Method for silicon carbide component plasma bombardment energy distribution detection and lifetime prediction

Abstract

The application provides a method for detecting and predicting the service life of a silicon carbide component plasma bombardment energy distribution, and belongs to the technical field of semiconductor processing based on deep learning; first, a magnetic probe array is laid out to obtain local electromagnetic parameters, and optical imaging and spectrum acquisition equipment are used to obtain plasma light emission characteristics; the obtained data is preprocessed to obtain multi-source fusion features; an energy distribution reconstruction framework based on a generative adversarial network is designed, wherein a generator is used to infer a three-dimensional energy distribution field of the plasma; a three-dimensional energy distribution field time sequence feature is input into a bombardment dynamic feature modeling time sequence deep network to establish a nonlinear mapping relationship between the energy distribution and material aging, further obtain a service life prediction parameter, and finally obtain the remaining service life of the silicon carbide component. The application can more accurately reflect the complex degradation mechanism and significantly improve the accuracy and robustness of the service life prediction.

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:

[0003] 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.

[0004] 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.

[0005] 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.

[0006] 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

[0007] 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:

[0008] 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;

[0009] S2, pre-process the acquired data using environmental processing parameters, map data from different sources to a unified coordinate frame and time reference, and extract key feature parameters closely related to energy input, obtaining multi-source fusion features, including electromagnetic and optical features under plasma bombardment and the aligned process condition sequence ;

[0010] S3, using a generative adversarial network (GAN), based on and , combining physical constraints and data-driven methods, reconstructing the three-dimensional energy deposition distribution inside the workpiece;

[0011] S4, based on the dynamic feature modeling time series deep network, first construct a learnable frame-level vector based on the three-dimensional energy deposition distribution, generate a low-dimensional feature vector for each frame , and splice it with the process condition sequence to form a joint feature vector ; input the joint frame-level feature sequence into a bidirectional LSTM to obtain the time-level hidden state, and build event detection and time attention on it to obtain the global time representation vector ; train a regression mapping to convert the energy time series features into material aging index prediction results ;

[0012] S5, based on the obtained features , , , , establish a life degradation model to calculate the remaining useful life of the silicon carbide component.

[0013] Preferably, the data collected by S1 includes magnetic field vectors obtained by discrete probes , magnetic probe output voltages , raw images recorded by cameras , raw spectral data recorded by spectrometers , process condition vectors ;

[0014] wherein, is the pixel coordinate, t is the absolute timestamp, is the wavelength, and the index represents the th spectral line;

[0015] The process condition vector includes radio frequency power, bias voltage, gas flow, cavity pressure, and applied magnetic field strength.

[0016] 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 ;

[0017] The specific preprocessing steps include:

[0018] 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. ;

[0019] 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. ;

[0020] 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. ;

[0021] 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 via clock jitter correction, ensuring optical and electromagnetic data are synchronized under the same time reference;

[0022] Set the sliding window length in time dimension as , select the number of frames in the time window for subsequent spatial resampling and feature construction, the step size is , resample to 128*128 grid in spatial dimension, and the final multi-source fusion feature tensor is constructed as , where T=100 represents the number of time frames, H=W=128 represents the spatial resolution, the number of channels C=6, and: , where is the product of the current density amplitude and the magnetic field change rate and the corrected luminous intensity distribution .

[0023] Preferably, the goal output of the generative adversarial network GAN in the training stage is the three-dimensional voxel energy deposition distribution , and the uncertainty estimate of each voxel is also predicted , on this basis, the surface energy flux sequence is obtained by numerical integration of the three-dimensional deposition distribution along the depth direction.

[0024] Where the three-dimensional voxel energy deposition distribution is obtained by acquiring the time sequence image of the arc through high-speed photography, recording the arc radiation intensity by combining the spectrum acquisition device, and measuring the material surface temperature rise distribution by using the infrared thermal imager, and based on the heat conduction inversion model, thereby obtaining the three-dimensional voxel energy distribution in a limited space range .

[0025] Preferably, in the generative adversarial network GAN, the core task of the generator is to map the spatio-temporal features and process conditions into a three-dimensional energy field, while considering randomness , through the mapping function parameterized by the generator , the output includes the predicted three-dimensional voxel energy deposition distribution and the corresponding uncertainty estimate ; in order to obtain a compact spatio-temporal feature representation, first encode the input tensor: , where is the spatio-temporal convolutional encoder, and the output is the spatio-temporal compressed feature tensor; in the generation process, the process conditions play a role in constraint and modulation, through the FiLM mechanism, the process conditions are converted into scaling and offset parameters , and used to modulate the features, the first layer feature map performing element-wise multiplication with a scaling parameter performing element-wise multiplication with a scaling parameter summing up to obtain

[0026] To extend the two-dimensional surface features to three-dimensional voxel space, a learnable depth extension kernel is introduced, which is specifically implemented by the numerical integration of the spatiotemporal compressed feature tensor and the weight matrix in the depth direction to learn how the surface features propagate into the material interior; finally, the non-negative energy distribution is output through the decoder network and the logarithmic variance .

[0027] Preferably, in the generative adversarial network (GAN), the core of the decoder is composed of a module combining multi-layer transposed convolution and up-sampling convolution; and respectively adopt a three-layer structure: the first layer is a 3*3*3 three-dimensional transposed convolution, which gradually restores the low-dimensional latent features to the voxel space while performing nonlinear mapping, the second layer is an up-sampling convolution, which further improves the spatial resolution, and the third layer is a 1*1*1 convolution layer, which is used to output the final voxel-level prediction result; wherein softplus is an activation function to ensure the non-negativity of the predicted energy, the predicted value of the energy deposition distribution, the logarithmic variance of the uncertainty corresponding to the voxel;

[0028] The generator outputs a three-dimensional voxel energy deposition distribution and simultaneously predicts the variance of each voxel to depict the uncertainty.

[0029] Preferably, in the 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 is extracted from the obtained three-dimensional energy deposition distribution , the surface energy flux sequence peak value and pulse width, pulse energy, short / long window cumulative amount, spatial centroid and migration velocity, and frequency energy; then through spatial aggregation and dimensionality reduction, the pixel-level and voxel-level quantities are merged into a low-dimensional feature vector for each frame, denoted as .

[0030] Preferably, in the S4, the joint frame-level feature sequence is input into the bidirectional LSTM to obtain the time-level hidden state, and the event detection and timing attention are constructed on it to obtain the global time representation vector , the specific process is as follows:

[0031] 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;

[0032] 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;

[0033] 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. .

[0034] 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.

[0035] 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. ;

[0036] where, is a two-layer fully connected network with hidden dimension 512 and activation function ReLU, and the output dimension is consistent with the aging index dimension.

[0037] Preferably, the life degradation model in S5 is defined as the cumulative degradation amount describes the performance degradation of the material over time,

[0038] Based on the defined degradation amount , the remaining life is calculated: when the cumulative degradation amount reaches the set failure threshold , the component is judged to be failed; at this time, the remaining life is estimated by the ratio of the degradation amount to the average degradation rate . represents the remaining service life of the silicon carbide component at the kth frame.

[0039] Compared with the prior art, the innovation points of the present application include:

[0040] (1) Multi-source sensing and GAN-driven energy field reconstruction method, which breaks through the technical bottleneck of difficult direct measurement of spatial distribution, and can obtain high-resolution three-dimensional energy distribution in a non-invasive manner.

[0041] (2) The introduction of time series deep network enables complete capture of transient dynamic characteristics, effectively making up for the shortcomings of traditional static models in peak and fluctuation feature extraction.

[0042] (3) Nonlinear modeling of the relationship between energy distribution and material aging not only improves the accuracy of life prediction, but also realizes the characterization of the aging process from both mechanism and data.

[0043] The beneficial effects brought by the innovation points of the present application include:

[0044] (1) The detection accuracy is greatly improved. Through multi-source data fusion and GAN-driven distribution reconstruction, high-resolution three-dimensional energy field can be obtained, which makes up for the lack of information of traditional single-point detection.

[0045] (2) The prediction reliability is enhanced. Combined with the dynamic modeling of time series deep network and the mapping of aging relationship, the present method can maintain a low prediction error in real service environment, which is significantly better than existing methods.

[0046] (3) Support real-time online application. The system can continuously collect probe and optical signals during actual operation, and output energy distribution and life prediction results in real time, which has strong engineering applicability.

[0047] (4) Reduce maintenance cost and risk. By predicting the remaining useful life (RUL) of silicon carbide components in advance, planned maintenance can be achieved to avoid economic losses and safety risks caused by sudden failure of the system.

[0048] The method has real-time and generalization, and can realize online life prediction in the actual operation process of the silicon carbide component, and can be applied to health assessment and life prediction of other plasma bombarding materials such as gallium nitride, quartz and metal electrode. By obtaining the remaining useful life information of the component in advance, the present application can effectively support planned maintenance, reduce operation and maintenance cost, and avoid safety hazards and economic losses caused by sudden failure. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 The overall technical route flowchart of the present application.

[0050] Figure 2 The structure diagram of the generated adversarial network for plasma bombarding energy distribution reconstruction of the present application.

[0051] Figure 3 The structure diagram of the timing deep network for bombarding dynamic characteristic modeling of the present application.

[0052] Figure 4 The global peak time series and pulse detection and energy centroid trajectory and cumulative energy diagram of the present application.

[0053] Figure 5 The life prediction and health assessment diagram of the present application. DETAILED DESCRIPTION

[0054] The present application provides a method for detecting and predicting the life of silicon carbide components bombarded by plasma, and the overall process is as shown in Figure 1

[0055] Plasma parameter acquisition and data processing: First, a magnetic probe array is laid out to obtain local electromagnetic parameters, and optical imaging and spectral acquisition equipment are used to obtain the light emission characteristics of the plasma. Synchronous acquisition of the probe and optical channel is realized through hardware triggering and unified clock, and unified marking is carried out in combination with process condition data. Secondly, after obtaining the original data, the magnetic probe signal is filtered, denoised and normalized, and the optical image is background subtracted, flat field corrected and geometrically corrected, so as to improve the reliability and consistency of the signal.

[0056] ​Three-dimensional energy distribution reconstruction based on generative adversarial network: The high-resolution reconstruction of plasma energy distribution is realized by using a generative adversarial network. The generator takes the feature data after multi-source fusion as input to infer the energy distribution field in three-dimensional space, and the discriminator improves the authenticity and accuracy of the reconstruction results by comparing with experimental or simulation data. Through this process, the complete energy distribution map can be obtained without direct contact with the silicon carbide component, and the three-dimensional spatial resolution that is difficult to achieve by traditional methods can be realized.

[0057] Timing modeling of bombardment dynamic characteristics and feature correlation: A timing deep network is introduced to model the evolution law of energy distribution over time. The timing data of energy distribution are taken as input, and the model can identify the peak value, pulse width, energy accumulation effect and spatial migration trend in the bombardment process, so as to completely capture the dynamic characteristics of plasma bombardment. On the basis of dynamic characteristic modeling, the energy distribution and the aging characteristics of silicon carbide material are correlated and modeled. Through experimental calibration and offline testing, the degradation law of the material under different energy conditions is obtained. The nonlinear mapping relationship between energy distribution features and aging indicators is established by using a deep learning model, so that the model can dynamically predict the degree of material performance degradation according to the actual energy input.

[0058] Lifetime prediction and online health assessment: When new plasma bombardment data is input, the complete energy features are obtained through energy distribution reconstruction and timing dynamic modeling, and these features are input into the lifetime prediction model to output the remaining useful life of the silicon carbide component. The prediction result can be fed back to the system in real time to realize online health assessment and early warning. When the life prediction value approaches the failure threshold, the system can trigger maintenance recommendations, so as to realize planned maintenance and avoid economic losses and safety risks caused by sudden component failure.

[0059] The invention is further described below in conjunction with specific embodiments.

[0060] I. Plasma parameter acquisition and data processing

[0061] Due to the strong transient nature, complex spatial distribution and nonlinear coupling of the plasma bombardment process, it is often difficult to completely reveal its physical laws by relying on a single measurement means. Therefore, the present invention adopts a multi-source fusion strategy of magnetic probe array and optical imaging / spectral diagnosis to synchronously acquire the electromagnetic characteristics and optical emission characteristics near the surface of the silicon carbide component, and ensures the consistency and comparability of various data through system calibration and time synchronization means.

[0062] After obtaining the original signals, the present invention further processes denoising, filtering, normalization and spatio-temporal registration, maps the data from different sources to a unified coordinate frame and time reference, and extracts key feature parameters closely related to energy input.

[0063] 1. Plasma parameter acquisition and system calibration:

[0064] Firstly, system calibration is needed to ensure that the output of multi-source sensors can be accurately mapped to the workpiece coordinate system under the plasma bombardment environment of silicon carbide components. Inside or outside the working cavity of the silicon carbide component, a magnetic probe array and an optical imaging / spectral acquisition system are arranged according to the geometric size of the surface area to be measured and the desired spatial resolution.

[0065] A magnetic probe array is arranged around the silicon carbide component in a grid-like manner according to the target bombarded surface, and the magnetic probe array is used to acquire local magnetic field strength (1 th probe corresponds to local magnetic field data), the probe spacing is 5 mm, which matches the desired spatial resolution of the target surface. The number of turns of a single probe is 50; the sampling frequency of the magnetic probe is fixed at , and the sampling time step resolution is , which can capture the rapid changes of the plasma magnetic field on the nanosecond scale.

[0066] The probe body is composed of an induction coil and an insulating support structure, and the transient magnetic field change in the plasma produces an induced electromotive force in the probe coil. This induced electromotive force is converted into a voltage signal by a preamplifier circuit, i.e. the magnetic probe output voltage . The magnetic probe output voltage is proportional to the magnetic field change rate , and the local current density is calculated by electromagnetic inversion.

[0067] Specifically, the optical part is composed of a high-speed camera and a spectrometer. The high-speed camera and the fiber-coupled spectrometer are arranged along the unobstructed line of sight, the frame rate of the high-speed camera is fixed at , the camera resolution is fixed at 1024*1024 pixels, and the camera exposure time is fixed at . The original image recorded by the camera is , where index k represents the kth camera, is the pixel coordinate, and t is the absolute timestamp. All acquisition devices are synchronized by a unified clock, and the system clock jitter is upper limited to , which can capture the spatial distribution of plasma light intensity on the millisecond time scale.

[0068] The spectrometer is used to record the emission spectrum, and the electron temperature is calculated by the Boltzmann spectrum method to reflect the local plasma energy state. The original spectral data recorded by the spectrometer is represented as , where is the wavelength, and index represents the The spectrum lines. All devices are triggered by a unified clock, and the clock jitter is less than , which ensures the accurate alignment of different signals in time. 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 denoted as a set of probe coordinates At the same time, the process condition vector is recorded during the acquisition process, where the process condition items include: radio frequency power, bias voltage, gas flow, cavity pressure, and applied magnetic field strength.

[0069] Further, in the hardware calibration link, in order to ensure the authenticity of the high-frequency signal, the frequency response function of each magnetic probe is measured by experiment, which is used for compensation in post-processing, where ω represents the angular frequency, and the function describes the gain and phase characteristics of the probe at different frequencies, which is used for response compensation of the original voltage to restore the true signal. The system noise is represented by the noise spectral density , which is used to suppress noise amplification in the filtering and deconvolution process. The camera obtains the intrinsic matrix K describing the relationship between the pixel coordinates and the imaging plane, the extrinsic matrix R, and the distortion parameter set , which realizes the mapping between the image pixels and the three-dimensional surface coordinates of the workpiece.

[0070] The set of magnetic field vectors obtained by the discrete probe is denoted as , and the local current density is estimated by the magnetic field . The kernel matrix is obtained by electromagnetic simulation in the simulation stage, and the regularization parameter is set to , is the unit matrix. After the calibration is completed, the following data and meta-information are written into the system database: time-synchronized original acquisition data, response function of each probe , camera intrinsic and extrinsic parameters K, R, t, mapping of the probe coordinate system to the workpiece coordinate system, system clock jitter measurement value, and noise spectrum estimation.

[0071] The above data is summarized to obtain the original input data ;

[0072] In addition, the environmental treatment parameters are used for pre-processing of the original input data .

[0073] 2. Data preprocessing and feature construction

[0074] In the data preprocessing stage, the parameters in are used to process in the following order.

[0075] 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. .

[0076] 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 set of flat-field images and the pixel dark current background. The difference is obtained by comparing the values.

[0077] 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. .

[0078] 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 are synchronized via clock jitter The correction ensures that the optical and electromagnetic data are synchronized under the same time reference.

[0079] Further, after the pre-processing is completed, in order to ensure the uniformity of subsequent modeling, a sliding window length of is set in the time dimension , the number of frames in the time window is selected , and the step size is , and the spatial dimension is uniformly resampled to 128*128 grid. The multi-source fusion feature tensor finally constructed is , where T=100 represents the number of time frames, H=W=128 represents the spatial resolution, the number of channels C=6, and: , where , and the corrected luminous intensity distribution is weighted and combined.

[0080] After data preprocessing and feature construction, the original input data is obtained . Among them, concentrates the electromagnetic and optical characteristics under plasma bombardment, is the aligned process condition sequence, which provides external constraints for subsequent energy distribution reconstruction.

[0081] II. Three-dimensional energy distribution reconstruction based on generative adversarial network

[0082] After the collection, preprocessing and feature construction of multi-source data are completed, the multi-source fusion feature tensor and the process condition sequence are obtained. These information depict the external performance of the process. However, only two-dimensional observation results and surface proxy variables are still insufficient to directly reflect the deposition distribution of energy in the material body. Therefore, the present application adopts a generative adversarial network, based on and , combines physical constraints and data-driven methods, and reconstructs the three-dimensional voxel-level energy deposition distribution inside the workpiece.

[0083] 1. Input and target

[0084] The goal of this process is to reconstruct the voxelized energy deposition distribution from , where is the time frame index; is the spatial grid index; is the depth layer index. ​

[0085] In this process, the input of the model is first from the S1 of the , which collectively contains 6 channels: magnetic field strength , magnetic field rate of change , the current density estimate obtained by inversion , the loop current after system correction , electron temperature , flux estimates based on optical proxies These quantities comprehensively characterize the spatial distribution and dynamic characteristics of the plasma. Secondly, the process condition sequence is used to introduce the timing prior of external process driving; finally, the random noise subject to normal distribution enhances the expression ability of the model to observation uncertainty.

[0086] The target output of the GAN model in the training stage is the three-dimensional voxel energy deposition distribution , and simultaneously predicts the uncertainty estimate of each voxel . This distribution field cannot be easily and completely directly observed by experiment, and can only be obtained under limited conditions with the help of small-scale calibration experiments. That is: under controlled conditions, a limited number of high-voltage breakdown discharge tests are performed on silicon carbide wafers. The high-speed camera acquires the timing image of the arc, combined with the spectrum acquisition device to record the arc radiation intensity, while the infrared thermal imager measures the material surface temperature rise distribution. Based on the heat conduction inversion model, the three-dimensional voxel energy distribution in a limited spatial range is obtained.

[0087] Therefore, in the present application, its main role is as a supervisory signal in the training and verification process. The system only needs to input , and the corresponding three-dimensional energy deposition distribution can be inferred using the trained generative model, while outputting the uncertainty estimate of each voxel .

[0088] 2. Condition construction and position embedding

[0089] In the GAN model, the generator is used to infer the three-dimensional distribution from the two-dimensional timing characteristics, and simultaneously outputs the predicted variance of each voxel to characterize the uncertainty. For this purpose, the present application adopts the strategy of “spatiotemporal coding → deep expansion → voxel decoding”, and uses the FiLM (Feature-wise Linear Modulation) conditional modulation mechanism in multiple layers to realize the control of the working condition on the generation process.

[0090] Specifically, the core task of the generator is to map the spatiotemporal characteristics and process conditions into a three-dimensional energy field, while considering 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.

[0091] 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 .

[0092] 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.

[0093] 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.

[0094] 3. Discriminator modeling

[0095] The role of the discriminator is to determine whether the input energy distribution is real or generated by the generator. By introducing adversarial training, the authenticity and detail quality of the generator output can be improved. In this invention, the discriminator adopts the Hinge loss form:

[0096] wherein, represents the real energy field distribution. The first term ensures that the real energy field is judged to be true, and the second term punishes the generated energy field for being judged to be true. The adversarial loss of the generator is: .

[0097] The PatchGAN structure is used to determine the authenticity of the input energy field in the local area. The output of the real generated adversarial probability, i.e. the discriminator probability , is used for adversarial training to update the generator and discriminator parameters.

[0098] Further, in order to ensure that the generated results are not only realistic but also physically reasonable, the loss function includes not only the adversarial term but also the reconstruction constraint, the energy conservation constraint, the proxy alignment, the smoothing constraint, and the uncertainty modeling. The overall optimization goal is:

[0099]

[0100] wherein, is the calculation of the adversarial loss, which makes the generated three-dimensional voxel energy deposition distribution approach the real distribution in a statistical sense through the game process of the generator and the discriminator, and improves the generation authenticity; is the reconstruction loss, which calculates the Euclidean distance between the generated result and the limited sparse reference data to ensure that the model can be strictly aligned on the existing observation points, thereby enhancing the reliability of the prediction; is the calculation of the energy conservation loss, which compares the difference between the total power corresponding to the predicted energy deposition distribution and the input power to ensure that the generated result follows the law of conservation of energy; 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 the interference of non-physical noise, and enhancing the spatiotemporal smoothness; represents the calculation of the uncertainty loss, which jointly models the prediction result and the corresponding variance, so that the output not only contains the point estimate value of the energy distribution, but also provides the uncertainty information of the prediction. are the weight coefficients, respectively.

[0101] 4. Training and optimization details and output

[0102] In the present application, the training of the GAN model is carried out, the optimizer adopts Adam, the generator learning rate is set to , the discriminator learning rate is ; the batch size B=2, and the total round is 200.

[0103] In the inference stage, a new time window and a noise sampling are given , and then the surface energy flux is calculated.

[0104] Finally, the reconstructed three-dimensional voxel energy deposition distribution , the surface energy flux sequence and the uncertainty estimation are obtained through the training of the GAN network. Among them as the driving term of the surface loading intensity, as the body domain driving of the defect evolution caused by the energy deposition in the body, and the uncertainty is used for confidence weighting and risk boundary discrimination. The structure diagram of the plasma bombardment energy distribution reconstruction generative adversarial network is shown in Figure 2 .

[0105] III. Construction of bombardment dynamic characteristic modeling time series deep network

[0106] 1. Construct the bombardment dynamic characteristic modeling time series deep network, as shown in Figure 3 : first, construct a learnable frame-level vector from the three-dimensional energy deposition distribution, generate a low-dimensional feature vector for each frame, and splice it with the process condition sequence to form a joint feature vector ; input the joint frame-level feature sequence into the bidirectional LSTM to obtain the time-level hidden state, and build event detection and time series attention on it to obtain the global time representation vector ; train a regression mapping to convert the energy time series features into material aging index prediction results .

[0107] 1. Construct a learnable frame-level vector from the time series energy field

[0108] Extract the peak value and pulse width, pulse energy, short / long window cumulative amount, spatial centroid and migration speed, and frequency energy from the obtained three-dimensional energy deposition distribution , the surface energy flux sequence ; then through spatial aggregation and dimension reduction, the pixel-level and voxel-level quantities are merged into a low-dimensional feature vector for each frame, denoted as .

[0109] Specifically, to suppress the distribution drift caused by each working condition, the normalized process condition sequence is linearly transformed to a similar scale with by incorporating frame-level features, and then concatenated to form the joint input , where A is a fixed diagonal scaling matrix. To obtain the peak criterion, the background baseline mean and standard deviation of each pixel are first estimated based on the previous frames.

[0110] Specifically, the predicted energy flux values of the previous 10 frames are taken at the same position. Then, they are arithmetically averaged, i.e., the values of the 10 frames are added and divided by 10, to obtain the mean of the background energy. Subsequently, the standard deviation of the same position is calculated to characterize the fluctuation size of the background energy flux. The specific method is: the square of the difference between the energy flux value of each frame and the mean is 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 , which is obtained by the weighted sum of the standard deviation and the mean .

[0111] Local maximum detection is performed on the time series of each pixel to determine the peak amplitude and occurrence time, and a peak list is recorded for subsequent pulse width and energy calculation. If a frame is smaller on both sides, it is recorded as and its time .

[0112] Secondly, for each peak, the full width at half maximum (FWHM) is used to quantify the pulse duration, which is specifically found by searching for the time index , reaching half the height on the left and right of the current peak, and converted to seconds. The pulse duration is calculated by the difference between the time indexes reaching half the height on the left and right, in time steps.

[0113] To measure the energy load of a single pulse, the surface flux in the pulse interval is discretely integrated to obtain the pulse energy: . To describe the energy accumulation effect, the window length frames is defined, and the surface flux is slidingly integrated: . To depict the migration trend of the energy hotspot, the energy centroid is calculated by taking the surface flux as the weight: , . And the migration speed is obtained by dividing the displacement of the centroid of adjacent frames by ​ .

[0114] Thus, by the above processing, a low-dimensional feature vector is generated for each frame , whose components include peak amplitude, pulse energy, cumulative energy, centroid position, and migration speed, etc. Then, the process condition sequence obtained in step two is spliced to form a joint feature vector .

[0115] 2. Time series deep encoding

[0116] Determine the peak-pulse-cumulative-migration evolution law. In the present invention, the joint frame-level feature sequence is input into a bidirectional LSTM to obtain the time-level hidden state, and an event detection and time series attention are constructed thereon to form a high-order representation of the dynamic of the bombardment process.

[0117] In order to obtain a time representation containing context, the sequence is input into a bidirectional LSTM, and the forward and backward hidden states and are calculated respectively and spliced in the feature dimension. In the forward propagation direction, the hidden state of the first frame is calculated by the forward LSTM module. Its input includes the feature representation of the current frame and the hidden state of the previous time In the backward propagation direction, the hidden state of the first frame is calculated by the backward LSTM module and the hidden state of the next time Finally, the complete hidden state of the first frame is spliced from the forward hidden state and the backward hidden state , thereby retaining the context information of the previous and subsequent time series. The unidirectional hidden dimension is 256.

[0118] In order to label important bombardment events (strong peaks and migration mutations) at the frame level, the hidden state is input into an MLP network, and the event probability is calculated by a layer of MLP and Sigmoid. A fixed threshold is used to determine whether it is an event frame.

[0119] Specifically, the hidden state is input into a layer of multi-layer perceptron (MLP) network, and the event score of each time frame is calculated by linear transformation, and then the score is mapped to an event probability value between 0 and 1 by Sigmoid activation function . The linear transformation corresponding to the formula includes weight matrix and bias vector , respectively, for adjusting the channel contribution and the overall offset of the hidden state.

[0120] To let the model focus on the key segments in the time series, the invention adopts dot-product attention to aggregate the context in the time dimension to obtain a weighted global time representation for subsequent correlation modeling. Dot-product attention mechanism is adopted in the time dimension. For the hidden state of the i-th frame, first calculate the dot product with the global context vector , and then map it through the weight matrix to obtain the attention score, and then perform exponential normalization on the scores of all frames to obtain the attention weight : weight the hidden state of each frame according to the corresponding attention weight , and sum to obtain the global time representation vector . Among them, is the average hidden state, is the attention projection matrix.

[0121] 3. Feature-aging correlation

[0122] Offline calibration data is used to train a nonlinear mapping. According to the obtained global time representation , pair it with the aging index label obtained by offline experimental calibration, and train a regression mapping . This mapping converts the energy time series features into material aging index prediction results (surface roughness increment, cracking density increment, and conductivity degradation rate).

[0123] The mapping relationship from the time series representation to the aging index is input into a multi-layer perceptron and outputs a multi-dimensional aging index prediction vector .

[0124] Among them, is a two-layer fully connected network with a hidden dimension of 512 and a ReLU activation function, and the output dimension is consistent with the aging index dimension. The above mapping is trained using experimental labels, and mean square error is used as the main loss, and L2 regularization is added to improve generalization:

[0125]

[0126] Among them, is the number of calibration samples, . is the true aging index of the i-th experimental sample.

[0127] ​​​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 according to the time sequence characteristics of energy input.

[0128] Finally, the time sequence feature sequence , the joint feature sequence , the time sequence global representation and the aging index prediction result are output as output, which are used for subsequent life prediction.

[0129] Four, life prediction and online health assessment

[0130] On the basis of the output features , , , , a life degradation model is established to further calculate the remaining useful life (RUL) of the silicon carbide component. In addition, the application also designs an online health assessment and early warning mechanism, which can timely trigger maintenance suggestions when the life approaches the threshold, thereby ensuring operation safety and economy.

[0131] First, the degradation degree of the material is quantified from the dynamic characteristics. Considering that the damage caused by plasma bombardment is not instantaneous, but a gradual accumulation process. The cumulative degradation amount is defined to describe the performance degradation of the material over time. Its calculation form is:

[0132]

[0133] Wherein, is the degradation rate function, which is used to comprehensively calculate the degradation speed at time , represents the discrete frame number, is a fixed sampling interval. The degradation amount continuously accumulates with the passage of time, and its value directly reflects the comprehensive degradation degree of the silicon carbide component at the kth frame.

[0134] After the degradation amount is defined, the remaining life is further calculated. When the cumulative degradation amount reaches the set failure threshold , the component is judged to be failed. At this time, the remaining life is estimated by the ratio of the degradation amount and the average degradation rate . represents the remaining useful life of the silicon carbide component at the kth frame.

[0135] To evaluate the component health status more intuitively, a health index is introduced , which is used to normalize the current degradation level of the material:

[0136]

[0137] where, when , the material is in a completely healthy state, and as the bombardment accumulates, increases, the health index will gradually decrease, and when approaches 0, it indicates that the health state is about to disappear.

[0138] Further, the system not only needs to evaluate the health index, but also needs to trigger maintenance recommendations when the life is approaching the limit. Set the initial design life as , the warning threshold is , when the remaining life satisfies: , the system will trigger a maintenance warning signal to prompt the operation and maintenance personnel to replace or repair the component within the specified time. Avoid economic losses and safety risks caused by sudden failure.

[0139] After life prediction and online health assessment, the degradation level of the component at any time , and the remaining life prediction of the component ; the health index reflecting the health status of the component ; warning signal, triggered when the life is lower than the set threshold.

[0140] Five, experimental analysis

[0141] As shown in Figure 4 , the global peak time series and pulse detection and energy centroid trajectory and cumulative energy map are shown. In the global peak time series and pulse detection, the blue curve represents the global energy peak value evolving over time, the red dot marks the identified pulse peak, and the vertical line gives the corresponding pulse occurrence time. In the energy centroid trajectory and cumulative energy distribution. The background color map shows the cumulative energy distribution in the entire observation period, and the brighter the color, the greater the cumulative energy. The white curve represents the evolution process of the centroid trajectory, and the green point is the starting position and the red point is the end position. From the trajectory, the migration law of the energy hot spot in space can be seen.

[0142] As shown in Figure 5 , in Figure 5The upper half of the figure shows the degradation process of the silicon carbide component. The horizontal axis is time, and the vertical axis is the degradation amount. The curve clearly reflects the continuous accumulation of the degradation amount of the component under the action of plasma bombardment over time. When the degradation amount gradually approaches the preset failure threshold, i.e., at the position of "approaching failure", it indicates that the component is about to enter the end-of-life stage. The figure directly reveals the degradation evolution law and provides a basis for subsequent life prediction. Figure 5 The lower half of the figure draws the residual life prediction curve. The curve gradually decreases over time, reflecting the continuous shortening of the remaining available life of the component. When the remaining life decreases to 20% of the initial life, the system will trigger a warning. It indicates that at this time, planned maintenance is needed to avoid the component entering a sudden failure state.

[0143] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included within the protection scope of the present application.

[0144] Although the specific embodiments of the present application have been described above, they are not intended to limit the scope of protection of the present application. Those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A method of silicon carbide component plasma bombardment energy distribution detection and lifetime prediction, characterized by, The process comprises the following steps: S1, arranging a magnetic probe array to obtain local electromagnetic parameters, and synchronously using optical imaging and optical spectrum acquisition equipment to obtain the light emission characteristics of the plasma; S2, using environmental processing parameters to preprocess the acquired data, mapping data of 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 the aligned process condition vector ; S3, using a generative adversarial network (GAN) to reconstruct a three-dimensional energy deposition distribution inside the workpiece based on and , combining physical constraints and data-driven methods, reconstruct a three-dimensional energy deposition distribution inside the workpiece; S4, based on the dynamic characteristics of the modeling timing depth network, first constructs a learnable frame-level vector through a three-dimensional energy deposition distribution, and generates a low-dimensional feature vector for each frame , and spliced with the process condition vector to form a joint frame-level feature sequence ; the joint frame-level feature sequence is input into the bidirectional LSTM to obtain the time-level hidden state, and the event detection and timing attention are constructed thereon to obtain the energy timing feature ; Training a regression mapping Converting energy temporal features Converting to material aging index prediction results ; S5, on the basis of the obtained , , , , a life degradation model is established to calculate the remaining service life of the silicon carbide component.

2. A method of detecting plasma strike energy distribution and predicting lifetime of a silicon carbide part as recited in claim 1, wherein: S1 data collected include magnetic field vectors from discrete probes , magnetic probe output voltages , raw images recorded by cameras are , raw spectral data recorded by spectrometers , process condition vectors ; wherein, is a pixel coordinate, t is an absolute time stamp, is a wavelength, index denotes the th spectral line; Process condition vector Including RF power, bias voltage, gas flow, chamber pressure and applied magnetic field strength.

3. A method of detecting plasma strike energy distribution and predicting lifetime of a silicon carbide part as recited in claim 2, wherein: The environment processing parameters described in S2 include a response function of each probe , a noise spectral density , a camera intrinsic matrix K, a camera extrinsic matrix R, a set of probe coordinates , a set of distortion parameters , a mapping matrix from the probe coordinate system to the workpiece coordinate system, denoted as M, system clock jitter , a local current density calculated by the magnetic probe through electromagnetic inversion ; The specific pre-processing process comprises: First, the magnetic probe output voltage Low frequency drift and high frequency noise are removed by a band pass filter; in the frequency domain each probe response is compensated and the noise spectrum is suppressed In the frequency domain representation, the Fourier transform of the original voltage is ; using the known number of turns and effective area of the coil of the single probe and the sampling time step the voltage integral is normalized back to the instantaneous magnetic field ; the time derivative of the magnetic field is then calculated This quantity is used as a dynamic proxy for the energy input; after obtaining the magnetic field field distribution, the local current density is estimated from the obtained , the kernel matrix and the regularization parameter , the current density estimate is obtained by a regularized linear inverse problem solution ; For optical image, the original image of optical channel is first corrected by dark field correction After dark field correction, the pixel dark current background is obtained , the average of flat field image set is obtained ; the corrected brightness distribution is obtained according to the pixel correction formula ; Subsequently, the pixel coordinates were determined using the camera intrinsic and extrinsic parameter matrices K and R. After coordinate transformation, the image is projected and back-projected onto the workpiece surface in three-dimensional coordinates (X1, Y1, Z1). Three-dimensional interpolation is then used to map the image brightness onto the same workpiece surface grid as the magnetic field probe, yielding the luminous intensity distribution in physical space. ; Raw spectral data After calibration and background subtraction, several diagnostic lines are selected and the electron temperature is estimated from the Boltzmann plot ; all data are synchronized on the same time reference by clock jitter correction ; The length of the sliding window is set in the time dimension as , the number of frames in the time window is selected as , the step is , and the resampling in the spatial dimension is unified to a 128*128 grid, so that a multi-source fusion feature tensor is finally constructed as , wherein T=100 represents the number of time frames, H=W=128 represents the spatial resolution, the number of channels C=6, and , wherein The product of the current density amplitude and the magnetic field change rate and the corrected luminous intensity distribution are combined to form the magnetic field change rate.

4. The method of claim 1, wherein the method further comprises: The generative adversarial network (GAN) has a target output of a three-dimensional voxel energy deposition distribution in a training phase , and simultaneously predicts an uncertainty estimate for each voxel , and a surface energy flux sequence is obtained by numerically integrating the three-dimensional deposition distribution along a depth direction ; The three-dimensional voxel energy deposition distribution is obtained by high-speed photography to acquire the timing image of the electric arc, recording the electric arc radiation intensity by a spectrum acquisition device, measuring the material surface temperature rise distribution by an infrared thermal imager, and based on a heat conduction inversion model, thereby obtaining the three-dimensional voxel energy distribution in a limited space range .

5. The method of claim 4, wherein the method further comprises: In the generative adversarial network (GAN), a core task of the generator is to map spatio-temporal features and process conditions into a three-dimensional energy field, while considering randomness , through a mapping function parameterized by the generator , to obtain an output including a predicted three-dimensional voxel energy deposition distribution and a corresponding uncertainty estimate ; To obtain a compact spatio-temporal feature representation, the input tensor is first encoded: wherein, is a spatio-temporal convolutional encoder, outputting is a spatio-temporal compressed feature tensor; During the generation process, the process conditions play a role of constraint and modulation, through the FiLM mechanism, the process conditions are converted into scaling and offset parameters , and used to modulate the features, for the first layer feature map , the scaling parameters are multiplied element-wise per channel, and summed with the offset parameters ; To extend the two-dimensional surface features to three-dimensional voxel space, a learnable deep extension kernel is introduced, which is implemented by numerical integration of the spatio-temporal compression feature tensor and a weight matrix in the depth direction to learn how the surface features propagate into the material interior; finally, a non-negative energy distribution is output by the decoder network with a log-variance .

6. The method of claim 5, wherein the method further comprises: In the generative adversarial network (GAN), the core of the decoder is composed of a module combining multi-layer transposed convolution and up-sampling convolution. and respectively adopts a three-layer structure: the first layer is a 3*3*3 three-dimensional transposed convolution, which gradually restores the low-dimensional latent feature to the voxel space while performing nonlinear mapping, the second layer is an up-sampling convolution, which further improves the spatial resolution, and the third layer is a 1*1*1 convolution layer, which is used to output the final voxel-level prediction result; wherein softplus is an activation function to ensure that the prediction energy is non-negative, output the prediction value of the energy deposition distribution, output the logarithmic variance of uncertainty corresponding to the voxel; The generator outputs a three-dimensional voxel energy deposition distribution and simultaneously predicts the variance of each voxel for characterizing uncertainty.

7. The method of claim 1, wherein the method further comprises: The S4 constructs a learnable frame-level vector based on the three-dimensional energy deposition distribution, and generates a low-dimensional feature vector for each frame , which is extracted from the obtained three-dimensional energy deposition distribution , the surface energy flux sequence The peak value and pulse width, pulse energy, short / long window cumulative amount, spatial centroid and migration velocity, and frequency energy description amount are extracted; then through spatial aggregation and dimension reduction, the pixel-level and voxel-level quantities are merged into a low-dimensional feature vector for each frame, denoted as .

8. The method of claim 7, wherein the method further comprises: The joint frame-level feature sequence in S4 The input bidirectional LSTM obtains the time-level hidden state, and constructs event detection and timing attention thereon to obtain the energy timing feature 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 reverse propagation, the first Hidden state of a frame The inputs, calculated by the inverse LSTM module, include... 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 the important bombardment events at frame level, the hidden state is passed through a layer of MLP with Sigmoid output event probability, and a fixed threshold is used to determine whether it is an event frame or not. determines whether it is an event frame or not. 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 energy time series characteristics. .

9. The method of claim 8, wherein the method further comprises: training a regression mapping in the S4 energy temporal features into material aging index prediction results the mapping converts energy temporal features into material aging index prediction results, the results including surface roughness increment, cracking density increment, and electrical conductivity degradation rate mapping relationship from the time series representation to the aging indicator, the energy time series feature inputting the multilayer perceptron and outputting a multi-dimensional aging indicator prediction vector ; wherein, is a two-layer fully connected network with hidden dimension 512 and ReLU activation function, and the output dimension is consistent with the dimension of the aging index.

10. A method of detecting plasma strike energy distribution and predicting lifetime of a silicon carbide part as recited in claim 1, wherein: The S5 lifetime degradation model first defines the cumulative degradation amount Describes the performance degradation of a material over time, based on the defined degradation quantity , the remaining life is calculated: when the accumulated degradation quantity reaches a set failure threshold , the component is determined to be failed; at this time, the remaining life is estimated by the ratio of the average degradation rate ; represents the remaining service life of the silicon carbide component at the kth frame.

Images