxel energy deposition distribution. And simultaneously predict the uncertainty estimate for each voxel. Based on this, the surface energy flux sequence is obtained by numerically integrating the three-dimensional deposition distribution along the depth direction. ; The three-dimensional voxel energy deposition distribution was obtained by acquiring time-series images of the electric arc through high-speed photography, recording the arc radiation intensity using a spectral acquisition device, and measuring the temperature rise distribution on the material surface using an infrared thermal imager. Based on a heat conduction inversion model, the three-dimensional voxel energy distribution within a limited spatial range was thus obtained. .

[0010] Preferably, in the Generative Adversarial Network (GAN), the core task of the generator is to map spatiotemporal features and process conditions into a three-dimensional energy field, while also considering randomness. The mapping function parameterized by the generator The output includes the predicted three-dimensional voxel energy deposition distribution. and the corresponding uncertainty estimate To obtain a compact spatiotemporal feature representation, the input tensor is first encoded: ,in, It is a spatiotemporal convolutional encoder, output This is a spatiotemporally compressed feature tensor; during the generation process, process conditions play a constraining and modulating role, and through the FiLM mechanism, the process conditions are transformed into scaling. and offset parameters And used for modulation features, for the first Layer feature map Execution and scaling parameters Multiply element-wise by channel and then with the offset parameter. Summation yields the result; To extend two-dimensional surface features to three-dimensional voxel space, a learnable depth extension kernel is introduced. This depth extension kernel specifically compresses the feature tensor in a spatiotemporal manner. Weight matrix in the depth direction The numerical integration is used to learn how surface features propagate into the material interior; finally, the non-negative energy distribution is obtained by the decoder network output. With log variance .

[0011] Preferably, in the Generative Adversarial Network (GAN), the core of the decoder is a module that combines multi-layer transposed convolutions with upsampling convolutions; and A three-layer structure is adopted: the first layer is a 3*3*3 transposed convolution, which gradually restores the low-dimensional latent features to voxel space and performs nonlinear mapping; the second layer is an upsampling convolution, which further improves the spatial resolution; and the third layer is a 1*1*1 convolutional layer, which is used to output the final voxel-level prediction result. The softplus activation function ensures that the prediction energy is non-negative. Output predicted values ​​for energy deposition distribution. Output the log-variance of the uncertainty of the corresponding voxel; The generator outputs a three-dimensional voxel energy deposition distribution. And simultaneously predict the variance of each voxel. It is used to characterize uncertainty.

[0012] Preferably, in step S4, a learnable frame-level vector is constructed based on the three-dimensional energy deposition distribution, and a low-dimensional feature vector is generated for each frame. It is from the obtained three-dimensional energy deposition distribution Surface energy flux sequence Descriptive parameters such as peak value and pulse width, pulse energy, short / long window accumulation, spatial centroid and migration velocity, and frequency domain energy are extracted. Then, through spatial summarization and dimensionality reduction, pixel-level and voxel-level parameters are merged into a low-dimensional feature vector for each frame, denoted as [missing information]. .

[0013] Preferably, in step S4, the frame-level feature sequence is combined. The time-level hidden states are obtained by inputting a bidirectional LSTM, and event detection and temporal attention are built on top of them to obtain a global temporal representation vector. The specific process is as follows: Joint frame-level feature sequences The data is fed into a bidirectional LSTM to calculate the forward and backward hidden states respectively. and And spliced ​​on the feature dimension, in the forward propagation direction, the first Hidden state of a frame It is computed by the feedforward LSTM module, and its input includes the feature representation of the current frame. And the hidden state of the previous moment. In the direction of backward propagation, the first Hidden state of a frame Calculated by the inverse LSTM module And the hidden state in the next moment. Finally, the first Full hidden state of the frame From forward hidden state With reverse hidden state It is pieced together; To annotate important bombardment events at the frame level, the hidden state is processed through an MLP layer and a sigmoid function to output event probabilities, with a fixed threshold applied. Determine if it is an event frame; Employing a dot product attention mechanism in the time dimension, for the ... Hidden state of a frame First, with the global context vector Calculate the dot product and use the weight matrix. Mapping, obtaining attention scores, and then applying them to all... The frame scores are exponentially normalized to obtain the attention weights. : Hidden state of each frame According to the corresponding attention weight Weighted summation yields the global time representation vector. .

[0014] Preferably, in S4, a regression mapping is trained. Energy time series characteristics Converted into material aging index prediction results This mapping incorporates energy temporal characteristics The results are converted into material aging index predictions, including surface roughness increment, cracking density increment, and electrical conductivity degradation rate. The mapping relationship from time series representation to aging metrics, representing the global time representation. Input a multilayer perceptron and output a multidimensional aging index prediction vector. ; in, It is a two-layer fully connected network with a hidden dimension of 512, an activation function of ReLU, and an output dimension that is consistent with the aging metric dimension.

[0015] Preferably, in the lifespan degradation model of S5, the cumulative degradation amount is first defined. Describe the degradation of material properties over time. Based on the defined degradation amount Calculate remaining lifespan: when the cumulative degradation amount Reaching the set failure threshold When this happens, the component is deemed to have failed; at this point, the remaining lifespan... Through degradation With average degradation rate Ratio estimation; This indicates the remaining lifespan of the silicon carbide component in frame k.

[0016] Compared with the prior art, the innovative points of this invention include: (1) The energy field reconstruction method driven by multi-source sensing and GAN breaks through the technical bottleneck that makes it difficult to directly measure spatial distribution, and can obtain high-resolution three-dimensional energy distribution in a non-invasive manner.

[0017] (2) The introduction of temporal deep networks enables the complete capture of transient dynamic features, effectively making up for the shortcomings of traditional static models in extracting peak and fluctuation features.

[0018] (3) Nonlinear modeling of the relationship between energy distribution and material aging not only improves the accuracy of lifetime prediction, but also enables the characterization of the aging process from both mechanism and data perspectives.

[0019] The beneficial effects brought about by the innovation of this invention include: (1) Detection accuracy is greatly improved. Through multi-source data fusion and GAN-driven distributed reconstruction, a high-resolution three-dimensional energy field can be obtained, which makes up for the lack of information in traditional single-point detection.

[0020] (2) Enhanced prediction reliability. By combining dynamic modeling and aging relationship mapping of time-series deep networks, this method can maintain a low prediction error in real service environments, which is significantly better than existing methods.

[0021] (3) Supports real-time online applications. The system can continuously collect probe and optical signals during actual operation and output energy distribution and lifetime prediction results in real time, which has strong engineering applicability.

[0022] (4) Reduce maintenance costs and risks. By predicting the remaining useful life (RUL) of silicon carbide components in advance, planned maintenance can be achieved, avoiding economic losses and safety risks to the system caused by sudden failures.

[0023] The method of this invention is real-time and scalable, enabling online lifetime prediction during the actual operation of silicon carbide components. It can also be applied to the health assessment and lifetime prediction of other plasma-resistant materials such as gallium nitride, quartz, and metal electrodes. By obtaining information on the remaining service life of components in advance, this invention effectively supports planned maintenance, reduces operating costs, and avoids safety hazards and economic losses caused by sudden failures. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the overall technical route of the present invention.

[0025] Figure 2 This is a diagram of the adversarial network structure for plasma bombardment energy distribution reconstruction in this invention.

[0026] Figure 3The temporal deep network structure diagram for modeling the dynamic features of bombardment in this invention is shown.

[0027] Figure 4 This invention provides global peak time series and pulse detection, as well as energy centroid trajectory and cumulative energy map.

[0028] Figure 5 This is a lifespan prediction and health assessment diagram for the present invention. Detailed Implementation

[0029] This invention proposes a method for detecting the energy distribution of plasma bombardment and predicting the lifetime of silicon carbide components. The overall process is as follows: Figure 1 As shown: Plasma parameter acquisition and data processing: First, a magnetic probe array is deployed to acquire local electromagnetic parameters, while optical imaging and spectral acquisition equipment are used to obtain the plasma's luminescence characteristics. Synchronous acquisition of the probe and optical channel is achieved through hardware triggering and a unified clock, and the data is uniformly labeled in conjunction with process condition data. Next, after obtaining the raw data, the magnetic probe signal is filtered, denoised, and normalized, and the optical image undergoes background subtraction, flat-field correction, and geometric correction to improve signal reliability and consistency.

[0030] 3D Energy Distribution Reconstruction Based on Generative Adversarial Networks: High-resolution reconstruction of plasma energy distribution is achieved using generative adversarial networks. The generator uses multi-source fused feature data as input to infer the energy distribution field in 3D space, while the discriminator continuously improves the realism and accuracy of the reconstruction results by comparing them with experimental or simulation data. Through this process, a complete energy distribution map can be obtained without direct contact with silicon carbide components, achieving a 3D spatial resolution capability that is difficult to achieve with traditional methods.

[0031] Temporal Modeling and Feature Correlation of Bombardment Dynamics: A temporal deep network is introduced to model the evolution of energy distribution over time. Using temporal data of energy distribution as input, the model can identify peak values, pulse widths, energy accumulation effects, and spatial migration trends during bombardment, thus comprehensively capturing the dynamic characteristics of plasma bombardment. Based on the dynamic feature modeling, the energy distribution is correlated with the aging characteristics of silicon carbide materials. Through experimental calibration and offline testing, the degradation patterns of the material under different energy conditions are obtained. A nonlinear mapping relationship between energy distribution characteristics and aging indices is established using a deep learning model, enabling the model to dynamically predict the degree of material performance degradation based on actual energy input.

[0032] Lifetime prediction and online health assessment: When new plasma bombardment data is input, complete energy characteristics are obtained through energy distribution reconstruction and temporal dynamic modeling. These characteristics are then input into the lifetime prediction model to output the remaining lifespan of the silicon carbide component. The prediction results can be fed back to the system in real time, enabling online health assessment and early warning. When the predicted lifespan value approaches the failure threshold, the system can trigger maintenance recommendations, thereby achieving planned maintenance and avoiding economic losses and safety risks caused by sudden component failure.

[0033] The invention will be further described below with reference to specific embodiments.

[0034] I. Plasma Parameter Acquisition and Data Processing Because plasma bombardment processes are characterized by strong transients, complex spatial distribution, and nonlinear coupling, relying solely on a single measurement method often fails to fully reveal their physical laws. Therefore, this invention employs a multi-source fusion strategy combining magnetic probe arrays and optical imaging / spectral diagnostics to simultaneously acquire the electromagnetic properties and optical luminescence characteristics near the surface of silicon carbide components. System calibration and time synchronization methods ensure the consistency and comparability of various data.

[0035] After obtaining the original signal, the present invention further performs denoising, filtering, normalization and spatiotemporal registration, etc., to map data from different sources to a unified coordinate frame and time reference, and extract key feature parameters closely related to energy input.

[0036] 1. Plasma parameter acquisition and system calibration: In a plasma bombardment environment for silicon carbide components, system calibration is required first to ensure that the output of the multi-source sensors can be accurately mapped to the workpiece surface coordinate system. Inside or outside the working chamber of the silicon carbide component, a magnetic probe array and an optical imaging / spectral acquisition system are first deployed according to the geometric dimensions of the surface area to be measured and the desired spatial resolution.

[0037] A magnetic probe array is deployed in a grid pattern around the target bombarded surface of the silicon carbide component. The magnetic probe array is used to collect local magnetic field strength. (No. (Local magnetic field data corresponding to each probe), with a probe spacing of 5 mm to match the desired spatial resolution of the target surface. Each probe has 50 coil turns; the sampling frequency of the magnetic probe is fixed at [frequency value missing]. The sampling time step resolution is It can capture rapid changes in plasma magnetic fields on the nanosecond scale.

[0038] The probe body consists of an induction coil and an insulating support structure. The transient magnetic field changes in the plasma induce an electromotive force (EMF) in the probe coil. This induced EMF is converted into a voltage signal by a preamplifier circuit, thus obtaining the magnetic probe output voltage. Magnetic probe output voltage With the rate of change of magnetic field Proportional, and then the local current density is calculated through electromagnetic inversion. .

[0039] Specifically, the optical component consists of a high-speed camera and a spectrometer. The high-speed camera and fiber-coupled spectrometer are positioned along an unobstructed line of sight, with the high-speed camera's frame rate fixed at [value missing]. The camera resolution is fixed at 1024*1024 pixels, and the camera exposure time is fixed at... The original images recorded by the camera are Where index k represents the k-th camera, Here, is the pixel coordinate, and t is the absolute timestamp. All acquisition devices are synchronized using a unified clock, and system clock jitter is negligible. The upper limit is It can capture the spatial distribution of plasma luminescence intensity on a millisecond timescale.

[0040] The spectrometer is used to record emission lines, and the electron temperature is calculated using the Boltzmann plot method. This reflects the local plasma energy state. The raw spectral data recorded by the spectrometer is represented as follows: ,in For wavelength, index Indicates the first Spectral lines. All devices are triggered by a unified clock, with clock jitter less than [missing information]. This ensures precise temporal alignment of different signals. The mapping matrix from the probe coordinate system to the workpiece coordinate system is denoted as M, and the physical position of the probe is determined by the probe coordinate set. Simultaneously, process condition vectors are recorded during the data acquisition process. The process conditions include: radio frequency power, bias voltage, gas flow rate, cavity pressure, and applied magnetic field strength.

[0041] Furthermore, in the hardware calibration stage, to ensure the authenticity of the high-frequency signal, the frequency response function of each magnetic probe is... Experiments have shown that this method is used for compensation in post-processing. This represents the angular frequency. This function describes the gain and phase characteristics of the probe at different frequencies and is used to compensate for the original voltage to recover the true signal. System noise is expressed as noise spectral density. This indicates that the spectrum is used to suppress noise amplification during filtering and deconvolution. The camera obtains the intrinsic parameter matrix K from 15 checkerboard calibration maps, which describes the relationship between pixel coordinates and the imaging plane, the extrinsic parameter matrix R, and stores the set of distortion parameters. This enables the mapping between image pixels and the coordinates of the workpiece's three-dimensional surface.

[0042] The set of magnetic field vectors obtained by the discrete probe is denoted as And estimate the local current density from the magnetic field. Among them, the kernel matrix The regularization parameter was obtained using electromagnetic simulation during the calibration simulation phase and was set to [value missing]. , This is an identity matrix. After calibration, the following data and metadata are written to the system database: the original acquisition data after time synchronization, and the response function of each probe. Camera intrinsic and extrinsic parameters K, R, t; mapping from probe coordinate system to workpiece coordinate system; system clock jitter measurement and noise spectrum estimation.

[0043] Summarize the above data to obtain the original input data. ; In addition, environmental treatment parameters Used for raw input data Perform preprocessing.

[0044] 2. Data Preprocessing and Feature Construction During the data preprocessing stage. Based on Parameters in Process in the following order.

[0045] First, the raw signal of the magnetic probe. Low-frequency drift and high-frequency noise are removed using a bandpass filter. For each probe in the frequency domain... Perform probe response compensation and suppress noise spectrum In the frequency domain representation, the Fourier transform of the original voltage is: This invention employs a stabilized Wiener-style compensation to obtain the compensated voltage. That is: the response function of the probe The complex conjugate of the response function and the square of the absolute value of the response function and the noise spectrum and seek to negotiate and with The product is obtained by quadrature. The number of coil turns (50) and the effective coil area are then determined using a known single probe. and sampling time step The voltage integral is normalized to restore the instantaneous magnetic field. Then the time derivative of the magnetic field was calculated. This quantity is used as a dynamic proxy for energy input. After obtaining the magnetic field distribution, the local current density is estimated by combining the obtained data and using the magnetic field. Kernel matrix With regularization parameters The current density estimate is obtained by solving a linear inverse problem with regularization. .

[0046] Furthermore, for optical images, the original image of the optical channel is first... Pixel dark current background obtained after dark field correction The mean value is obtained by calculating the set of flat field images. The corrected brightness distribution is obtained based on the pixel correction formula. It calculates the original image With pixel dark current background The difference, multiplied by the pixel mean Calculate the mean of the flat-field image set and the pixel dark current background. The difference is obtained by comparing the values.

[0047] Subsequently, the pixel coordinates were projected using the camera's intrinsic and extrinsic parameter projection matrices K and R. The coordinate transformation is performed based on the projection matrix K. The image is projected and back-projected onto the three-dimensional coordinates (X1, Y1, Z1) of the workpiece surface. Then, three-dimensional interpolation is used to map the image brightness onto the same workpiece surface grid as the magnetic field probe, thus obtaining the luminous intensity distribution in physical space. .

[0048] Spectral data After calibration and background subtraction at the wavelength, several spectral lines suitable for diagnosis are selected, and the electron temperature is estimated according to the Boltzmann relation. All data is jittered by the clock. Corrections were made to ensure that optical and electromagnetic data were synchronized on the same time reference.

[0049] Furthermore, after preprocessing, to ensure consistency in subsequent modeling, the sliding window length is set to [value] in the time dimension. Select the number of frames within the time window Subsequent spatial resampling and feature construction are performed with a step size of [missing value]. Spatially, the data is uniformly resampled to a 128*128 grid. The final constructed multi-source fusion feature tensor is... Where T=100 represents the number of time frames, H=W=128 represents the spatial resolution, and the number of channels C=6, and: ,in, From the current density amplitude With the rate of change of magnetic field The product of and the corrected luminous intensity distribution It is composed of weighted combinations.

[0050] After data preprocessing and feature construction, the original input data is obtained. Complete output after preprocessing .in, It focuses on characterizing the electromagnetic and optical features under plasma bombardment. This is the aligned sequence of process conditions, which provides external constraints for subsequent energy distribution reconstruction.

[0051] II. 3D Energy Distribution Reconstruction Based on Generative Adversarial Networks After completing the acquisition, preprocessing, and feature construction of multi-source data, the resulting multi-source fusion feature tensor is... With process condition sequence This information characterizes the external manifestations of the process. However, two-dimensional observations and surface proxy quantities alone are insufficient to directly reflect the energy deposition distribution within the material. Therefore, this invention employs generative adversarial networks, based on... and By combining physical constraints and data-driven methods, the three-dimensional voxel-level energy deposition distribution inside the workpiece was reconstructed.

[0052] 1. Input and Target The goal of this process is to... Reconstructing the energy deposition distribution of voxels ,in For time frame indexing; For spatial grid indexing; This is the index for the depth layer.

[0053] In this process, the model's input first comes from S1. It contains a total of 6 channels: magnetic field strength Magnetic field change rate Current density estimate obtained from inversion The corrected loop current Electronic temperature Flux estimation based on optical proxy These quantities collectively characterize the spatial distribution and dynamic features of the plasma. Secondly, there is the process condition sequence. This is used to introduce timing priors driven by external processes; finally, there is random noise that follows a normal distribution. This enhances the model's ability to express observational uncertainties.

[0054] The target output of the GAN model during the training phase is a three-dimensional voxel energy deposition distribution. And simultaneously predict the uncertainty estimate for each voxel. This energy distribution field cannot be easily and completely obtained through direct observation via experiments; it can only be obtained under limited conditions using small-scale calibration experiments. Specifically, under controlled conditions, a limited number of high-voltage breakdown discharge tests are conducted on a silicon carbide wafer. High-speed photography is used to acquire arc timing images, and a spectral acquisition device is used to record the arc radiation intensity. Simultaneously, an infrared thermal imager is used to measure the temperature rise distribution on the material surface. Based on a heat conduction inversion model, the three-dimensional voxel energy distribution within a limited spatial range is obtained. .

[0055] Therefore, its main function in this invention is as a supervisory signal during the training and verification process. The system only needs to input... The corresponding three-dimensional energy deposition distribution can then be inferred using the trained generative model. It also outputs the uncertainty estimate for each voxel. .

[0056] 2. Conditional Construction and Position Embedding In GAN models, the generator This invention utilizes two-dimensional temporal features to infer three-dimensional distributions and simultaneously outputs the prediction variance of each voxel to characterize uncertainty. To this end, the invention employs a strategy of "spatiotemporal encoding → depth extension → voxel decoding," and uses a FiLM (Feature-wise Linear Modulation) conditional modulation mechanism in multiple layers to control the generation process under varying operating conditions.

[0057] Specifically, the core task of the generator is to map spatiotemporal features and process conditions into a three-dimensional energy field, while taking into account randomness. Mapping functions parameterized by the generator. The output includes the predicted three-dimensional voxel energy deposition distribution. and the corresponding uncertainty estimate To obtain a compact spatiotemporal feature representation, the input tensor is first encoded: .in, It is a spatiotemporal convolutional encoder, output This is a spatiotemporally compressed feature tensor, facilitating subsequent unfolding into a three-dimensional representation. During the generation process, process conditions play a constraining and modulating role. Through the FiLM mechanism, these process conditions are transformed into scaling... and offset parameters And used for modulation features, for the first Layer feature map Execution and scaling parameters Multiply element-wise by channel and then with the offset parameter. The summation is then applied to the convolutional output, incorporating scaling and offset calculated from process parameters. Under varying power, voltage, or gas conditions, the network can automatically adjust its feature representation to generate an energy deposition distribution that conforms to the current operating conditions.

[0058] To extend two-dimensional surface features to three-dimensional voxel space, a learnable depth extension kernel is introduced. This depth extension kernel specifically compresses the feature tensor in a spatiotemporal manner. Weight matrix in the depth direction The numerical integration is used to learn how surface features propagate into the material interior; finally, the non-negative energy distribution is obtained by the decoder network output. With log variance .

[0059] The core of the decoder in this invention is a module combining multi-layer transposed convolution and upsampling convolution. Specifically, and A three-layer structure is employed: the first layer is a 3x3x3 transposed convolution, which progressively restores low-dimensional latent features to voxel space while performing nonlinear mapping. The second layer is an upsampling convolution, further improving spatial resolution. The third layer is a 1x1x1 convolutional layer, used to output the final voxel-level prediction result. Softplus is used as the activation function to ensure non-negative prediction energy. Output predicted values ​​for energy deposition distribution. Output the logarithmic variance of the uncertainty of the corresponding voxel.

[0060] Ultimately, the generator outputs a three-dimensional voxel energy deposition distribution. And simultaneously predict the variance of each voxel. It is used to characterize uncertainty.

[0061] 3. Discriminator Modeling The discriminator's role is to determine whether the input energy distribution is real or generated by the generator. Introducing adversarial training can improve the realism and detail quality of the generator's output. In this invention, the discriminator uses a Hinge loss form:

[0062] in, This represents the true energy field distribution. The first term guarantees that the true energy field is judged as true, and the second term penalizes the generated energy field for being judged as true. The generator's adversarial loss is: .

[0063] The PatchGAN structure is used to determine the authenticity of the input energy field in a local region. The output is the adversarial probability of the actual generation, i.e., the discriminator probability. It is used for adversarial training to update the parameters of the generator and discriminator.

[0064] Furthermore, to ensure that the generated results are not only realistic but also physically plausible, the loss function includes, in addition to the adversarial term, reconstruction constraints, energy conservation constraints, surrogate alignment, smoothing constraints, and uncertainty modeling. The overall optimization objective is:

[0065] in, To calculate the adversarial loss, a game process between the generator and the discriminator is used to make the generated 3D voxel energy deposition distribution statistically approximate the real distribution, thereby improving the realism of the generation. To reconstruct the loss, the Euclidean distance between the generated results and the finite sparse reference data is calculated to ensure that the model can be strictly aligned on the existing observation points, thereby enhancing the reliability of the prediction. To calculate the energy conservation loss, the difference between the total power corresponding to the predicted energy deposition distribution and the input power is compared to ensure that the generated result follows the law of energy conservation. This represents the total variation loss, which suppresses excessive gradient changes in the spatial and temporal dimensions, thereby enhancing the continuity and smoothness of the prediction results, reducing interference from non-physical noise, and enhancing spatiotemporal smoothness. This indicates the calculation of uncertainty loss, which involves joint modeling of the prediction results and corresponding variances, so that the output not only includes point estimates of the energy distribution, but also provides information on the uncertainty of the prediction. These are the weighting coefficients.

[0066] 4. Training and optimization details and output In this invention, the GAN model is trained using Adam as the optimizer and the generator learning rate is set to... Discriminator learning rate Batch size B=2, total number of rounds 200.

[0067] During the inference phase, a new time window is given and a noise sample is taken. : Then the surface energy flux was calculated. .

[0068] Finally, the reconstructed three-dimensional voxel energy deposition distribution is obtained after training the GAN network. Surface energy flux sequence and uncertainty estimation .in As a driving factor of surface loading intensity, The evolution of defects is driven by bulk energy deposition, and uncertainty is used for confidence weighting and risk boundary discrimination. The plasma bombardment energy distribution reconstruction of this process generates an adversarial network structure as shown in the diagram. Figure 2 As shown.

[0069] III. Modeling Dynamic Features of Bombardment and Constructing Temporal Deep Networks 1. Construct a temporal deep network for bombardment dynamic feature modeling, such as... Figure 3 As shown: First, a learnable frame-level vector is constructed using the three-dimensional energy deposition distribution, and a low-dimensional feature vector is generated for each frame. and with process condition sequence Concatenate to form a joint feature vector ; Combine frame-level feature sequences The time-level hidden states are obtained by inputting a bidirectional LSTM, and event detection and temporal attention are built on top of them to obtain a global temporal representation vector. Train a regression mapping Energy time series characteristics Converted into material aging index prediction results .

[0070] 1. Constructing learnable frame-level vectors from temporal energy fields From the obtained three-dimensional energy deposition distribution Surface energy flux sequence Descriptive parameters such as peak value and pulse width, pulse energy, short / long window accumulation, spatial centroid and migration velocity, and frequency domain energy are extracted. Then, through spatial summarization and dimensionality reduction, pixel-level and voxel-level parameters are merged into a low-dimensional feature vector for each frame, denoted as [missing information]. .

[0071] Specifically, to suppress distribution drift caused by various operating conditions, the normalized process condition sequence is... Incorporating frame-level features, the working condition vector is linearly transformed to... Similar scales are then spliced ​​together to form a joint input. , where A is a fixed diagonal scaling matrix. To obtain the peak criterion, we first base it on the previous... The frame estimates the mean and standard deviation of the background baseline for each pixel.

[0072] Specifically, in The predicted energy flux values ​​are taken from 10 consecutive frames at the location. Then, the values ​​are averaged, that is, the sum of the values ​​of these 10 frames is divided by 10 to obtain the mean of the background energy. Subsequently, the standard deviation at the same location is calculated. This is used to characterize the magnitude of fluctuations in background energy flux. Specifically, it involves taking the energy flux value of each frame... with the mean The squares of the differences are summed, then divided by 9, and finally the square root is taken. After obtaining the background statistics, a peak determination threshold is constructed for each pixel to identify strong bombardment events. Through standard deviation and mean The weighted sum is obtained.

[0073] Local maximum detection is performed on the time series of each pixel to determine the peak amplitude and occurrence time, and a list of peaks is recorded for subsequent pulse width and energy calculations. If a certain frame... If both sides are smaller, then it is denoted as and its time .

[0074] Secondly, for each peak, the full width at half maximum (FWHM) is used to quantize the pulse duration, specifically by finding the time index to reach the half-width around the current peak. , In seconds. Pulse duration. The difference between the time indices at half height on the left and right sides is calculated at the time step.

[0075] To measure the energy load of a single pulse, the pulse energy is obtained by discretely integrating the surface flux within the pulse interval: To describe the energy accumulation effect, a window length is defined. Frame, sliding integral of surface flux: To depict the migration trend of energy hotspots, surface flux is used as a weight to calculate the energy centroid: , And remove the centroid bit of the adjacent frame. Obtain migration speed .

[0076] Therefore, through the above processing, a low-dimensional feature vector can be generated for each frame. Its components include peak amplitude, pulse energy, cumulative energy, centroid position, and migration velocity. This is then compared with the process condition sequence obtained in step two. Concatenate to form a joint feature vector .

[0077] 2. Temporal depth coding The peak-pulse-accumulation-migration evolution law is determined. In this invention, the joint frame-level feature sequence is used. The time-level hidden states are obtained by inputting a bidirectional LSTM, and event detection and temporal attention are built on it to form a high-order representation of the bombardment process dynamics.

[0078] To obtain a temporal representation including context, the sequence is fed into a bidirectional LSTM, and the forward and backward hidden states are calculated separately. and And concatenate them along the feature dimension. In the forward propagation direction, the first... Hidden state of a frame It is computed by the feedforward LSTM module. Its input includes the feature representation of the current frame. And the hidden state of the previous moment. In the direction of backward propagation, the first Hidden state of a frame Calculated by the inverse LSTM module And the hidden state in the next moment. Finally, the first Full hidden state of the frame From forward hidden state With reverse hidden state It is pieced together, thus preserving the contextual information of the preceding and following timelines. The one-way hidden dimension is 256.

[0079] To annotate important bombardment events (strong peaks and migration mutations) at the frame level, the hidden state is processed through an MLP layer and a sigmoid function to output the event probabilities, with a fixed threshold applied. Determine whether it is an event frame.

[0080] Specifically, the hidden state is input into a multilayer perceptron (MLP) network, and the event score for each time frame is calculated through a linear transformation. Then, the score is mapped to an event probability value between 0 and 1 by a sigmoid activation function. The linear transformation corresponding to the formula includes the weight matrix. and bias vector These are used to adjust the contribution of each channel and the overall offset in the hidden state, respectively.

[0081] To enable the model to focus on key segments in the time series, this invention employs dot product attention to aggregate context in the time dimension, resulting in a weighted global temporal representation. This is used for subsequent association modeling. A dot product attention mechanism is employed in the time dimension. For the... Hidden state of a frame First, with the global context vector Calculate the dot product and use the weight matrix. Mapping, obtaining attention scores, and then applying them to all... The frame scores are exponentially normalized to obtain the attention weights. : Hidden state of each frame According to the corresponding attention weight Weighted summation yields the global time representation vector. .in, It is the average hidden state. This is the attention projection matrix.

[0082] 3. Characteristics-Aging Association Offline calibration data is used to train a nonlinear mapping. Based on the obtained temporal global representation... Pair the aging index labels obtained from offline experiments with the labels obtained from the calibration to train a regression mapping. This mapping incorporates energy temporal characteristics. Converted into predicted results of material aging indices (surface roughness increment, cracking density increment, electrical conductivity degradation rate).

[0083] The mapping relationship from time series representation to aging metrics, representing the global time representation. Input a multilayer perceptron and output a multidimensional aging index prediction vector. .

[0084] in, This is a two-layer fully connected network with 512 hidden dimensions and ReLU activation function. The output dimension matches the aging metric dimension. The above mapping is trained using experimental labels, with mean squared error as the main loss and L2 regularization added to improve generalization.

[0085] in, To determine the sample size, . For the first The true aging index of the experimental samples.

[0086] In this way, the model can establish the correspondence between the dynamic characteristics of the three-dimensional voxel energy deposition distribution and the material aging process, and predict the degradation trend of the material based on the temporal characteristics of energy input.

[0087] Finally, the time-series feature sequence Joint feature sequences Global representation of timing and aging index prediction results As output, it is used for subsequent lifetime prediction.

[0088] IV. Life Prediction and Online Health Assessment In output features , , , Based on this, a life degradation model is established to further estimate the remaining useful life (RUL) of silicon carbide components. Furthermore, this invention also designs an online health assessment and early warning mechanism, which can promptly trigger maintenance recommendations when the lifespan approaches the threshold, thereby ensuring operational safety and economy.

[0089] First, the degree of material degradation is quantified from dynamic characteristics. This is because damage caused by plasma bombardment is not instantaneous but rather a gradual, cumulative process. The cumulative degradation amount is defined. This describes the degradation of material properties over time. Its calculation form is:

[0090] in, This is the degradation rate function, used for comprehensive calculation at time [time]. The rate of degradation, Indicates the discrete frame number. Fixed sampling interval. Degradation amount. As time progresses and the values ​​accumulate, they directly reflect the overall degradation of the silicon carbide component in the k-th frame.

[0091] After defining the degradation amount Next, the remaining lifetime is calculated. This is based on the cumulative degradation. Reaching the set failure threshold At this point, the component is determined to have failed. At this time, the remaining service life... Through degradation With average degradation rate The ratio estimate. This indicates the remaining lifespan of the silicon carbide component in frame k.

[0092] To more intuitively assess the health status of components, a health index is introduced. This is used to normalize the current degree of material degradation.

[0093] Among them, when At that time, it indicates that the material is in a completely healthy state; as bombardment accumulates, Increase, It will gradually decrease, when When the value is close to 0, it indicates that the healthy state is about to disappear.

[0094] Furthermore, the system not only needs to assess health indicators but also needs to trigger maintenance recommendations when its lifespan approaches its limit. The initial design life is set as... The warning threshold is When the remaining lifespan satisfies: The system will trigger a maintenance warning signal, prompting maintenance personnel to replace or repair components within a specified time. This helps avoid economic losses and safety risks caused by sudden failures.

[0095] Through lifespan prediction and online health assessment, the degree of degradation of a component at any given time can be obtained. and component remaining life prediction Health index reflects the health status of components. A warning signal is triggered when the lifespan falls below a set threshold.

[0096] V. Experimental Analysis like Figure 4 As shown, this diagram illustrates the global peak time series and pulse detection, as well as the energy centroid trajectory and cumulative energy map. In the global peak time series and pulse detection, the blue curve represents the global energy peak evolving over time, the red dots mark the identified pulse peaks, and the vertical lines indicate the corresponding pulse occurrence times. In the energy centroid trajectory and cumulative energy distribution diagram, the background color plot shows the cumulative energy distribution over the entire observation period; brighter colors indicate greater cumulative energy. The white curve represents the evolution of the centroid trajectory, with green dots indicating the starting position and red dots indicating the ending position. The trajectory reveals the spatial migration pattern of energy hotspots.

[0097] like Figure 5 As shown, in Figure 5 The upper part of the graph illustrates the degradation process of the silicon carbide component. The horizontal axis represents time, and the vertical axis represents the degradation amount. The curve clearly reflects the continuous accumulation of degradation over time under plasma bombardment. When the degradation amount gradually approaches the preset failure threshold, i.e., at the "near failure" position, it indicates that the component is about to enter the end-of-life stage. This graph visually reveals the degradation evolution law and provides a basis for subsequent lifetime prediction. Figure 5 The lower half of the diagram shows the remaining life prediction curve. The curve gradually declines over time, reflecting the continuous reduction in the component's remaining usable life. When the remaining life decreases to 20% of the initial life, the system will trigger an early warning. This indicates that planned maintenance is required at this point to prevent the component from entering a sudden failure state.

[0098] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0099] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for detecting the energy distribution of plasma bombardment and predicting the lifetime of silicon carbide components, characterized in that, Includes the following processes: S1, deploy a magnetic probe array to obtain local electromagnetic parameters, and simultaneously use optical imaging and spectral acquisition equipment to obtain the luminescence characteristics of the plasma; S2 uses environmental processing parameters to preprocess the acquired data, mapping data from different sources to a unified coordinate frame and time reference, and extracting key feature parameters closely related to energy input to obtain multi-source fusion features, including electromagnetic and optical features under plasma bombardment. and aligned process condition sequence ; S3, utilizing Generative Adversarial Networks (GANs), based on and By combining physical constraints and data-driven methods, the three-dimensional energy deposition distribution inside the workpiece can be reconstructed. S4, a temporal deep network based on bombardment dynamic feature modeling, first constructs a learnable frame-level vector through the three-dimensional energy deposition distribution, and generates a low-dimensional feature vector for each frame. and with process condition sequence Concatenate to form a joint feature vector ; Combine frame-level feature sequences The time-level hidden states are obtained by inputting a bidirectional LSTM, and event detection and temporal attention are built on top of them to obtain a global temporal representation vector. ; Training a regression mapping Energy time series characteristics Converted into material aging index prediction results ; S5, in the obtained stigma , , , Based on this, a life degradation model is established to estimate the remaining service life of silicon carbide components.

2. The method for detecting plasma bombardment energy distribution and predicting lifetime of silicon carbide components as described in claim 1, characterized in that: The data collected by S1 includes the magnetic field vector obtained from the discrete probe. Magnetic probe output voltage The original images recorded by the camera are Raw spectral data recorded by the spectrometer Process condition vector ; in, Here, t represents the pixel coordinates, and t represents the absolute timestamp. For wavelength, index Indicates the first Spectral lines; Process condition vector This includes radio frequency power, bias voltage, gas flow rate, cavity pressure, and applied magnetic field strength.

3. The method for detecting the energy distribution and predicting the lifetime of a silicon carbide component by plasma bombardment as described in claim 2, characterized in that: The environmental processing parameters described in S2 include the response function of each probe. noise spectral density Camera intrinsic matrix K, camera extrinsic matrix R, probe coordinate set Save the set of distortion parameters The mapping matrix from the probe coordinate system to the workpiece coordinate system is denoted as M, and the system clock jitter is... The local current density obtained by the magnetic probe through electromagnetic inversion calculation ; The specific preprocessing steps include: First, the magnetic probe output voltage Low-frequency drift and high-frequency noise are removed using a bandpass filter; in the frequency domain, for each Perform probe response compensation and suppress noise spectrum In the frequency domain representation, the Fourier transform of the original voltage is: Using the known number of coil turns and effective coil area of ​​a single probe, as well as the sampling time step. The voltage integral is normalized to restore the instantaneous magnetic field. Then the time derivative of the magnetic field was calculated. This quantity is used as a dynamic proxy for energy input; after obtaining the magnetic field distribution, it is combined with the obtained local current density estimated from the magnetic field. Kernel matrix With regularization parameters The current density estimate is obtained by solving a linear inverse problem with regularization. ; For optical images, the original image of the optical channel is first... Pixel dark current background obtained after dark field correction The mean value is obtained by calculating the set of flat field images. The corrected brightness distribution is obtained based on the pixel correction formula. ; Subsequently, the pixel coordinates were determined using the camera intrinsic and extrinsic parameter matrices K and R. Based on the coordinate transformation performed by K, the image is projected and back-projected onto the three-dimensional coordinates (X1, Y1, Z1) of the workpiece surface. Then, three-dimensional interpolation is used to map the image brightness onto the same workpiece surface grid as the magnetic field probe, obtaining the luminous intensity distribution in physical space. ; Raw spectral data After calibration and background subtraction at the wavelength, several spectral lines suitable for diagnosis are selected, and the electron temperature is estimated according to the Boltzmann relation. All data is jittered by the clock. Corrections were made to ensure that optical and electromagnetic data were synchronized on the same time reference. Set the sliding window length to [value] in the time dimension. Select the number of frames within the time window Subsequent spatial resampling and feature construction are performed with a step size of [missing value]. The feature tensor is uniformly resampled to a 128*128 grid in the spatial dimension, and the final multi-source fusion feature tensor is: Where T=100 represents the number of time frames, H=W=128 represents the spatial resolution, and the number of channels C=6, and: ,in, From the current density amplitude With the rate of change of magnetic field The product of and the corrected luminous intensity distribution It is composed of weighted combinations.

4. The method for detecting plasma bombardment energy distribution and predicting lifetime of silicon carbide components as described in claim 1, characterized in that: The target output of the Generative Adversarial Network (GAN) during the training phase is a three-dimensional voxel energy deposition distribution. And simultaneously predict the uncertainty estimate for each voxel. Based on this, the surface energy flux sequence is obtained by numerically integrating the three-dimensional deposition distribution along the depth direction. ; The three-dimensional voxel energy deposition distribution was obtained by acquiring time-series images of the electric arc through high-speed photography, recording the arc radiation intensity using a spectral acquisition device, and measuring the temperature rise distribution on the material surface using an infrared thermal imager. Based on a heat conduction inversion model, the three-dimensional voxel energy distribution within a limited spatial range was thus obtained. .

5. The method for detecting the energy distribution of plasma bombardment and predicting the lifetime of a silicon carbide component as described in claim 4, characterized in that: In the Generative Adversarial Network (GAN), the core task of the generator is to map spatiotemporal features and process conditions into a three-dimensional energy field, while also considering randomness. The mapping function parameterized by the generator The output includes the predicted three-dimensional voxel energy deposition distribution. and the corresponding uncertainty estimate To obtain a compact spatiotemporal feature representation, the input tensor is first encoded: ,in, It is a spatiotemporal convolutional encoder, output This is a spatiotemporally compressed feature tensor; during the generation process, process conditions play a constraining and modulating role, and through the FiLM mechanism, the process conditions are transformed into scaling. and offset parameters And used for modulation features, for the first Layer feature map Execution and scaling parameters Multiply element-wise by channel and then with the offset parameter. Summation yields the result; To extend two-dimensional surface features to three-dimensional voxel space, a learnable depth extension kernel is introduced. This depth extension kernel specifically uses a spatiotemporally compressed feature tensor. Weight matrix in the depth direction The numerical integration is used to learn how surface features propagate into the material interior; finally, the non-negative energy distribution is obtained by the decoder network output. With log variance .

6. The method for detecting the energy distribution of plasma bombardment and predicting the lifetime of a silicon carbide component as described in claim 5, characterized in that: In the Generative Adversarial Network (GAN), the core of the decoder is a module that combines multi-layer transposed convolutions with upsampling convolutions. and A three-layer structure is adopted: the first layer is a 3*3*3 transposed convolution, which gradually restores the low-dimensional latent features to voxel space and performs nonlinear mapping; the second layer is an upsampling convolution, which further improves the spatial resolution; and the third layer is a 1*1*1 convolutional layer, which is used to output the final voxel-level prediction result. The softplus activation function ensures that the prediction energy is non-negative. Output predicted values ​​for energy deposition distribution. Output the log-variance of the uncertainty of the corresponding voxel; The generator outputs a three-dimensional voxel energy deposition distribution. And simultaneously predict the variance of each voxel. It is used to characterize uncertainty.

7. The method for detecting the energy distribution of plasma bombardment and predicting the lifetime of a silicon carbide component as described in claim 1, characterized in that: In step S4, a learnable frame-level vector is constructed based on the three-dimensional energy deposition distribution, and a low-dimensional feature vector is generated for each frame. It is from the obtained three-dimensional energy deposition distribution Surface energy flux sequence Descriptive parameters such as peak value and pulse width, pulse energy, short / long window accumulation, spatial centroid and migration velocity, and frequency domain energy are extracted. Then, through spatial summarization and dimensionality reduction, pixel-level and voxel-level parameters are merged into a low-dimensional feature vector for each frame, denoted as [missing information]. .

8. The method for detecting plasma bombardment energy distribution and predicting lifetime of silicon carbide components as described in claim 7, characterized in that: In S4, the combined frame-level feature sequence will be... The time-level hidden states are obtained by inputting a bidirectional LSTM, and event detection and temporal attention are built on top of them to obtain a global temporal representation vector. The specific process is as follows: Joint frame-level feature sequences The data is fed into a bidirectional LSTM to calculate the forward and backward hidden states respectively. and And spliced ​​on the feature dimension, in the forward propagation direction, the first Hidden state of a frame It is computed by the feedforward LSTM module, and its input includes the feature representation of the current frame. And the hidden state of the previous moment. In the direction of backward propagation, the first Hidden state of a frame Calculated by the inverse LSTM module And the hidden state in the next moment. Finally, the first Full hidden state of the frame From forward hidden state With reverse hidden state It is pieced together; To annotate important bombardment events at the frame level, the hidden state is processed through an MLP layer and a sigmoid function to output event probabilities, with a fixed threshold applied. Determine if it is an event frame; Employing a dot product attention mechanism in the time dimension, for the ... Hidden state of a frame First, with the global context vector Calculate the dot product and use the weight matrix. Mapping, obtaining attention scores, and then applying them to all... The frame scores are exponentially normalized to obtain the attention weights. : Hidden state of each frame According to the corresponding attention weight Weighted summation yields the global time representation vector. .

9. The method for detecting the energy distribution of plasma bombardment and predicting the lifetime of a silicon carbide component as described in claim 8, characterized in that: In S4, a regression mapping is trained. Energy time series characteristics Converted into material aging index prediction results This mapping incorporates energy temporal characteristics The results are converted into material aging index predictions, including surface roughness increment, cracking density increment, and electrical conductivity degradation rate. The mapping relationship from time series representation to aging metrics, representing the global time representation. Input a multilayer perceptron and output a multidimensional aging index prediction vector. ; in, It is a two-layer fully connected network with a hidden dimension of 512, an activation function of ReLU, and an output dimension that is consistent with the aging metric dimension.

10. The method for detecting the energy distribution of plasma bombardment and predicting the lifetime of a silicon carbide component as described in claim 1, characterized in that: The lifespan degradation model in S5 first defines the cumulative degradation amount. Describes the degradation of material properties over time. Based on the defined degradation amount Calculate remaining lifespan: when the cumulative degradation amount Reaching the set failure threshold When this happens, the component is deemed to have failed; at this point, the remaining lifespan... Through degradation With average degradation rate The ratio estimate; This indicates the remaining lifespan of the silicon carbide component in frame k.

Images