A method for receiving and processing a coherent Thomson scattering diagnostic signal

Abstract

The application belongs to the technical field of magnetic confinement thermonuclear fusion, and discloses a coherent Thomson scattering diagnostic signal receiving and processing method. The application constructs an ultra-low noise amplification front-end hardware circuit including four core function links of front-end protection, high-frequency amplification and isolation, mixing and intermediate frequency conditioning, channelization and detection quantization, so as to realize conversion of continuous high-frequency power spectrum into high-quality 32-channel discrete digital features. Based on the design method, a backend intelligent spectrum cleaning method based on a one-dimensional residual convolution deep learning network is further provided. The application can provide a systematic coherent Thomson scattering diagnostic signal receiving and processing process and method from three aspects of data set generation, front-end hardware and backend software for the typical application scene of a thermonuclear fusion device, and solves technical problems such as lack of real and reliable experimental data source under strong interference background in the existing coherent Thomson scattering diagnosis, weak signal anti-blocking ability and sensitivity in the nanowatt level, and difficulty in adaptive extraction of fine spectral line features.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of magnetic confinement thermonuclear fusion technology, and particularly relates to a method for receiving and processing coherent Thomson scattering diagnostic signals. Background Technology

[0002] Thomson scattering diagnostics is widely recognized as one of the most direct and accurate methods for measuring plasma electron temperature and density in magnetically confined thermonuclear fusion (MFC) research, and has been widely applied in typical thermonuclear fusion reactors both domestically and internationally (such as KSTAR and ITER). Coherent Thomson scattering (CTS), by measuring and analyzing the scattering power spectrum of incident millimeter waves on collective plasma fluctuations, can obtain key physical parameters and information such as plasma velocity distribution, temperature, and fast ion behavior, making it particularly suitable for diagnosing high-density plasmas in current and future MFC devices. However, in practical engineering applications, existing coherent Thomson scattering diagnostic systems suffer from a core contradiction between extremely weak scattering target signals and extremely strong background interference from spontaneous plasma emission.

[0003] The real-time diagnostic requirements of modern high-parameter thermonuclear fusion devices place high demands on signal acquisition and processing methods, necessitating a systematic design from the front-end hardware link to the back-end spectral reconstruction. Currently, most mainstream coherent Thomson scattering diagnostic theories are based on electrostatic approximations. These models, by neglecting the contributions of electromagnetic field fluctuations and current density, inherently bias the predictions of ion acoustic resonance structures and fast ion phase spatial distributions. Furthermore, existing simulation studies often lack systematic modeling of the complex noise environment of real devices, resulting in a significant simulation-reality gap between theoretical spectral lines and experimental data.

[0004] At the signal receiving hardware level, existing receiver architectures generally face the extreme contradiction between strong interference and weak signals: the central spurious spikes generated by megawatt-level incident waves can easily cause low-noise amplifiers to saturate or even burn out, while extremely deep notch filters can cause irreversible loss of physical information in the center frequency band; in addition, traditional designs have difficulty balancing the noise figure and linearity of cascaded links, and due to the limitation of the quantization bits of analog-to-digital conversion, weak scattered signals are often submerged in the thermal noise and electron cyclotron radiation background fluctuations of the system, resulting in a low signal-to-noise ratio at the front-end output, which cannot meet the requirements of subsequent high-precision inversion.

[0005] In terms of backend data processing, traditional ON-OFF background cancellation methods can only subtract persistent noise, making it difficult to handle rapid time-varying fluctuations in electron cyclotron radiation and sudden pulses of parametric decay instability. Residual noise severely interferes with physical feature extraction. Existing spectral inversion algorithms mostly rely on iterative fitting of pure physical models, which is computationally time-consuming and difficult to achieve online real-time diagnosis. If general AI models are directly applied, there is often a lack of constraints on the physical characteristics of coherent Thomson scattering power spectrum, resulting in poor generalization ability of the network when facing notch blind zones or low signal-to-noise ratio samples. This makes it easy to generate false features that violate physical laws, and the physical reliability of the diagnostic results cannot be guaranteed.

[0006] Based on the above analysis, the urgent technical problems that need to be solved in the existing technology are:

[0007] (1) The traditional coherent Thomson scattering signal receiver architecture is easily affected by the central spurious spikes generated by megawatt-level incident waves, resulting in irreversible loss of physical information in the center frequency band. Moreover, it is difficult to balance the noise figure and linearity of the cascaded link. The weak scattering signal is often submerged in the thermal noise and electron cyclotron radiation background fluctuations of the system, resulting in a low signal-to-noise ratio at the front-end output, which cannot meet the requirements of subsequent high-precision inversion.

[0008] (2) The traditional ON-OFF background cancellation method can only subtract constant noise. It is difficult to handle the rapid time-varying fluctuations of electron cyclotron radiation and sudden pulses of parametric decay instability. The residual noise seriously interferes with the extraction of physical features.

[0009] (3) Existing spectral inversion algorithms mostly rely on iterative fitting of pure physical models, which takes a lot of time to calculate and makes it difficult to achieve online real-time diagnosis. If a general AI model is directly applied, it often lacks constraints on the physical characteristics of the coherent Thomson scattering power spectrum, resulting in poor generalization ability of the network when facing notch blind zones or low signal-to-noise ratio samples. It is very easy to generate false features that violate physical laws and cannot guarantee the physical credibility of the diagnostic results. Summary of the Invention

[0010] To address the problems existing in the prior art, this invention provides a method for receiving and processing coherent Thomson scattering diagnostic signals. Its purpose is to achieve high-precision signal processing and online analysis of the coherent Thomson scattering diagnostic system in a magnetically confined thermonuclear fusion device under strong local noise.

[0011] This invention is implemented as follows: A method for receiving and processing coherent Thomson scattering diagnostic signals includes:

[0012] S1. Introduce the code for calculating the pure power spectrum of coherent Thomson scattering based on the full electromagnetic kinetic model, and consider the typical operating parameters of thermonuclear fusion devices to generate a dataset containing 2500 sets of high-fidelity pure power spectrum data.

[0013] S2. Inject four types of physical noise with non-stationary and time-varying characteristics: electron cyclotron radiation background fluctuations, random bursts of parametric decay instability narrowband spikes, incident wave stray radiation leakage, and statistical thermal fluctuations that obey the radiometer equation. Construct a coherent Thomson scattering band-noise power spectrum dataset with 10,000 data points per group to reconstruct the real diagnostic environment of the thermonuclear fusion device described in S1.

[0014] Furthermore, in step S1, generating a dataset containing 2500 sets of high-fidelity, clean power spectrum data includes:

[0015] Based on the operating parameter range of the thermonuclear fusion device, the plasma electron temperature, ion temperature, electron density, ion velocity distribution function parameters, scattering angle, and incident microwave frequency are sampled in layers.

[0016] Input the operating parameters obtained from each set of samples into the coherent Thomson scattering pure power spectrum calculation code based on the all-electromagnetic kinetic model to obtain the corresponding coherent Thomson scattering pure power spectrum.

[0017] The coherent Thomson scattering pure power spectrum is labeled with peak position, linewidth, peak intensity and line broadening shape. The operating parameters, coherent Thomson scattering pure power spectrum and corresponding labeling information are associated and stored to form a pure power spectrum dataset with traceable physical parameters.

[0018] Furthermore, in S2, constructing the noisy power spectrum dataset of coherent Thomson scattering includes:

[0019] The noise components corresponding to the background fluctuations of electron cyclotron radiation, the narrowband spikes of parametric decay instability, the incident wave stray radiation leakage, and the statistical thermal fluctuations are generated respectively.

[0020] The noise components are superimposed onto the coherent Thomson scattering pure power spectrum data in two ways: single noise injection and multi-noise coupled injection.

[0021] Record the noise type, noise intensity, frequency position, duration, and injection timing parameters for each noise component;

[0022] A mapping relationship is established between the pure power spectrum of coherent Thomson scattering before injection, the power spectrum after single noise injection, the power spectrum after multi-noise coupled injection, and the corresponding noise parameters to form a dataset of noisy power spectrum of coherent Thomson scattering with interpretable noise sources.

[0023] Furthermore, in S2, the background noise of electron cyclotron radiation is subjected to four-fold superposition processing, including: secondary base profile to simulate basic energy distribution, linear baseline drift to simulate global baseline tilt, periodic low-frequency fluctuations to simulate frequency domain fluctuations caused by magnetohydrodynamic instability or density perturbation, and low-pass smooth Gaussian random envelopes with a relative amplitude of 0.5% being independently superimposed in the ON and OFF frames during rapid time-varying fluctuations.

[0024] Another objective of this invention is to provide an ultra-low noise amplifier front-end hardware circuit structure comprising four core functional components, the design method of which includes:

[0025] (1) Front-end protection design: A notch filter with extremely high suppression is deployed at the front end of the radio frequency receiving link to filter out the central spike.

[0026] (2) High-frequency amplification and isolation design: First, the extremely weak signal output from the (1) functional link is input into the high-frequency low-noise amplifier to achieve power boost and avoid the effective signal being submerged by noise. Then, a ferrite microwave isolator is introduced to achieve unidirectional electromagnetic wave conduction and avoid ripple misjudgment caused by output return loss.

[0027] (3) The design of the mixing and intermediate frequency conditioning stage: the front part introduces a local oscillator with extremely low phase noise and a broadband microwave mixer to realize the efficient power spectrum shift of the micro-watt-level high-frequency scattered signal to the intermediate frequency band; the middle part uses an intermediate frequency amplifier to realize the last stage of active conditioning and gain matching of the analog signal before entering the analog-to-digital conversion; the rear part is connected to a microwave attenuator to adjust the signal power level and optimize the impedance matching characteristics of the cascaded circuit.

[0028] (4) The design of the channelization and detection quantization stage is as follows: the front part is set with a high-order anti-aliasing 32-channel filter bank, the middle part uses a logarithmic detector to accurately convert the RF power fluctuations of the front-end microwave band into low-frequency video baseband voltage signals, and the rear part uses a high-performance 24-bit synchronous sampling analog-to-digital converter to complete the synchronous high-precision acquisition of 32 analog baseband signals in the same nanosecond time window, and sends them to the back-end digital signal processor for storage and operation.

[0029] Furthermore, in (2), the high-frequency amplifier has strict requirements of high gain and low noise, and the selection of its device model needs to be compared and matched with the full-link gain allocation and component query.

[0030] In (3), the mixing and intermediate frequency conditioning stage includes:

[0031] S31. The microwave mixer accurately shifts the high-frequency microwave signal downwards to its respective standard intermediate frequency range according to the different power spectra of the high-frequency microwave signal, falling into the optimal sampling aperture of the broadband high-speed analog-to-digital converter described in (4).

[0032] S32. The intermediate frequency amplifier needs to be set with different target gain and specific component selection based on different power spectrum intensity characteristics;

[0033] S33. One microwave attenuator is placed before and after the intermediate frequency amplifier described in S32. The signal in the transmission line undergoes bidirectional attenuation and is weakened by a total of 6dB.

[0034] Furthermore, in (4), the channelization and detection quantization steps include:

[0035] The S41 anti-aliasing 32-channel filter bank adopts a highly integrated channelized architecture, which divides the effective intermediate frequency bandwidth into 32 independent physical channels in an equal proportion, so that the effective detection bandwidth of a single channel is divided into 32 equally divided fine intervals. Each channel unit uses a 5th order Chebyshev type I bandpass filter.

[0036] S42. After receiving the voltage acquisition data from the analog-to-digital converter module, the digital signal processor performs physical inverse mapping to perform exponential operation on the 24-bit quantized voltage value to restore it to the linear power domain value; and performs ON-OFF digital cancellation algorithm to remove most of the electron cyclotron radiation background and hardware standing wave ripple noise.

[0037] Another objective of this invention is to provide a backend intelligent spectral cleaning method based on a one-dimensional residual convolutional deep learning network, including: dataset construction and physical constraint cleaning mechanism, hybrid neural network architecture design, loss function design, training strategy and process, and reconstructed spectral line evaluation and post-processing process;

[0038] The dataset construction and physical constraint cleanup mechanism is used to construct a highly robust training dataset through physical criteria interception and logarithmic domain transformation.

[0039] The hybrid neural network architecture is designed to enhance the original discrete power spectrum information with severe sparsity defects output by the front-end circuit module of the second aspect of the present invention with higher-dimensional feature data and smooth spectral evolution.

[0040] The loss function is designed to integrate frequency adaptive weighting and smoothness regularization in the logarithmic domain, enabling the network to achieve high-fidelity extrapolation in the notch region.

[0041] The training strategy and process are used to fit the logarithmic spectrum of the training data, integrating dynamic noise injection, periodic learning rate restart and implicit early stopping mechanism, so that the model avoids getting trapped in local optima and gradually converges to the optimal weights within 300 training cycles.

[0042] The reconstructed spectral line evaluation and post-processing process is used to restore the normalized prediction in the logarithmic domain to the absolute power spectrum with physical units, and to realize the quantitative evaluation of the reconstruction accuracy of the model on an independent validation set.

[0043] The hybrid neural network architecture design includes:

[0044] 1) The global feature mapping module adopts a progressively expanded fully connected structure. By introducing a multi-layer nonlinear transformation of the LeakyReLU activation function and embedding the Dropout mechanism, the enhanced 79-dimensional input vector discrete signal is transformed into a 400-point feature sequence with a preliminary physical form.

[0045] 2) The deep one-dimensional residual 1D-ResNet decoding module adopts a cascaded structure of feature expansion-residual purification-feature compression, and uses an edge copying and padding strategy in all convolution operations to achieve spectral smoothing without losing physical information;

[0046] The loss function design adopts a frequency-adaptive weighting strategy of blind zone exemption and bilateral weighting, and integrates smoothing L1 error, first-order total variation regularization and second-order gradient consistency constraint to construct a composite loss function with multi-objective joint optimization to achieve the purpose of spectral restoration.

[0047] The parameter update of the training strategy model adopts the Adam optimizer and uses a cosine annealing hot restart strategy to schedule the learning rate to improve convergence efficiency and avoid getting trapped in local optima.

[0048] The reconstructed spectral line evaluation uses a validation set that is completely identical to that used in the training phase, and adopts formula (1) as the core evaluation metric:

[0049] (1)

[0050] Where Rel-RMSE represents the relative root mean square error; N is the number of validation samples, and i is an integer value between 1 and N. The peak value of the true spectral line. To reconstruct spectral values, These are the true spectral line values;

[0051] After calculating the Rel-RMSE of all validation samples, the average error, median error, worst and best sample errors are output to reflect the model performance from the perspective of overall distribution.

[0052] Another object of the present invention is to provide a coherent Thomson scattering diagnostic signal receiving and processing system comprising:

[0053] The module is used to import the code for calculating the coherent Thomson scattering pure power spectrum based on the full electromagnetic kinetic model, and to generate a dataset containing 2,500 sets of high-fidelity pure power spectrum data, taking into account the typical operating parameters of thermonuclear fusion devices.

[0054] The injection module is used to inject four types of physical noise with non-stationary and time-varying characteristics: electron cyclotron radiation background fluctuations, random bursts of parametric decay instability narrowband spikes, incident wave stray radiation leakage, and statistical thermal fluctuations that obey the radiometer equation. This constructs a coherent Thomson scattering band-noise power spectrum dataset with 10,000 data points per group, which can reconstruct the real diagnostic environment of the thermonuclear fusion device described in S1.

[0055] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the coherent Thomson scattering diagnostic signal receiving and processing method.

[0056] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the coherent Thomson scattering diagnostic signal receiving and processing method.

[0057] Another objective of this invention is to provide an information data processing terminal for implementing the coherent Thomson scattering diagnostic signal receiving and processing system.

[0058] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0059] This invention constructs a complete technical solution for receiving and processing coherent Thomson scattering diagnostic signals from a controlled thermonuclear fusion device, encompassing high-fidelity physical modeling and simulation, ultra-low noise RF front-end circuit design, and back-end artificial intelligence spectral cleaning algorithm design. This provides a systematic solution to the problem of effectively extracting weak nanowatt-level signals under strong interference.

[0060] In the physical modeling and dataset construction, this invention introduces code for calculating the pure power spectrum of coherent Thomson scattering based on a fully electromagnetic kinetic model. Considering the typical operating parameters of a thermonuclear fusion device, it generates a high-fidelity pure coherent Thomson scattering power spectrum dataset. On this basis, by injecting four types of physical noise—electron cyclotron radiation fluctuations, parametric decay instability narrowband spikes, incident wave stray radiation leakage, and statistical thermal fluctuations—it constructs a noisy coherent Thomson scattering power spectrum dataset that can recreate the real diagnostic environment of a thermonuclear fusion device. This provides a solid data foundation for subsequent hardware design and algorithm verification.

[0061] In its front-end hardware design, this invention employs an ultra-low noise receiver front-end design based on an integrated high-performance notch filter, broadband low-noise amplifier, ferrite microwave isolator, superheterodyne mixer, and high-order Chebyshev filter bank. Addressing the differences in detection center frequency, signal strength, and background noise levels in thermonuclear fusion devices, the invention features independent parameter matching design for key components. This significantly improves the extremely low signal-to-noise ratio at the antenna port while effectively suppressing central spurious spikes, most electron cyclotron radiation background, and hardware standing wave ripple interference. It can convert the continuous high-frequency power spectrum into high-quality 32-channel discrete digital features, providing reliable input for subsequent intelligent processing.

[0062] In the backend intelligent spectral cleansing method, this invention employs a one-dimensional residual convolutional deep learning network model based on a global mapping-residual decoding architecture, using direct mapping in the logarithmic domain as the data representation method for spectral lines. To address the problem that a single uniform weighted loss function cannot effectively handle information blind spots and residual non-stationary interference, a blind spot exemption-both-side weighting frequency adaptive weighting strategy is introduced to construct a multi-objective joint optimization composite loss function to achieve spectral restoration. Simultaneously, through multiple training regularization methods such as dynamic noise injection, cosine annealing hot restart, snapshot integration, and early stopping strategy, the risk of overfitting to training noise can be effectively suppressed, improving the model's generalization robustness to unknown diagnostic conditions.

[0063] The technical solution of this invention solves a long-standing technical problem that people have long desired to solve but have been unable to achieve: Faced with huge background noise, existing coherent Thomson scattering diagnostic systems cannot effectively resolve the core contradiction between extremely weak scattering target signals and extremely strong plasma spontaneous emission background interference, and there are hardware and software technical challenges in effectively extracting, receiving and processing nanowatt-level weak signals under strong interference background.

[0064] (1) At present, there is no mature experimental database for coherent Thomson scattering diagnostic research: The current mainstream coherent Thomson scattering diagnostic theoretical modeling is mostly based on the electrostatic approximation assumption. This model ignores the contribution of electromagnetic field fluctuations and current density, resulting in inherent biases in the prediction of ion acoustic resonance structure and fast ion phase spatial distribution. At the same time, existing simulation studies often lack systematic modeling of the complex noise environment of real devices, resulting in a large simulation-reality gap between theoretical spectral lines and experimental data.

[0065] (2) At the signal receiving hardware level, the existing receiver architecture generally faces the extreme contradiction between strong interference and weak signal: the central spurious spikes generated by megawatt-level incident waves can easily cause the low-noise amplifier to saturate or even burn out, while the extremely deep notch filter will cause irreversible loss of physical information in the center frequency band; in addition, traditional designs are difficult to balance between the noise figure and linearity of the cascaded link, and due to the limitation of the quantization bits of the analog-to-digital conversion, the weak scattered signal is often submerged in the thermal noise and electron cyclotron radiation background fluctuations of the system, resulting in a low signal-to-noise ratio at the front-end output, which cannot meet the requirements of subsequent high-precision inversion;

[0066] (3) In terms of back-end data processing, the traditional ON-OFF background cancellation method can only deduct constant noise. It is difficult to handle the rapid time-varying fluctuations of electron cyclotron radiation and the sudden pulses of parametric decay instability. The residual noise seriously interferes with the extraction of physical features. The existing spectral inversion algorithms mostly rely on the iterative fitting of pure physical models, which consumes a lot of computation time and is difficult to achieve online real-time diagnosis. If a general AI model is directly applied, it often lacks constraints on the physical characteristics of the coherent Thomson scattering power spectrum, resulting in poor generalization ability of the network when facing notch blind zones or low signal-to-noise ratio samples. It is very easy to generate false features that violate physical laws and cannot guarantee the physical credibility of the diagnostic results. Attached Figure Description

[0067] Figure 1 This is a flowchart of the method for constructing a noisy power spectrum dataset of coherent Thomson scattering provided in an embodiment of the present invention.

[0068] Figure 2 This is a system block diagram of the coherent Thomson scattering diagnostic signal receiving and processing method provided in the embodiments of the present invention.

[0069] Figure 3 This is a flowchart of the overall process for receiving and processing coherent Thomson scattering diagnostic signals provided in this embodiment of the invention.

[0070] Figure 4 This is a schematic diagram of the calculation results of the pure power spectrum of coherent Thomson scattering provided in the embodiments of the present invention.

[0071] Figure 5 This is a structural diagram of the ultra-low noise amplifier front-end circuit provided in an embodiment of the present invention.

[0072] Figure 6 This is a diagram showing the processing procedure and results of the ITER coherent Thomson scattering power spectrum provided in an embodiment of the present invention.

[0073] Figure 7 This is a diagram showing the processing procedure and results of the KSTAR group coherent Thomson scattering power spectrum provided in an embodiment of the present invention.

[0074] Figure 8This is a schematic diagram of the deep learning model structure based on the one-dimensional residual convolutional network 1D-ResNet provided in an embodiment of the present invention.

[0075] Figure 9 This is a comparison diagram of the logarithmic and linear domains of the ITER group coherent Thomson scattering power spectrum cleaning random verification samples provided in the embodiments of the present invention.

[0076] Figure 10 This is a comparison diagram of the logarithmic and linear domains of the KSTAR group coherent Thomson scattering power spectrum cleaning random verification samples provided in the embodiments of the present invention. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0078] like Figure 1 As shown in the figure, the method for receiving and processing coherent Thomson scattering diagnostic signals provided by an embodiment of the present invention includes the following steps:

[0079] S1. Introduce the code for calculating the pure power spectrum of coherent Thomson scattering based on the full electromagnetic kinetic model, and consider the typical operating parameters of thermonuclear fusion devices to generate a dataset containing 2500 sets of high-fidelity pure power spectrum data.

[0080] S2. Inject four types of physical noise with non-stationary and time-varying characteristics: electron cyclotron radiation background fluctuations, random bursts of parametric decay instability narrowband spikes, incident wave stray radiation leakage, and statistical thermal fluctuations that obey the radiometer equation. Construct a coherent Thomson scattering band-noise power spectrum dataset with 10,000 data points per group to reconstruct the real diagnostic environment of the thermonuclear fusion device described in S1.

[0081] like Figure 2 As shown, an embodiment of the present invention provides a coherent Thomson scattering diagnostic signal receiving and processing system, comprising:

[0082] The module is used to import the code for calculating the coherent Thomson scattering pure power spectrum based on the full electromagnetic kinetic model, and to generate a dataset containing 2,500 sets of high-fidelity pure power spectrum data, taking into account the typical operating parameters of thermonuclear fusion devices.

[0083] The injection module is used to inject four types of physical noise with non-stationary and time-varying characteristics: electron cyclotron radiation background fluctuations, random bursts of parametric decay instability narrowband spikes, incident wave stray radiation leakage, and statistical thermal fluctuations that obey the radiometer equation. This constructs a coherent Thomson scattering band-noise power spectrum dataset with 10,000 data points per group, which can reconstruct the real diagnostic environment of the thermonuclear fusion device described in S1.

[0084] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the coherent Thomson scattering diagnostic signal receiving and processing method.

[0085] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the coherent Thomson scattering diagnostic signal receiving and processing method.

[0086] Another objective of this invention is to provide an information data processing terminal for implementing the coherent Thomson scattering diagnostic signal receiving and processing system.

[0087] The overall block diagram of the method for receiving and processing coherent Thomson scattering diagnostic signals from a controlled thermonuclear fusion device in this invention embodiment is attached. Figure 3 As shown, it includes the following steps:

[0088] In a first aspect, the present invention provides a method for constructing a noisy power spectrum dataset of coherent Thomson scattering, the design method comprising:

[0089] S1. Introduce the code for calculating the pure power spectrum of coherent Thomson scattering based on the full electromagnetic kinetic model, and consider the typical operating parameters of thermonuclear fusion devices to generate a dataset containing 2500 sets of high-fidelity pure power spectrum data.

[0090] In the embodiments, the all-electromagnetic kinetic model divides plasma perturbations into four core variables: electron density (n), current density (j), electric field strength (E), and magnetic field strength (B). This allows for the calculation of the joint contribution of all these perturbations to the coherent Thomson scattering signal, and also reflects the Bernstein wave structure, indicating changes in the ion ratio. When generating high-fidelity pure power spectrum datasets for two typical thermonuclear fusion devices, ITER and KSTAR, using the coherent Thomson scattering pure power spectrum calculation code based on the all-electromagnetic kinetic model, the typical parameters used for each calculation are shown in Table 1.

[0091] Table 1 Typical parameters used for calculating the pure power spectrum of coherent Thomson scattering

[0092]

[0093] Appendix Figure 4 An example of the results of calculating the pure power spectrum of coherent Thomson scattering using all-electromagnetic code is shown, which can intuitively present the important structural features of the pure power spectrum of coherent Thomson scattering.

[0094] S2. Inject four types of physical noise with non-stationary and time-varying characteristics: electron cyclotron radiation background fluctuations, random bursts of parametric decay instability narrowband spikes, incident wave stray radiation leakage, and statistical thermal fluctuations that obey the radiometer equation. Construct a coherent Thomson scattering band-noise power spectrum dataset with 10,000 data points per group to reconstruct the real diagnostic environment of the thermonuclear fusion device described in S1.

[0095] In S2, the background noise of electron cyclotron radiation is processed by four superpositions, including: secondary base profile to simulate the basic energy distribution, linear baseline drift to simulate the global baseline tilt, periodic low-frequency fluctuations to simulate the frequency domain fluctuations caused by magnetohydrodynamic instability or density perturbation, and low-pass smooth Gaussian random envelopes with a relative amplitude of 0.5% being independently superimposed in the ON and OFF frames during rapid time-varying fluctuations.

[0096] A second aspect of the present invention provides an ultra-low noise amplifier front-end hardware circuit structure comprising four core functional components, the overall design block diagram of which is attached. Figure 5 As shown; the design method includes:

[0097] (1) Front-end protection design: A notch filter with extremely high suppression is deployed at the front end of the radio frequency receiving link to filter out the central spike.

[0098] In the embodiments, the selected notch filters all have extremely deep stopband suppression, extremely high quality factor, and excellent passband insertion loss; among them, the quality factor Q used in the ITER device is approximately 153.85, and the passband insertion loss fluctuates only between -0.01dB and -0.66dB; the quality factor Q used in the KSTAR device is approximately 262.5, and the passband insertion loss fluctuates only between -0.9dB and -1.5dB.

[0099] (2) High-frequency amplification and isolation design: First, the extremely weak signal output from the (1) functional link is input into the high-frequency low-noise amplifier to achieve power boost and avoid the effective signal being submerged by noise. Then, a ferrite microwave isolator is introduced to achieve unidirectional electromagnetic wave conduction and avoid ripple misjudgment caused by output return loss.

[0100] (3) The design of the mixing and intermediate frequency conditioning stage: the front part introduces a local oscillator with extremely low phase noise and a broadband microwave mixer to realize the efficient power spectrum shift of the micro-watt-level high-frequency scattered signal to the intermediate frequency band; the middle part uses an intermediate frequency amplifier to realize the last stage of active conditioning and gain matching of the analog signal before entering the analog-to-digital conversion; the rear part is connected to a microwave attenuator to adjust the signal power level and optimize the impedance matching characteristics of the cascaded circuit.

[0101] (4) The design of the channelization and detection quantization stage is as follows: the front part is set with a high-order anti-aliasing 32-channel filter bank, the middle part uses a logarithmic detector to accurately convert the RF power fluctuations of the front-end microwave band into low-frequency video baseband voltage signals, and the rear part uses a high-performance 24-bit synchronous sampling analog-to-digital converter to complete the synchronous high-precision acquisition of 32 analog baseband signals in the same nanosecond time window, and sends them to the back-end digital signal processor for storage and operation.

[0102] (2) The high-frequency amplifier has strict requirements of high gain and low noise, and the selection of its device model needs to be compared and matched with the full-link gain allocation and component query.

[0103] In the embodiment, an SBL-5037534025-1515-E1 amplifier with a gain of 45dB and a noise of 3dB is selected for the ITER device with a center frequency of 60GHz, and a W-LNA 75-110403 amplifier with a gain of 40dB and a noise of 3.5dB is selected for the KSTAR device with a center frequency of 105GHz.

[0104] In (3), the mixing and intermediate frequency conditioning stage includes:

[0105] S31. The microwave mixer accurately shifts the high-frequency microwave signal down to its respective standard intermediate frequency range according to the different power spectra of the high-frequency microwave signal, so that it falls into the optimal sampling aperture of the broadband high-speed analog-to-digital converter described in S4.

[0106] In the embodiment, the power spectrum of the ITER group is shifted downward to 2.5~4.5 GHz, and the power spectrum of the KSTAR group is shifted downward to 2~6 GHz;

[0107] S32. The intermediate frequency amplifier needs to be set with different target gain and specific component selection based on different power spectrum intensity characteristics;

[0108] In this embodiment, the ITER group with a total gain of 40dB uses the PE15A1010 model, and the KSTAR group with a total gain of 70dB uses the PE15A1010 model.

[0109] S33. One microwave attenuator is placed before and after the intermediate frequency amplifier described in S32. The signal in the transmission line undergoes bidirectional attenuation and is weakened by a total of 6dB.

[0110] In (4), the channelization and detection quantization process includes:

[0111] The S41 anti-aliasing 32-channel filter bank adopts a highly integrated channelized architecture, which divides the effective intermediate frequency bandwidth into 32 independent physical channels in an equal proportion, so that the effective detection bandwidth of a single channel is divided into 32 equally divided fine intervals. Each channel unit uses a 5th order Chebyshev type I bandpass filter.

[0112] S42. After receiving the voltage acquisition data from the analog-to-digital converter module, the digital signal processor performs physical inverse mapping to perform exponential operation on the 24-bit quantized voltage value to restore it to the linear power domain value; and performs ON-OFF digital cancellation algorithm to remove most of the electron cyclotron radiation background and hardware standing wave ripple noise.

[0113] In this embodiment, the effective bandwidth of the first part of the intermediate frequency is cut into equal parts by 32. For the coherent Thomson scattering power spectrum of the ITER group with a total length of 2 GHz, the width of each independent physical channel is about 62.5 MHz. For the coherent Thomson scattering power spectrum of the KSTAR group with a total length of 4 GHz, the width of each independent physical channel is about 125 MHz. The logarithmic detector in the middle part is based on the true RMS response and uses the ADL5902. The rear part uses the high-performance 24-bit synchronous sampling analog-to-digital converter AD7768.

[0114] Based on the appendix Figure 5 The overall circuit design block diagram is shown below. Using the selected component models, simulations were performed to obtain the processing results of the ITER and KSTAR coherent Thomson scattering power spectra from the ultra-low noise amplification front end. Random sampling examples are attached. Figure 6 Appendix Figure 7 As shown.

[0115] A third aspect of the present invention provides an overall model structure as shown in the attached figure. Figure 8 The back-end intelligent spectral cleaning method based on a one-dimensional residual convolutional network shown includes: dataset construction and physical constraint cleaning mechanism, hybrid neural network architecture design, loss function design, training strategy and process, and reconstructed spectral line evaluation and post-processing process;

[0116] The dataset construction and physical constraint cleanup mechanism is used to construct a highly robust training dataset through physical criteria interception and logarithmic domain transformation.

[0117] In this embodiment, to enhance the model's ability to capture weak physical features, a direct mapping strategy in the logarithmic domain is adopted; a data filtering mechanism based on physical criteria is embedded to intercept samples with a global power mean below 10⁻¹², samples with edge energy exceeding 60% of the peak value, and strong pulse samples with a first-order total variation and peak-to-mass ratio exceeding the threshold by 30 times; in the grouping compliance partitioning strategy, the training / validation set is divided into 8000:2000 units based on sample blocks; the original 32-channel discrete sampling data with information sparsity is enhanced by 31-dimensional first-order difference features and 16-dimensional symmetry features, expanding the input vector dimension from 32 dimensions to 79 dimensions;

[0118] The hybrid neural network architecture is designed to enhance the original discrete power spectrum information with severe sparsity defects output by the front-end circuit module of the second aspect of the present invention with higher-dimensional feature data and smooth spectral evolution.

[0119] The loss function is designed to integrate frequency adaptive weighting and smoothness regularization in the logarithmic domain, enabling the network to achieve high-fidelity extrapolation in the notch region.

[0120] The training strategy and process are used to fit the logarithmic spectrum of the training data, integrating dynamic noise injection, periodic learning rate restart and implicit early stopping mechanism, so that the model avoids getting trapped in local optima and gradually converges to the optimal weights within 300 training cycles.

[0121] The reconstructed spectral line evaluation and post-processing process is used to restore the normalized prediction in the logarithmic domain to the absolute power spectrum with physical units, and to realize the quantitative evaluation of the reconstruction accuracy of the model on an independent validation set.

[0122] The hybrid neural network architecture design includes:

[0123] 1) The global feature mapping module adopts a progressively expanded fully connected structure. By introducing a multi-layer nonlinear transformation of the LeakyReLU activation function and embedding the Dropout mechanism, the enhanced 79-dimensional input vector discrete signal is transformed into a 400-point feature sequence with a preliminary physical form.

[0124] In this embodiment, the slope of the introduced LeakyReLU activation function is fixed at 0.1; the dropout rate of the embedded Dropout mechanism is set to 10%.

[0125] 2) The deep one-dimensional residual 1D-ResNet decoding module adopts a cascaded structure of feature expansion-residual purification-feature compression, and uses an edge copying and padding strategy in all convolution operations to achieve spectral smoothing without losing physical information.

[0126] In this embodiment, three consecutive nested one-dimensional residual blocks are designed for deep purification, wherein the dropout rate of the third residual block is increased to 15%, and residual system thermal noise and time-varying fluctuations of electron cyclotron radiation are stripped away layer by layer, and nonlinear interpolation is used to repair the power pit caused by the notch filter.

[0127] The loss function design adopts a frequency-adaptive weighting strategy of blind zone exemption and bilateral weighting, and integrates smoothing L1 error, first-order total variation regularization and second-order gradient consistency constraint to construct a composite loss function with multi-objective joint optimization to achieve the purpose of spectral restoration.

[0128] The parameter update of the training strategy model adopts the Adam optimizer and uses a cosine annealing hot restart strategy to schedule the learning rate to improve convergence efficiency and avoid getting trapped in local optima.

[0129] In this embodiment, the Adam optimizer is set to an initial learning rate of 0.001 and a weight decay coefficient of 10⁻⁵.

[0130] The reconstructed spectral line evaluation uses a validation set that is completely identical to that used in the training phase, and adopts formula (1) as the core evaluation metric:

[0131] (1)

[0132] Where Rel-RMSE represents the relative root mean square error; N is the number of validation samples, and i is an integer value between 1 and N. The peak value of the true spectral line. To reconstruct spectral values, These are the true spectral line values;

[0133] After calculating the Rel-RMSE of all validation samples, the average error, median error, worst and best sample errors are output to reflect the model performance from the perspective of overall distribution.

[0134] In the embodiments, the linear domain comparison diagrams and logarithmic domain comparison diagrams of the cleaned random verification samples of the ITER and KSTAR coherent Thomson scattering power spectra are respectively attached as follows. Figure 9 Appendix Figure 10 As shown in Table 2, the mean error, median error, worst and best sample errors of the two sets of coherent Thomson scattering power spectra are as follows.

[0135] Table 2. Mean error, median error, worst and best sample errors of the cleaned verification samples of coherent Thomson scattering power spectrum.

[0136]

[0137] Table 2 shows the error statistics of the coherent Thomson scattering spectra of the two thermonuclear fusion devices on the independent validation set. The results indicate that the overall reconstruction accuracy of the coherent Thomson scattering diagnostic signal receiving and processing method designed in this invention is good; the median error is close to the average error, and the model has good robustness.

[0138] In the design of constructing coherent Thomson scattering noisy power spectrum datasets for the ITER and KSTAR thermonuclear fusion devices, this invention introduces coherent Thomson scattering pure power spectrum calculation code based on a fully electromagnetic kinetic model. Referring to the typical experimental or design parameters of each device listed in Table 1, and batch modifying some structural characteristic parameters, the resulting datasets are generated as shown in Table 1. Figure 4 The electromagnetic calculation results of the coherent Thomson scattering pure power spectrum shown intuitively present the important structural characteristics of the coherent Thomson scattering pure power spectrum corresponding to the two different fusion devices: In terms of frequency band distribution, ITER uses a fundamental detection frequency around 60 GHz, while KSTAR is located in the band around 105 GHz; in terms of physical morphology, most of the ITER power spectrum exhibits a typical bimodal structure dominated by ion acoustic resonance, while most of the KSTAR power spectrum shows a broad single peak, reflecting more the Doppler broadening effect of thermal motion under specific observation conditions; in terms of signal intensity, the coherent Thomson scattering power spectrum of the ITER group is relatively strong, with peak values ​​reaching tens or even hundreds of electron volts, while the KSTAR group is relatively weak, with peak values ​​generally at the level of a few electron volts; these differences provide important basis for subsequent amplification front-end design and component selection.

[0139] The overall design block diagram of the ultra-low noise amplifier front-end hardware circuit design in this embodiment of the invention is shown in the attached figure. Figure 5 As shown, further independent parameter matching design of key components can be performed to address the differences in detection center frequency, signal strength, and background noise levels between the ITER and KSTAR devices. (See attached image) Figure 6 Appendix Figure 7 The figure shows a random sampling example of the coherent Thomson scattering power spectrum processing results of the two sets of devices corresponding to the ultra-low noise amplification front end obtained by simulation. It shows that the design has successfully achieved the expected goals: filtering out most of the in-situ noise such as electron cyclotron radiation and hardware ripple, and removing the central spurious peak; the signal-to-noise ratio of each set of coherent Thomson scattering power spectra has been significantly improved, reaching 8~20dB; while ensuring the working stability and safety of each component, the gain allocation has been completed, making the best use of the ideal working range of each component; the conversion of the power spectrum to a 32-dimensional discrete feature point set and the conversion of the analog domain to the digital domain have been realized, and the construction of the input end of the artificial intelligence neural network has been completed.

[0140] The schematic diagram of the one-dimensional residual convolutional deep learning network structure based on the global mapping-residual decoding architecture designed in the artificial intelligence neural network design for spectral cleaning in this embodiment of the invention is shown in the attached figure. Figure 8 As shown in the attached figure, the weighted mask is further fine-tuned using the designed artificial intelligence neural network based on the characteristics of the ITER and KSTAR devices. Figure 9 Appendix Figure 10 The logarithmic and linear domain comparison diagrams of the random verification samples of coherent Thomson scattering power spectrum cleaning are shown in Table 2. Based on these, the average error, median error, worst-case error, and best-case error statistics of the corresponding coherent Thomson scattering power spectrum cleaning verification samples are calculated. Simulation results show that the average relative root mean square errors of the two sets of devices in this spectral cleaning network are approximately 5.42% and 4.67%, respectively. The median error is highly consistent with the average error, demonstrating good robustness. Even in the central notch blind zone, the network can still achieve smooth and reasonable extrapolation reconstruction based on the dispersion relationship of the physical boundaries on both sides. It exhibits good effectiveness and adaptability in strong stray noise environments, solving the failure problem of traditional methods in such scenarios and providing a feasible signal processing scheme for high-precision online analysis of coherent Thomson scattering diagnostic systems for future fusion devices.

[0141] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0142] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for receiving and processing coherent Thomson scattering diagnostic signals, characterized in that, The method includes the following steps: S1. Using the coherent Thomson scattering pure power spectrum calculation code based on the all-electromagnetic kinetic model, and considering the typical operating parameters of thermonuclear fusion devices, a dataset containing 2500 sets of high-fidelity pure power spectrum data is generated. S2. Inject four types of physical noise with non-stationary and time-varying characteristics: electron cyclotron radiation background fluctuations, random bursts of parametric decay instability narrowband spikes, incident wave stray radiation leakage, and statistical thermal fluctuations that obey the radiometer equation. Construct a coherent Thomson scattering band-noise power spectrum dataset with 10,000 data points per group to reconstruct the real diagnostic environment of the thermonuclear fusion device described in S1.

2. The method for receiving and processing coherent Thomson scattering diagnostic signals as described in claim 1, characterized in that, In S2, the background noise of electron cyclotron radiation is subjected to four-fold superposition processing, including: secondary base profile to simulate basic energy distribution, linear baseline drift to simulate global baseline tilt, periodic low-frequency fluctuations to simulate frequency domain fluctuations caused by magnetohydrodynamic instability or density perturbation, and low-pass smooth Gaussian random envelopes with a relative amplitude of 0.5% being independently superimposed in the ON and OFF frames during rapid time-varying fluctuations.

3. The method for receiving and processing coherent Thomson scattering diagnostic signals according to claim 1, characterized in that, In step S1, generating a dataset containing 2500 sets of high-fidelity, clean power spectrum data includes: Based on the operating parameter range of the thermonuclear fusion device, the plasma electron temperature, ion temperature, electron density, ion velocity distribution function parameters, scattering angle, and incident microwave frequency are sampled in layers. Input the operating parameters obtained from each set of samples into the coherent Thomson scattering pure power spectrum calculation code based on the all-electromagnetic kinetic model to obtain the corresponding coherent Thomson scattering pure power spectrum. The coherent Thomson scattering pure power spectrum is labeled with peak position, linewidth, peak intensity and line broadening shape. The operating parameters, coherent Thomson scattering pure power spectrum and corresponding labeling information are associated and stored to form a pure power spectrum dataset with traceable physical parameters.

4. The method for receiving and processing coherent Thomson scattering diagnostic signals according to claim 1, characterized in that, In S2, constructing the noisy coherent Thomson scattering power spectrum dataset includes: The noise components corresponding to the background fluctuations of electron cyclotron radiation, the narrowband spikes of parametric decay instability, the incident wave stray radiation leakage, and the statistical thermal fluctuations are generated respectively. The noise components are superimposed onto the coherent Thomson scattering pure power spectrum data in two ways: single noise injection and multi-noise coupled injection. Record the noise type, noise intensity, frequency position, duration, and injection timing parameters for each noise component; A mapping relationship is established between the pure power spectrum of coherent Thomson scattering before injection, the power spectrum after single noise injection, the power spectrum after multi-noise coupled injection, and the corresponding noise parameters to form a dataset of noisy power spectrum of coherent Thomson scattering with interpretable noise sources.

5. An ultra-low noise amplification front-end hardware circuit structure comprising four core functional components, implementing the coherent Thomson scattering diagnostic signal receiving and processing method as described in any one of claims 1-4, wherein the design method includes: (1) Front-end protection design: A notch filter with extremely high suppression is deployed at the front end of the radio frequency receiving link to filter out the central spike. (2) High-frequency amplification and isolation design: First, the extremely weak signal output from (1) functional stage is input into the high-frequency low-noise amplifier to achieve power boost and avoid the effective signal being submerged by noise. Then, a ferrite microwave isolator is introduced to achieve unidirectional electromagnetic wave conduction and avoid ripple misjudgment caused by output return loss. (3) The mixing and intermediate frequency conditioning stage design: the front part introduces a local oscillator with extremely low phase noise and a broadband microwave mixer to realize the efficient power spectrum shift of the micro-watt-level high-frequency scattered signal to the intermediate frequency band; the middle part uses an intermediate frequency amplifier to realize the last stage of active conditioning and gain matching of the analog signal before entering the analog-to-digital conversion; the rear part is connected to a microwave attenuator to adjust the signal power level and optimize the impedance matching characteristics of the cascaded circuit. (4) The channelization and detection quantization stage design is as follows: the front part is set with a high-order anti-aliasing 32-channel filter bank, the middle part uses a logarithmic detector to accurately convert the RF power fluctuations of the front-end microwave band into low-frequency video baseband voltage signals, and the rear part uses a high-performance 24-bit synchronous sampling analog-to-digital converter to complete the synchronous high-precision acquisition of 32 analog baseband signals in the same nanosecond time window, and sends them to the back-end digital signal processor for storage and operation.

6. The design method according to claim 5, characterized in that, In (2), the high-frequency amplifier has strict requirements of high gain and low noise, and the selection of its device model needs to be carried out through full-link gain allocation and component query comparison; In (3), the mixing and intermediate frequency conditioning stage includes: S31. The microwave mixer accurately shifts the high-frequency microwave signal downwards to its respective standard intermediate frequency range according to the different power spectra of the high-frequency microwave signal, falling into the optimal sampling aperture of the broadband high-speed analog-to-digital converter described in (4). S32. The intermediate frequency amplifier needs to be set with different target gain and specific component selection based on different power spectrum intensity characteristics; S33. One microwave attenuator is placed before and after the intermediate frequency amplifier described in S32. The signal in the transmission line undergoes bidirectional attenuation and is weakened by a total of 6dB.

7. The design method according to claim 5, characterized in that, In (4), the channelization and detection quantization steps include: The S41 anti-aliasing 32-channel filter bank adopts a highly integrated channelized architecture, which divides the effective intermediate frequency bandwidth into 32 independent physical channels in an equal proportion, so that the effective detection bandwidth of a single channel is divided into 32 equally divided fine intervals. Each channel unit uses a 5th order Chebyshev type I bandpass filter. S42. After receiving the voltage acquisition data from the analog-to-digital converter module, the digital signal processor performs physical inverse mapping to perform exponential operation on the 24-bit quantized voltage value to restore it to the linear power domain value; and performs ON-OFF digital cancellation algorithm to remove most of the electron cyclotron radiation background and hardware standing wave ripple noise.

8. A backend intelligent spectral cleaning method based on a one-dimensional residual convolutional deep learning network, implementing the method as described in any one of claims 1-4, comprising: Dataset construction and physical constraint cleanup mechanism, hybrid neural network architecture design, loss function design, training strategy and process, as well as reconstruction spectral line evaluation and post-processing process; The dataset construction and physical constraint cleanup mechanism is used to construct a highly robust training dataset through physical criteria interception and logarithmic domain transformation. The hybrid neural network architecture is designed to enhance the original discrete power spectrum information with severe sparsity defects output by the front-end circuit module of the second aspect of the present invention with higher-dimensional feature data and smooth spectral evolution. The loss function is designed to integrate frequency adaptive weighting and smoothness regularization in the logarithmic domain, enabling the network to achieve high-fidelity extrapolation in the notch region. The training strategy and process are used to fit the logarithmic spectrum of the training data, integrating dynamic noise injection, periodic learning rate restart and implicit early stopping mechanism, so that the model avoids getting trapped in local optima and gradually converges to the optimal weights within 300 training cycles. The reconstructed spectral line evaluation and post-processing process is used to restore the normalized prediction in the logarithmic domain to the absolute power spectrum with physical units, and to realize the quantitative evaluation of the reconstruction accuracy of the model on an independent validation set. The hybrid neural network architecture design includes: 1) A global feature mapping module with a progressively expanded fully connected structure is adopted. By introducing a multi-layer nonlinear transformation of the LeakyReLU activation function and embedding the Dropout mechanism, the enhanced 79-dimensional input vector discrete signal is transformed into a 400-point feature sequence with a preliminary physical form. 2) The deep one-dimensional residual 1D-ResNet decoding module adopts a cascaded structure of feature expansion-residual purification-feature compression, and uses an edge copying and padding strategy in all convolution operations to achieve spectral smoothing without losing physical information; The loss function design adopts a frequency-adaptive weighting strategy of blind zone exemption and bilateral weighting, and integrates smoothing L1 error, first-order total variation regularization and second-order gradient consistency constraint to construct a composite loss function with multi-objective joint optimization to achieve the purpose of spectral restoration. The parameter update of the training strategy model adopts the Adam optimizer and uses a cosine annealing hot restart strategy to schedule the learning rate to improve convergence efficiency and avoid getting trapped in local optima. The reconstructed spectral line evaluation uses a validation set that is completely identical to that used in the training phase, and adopts formula (1) as the core evaluation metric: （1） Where Rel-RMSE represents the relative root mean square error; N is the number of validation samples, and i is an integer value between 1 and N. The peak value of the true spectral line. To reconstruct spectral values, These are the true spectral line values; After calculating the Rel-RMSE of all validation samples, the average error, median error, worst and best sample errors are output to reflect the model performance from the perspective of overall distribution.

9. A coherent Thomson scattering diagnostic signal receiving and processing system implementing the coherent Thomson scattering diagnostic signal receiving and processing method as described in any one of claims 1-4, characterized in that, The coherent Thomson scattering diagnostic signal receiving and processing system includes: The module is used to import the code for calculating the coherent Thomson scattering pure power spectrum based on the full electromagnetic kinetic model, and to generate a dataset containing 2,500 sets of high-fidelity pure power spectrum data, taking into account the typical operating parameters of thermonuclear fusion devices. The injection module is used to inject four types of physical noise with non-stationary and time-varying characteristics: electron cyclotron radiation background fluctuations, random bursts of parametric decay instability narrowband spikes, incident wave stray radiation leakage, and statistical thermal fluctuations that obey the radiometer equation. This constructs a coherent Thomson scattering band-noise power spectrum dataset with 10,000 data points per group, which can reconstruct the real diagnostic environment of the thermonuclear fusion device described in S1.

10. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the coherent Thomson scattering diagnostic signal receiving and processing system as described in claim 9.

Images