A method and system for detecting the operating status of a controlled nuclear fusion device

Abstract

This invention proposes a method and system for detecting the operational status of a controlled fusion device, belonging to the field of controlled fusion device operational status detection technology. The method includes: obtaining a pulse height spectrum based on measurement results from a liquid scintillator detector; normalizing the count values ​​for each channel address; inputting the normalized pulse height spectrum data into a Bayesian neural network model for forward propagation calculation. The training process includes: constructing a neutron response matrix by combining Monte Carlo simulation and experimental calibration; generating a set of simulated neutron energy spectra based on preset basis functions; generating a set of simulated pulse height spectra by convolving the neutron response matrix with the set of simulated neutron energy spectra; training a neural network model with weight parameters following a probability distribution using Bayesian inference based on data pairs; and performing an inverse transform on the model output to obtain the incident neutron energy spectrum distribution. This invention improves the stability and accuracy of neutron energy spectrum interpretation and provides quantitative information on the uncertainty of the inversion results.

Description

Technical Field

[0001] This invention relates to the field of controlled nuclear fusion device operation status detection technology, and in particular to a method and system for detecting the operation status of a controlled nuclear fusion device. Background Technology

[0002] Neutron spectrum is one of the key physical quantities for diagnosing the internal plasma behavior of controlled fusion devices, assessing the load on the first wall material, and monitoring fusion power. Liquid scintillator detectors are widely used in neutron spectrum measurements in fusion devices due to their strong ability to distinguish neutrons from gamma rays and their high detection efficiency. However, retrieving the true neutron spectrum from the pulse height spectrum measured by the detector (i.e., "spectral decomposition") is a typical ill-posed inverse problem, inherently difficult due to the non-uniqueness of the solution and sensitivity to measurement noise. Traditional spectral decomposition methods, such as iterative deconvolution, often heavily rely on initial guesses and manually adjusted parameters, resulting in unstable solutions and difficulty in quantifying the uncertainty of the inversion results. In recent years, spectral decomposition methods based on deterministic weighted neural networks have improved the solution efficiency to some extent, but their output is a single-point estimate, failing to provide a reliability assessment of the spectral decomposition results and thus failing to meet the urgent need for uncertainty information for accurate diagnosis and safety decision-making in fusion devices.

[0003] Therefore, there is an urgent need for a detection method that can stably and accurately invert the neutron energy spectrum from the pulse height spectrum of a neutron detector and provide a reliable quantitative assessment of the uncertainty of the inversion results. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, the first aspect of this invention proposes a method for detecting the operating status of a controlled nuclear fusion device, comprising:

[0005] S1: Obtain the pulse height spectrum based on the measurement results of the liquid scintillator detector deployed at the site of the controlled fusion device;

[0006] S2: Based on the pulse height spectrum, the count values ​​of each channel address are normalized to generate normalized pulse height spectrum data;

[0007] S3: Input the normalized pulse height spectrum data into a pre-trained Bayesian neural network model for forward propagation calculation. The network weight parameters of the Bayesian neural network model are represented in the form of a probability distribution. The construction and training of the Bayesian neural network model includes:

[0008] Sa: Based on the physical characteristics of the liquid scintillator detector, the neutron response matrix is ​​constructed by combining Monte Carlo simulation with experimental calibration.

[0009] Sb: Based on a variety of preset basis function neutron energy spectra, a set of simulated neutron energy spectra is generated through random combination and parameter adjustment;

[0010] Sc: Based on the simulated neutron energy spectrum set and the neutron response matrix, a convolution operation is performed to generate the corresponding simulated pulse height spectrum set;

[0011] Sd: Based on the training data pair consisting of the simulated pulse height spectrum set and the simulated neutron energy spectrum set, the initialized neural network is trained using the Bayesian inference method to obtain a Bayesian neural network model in which the network weight parameters follow a probability distribution.

[0012] S4: Perform inverse transformation on the output of the forward propagation calculation of the Bayesian neural network model to obtain the incident neutron energy spectrum distribution.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0014] Through the synergistic effect of steps S1 to S4, the shortcomings of traditional spectral interpretation methods are effectively overcome, achieving high-precision inversion of the incident neutron energy spectrum with uncertainty information. First, step S1 directly acquires the pulse height spectrum based on in-situ measurements of the liquid scintillator detector, ensuring the authenticity and timeliness of the input data. Subsequently, step S2 normalizes the pulse height spectrum, eliminating the influence of absolute count rate fluctuations, making the data format suitable for the input requirements of subsequent models, and improving the robustness of the method. The core and significant effect of the method stems from step S3 and the process of constructing and training the Bayesian neural network model. Specifically, step Sa constructs a neutron response matrix by combining Monte Carlo simulation with experimental calibration, accurately characterizing the physical response characteristics of the detector to neutron energy, laying the physical foundation for generating realistic training data. Step Sb utilizes various preset basis function neutron energy spectra to generate a set of simulated neutron energy spectra through random combination and parameter adjustment, greatly enriching the diversity of training samples and covering a wide range of potential energy spectrum forms. Step Sc generates a corresponding set of simulated pulse height spectra based on the convolution of the simulated neutron energy spectrum set and the neutron response matrix, thus constructing a large number of paired "energy spectrum-pulse height spectrum" training data pairs. Step Sd trains the neural network using Bayesian inference, representing the network weight parameters in the form of a probability distribution. This key design enables the final Bayesian neural network model to not only learn the complex mapping relationship from the pulse height spectrum to the neutron energy spectrum, but also intrinsically encapsulates the uncertainties caused by the uncertainty of the training data and the model structure. Therefore, in the forward propagation calculation of step S3, the model output is no longer a deterministic value, but a predicted distribution considering the uncertainty. Finally, step S4 directly obtains the incident neutron energy spectrum distribution and its uncertainty information by performing an inverse transformation on the output. The entire process constructs training data based on physical principles, solves the problem using a probabilistic deep learning model, and finally outputs an inversion result with reliable uncertainty quantification, thus providing key technical support for accurately and reliably diagnosing the operating status of controlled nuclear fusion devices. Attached Figure Description

[0015] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 The diagram shown is a flowchart illustrating a method for detecting the operating status of a controllable fusion device according to an embodiment of the present invention.

[0017] Figure 2The diagram shown is a structural schematic of a controllable fusion device operation status detection system provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0019] The specific embodiments of the present invention will be described below.

[0020] Example 1

[0021] like Figure 1 As shown, the first aspect of the present invention proposes a method for detecting the operating status of a controllable fusion device, comprising:

[0022] S1: Obtain the pulse height spectrum based on the measurement results of the liquid scintillator detector deployed at the site of the controlled fusion device;

[0023] S2: Based on the pulse height spectrum, the count values ​​of each channel address are normalized to generate normalized pulse height spectrum data;

[0024] S3: Input the normalized pulse height spectrum data into a pre-trained Bayesian neural network model for forward propagation calculation. The network weight parameters of the Bayesian neural network model are represented in the form of a probability distribution. The construction and training of the Bayesian neural network model includes:

[0025] Sa: Based on the physical characteristics of the liquid scintillator detector, the neutron response matrix is ​​constructed by combining Monte Carlo simulation with experimental calibration.

[0026] Sb: Based on a variety of preset basis function neutron energy spectra, a set of simulated neutron energy spectra is generated through random combination and parameter adjustment;

[0027] Sc: Based on the simulated neutron energy spectrum set and the neutron response matrix, a convolution operation is performed to generate the corresponding simulated pulse height spectrum set;

[0028] Sd: Based on the training data pair consisting of the simulated pulse height spectrum set and the simulated neutron energy spectrum set, the initialized neural network is trained using the Bayesian inference method to obtain a Bayesian neural network model in which the network weight parameters follow a probability distribution.

[0029] S4: Perform inverse transformation on the output of the forward propagation calculation of the Bayesian neural network model to obtain the incident neutron energy spectrum distribution.

[0030] This method aims to solve the technical problem of stably and accurately retrieving the incident neutron energy spectrum from detector measurement signals and quantifying the uncertainty of the inversion results. It begins with data acquisition, specifically step S1: obtaining the pulse height spectrum based on the measurement results of a liquid scintillator detector deployed at the controlled fusion device site. A liquid scintillator detector is a radiation detection device that uses organic or inorganic scintillating materials to convert the energy of incident particles into light signals, which are then converted into electrical pulses by devices such as photomultiplier tubes. The pulse height spectrum obtained in real-time is a distribution graph with the "channel address" as the x-axis and the "count value" of the pulse signal recorded at that channel address as the y-axis. The channel address corresponds to the amplitude of the pulse signal, which is related to the energy deposited by the particle in the detector; the count value reflects the number of signals of that amplitude recorded. This pulse height spectrum directly reflects the energy deposited by secondary particles generated after the interaction between neutrons and the detector material, forming the original data basis for all subsequent analyses and ensuring the real-time and in-situ nature of state detection.

[0031] After obtaining the raw pulse height spectrum, normalization processing in step S2 is required. This step first calculates the sum of the counts for all channels to obtain the total count. Then, the raw count for each channel is divided by this total count to obtain the normalized count for each channel. This process no longer generates raw counts, but rather normalized pulse height spectrum data, whose physical meaning is transformed into the proportion or probability of each channel's count relative to the total count. The core effect of normalization is to eliminate the influence of absolute flux or measurement time differences on the data. For example, when the neutron yield increases, the total count increases, but the proportion of each channel's count may remain stable. By uniformly transforming the data into a probability distribution, subsequent models can focus on the shape characteristics of the energy spectrum itself, rather than the overall strength of the signal, greatly improving the adaptability and robustness of the method to measurements of neutron sources with different intensities, creating conditions for the model's general application.

[0032] The core and significant advancement of this method lies in step S3: inputting normalized pulse height spectrum data into a pre-trained Bayesian neural network model for forward propagation calculation. The key here is the "Bayesian neural network model" and its "network weight parameters represented in the form of a probability distribution." In ordinary neural networks, the weights are fixed values ​​after training, while Bayesian neural networks treat each weight as a random variable, describing its possible values ​​using a probability distribution (such as a Gaussian distribution). This probabilistic representation implies that the model itself acknowledges the uncertainty of its parameters. The construction and training of this model is a carefully designed process, starting with step Sa: constructing the neutron response matrix. The neutron response matrix is ​​a mathematical transformation matrix, where rows typically represent the detector's output address and columns represent the energy of the incident neutron. Each element in the matrix quantitatively describes the probability that a signal will be generated and fall on a specific address when a unit flux of neutrons of a specific energy is incident. Building it requires combining an accurate physical model of the detector (achieved through Monte Carlo simulation) with the performance calibration of the actual detector (achieved through experimental calibration) to ensure that the matrix reflects both the complex physical processes and the response characteristics of the real device.

[0033] After obtaining an accurate detector response model, step Sb focuses on generating a large and diverse set of simulated neutron energy spectra as the "standard answer" for training. This is achieved by pre-setting a set of mathematical functions (basis function neutron energy spectra, such as Gaussian peaks, exponential decay spectra, etc.) that can describe typical energy spectrum shapes, and then randomly combining and intensifying these basis functions, and adding random energy shifts and broadening. This randomized generation method can systematically create energy spectra covering various possible forms, simulating the diversity of neutron energy spectra that real fusion devices may produce under different operating conditions, providing rich learning samples for the model. Next, step Sc uses the neutron response matrix obtained in step Sa to perform a convolution operation with each simulated neutron energy spectrum generated in step Sb. From a mathematical physics perspective, this simulates the pulse height spectrum that should theoretically be produced when this hypothetical neutron energy spectrum is measured by a real liquid scintillator detector. Through this step, we obtain "simulated neutron energy spectra" and their corresponding "simulated pulse height spectra" in pairs, forming a massive amount of training data pairs.

[0034] With these data pairs, step Sd uses Bayesian inference to train the neural network. Training does not involve finding an optimal set of fixed weights, but rather inferring the posterior probability distribution of all weight parameters. This process begins by setting a prior distribution (e.g., a Gaussian distribution with a mean of 0) for all weights, representing the initial belief before training. Then, by inputting a simulated pulse height spectrum into the network, its predicted energy spectrum is calculated and compared with the actual simulated neutron energy spectrum in the data pair, constructing a likelihood function to measure the probability of data occurrence under the current weights. Finally, using Bayes' theorem, the prior distribution and the likelihood function are combined, and the posterior distribution of the weights is inferred through computation (e.g., Markov chain Monte Carlo sampling). After training, the weights of the resulting model are no longer numbers, but a distribution. Therefore, in the practical application stage of step S3, when a real normalized pulse height spectrum is input, the model performs forward propagation calculations. Because the weights are probabilistic, its output is no longer a deterministic neutron energy spectrum curve, but a distribution of the energy spectrum, which naturally contains inversion uncertainty information caused by model uncertainty and the ill-posedness of the problem itself.

[0035] Finally, step S4 performs an inverse transformation on the output of the model's forward propagation calculation to obtain a physically interpretable incident neutron energy spectrum distribution. The inverse transformation may include restoring the normalized values ​​of the model output to absolute flux based on the total count, or converting the output vector indices into specific energy values. Due to the probabilistic nature of the model, the final output incident neutron energy spectrum distribution typically represents an optimal estimation curve (e.g., the posterior mean spectrum) and a confidence interval around that curve (e.g., the 95% confidence band). This creates a closed loop in the entire process: starting from the original measurement data, a probabilistic deep learning model is trained using massive amounts of simulated data generated based on a precise physical model. This model not only provides the optimal estimate of the neutron energy spectrum but also simultaneously offers a quantifiable measure of the confidence level of that estimate. This solves the fundamental problem of unstable solutions and the inability to assess the reliability of results in traditional methods, elevating the diagnosis of fusion device operation status from simply "having results" to "having reliable results," providing more valuable decision-making basis for safe operation and physical analysis.

[0036] In some implementations, Sa includes:

[0037] Sa1: Based on the standard radioactive source, the liquid scintillator detector is calibrated with energy to obtain the correspondence between the channel address and the deposited energy;

[0038] Sa2: Based on Monte Carlo simulation software, the energy deposition process of monoenergetic neutrons in a liquid scintillator detector was simulated to obtain the energy deposition spectrum;

[0039] Sa3: Based on the correspondence between channel location and deposition energy and the energy deposition spectrum, the neutron response matrix is ​​calculated and arranged.

[0040] This embodiment elaborates on step Sa, which is fundamental to ensuring the physical accuracy of the entire method. Step Sa1 is the experimental calibration step, aiming to establish a quantitative relationship between the detector output signal address and the deposition energy. In practice, a standard radioactive source with known energy is required, such as a radioactive isotope emitting gamma rays at a specific energy (e.g., Cs-137, Co-60). This source is used to irradiate the liquid scintillator detector, and the pulse signal output by the detector is collected to form an energy spectrum. Since the energy of the particles emitted by the standard source is known, the peak position of the pulse amplitude (i.e., the address) corresponding to the light output produced by it in the detector is also measurable. By measuring the peak position addresses produced by multiple standard sources with different energies, a calibration curve or functional relationship between the address value and the deposition energy can be fitted. This step is crucial because it accurately correlates the detector's electronic readings (address) with the physical process (deposition energy), providing a benchmark for subsequently mapping the simulated energy information to the actual address response.

[0041] Step Sa2 involves simulation calculations based on physical principles, aiming to obtain the energy deposition response of monoenergetic neutrons in the detector. This requires the use of Monte Carlo simulation software such as Geant4, MCNP, or FLUKA. First, an accurate geometric model is established in the software based on the actual size, structure, and material composition of the liquid scintillator detector. Then, a list of physical processes is set, including cross-sectional data of various reactions between neutrons and the atomic nuclei of the detector material. During the simulation, the neutron source is set as a monoenergetic, unidirectional, or isotropic parallel beam, with energy values ​​ranging from low to high, covering the range of interest (e.g., from thermal energy to tens of MeV). For each set monoenergetic energy E, a large number (e.g., millions) of neutron incident events are simulated, tracking the transport process of each neutron and its induced secondary particles (such as recoil protons, alpha particles, etc.) in the scintillator until the energy is completely deposited or the particles escape. The total energy deposited in the scintillator in each event is recorded, and the energy deposited in all events is statistically analyzed and plotted to obtain the energy deposition spectrum corresponding to that monoenergetic neutron. The spectrum shows a distribution rather than a vertical line because the process of neutrons producing secondary particles through nuclear reactions is random, and the range and energy loss of secondary particles fluctuate.

[0042] Step Sa3 is the calculation process of synthesizing the complete neutron response matrix by integrating the information from the first two steps. First, using the channel-energy scale relationship obtained in step Sa1, the energy deposition spectrum simulated in step Sa2, with energy (MeV) as the abscissa, is transformed to a scale with channel address as the abscissa. This is usually achieved through interpolation or resampling. For a simulated monoenergetic neutron energy Ei, its energy deposition spectrum, after transformation, becomes a vector Ri. The length of vector Ri is equal to the total number of channels, and the value of its j-th element R_ij represents the probability (or normalized count) that the signal generated by the deposition energy event of a neutron with energy Ei falls on channel j. Then, this process is repeated for all covered monoenergetic neutron energy points, resulting in a series of vectors R1, R2, ..., Rn. Finally, these vectors are arranged into a matrix in energy order, where the i-th column is the vector Ri. This matrix is ​​the neutron response matrix R, and the matrix element R_ij is the response value. By combining "experimental calibration to determine the energy-channel correspondence" with "Monte Carlo simulation to determine the monoenergetic neutron-energy deposition relationship", the constructed neutron response matrix is ​​both faithful to the complex physical response mechanism of the detector and closely matches the actual response characteristics of the actual detector. Its accuracy directly determines the physical reliability of the subsequently generated training data and the final inversion results, and is the physical foundation upon which the entire method is established.

[0043] In some implementations, Sb includes:

[0044] Sb1: Provides a basis function library containing Gaussian peak spectra and exponential decay spectra;

[0045] Sb2: Based on random numbers, select at least one basis function from the basis function library and assign a random strength coefficient to each selected basis function;

[0046] Sb3: Based on the selected basis functions and their assigned intensity coefficients, a weighted summation operation is performed to generate a simulated neutron energy spectrum; during the weighted summation operation, a random energy shift and broadening are applied to the summation result to increase the diversity of the simulated neutron energy spectrum set;

[0047] Sb4: Repeat Sb2 to Sb3 to generate a set of simulated neutron energy spectra containing a preset number of spectra.

[0048] This embodiment refines step Sb of generating a set of simulated neutron energy spectra. The goal of this step is to create a sufficiently diverse "answer library" of training samples. Step Sb1 first establishes a basis function library. Basis functions here refer to the basic mathematical function forms used to construct complex neutron energy spectra. Based on the possible characteristics of fusion neutron energy spectra, the library typically includes Gaussian peak spectra and exponential decay spectra. The mathematical form of a Gaussian peak spectrum resembles a bell curve and can be used to simulate monoenergetic neutron peaks or characteristic energy peaks produced by nuclear reactions; exponential decay spectra exhibit a rapidly decreasing count rate with increasing energy and can be used to simulate the continuous background of high-energy regions or certain fission spectrum features. These basis functions constitute the "basic components" for piecing together various complex energy spectrum shapes, and their selection is based on the physical understanding of the neutron production mechanisms in fusion plasmas, such as the 14 MeV neutron peak from deuterium-tritium reactions and broad-spectrum neutrons from various nuclear reactions.

[0049] Step Sb2 is crucial for introducing randomness to generate different energy spectrum combinations. When generating each simulated neutron spectrum, a pseudo-random number generator determines the combination method. First, one or more basis functions are randomly selected from the basis function library. For example, only a Gaussian function might be selected, or both a Gaussian function and an exponential function might be chosen simultaneously. Next, an intensity coefficient is independently and randomly assigned to each selected basis function. This coefficient is a positive real number that determines the contribution weight or magnitude of that basis function in the final superimposed energy spectrum. This random selection and weighting method avoids manually pre-setting fixed energy spectrum patterns and can automatically explore various possible combinations and relative strengths between basis functions, laying the foundation for generating diverse energy spectra.

[0050] Step Sb3 performs the specific energy spectrum synthesis and enhancement. Its core operation involves weighted summation of the basis functions selected in step Sb2 according to their assigned random intensity coefficients. If multiple basis functions are selected, the summation result is a composite energy spectrum containing multiple characteristic components. However, to further approximate the complexity of the energy spectrum measured in actual experiments, two additional random perturbations need to be applied to the summation result after the weighted summation: random energy shift and random broadening. Energy shift refers to shifting the entire energy spectrum curve to the left or right along the energy axis by a small random amount, simulating small systematic biases or drifts that may exist in the energy scale. Broadening is achieved by convolving the energy spectrum curve with a Gaussian function with a certain width (such as random variance), simulating changes in detector energy resolution or physical broadening effects. These perturbation operations significantly enhance the diversity of the simulated neutron energy spectrum set, so that the generated training data not only includes ideal spectral shapes, but also various near-realistic, imperfect, and distorted spectral shapes, greatly expanding the coverage of the training data.

[0051] Step Sb4 generates data in batches through a loop mechanism. Steps Sb2 and Sb3 are repeated thousands or even millions of times, with each loop generating a unique simulated neutron spectrum based on different random numbers. All these spectra are aggregated to form a vast set of simulated neutron spectra. This set, as a whole, should have statistical properties that comprehensively cover the neutron spectrum morphology space generated by the target controlled fusion device under all possible operating conditions. This systematic data generation strategy based on random combination and perturbation provides a near-infinite and diverse set of learning paradigms for subsequent Bayesian neural networks. This ensures that the trained model does not merely memorize a few specific spectral shapes, but truly learns to extract essential features related to the spectrum shape from the pulse height spectrum. Thus, when faced with unknown or potentially unprecedented spectrum morphologies in actual measurements, it still exhibits strong generalization ability and robust inversion performance, which is a key guarantee for achieving practical and robust state detection.

[0052] In some implementations, Bayesian inference methods are used to train the initialized neural network in Sd, including:

[0053] Sd1: Sets the prior probability distribution for the weight parameters of the initialized neural network;

[0054] Sd2: Input the simulated pulse height spectrum from the training data pair into the initialized neural network to calculate the predicted neutron energy spectrum;

[0055] Sd3: Calculate the likelihood function based on the predicted neutron energy spectrum and the simulated neutron energy spectrum corresponding to the training data pair;

[0056] Sd4: Based on the prior probability distribution and likelihood function, the posterior probability distribution of the weight parameters is inferred using the Markov chain Monte Carlo sampling method;

[0057] Sd5: Based on the posterior probability distribution obtained from sampling, determine the final probabilistic weight parameters of the Bayesian neural network model.

[0058] This embodiment elaborates on the steps Sd of training a neural network using the Bayesian inference method, which is the core computational process for achieving the probabilistic output of the model. Step Sd1 is to set the prior probability distribution. Before training begins, a prior distribution p(w) needs to be specified for each weight parameter and bias parameter (collectively referred to as weight parameter w) to be learned in the neural network. A common and general choice is the standard normal distribution (mean of 0, variance of 1). This choice expresses a "prior belief": before observing any data, we believe that the weight parameters are more likely to take values ​​close to zero, which corresponds to a simple model with an initial output close to zero. Although the choice of prior distribution has a certain degree of subjectivity, it provides a reasonable starting point and regularization constraint for subsequent Bayesian updates, helping to prevent the model from overfitting when data is limited.

[0059] Steps Sd2 and Sd3 constitute the data evidence injection process during training iterations. During training, a small batch of samples is taken from the data pairs consisting of the simulated pulse height spectrum set and the simulated neutron energy spectrum set. This batch of simulated pulse height spectrum data is input into the neural network under the current parameter state. Through forward computation of the network layers, the network's predicted neutron energy spectrum is obtained. Then, step Sd3 needs to evaluate the goodness of this prediction, which is achieved by constructing a likelihood function p(D|w), where D represents this batch of training data. Assuming that for a single data pair, the residual between the network's predicted energy spectrum and the actual simulated energy spectrum follows a Gaussian distribution with zero mean and fixed variance, then for the entire batch of data, its likelihood function is the product of the individual likelihoods of all data pairs. The level of the likelihood function quantitatively reflects the probability that the current network model "generates" or "predicts" this batch of observations under the current weight parameters w. The more accurate the prediction, the smaller the residual, and the higher the likelihood function value.

[0060] Step Sd4 is the core step of posterior inference based on Bayes' theorem. Bayes' theorem tells us that after observing data D, the posterior probability distribution p(w|D) of the weight parameters is proportional to the product of the prior distribution p(w) and the likelihood function p(D|w): p(w|D) ∝ p(w)×p(D|w). However, for complex neural network models, this posterior distribution p(w|D) usually does not have an analytical solution and cannot be directly written in its specific form. Therefore, numerical methods are needed for approximate inference. Here, the Markov chain Monte Carlo sampling method is explicitly adopted. The MCMC method constructs a special Markov chain such that, after the chain runs for a sufficiently long time, the probability distribution of its access to different states (i.e., different combinations of weight parameters w) stabilizes exactly on the desired posterior distribution p(w|D). By running this chain and collecting a large number of state samples, we can use these samples to approximate the posterior distribution that is difficult to calculate directly.

[0061] Step Sd5 is the completion stage of training. After running MCMC sampling to reach the preset number of iterations and ensuring the chain has converged, we collect all weight parameter state samples generated after the chain stabilizes. The set of these samples {w^(1), w^(2),..., w^(N)} is an approximation of the posterior distribution p(w|D). The final form of the trained Bayesian neural network model is not a fixed set of weight values, but rather the posterior probability distribution of the weight parameters represented by a large number of samples. When using this model to predict (determine the spectrum) new data, multiple forward propagation calculations are required: each time, a set of weight samples w^(i) is randomly drawn from the posterior distribution, and a forward calculation is performed using this set of weights to obtain a predicted energy spectrum; this process is repeated multiple times, and finally, all predicted energy spectra are statistically analyzed (such as calculating the mean, variance, and quantiles) to obtain the final neutron energy spectrum prediction distribution and its uncertainty interval. This training method explicitly models and preserves the uncertainty of model parameters, enabling the model to reflect its cognitive limitations by expanding the uncertainty of the prediction distribution when faced with boundary situations not fully covered by the training data or noisy measured data. This provides more honest and reliable prediction results and significantly improves the credibility of state diagnosis conclusions.

[0062] In some implementations, Sd4 includes:

[0063] Sd41: A set of values ​​for the initial weight parameters is used as the initial state of the Markov chain;

[0064] Sd42: Generate a new set of candidate weight parameter values ​​based on the current chain state and the preset proposal distribution;

[0065] Sd43: Calculate the transition probability of accepting candidate weight parameter values ​​based on prior probability distribution and likelihood function;

[0066] Sd44: Based on the transition probability, decide whether to update the current chain state to the candidate weight parameter value, and repeat Sd42 to Sd44 for iteration;

[0067] Sd45: After completing a preset number of iterations, collect the chain state sequence and use the statistical characteristics of the chain state sequence to characterize the posterior probability distribution of the weight parameters.

[0068] This embodiment further refines step Sd4, which uses the Markov chain Monte Carlo sampling method to infer the posterior distribution, describing how the algorithmic exploration from the prior to the posterior is implemented. Step Sd41 is to initialize the Markov chain. A set of initial values, denoted as w^(0), needs to be set for the weight parameters of the neural network. This set of values ​​can be randomly drawn from the prior distribution p(w) or can be manually set small random numbers. This set of initial values ​​constitutes the initial state of the Markov chain. Although theoretically, after the chain runs infinitely long, its distribution will converge to the posterior distribution and become independent of the initial state, a reasonable initial value helps to shorten the "pre-burn-in period" required for the chain to reach a stable distribution (i.e., "convergence").

[0069] Step Sd42 involves proposing new candidate states, which is crucial for the chain's movement. At the current time t, the chain's state is w^(t). To propose a new state, we need to sample from a "proposal distribution" q(w* | w^(t)) to obtain candidate states w*. The design of the proposal distribution is critical, as it determines the chain's exploration method and efficiency in the parameter space. A simple and commonly used choice is random walk proposals, where q(w* | w^(t)) is a Gaussian distribution centered at the current state w^(t) with a covariance matrix Σ. Here, Σ needs to be optimized based on the characteristics of the parameter space. If Σ is too large, candidate states are prone to jumping to low-probability regions and being rejected; if Σ is too small, the chain moves slowly, resulting in inefficient exploration. The proposal distribution is responsible for "suggesting" a possible new position near the current position.

[0070] Step Sd43 calculates the probability of accepting a candidate point, i.e., the transition probability. According to the rules of the Metropolis-Hastings et al. MCMC algorithm, the probability α of accepting a candidate point w* is equal to min(1, A), where A = [p(w*) p(D|w*)] / [p(w^(t)) p(D|w^(t))]. Here, p(w*) and p(w^(t)) are the prior probability densities of the candidate point and the current point, respectively, and p(D|w*) and p(D|w^(t)) are the likelihood function values ​​of the candidate point and the current point, respectively. The ratio A is actually the ratio of the posterior probability densities. If the posterior probability density of the candidate point is higher than that of the current point (A>1), it will necessarily be accepted (α=1); if it is lower than that of the current point, it will be accepted with probability A. This calculation ensures that the chain tends to move towards regions with high posterior probability, but does not completely reject low-probability regions, thus making the long-term visit frequency of the chain proportional to the posterior probability density.

[0071] Step Sd44 determines whether to perform a state transition based on the calculated acceptance probability α. Specifically, a random number u is generated uniformly distributed in the interval [0,1]. If u ≤ α, the candidate point is accepted, and the next state of the chain is w^(t+1) = w*. If u > α, the candidate point is rejected, and the chain remains in place, i.e., w^(t+1) = w^(t). However, it's important to note that even if rejected, the current state w^(t) is recorded again in the sample sequence to accurately reflect its probability quality. Then, the algorithm repeats steps Sd42 to Sd44 based on the new current state w^(t+1) for the next iteration. This process is repeated to form a state sequence {w^(0), w^(1), w^(2), ...}.

[0072] Step Sd45 is the end of sampling and sample collection. After repeating the above iterations to a preset total number (usually tens of thousands to millions, depending on the model complexity), the chain is considered to have fully explored the parameter space and reached stability. At this point, a portion of the initial samples of the chain (e.g., the first 20% of iterations) needs to be discarded; this part is called the "pre-burning period" because it may still be overly influenced by the initial state and cannot represent a stationary distribution. After discarding the pre-burning period, the remaining chain state sequence {w^(B), w^(B+1), ..., w^(N)} is an approximately independent sample from the posterior distribution p(w|D) (in reality, there may be autocorrelation between adjacent samples, but this can be handled through sparse sampling or specialized methods). This sample set fully characterizes the uncertainty of the neural network weight parameters given the training data. We can calculate the mean of these samples as a point estimate of the weights, calculate their variance or quantiles to quantify the uncertainty, or directly use these sample sets to represent the model for probability prediction. Through this series of systematic proposal, computation, and acceptance / rejection steps, the MCMC sampling method transforms the complex Bayesian posterior integral problem into a feasibility problem that can be solved iteratively by computer. It is the core algorithmic guarantee for training Bayesian neural networks and obtaining probabilistic weight parameters.

[0073] In some implementations, S2 includes:

[0074] S21: Based on the pulse height spectrum, calculate the sum of the count values ​​for all addresses to obtain the total count;

[0075] S22: Based on the original count value and total count for each channel address, calculate the normalized count value for each channel address and generate normalized pulse height spectrum data.

[0076] This embodiment provides a clear operational procedure for the normalization process in step S2. Here, "normalization" is a data preprocessing method aimed at eliminating the influence of absolute dimensions in the data and highlighting its inherent relative distribution. In a controlled fusion device, the raw pulse height spectrum measured by a liquid scintillator detector has a vertical axis representing the "count value," which directly reflects the number of electrical pulse events recorded by the detector within a specific measurement time, with amplitudes falling within each "address" interval. An address is a number assigned by the analog-to-digital converter after discretizing the pulse amplitude; it is typically positively correlated with the deposition energy. The magnitude of the raw count value is strongly influenced by numerous non-spectral factors, such as the neutron yield of the fusion reaction (directly related to plasma density and temperature), the geometric distance between the detector and the plasma, the liveness time of data acquisition, and the detector's own absolute detection efficiency. If these raw data, heavily influenced by absolute intensity, are directly input into subsequent neural network models, the models will have to expend considerable effort to learn and adapt to these intensity variations unrelated to the physical nature of the energy spectrum. This not only increases the complexity and training difficulty of the model but may also lead to overfitting to specific intensity ranges, resulting in performance degradation when faced with measured data that differs significantly from the training data in intensity. Therefore, the design goal of step S2 is to transform the raw spectrum with absolute intensity information into a pure "shape" spectrum that represents the probability distribution.

[0077] The implementation consists of two sub-steps. First, step S21 is executed: based on the pulse height spectrum, the sum of the count values ​​for all addresses is calculated to obtain the total count. This means that after obtaining a pulse height spectrum array, it is necessary to traverse this array and sum the count values ​​corresponding to all addresses. This total count is a single scalar value that represents the sum of all valid events recorded by the detector across the entire effective energy response range during this measurement. It is a macroscopic representation of the entire measurement scale, but its value itself does not carry information about the shape of the energy spectrum.

[0078] Next, step S22 is executed: based on the original count value and the total count for each channel address, the normalized count value for each channel address is calculated. Specifically, for the i-th channel in the pulse height spectrum (i being the channel address number), its original count value C_i is taken, and it is divided by the total count C_total obtained in step S21, thus obtaining the normalized count value P_i = C_i / C_total for that channel. After performing this division operation on all channels sequentially, a new data sequence is generated, namely, the normalized pulse height spectrum data. At this point, the physical meaning of each data point P_i undergoes a fundamental change: it is no longer a "count," but rather represents "the proportion of events recorded at that channel address to the total number of events." From a probabilistic perspective, if each detection event is considered a random experiment, then P_i can be interpreted as "the estimated probability that the signal amplitude of a random detection event falls within the amplitude interval corresponding to the i-th channel." The vertical axis of the entire spectrum changes from absolute frequency to relative frequency or probability density.

[0079] This transformation brings crucial technical benefits. Normalization ensures that any measured spectrum, regardless of neutron flux or measurement time, is uniformly mapped to a standard probability distribution space. In this space, the integral (or summation) of the spectral data is always 1, and all spectra are "pulled" to the same scale for comparison. For subsequent Bayesian neural network models, the set of simulated pulse height spectra used in the training phase is either generated or has undergone the same normalization process. Therefore, from the beginning of training, the model learns the mapping relationship from one "probability distribution shape" (input spectrum) to another "probability distribution shape" (output energy spectrum). The model doesn't need to care about the total count of the input spectrum; it only focuses on resolving the distribution shape of the input. This greatly reduces the complexity of the problem the model needs to learn and enhances its generalization ability. For example, for the same physical state, two original spectra measured at high and low neutron yields may have absolute counts that differ by several orders of magnitude, but after normalization, their shapes will be almost identical. When these two normalized spectra are input into the trained model, the model will recognize the same shape features, thus providing highly consistent neutron energy spectrum inversion results. Therefore, the normalization process in step S2 is a fundamental technical step to ensure the robustness, reliability, and adaptability of the entire detection method to different measurement conditions. It effectively decouples signal intensity and spectral shape information, creating clean and standardized input conditions for subsequent intelligent spectral interpretation.

[0080] In some implementations, the Bayesian neural network model is a multi-layer feedforward structure comprising an input layer, at least one hidden layer, and an output layer; S3 includes:

[0081] S31: The input layer receives normalized pulse height spectrum data;

[0082] S32: At least one hidden layer includes a bottleneck layer with fewer nodes than the input layer, used for feature extraction and dimensionality reduction of normalized pulse height spectrum data;

[0083] S33: Output vector of the output layer. The dimension of the vector is the same as the number of energy ranges of the neutron energy spectrum to be solved.

[0084] This embodiment clarifies the specific network architecture of the Bayesian neural network model used, an architecture designed for the specific inverse problem of spectrum decomposition. The model employs a "multi-layer feedforward structure," a classic type of neural network where data flows unidirectionally from the input layer, through one or more hidden layers, and finally reaches the output layer, with all nodes between adjacent layers interconnected. This structure possesses a powerful ability to approximate nonlinear functions. The forward propagation calculation described in step S3 unfolds on this specific network structure, with its process tightly coupled to its structure.

[0085] First, in step S31, the input layer receives normalized pulse height spectrum data. The input layer is the network's entry point, and the number of its nodes must strictly match the dimension of the input data. The normalized pulse height spectrum data is a one-dimensional vector; assuming the spectrometer has N channels, this vector will have N elements. Accordingly, the input layer needs to have N neurons. Each neuron simply receives and transmits the normalized count value for the corresponding channel address without performing any calculations. The function of the input layer is to convert the external data format into a tensor form that the network can process internally.

[0086] The core computation of the network occurs in the hidden layers. At least one hidden layer includes a "bottleneck layer" with fewer nodes than the input layer. Hidden layers are layers between the input and output layers that perform feature transformations. A network can have one or more hidden layers. The "bottleneck layer" is a hidden layer intentionally designed to have fewer neurons (nodes) than the previous layer, especially significantly fewer than the input layer. For example, the input layer might have 512 nodes, the first hidden layer might have 256 nodes, and the second hidden layer (the bottleneck layer) might only have 64 nodes. As data flows from the wider layers to this narrower bottleneck layer, the network is forced to "compress" the data. It must learn how to use fewer-dimensional feature vectors to retain and express as much of the information in the input data as possible that is crucial for completing the final task (i.e., predicting the energy spectrum), while discarding redundant information and irrelevant noise. This process essentially achieves "feature extraction and dimensionality reduction" of high-dimensional pulse height spectrum data. Feature extraction refers to the network automatically learning discriminative patterns in spectral data, such as peak position, peak width, and continuous background slope; dimensionality reduction represents these patterns as low-dimensional vectors. The existence of the bottleneck layer imposes a structural constraint, which is a very effective regularization method. It prevents the network from simply memorizing all the details (including noise) of the high-dimensional input to fit the training data, thereby forcing the network to learn a more essential and universal data representation. This directly improves the model's generalization ability and stability when facing new and unseen test data.

[0087] After passing through a series of weighted sums and nonlinear activation function transformations in the hidden layers (including the bottleneck layer), the data finally reaches the output layer in step S33. The output layer is responsible for generating the final solution. Its design must be aligned with the requirements of the physics problem: it outputs a vector whose dimension is the same as the number of energy ranges in the neutron spectrum to be solved. In the spectrum solving problem, we need to discretize the continuous neutron energy range. For example, to invert the neutron spectrum from 0.1 MeV to 15 MeV, this range can be divided into M energy ranges of equal or unequal width (also called energy groups). Then, the output layer needs to have M neurons. The output value of the j-th neuron represents the normalized fluence rate or probability density of the incident neutron spectrum predicted by the neural network in the j-th energy range. In this way, the output of the entire network is a directly usable, discretized neutron spectrum estimate. From the N-dimensional spectral data in the input layer, to the compressed features in the hidden layers (especially the bottleneck layer), and then to the M-dimensional energy spectrum vector in the output layer, the entire network architecture constitutes a complete end-to-end "spectrum-spectrum" inversion computation pipeline. This specific structural design with a bottleneck layer is not a general-purpose network, but an optimization made to address the characteristics of data redundancy and the need for strong generalization ability in inverse problems. It ensures that the network can efficiently and stably extract the most relevant features from the measurement data and accurately map them to the physical parameter space.

[0088] In some implementations, after S4, the following is also included:

[0089] S5: Based on the obtained incident neutron energy spectrum distribution, perform a similarity comparison with the pre-stored characteristic neutron energy spectrum database;

[0090] S6: Based on the similarity comparison results, determine the current operating state category of the controlled fusion device.

[0091] This embodiment, after completing the core neutron energy spectrum inversion, further expands the application boundaries of the method, directly linking it to the ultimate goal in engineering practice—operational status diagnosis. The incident neutron energy spectrum distribution output in step S4 is a crucial intermediate physical quantity, quantitatively describing the energy distribution of neutrons leaking from the fusion device. However, for device operators or control systems, a semantically clear "state" conclusion is needed, such as "normal operation," "an instability has occurred," or "an anomaly has occurred." Steps S5 and S6 are the key steps in realizing this leap from "physical quantity" to "state category."

[0092] Step S5 specifies that, based on the obtained incident neutron energy spectrum distribution, a similarity comparison is performed with a pre-stored characteristic neutron energy spectrum database. Here, a new component is introduced—the "characteristic neutron energy spectrum database." This database is a pre-built knowledge base containing "standard" or "typical" neutron energy spectra that correspond one-to-one with specific operating states of controlled fusion devices, accumulated historically or derived theoretically. For example, for a tokamak device, the database may contain the following categories of characteristic spectra: 1) Energy spectrum under the pure deuterium-tritium thermonuclear fusion-dominated state, characterized by a significant Gaussian peak at 14.1 MeV; 2) Energy spectrum under the deuterium-deuterium reaction-dominated state, characterized by characteristic peaks at 2.5 MeV, etc.; 3) High-energy bremsstrahlung neutron spectra generated by the interaction of numerous high-energy escape electrons with background materials, exhibiting a continuous bulge at a high-energy end; 4) Characteristic radioactive decay neutron spectra generated after the first wall material (such as beryllium or tungsten) is activated by neutrons, exhibiting a series of narrow peaks at specific energies. Each characteristic energy spectrum serves as a "template" or "fingerprint," bound to a specific state label.

[0093] During the comparison, the object of the operation is the best estimated incident neutron energy spectrum distribution obtained from the inversion in step S4 (usually the posterior mean spectrum output by the Bayesian neural network model). The system compares this current energy spectrum with each (or a relevant subset) of feature energy spectra in the database one by one. This "similarity comparison" is not a simple numerical equality judgment, but a quantitative evaluation of the matching degree of morphological features such as the overall shape of the two curves, the position of the characteristic peaks, and the relative intensity distribution. The core idea is that if the current device is in a certain known state, then the neutron energy spectrum it produces should be highly similar in morphology to the feature energy spectrum corresponding to that state. By calculating a quantitative "similarity" score, this degree of similarity can be objectively measured.

[0094] Based on the similarity results calculated in step S5, step S6 performs a logical decision: determining the current operating state category of the controlled fusion device based on the similarity comparison results. The simplest decision rule is the "nearest neighbor" rule, which selects the feature spectrum with the highest similarity score to the current energy spectrum and directly determines the state category bound to it as the current state of the device. For example, if the current energy spectrum has a similarity of 0.95 with the "normal deuterium-tritium combustion" template in the database, while the similarity with other templates is all below 0.6, then the system has sufficient confidence to determine that the device is in a normal deuterium-tritium combustion state. More complex decisions can introduce threshold judgments, for example, only making a definite judgment when the highest similarity exceeds a certain confidence threshold (such as 0.8), otherwise outputting "unknown" or "state ambiguous". The decision result can be a single definite state, or a distribution containing multiple possible states and their probabilities (if a probabilistic comparison method is used). Through steps S5 and S6, the entire technical solution realizes a complete and automated state monitoring chain: from the original detector pulse signal to quantitative neutron energy spectrum physical parameters, and then to qualitative and operable operational status conclusions. This greatly enhances the engineering practical value of the method, transforming it from a physical diagnostic tool into an intelligent monitoring system that can directly provide immediate and intuitive criteria for the safe operation of the device, the adjustment of physical experiments, and the early warning of anomalies.

[0095] In some implementations, S5 includes:

[0096] S51: Retrieve multiple sets of reference neutron energy spectra related to the current device type and operating conditions from the pre-stored characteristic neutron energy spectrum database;

[0097] S52: Calculate the spectral similarity between the incident neutron energy spectrum distribution and each set of reference neutron energy spectra;

[0098] S53: Based on all the calculated spectral similarities, select the reference neutron energy spectrum with the highest spectral similarity as the comparison result.

[0099] This embodiment further breaks down and standardizes the similarity comparison process in step S5, aiming to make the process more operable and robust. Step S51 is the entry point and preprocessing of the comparison process: retrieving multiple sets of reference neutron spectra related to the current device type and operating conditions from a pre-stored characteristic neutron spectrum database. A complete characteristic neutron spectrum database may be enormous, containing spectral data from different types of fusion devices (such as large tokamaks, small experimental reactors, and stellarators), different operating modes (such as ohmic heating, neutral beam injection, and ion cyclotron heating), and different plasma parameters. If a global search is performed across the entire database for each comparison, not only is the computational efficiency low, but more seriously, it may introduce incomparable "noise" templates, leading to misjudgments. For example, the typical energy spectrum of a tokamak and the typical energy spectrum of a stellarator may be drastically different due to fundamental differences in their physical principles and geometric structures, making direct comparison meaningless. Therefore, targeted retrieval is necessary. The retrieval is based on the "current device type and operating conditions." The device type is known and fixed information. Operating conditions can include known or predictable macroscopic parameters such as the current experimental target, the range of applied heating power, and the type of filling gas. The system uses this information as query keywords or filtering conditions to quickly select the most relevant subset of reference spectra from the overall database. This subset constitutes the "candidate template library" for this comparison, ensuring that subsequent similarity calculations are performed within a physically reasonable and comparable range. This optimizes efficiency and is a crucial guarantee of accuracy.

[0100] After obtaining a relevant subset of reference spectra, step S52 performs the core quantization calculation: calculating the spectral shape similarity between the incident neutron energy spectrum distribution and each set of reference neutron energy spectra. "Spectral shape similarity" is a precisely defined metric; it must be insensitive to the absolute intensity of the energy spectrum, but only to its shape. Commonly, before calculation, the current incident neutron energy spectrum distribution and each set of reference neutron energy spectra are normalized (e.g., their integrals are normalized to 1) to completely remove intensity information. Then, various mathematical tools can be used to calculate the similarity between the two normalized vectors (i.e., the two energy spectrum curves). A common and intuitive method is to calculate the Pearson correlation coefficient, which measures the degree of linear correlation between two sequences, ranging from -1 to 1; the closer the value is to 1, the more similar the shapes. Another method is to calculate the cosine similarity, i.e., the cosine of the angle between the two vectors, which focuses on the consistency of direction. Distance functions based on specific features (such as main peak position, full width at half maximum, and spectral moments) can also be used. The system will independently and repeatedly perform this calculation process for each set of reference neutron energy spectra in the subset, and finally obtain a set of similarity numerical sequences that correspond one-to-one with the reference spectra.

[0101] After obtaining all similarity values, step S53 makes an intermediate conclusion based on a clear rule: based on all calculated spectral similarities, the set of reference neutron spectra with the highest spectral similarity is selected as the comparison result. The system iterates through and compares all similarity values ​​generated in step S52, finding the maximum value. The set of reference neutron spectra corresponding to this maximum value is identified as the template that best matches the current measured spectrum. This "highest similarity" rule is simple, clear, and easy to implement. Its output is singular: a specific set of reference spectra and their metadata (including their associated state labels). This result serves as the direct input for the final state determination in step S6. Through the context-aware retrieval in step S51, the standardized quantitative calculation in step S52, and the deterministic selection in step S53, the entire similarity comparison process is constructed as a clearly structured, step-by-step, repeatable, and interpretable automated process. It minimizes subjective arbitrariness, transforming an experience-based pattern recognition task into an objective computational task based on data and algorithms, thereby ensuring the reliability, consistency, and transparency of this crucial link from physical inversion results to engineering state determination.

[0102] Example 2

[0103] like Figure 2 As shown, in a second aspect, the present invention proposes a controllable fusion device operation status detection system. The system employs a controllable fusion device operation status detection method proposed in any of the above embodiments, and the system includes:

[0104] The pulse height spectrum acquisition module is used to perform step S1: acquire the pulse height spectrum based on the measurement results of the liquid scintillator detector deployed at the site of the controlled fusion device;

[0105] The spectrum data normalization processing module is used to perform step S2: based on the pulse height spectrum, normalize the count values ​​of each channel address to generate normalized pulse height spectrum data;

[0106] The Bayesian neural network spectrum calculation module is used to execute step S3: inputting the normalized pulse height spectrum data into a pre-trained Bayesian neural network model for forward propagation calculation. The network weight parameters of the Bayesian neural network model are represented in the form of a probability distribution. The construction and training of the Bayesian neural network model includes:

[0107] Sa: Based on the physical characteristics of the liquid scintillator detector, the neutron response matrix is ​​constructed by combining Monte Carlo simulation with experimental calibration.

[0108] Sb: Based on a variety of preset basis function neutron energy spectra, a set of simulated neutron energy spectra is generated through random combination and parameter adjustment;

[0109] Sc: Based on the simulated neutron energy spectrum set and the neutron response matrix, a convolution operation is performed to generate the corresponding simulated pulse height spectrum set;

[0110] Sd: Based on the training data pair consisting of the simulated pulse height spectrum set and the simulated neutron energy spectrum set, the initialized neural network is trained using the Bayesian inference method to obtain a Bayesian neural network model in which the network weight parameters follow a probability distribution.

[0111] The neutron energy spectrum inversion output module is used to perform step S4: perform inverse transformation processing on the output of the forward propagation calculation of the Bayesian neural network model to obtain the incident neutron energy spectrum distribution.

[0112] This system corresponds to the method proposed in Example 1, and will not be described in detail here.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting the operating status of a controllable fusion device, characterized in that, include: S1: Obtain the pulse height spectrum based on the measurement results of the liquid scintillator detector deployed at the site of the controlled fusion device; S2: Based on the pulse height spectrum, the count values ​​of each channel address are normalized to generate normalized pulse height spectrum data; S3: Input the normalized pulse height spectrum data into a pre-trained Bayesian neural network model for forward propagation calculation. The network weight parameters of the Bayesian neural network model are represented in the form of a probability distribution. The construction and training of the Bayesian neural network model includes: Sa: Based on the physical characteristics of the liquid scintillator detector, a neutron response matrix is ​​constructed by combining Monte Carlo simulation with experimental calibration. Sb: Based on a variety of preset basis function neutron energy spectra, a set of simulated neutron energy spectra is generated through random combination and parameter adjustment; wherein, Sb includes: Sb1: Provides a basis function library containing Gaussian peak spectra and exponential decay spectra; Sb2: Based on random numbers, select at least one basis function from the basis function library and assign a random strength coefficient to each selected basis function; Sb3: Based on the selected basis functions and their assigned intensity coefficients, a weighted summation operation is performed to generate a simulated neutron energy spectrum; during the weighted summation operation, a random energy shift and broadening are applied to the summation result to increase the diversity of the simulated neutron energy spectrum set; Sb4: Repeat Sb2 to Sb3 to generate a set of simulated neutron energy spectra containing a preset number of spectra; Sc: Based on the simulated neutron energy spectrum set and the neutron response matrix, a convolution operation is performed to generate the corresponding simulated pulse height spectrum set; Sd: Based on training data pairs consisting of simulated pulse height spectrum sets and simulated neutron energy spectrum sets, a Bayesian inference method is used to train the initialized neural network, resulting in a Bayesian neural network model where the network weight parameters follow a probability distribution; In Sd, the Bayesian inference method is used to train the initialized neural network, including: Sd1: Sets the prior probability distribution for the weight parameters of the initialized neural network; Sd2: Input the simulated pulse height spectrum from the training data pair into the initialized neural network to calculate the predicted neutron energy spectrum; Sd3: Calculate the likelihood function based on the predicted neutron energy spectrum and the simulated neutron energy spectrum corresponding to the training data pair; Sd4: Based on the prior probability distribution and likelihood function, the posterior probability distribution of the weight parameters is inferred using the Markov chain Monte Carlo sampling method; Sd5: Based on the posterior probability distribution obtained from sampling, determine the final probabilistic weight parameters of the Bayesian neural network model; S4: Perform inverse transformation on the output of the forward propagation calculation of the Bayesian neural network model to obtain the incident neutron energy spectrum distribution.

2. The method for detecting the operating status of a controllable fusion device according to claim 1, characterized in that, Sa includes: Sa1: Based on the standard radioactive source, the liquid scintillator detector is calibrated with energy to obtain the correspondence between the channel address and the deposited energy; Sa2: Based on Monte Carlo simulation software, the energy deposition process of monoenergetic neutrons in a liquid scintillator detector was simulated to obtain the energy deposition spectrum; Sa3: Based on the correspondence between channel location and deposition energy and the energy deposition spectrum, the neutron response matrix is ​​calculated and arranged.

3. The method for detecting the operating status of a controllable fusion device according to claim 1, characterized in that, Sd4 includes: Sd41: A set of values ​​for the initial weight parameters is used as the initial state of the Markov chain; Sd42: Generate a new set of candidate weight parameter values ​​based on the current chain state and the preset proposal distribution; Sd43: Calculate the transition probability of accepting candidate weight parameter values ​​based on prior probability distribution and likelihood function; Sd44: Based on the transition probability, decide whether to update the current chain state to the candidate weight parameter value, and repeat Sd42 to Sd44 for iteration; Sd45: After completing a preset number of iterations, collect the chain state sequence and use the statistical characteristics of the chain state sequence to characterize the posterior probability distribution of the weight parameters.

4. The method for detecting the operating status of a controllable fusion device according to claim 1, characterized in that, S2 include: S21: Based on the pulse height spectrum, calculate the sum of the count values ​​for all addresses to obtain the total count; S22: Based on the original count value and total count for each channel address, calculate the normalized count value for each channel address and generate normalized pulse height spectrum data.

5. The method for detecting the operating status of a controllable fusion device according to claim 1, characterized in that, The Bayesian neural network model is a multi-layer feedforward structure containing an input layer, at least one hidden layer, and an output layer; S3 includes: S31: The input layer receives normalized pulse height spectrum data; S32: At least one hidden layer includes a bottleneck layer with fewer nodes than the input layer, used for feature extraction and dimensionality reduction of normalized pulse height spectrum data; S33: Output vector of the output layer. The dimension of the vector is the same as the number of energy ranges of the neutron energy spectrum to be solved.

6. The method for detecting the operating status of a controllable fusion device according to claim 1, characterized in that, Following S4, it also includes: S5: Based on the obtained incident neutron energy spectrum distribution, perform a similarity comparison with the pre-stored characteristic neutron energy spectrum database; S6: Based on the similarity comparison results, determine the current operating state category of the controlled fusion device.

7. The method for detecting the operating status of a controllable fusion device according to claim 6, characterized in that, S5 include: S51: Retrieve multiple sets of reference neutron energy spectra related to the current device type and operating conditions from the pre-stored characteristic neutron energy spectrum database; S52: Calculate the spectral similarity between the incident neutron energy spectrum distribution and each set of reference neutron energy spectra; S53: Based on all the calculated spectral similarities, select the reference neutron energy spectrum with the highest spectral similarity as the comparison result.

8. A system for detecting the operating status of a controllable fusion device, characterized in that, The system employs a method for detecting the operating status of a controlled nuclear fusion device as described in any one of claims 1 to 7, and the system comprises: The pulse height spectrum acquisition module is used to perform step S1: acquire the pulse height spectrum based on the measurement results of the liquid scintillator detector deployed at the site of the controlled fusion device; The spectrum data normalization processing module is used to perform step S2: based on the pulse height spectrum, normalize the count values ​​of each channel address to generate normalized pulse height spectrum data; The Bayesian neural network spectrum calculation module is used to execute step S3: inputting the normalized pulse height spectrum data into a pre-trained Bayesian neural network model for forward propagation calculation. The network weight parameters of the Bayesian neural network model are represented in the form of a probability distribution. The construction and training of the Bayesian neural network model includes: Sa: Based on the physical characteristics of the liquid scintillator detector, a neutron response matrix is ​​constructed by combining Monte Carlo simulation with experimental calibration. Sb: Based on a variety of preset basis function neutron energy spectra, a set of simulated neutron energy spectra is generated through random combination and parameter adjustment; Sc: Based on the simulated neutron energy spectrum set and the neutron response matrix, a convolution operation is performed to generate the corresponding simulated pulse height spectrum set; Sd: Based on the training data pair consisting of the simulated pulse height spectrum set and the simulated neutron energy spectrum set, the initialized neural network is trained using the Bayesian inference method to obtain a Bayesian neural network model in which the network weight parameters follow a probability distribution. The neutron energy spectrum inversion output module is used to perform step S4: perform inverse transformation processing on the output of the forward propagation calculation of the Bayesian neural network model to obtain the incident neutron energy spectrum distribution.

Images