A method for classifying the oxygen state in liquid lead-bismuth

Abstract

The present invention discloses a method for classifying the oxygen state in liquid lead and bismuth, and relates to the technical field of intelligent classification and identification. The liquid lead and bismuth oxygen state classification model comprises a feature embedding network, a position encoding layer, an attention network, a dynamic feature fusion gating network, and an output layer. The attention network is composed of a plurality of stacked Transformer networks, and each Transformer network comprises a sequentially connected intra-feature self-attention layer, an inter-feature cross-attention layer, and a feedforward network layer. The liquid lead and bismuth oxygen state classification method comprises: obtaining approximate coefficient mean features from an oxygen potential time subsequence and a temperature time subsequence through wavelet transform, and sequentially passing the approximate coefficient mean features through the feature embedding network, the position encoding layer, the attention network, the dynamic feature fusion gating network, and the output layer in the liquid lead and bismuth oxygen state classification model to obtain a classification prediction result of the liquid lead and bismuth oxygen state. The method realizes classification of state transition regions and complex state boundaries, and simultaneously improves classification accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent classification and identification, in particular to a method for classifying oxygen states in liquid lead and bismuth. Background Art

[0002] Lead-bismuth reactors (LBEs), as fourth-generation advanced nuclear energy systems, have attracted considerable attention for their high safety, stability, and cooling performance. However, they suffer from serious material corrosion issues. Maintaining a stable oxygen state is crucial for their long-term safe operation. Therefore, identifying and monitoring the oxygen state in liquid lead-bismuth is crucial for the intelligent operation and maintenance of nuclear energy systems.

[0003] In existing technologies, hard labels are typically used to classify the oxygen state in liquid lead and bismuth into three states: saturation, unsaturation, and sensor anomaly. Traditional hard label classification methods mainly rely on discrete labels for state identification, but this method has difficulty capturing the ambiguity during state transitions. Especially when the oxygen state in liquid lead and bismuth is in a transitional stage, classification accuracy is often reduced due to an imbalance in data categories (saturation, unsaturation, sensor anomaly), which cannot fully reflect the actual operating conditions, resulting in low classification accuracy. Summary of the Invention

[0004] In order to overcome the defects of the above-mentioned prior art in low accuracy in classifying the oxygen state in liquid lead bismuth and difficulty in capturing the ambiguity during the transition of the oxygen state in liquid lead bismuth, the present invention proposes a method for classifying the oxygen state in liquid lead bismuth.

[0005] To achieve the above object, the present invention adopts the following technical solutions, including:

[0006] A method for classifying oxygen states in liquid lead and bismuth is characterized by utilizing a liquid lead and bismuth oxygen state classification model, the model comprising a feature embedding network, a position encoding layer, an attention network, a dynamic feature fusion gating network, and an output layer; wherein the attention network is composed of a stack of multiple Transformer networks, and between two adjacent Transformer networks, the output of the previous Transformer network serves as the input of the next Transformer network, and each Transformer network comprises a sequentially connected intra-feature self-attention layer, an inter-feature cross-attention layer, and a feedforward network layer. The method for classifying oxygen states in liquid lead and bismuth comprises:

[0007] Obtain oxygen potential time subseries and temperature time subseries, and obtain approximate coefficient mean characteristics through wavelet transform;

[0008] The approximate coefficient mean feature is nonlinearly projected through the feature embedding network to obtain the oxygen potential projection feature and temperature projection feature;

[0009] The oxygen potential projection feature and the temperature projection feature are encoded through the position encoding layer to obtain the encoded projection feature E;

[0010] The encoded projection feature E is extracted through the attention network to obtain the nonlinear transformation enhancement feature; the nonlinear transformation enhancement feature is obtained by the oxygen potential nonlinear transformation enhancement feature F v and temperature nonlinear transformation enhancement feature F t composition;

[0011] The nonlinear transformation enhancement features obtained by the attention network are input into the dynamic feature fusion gating network for feature fusion to obtain dynamic weighted fusion features;

[0012] The dynamically weighted fusion features are input into the output layer to obtain the classification prediction results of oxygen state in liquid lead and bismuth.

[0013] Preferably, before obtaining the oxygen potential time subsequence and the temperature time subsequence, the method further includes:

[0014] Obtaining oxygen potential time series data and temperature time series data in a lead-bismuth alloy and adding noise disturbance to obtain enhanced oxygen potential time series data and temperature time series data;

[0015] dividing the enhanced oxygen potential time series data and temperature time series data into oxygen potential time subseries and temperature time subseries with time continuity;

[0016] Based on the oxygen potential time subsequence and temperature time subsequence, the hard labels describing the oxygen state classification are converted into oxygen state classification soft labels with fuzziness and continuity. The oxygen state classification results are marked to obtain an oxygen state classification dataset; the oxygen state classification soft labels are the probability distribution of several oxygen state categories.

[0017] Preferably, obtaining the oxygen potential time subsequence and the temperature time subsequence, and obtaining the approximate coefficient mean value feature by wavelet transform includes:

[0018] First, wavelet transform is used to decompose the oxygen potential time subseries and temperature time subseries to obtain the approximate coefficients. The calculation formula is:

[0019] ;

[0020] Where a is the decomposition layer number; m is the position index in each decomposition layer; is the approximate coefficient of the mth position in the ath decomposition layer; k is the position index of the filter coefficient, K is the total index number of the filter coefficient position; h k is the filter coefficient at the kth position; is the first in the a-1th decomposition layer The approximate coefficients of the positions;

[0021] Then, based on the approximate coefficient, we get the approximate coefficient mean feature. , the calculation formula is:

[0022] ;

[0023] Where M is the number of approximation coefficients, , U is the length of the oxygen potential time subsequence or temperature time subsequence.

[0024] Preferably, extracting the encoded projection feature E through an attention network to obtain nonlinear transformation enhancement features includes:

[0025] First, the encoded projection feature is extracted through the self-attention layer within the feature to obtain the self-attention feature E intra , the self-attention update feature E intra Update features by oxygen potential self-attention and temperature self-attention update features composition;

[0026] Then the self-attention update feature is passed through the cross-attention layer between features to obtain the cross-attention update feature E inter ;

[0027] Then the cross attention is used to update the feature E inter Input to the feedforward network for nonlinear feature enhancement to obtain nonlinear transformation enhancement features; among them, nonlinear transformation enhancement features Enhanced characteristic F by nonlinear transformation of oxygen potential v and temperature nonlinear transformation enhancement feature F t composition;

[0028] Pass the self-attention update feature through the cross-attention layer between features to obtain the cross-attention update feature E inter include:

[0029] First, the feature is updated based on oxygen potential self-attention and temperature self-attention update features , calculate the cross attention of oxygen potential to temperature ;

[0030] Secondly, based on the oxygen potential self-attention update feature and temperature self-attention update features , calculate the cross attention from temperature to oxygen potential ;

[0031] Finally, the cross-attention based on oxygen potential to temperature and temperature to oxygen potential , get the cross attention update feature E inter ;

[0032] Among them, the intra-feature self-attention layer is used to extract different features separately; the inter-feature cross-attention layer is used to extract different features in a mixed manner; the feedforward network layer is used to enhance the features;

[0033] Preferably, the nonlinear transformation enhancement features obtained by the attention network are input into the dynamic feature fusion gating network for feature fusion, and obtaining the dynamic weighted fusion features includes:

[0034] First, the oxygen potential nonlinear transformation enhancement feature F obtained by the last Transformer network is v and temperature nonlinear transformation enhancement feature F t Input into the dynamic feature fusion gating network for feature fusion to obtain the dynamic weight matrix , the calculation formula is:

[0035] ;

[0036] Then, based on the dynamic weight matrix, the dynamic weighted fusion feature f is obtained. The calculation formula is:

[0037] ;

[0038] in, is the mapping probability function; ⊙ represents element-by-element multiplication; Represents the oxygen potential nonlinear transformation fusion feature F v and temperature nonlinear transformation fusion feature F t Perform feature vector splicing; g is the dynamic weight matrix generated by the Sigmoid function.

[0039] Preferably, the classification prediction results of oxygen state in liquid lead bismuth The calculation formula is:

[0040] ;

[0041] in, It is a dynamic weighted fusion feature; is the weight matrix of the classification head; The bias vector of the classification head; z is the unnormalized output of the classification head; LayerNorm(.) is layer normalization; Softmax(.) is the normalized exponential function.

[0042] Preferably, the oxygen state classification data set is obtained and a loss function L is used to train the oxygen state classification model in liquid lead and bismuth. The calculation formula of the loss function L is:

[0043] ;

[0044] Among them, α is a hyperparameter; L CE is the cross entropy loss; L KL is the KL divergence loss; δ is the equilibrium constant; N is the total number of samples; i is the sample number; C is the total number of categories; j is the category number; is the smoothed target probability of oxygen state classification in liquid lead bismuth of the jth category in the i-th sample, , is the label smoothing factor; is the predicted probability of classifying the jth category of oxygen state in liquid lead bismuth in the i-th sample; is the true probability of classifying the j-th category for the oxygen state in liquid lead bismuth in the i-th sample.

[0045] Preferably, the oxygen potential to temperature cross-attention The calculation formula is:

[0046] ;

[0047] in, represent the query, key, and value of the cross-attention mechanism from oxygen potential to temperature, respectively; is the oxygen potential to temperature cross-attention learnable weight matrix, used to generate query, key and value respectively; is the cross-attention bias from oxygen potential to temperature; Softmax(.) is the normalized exponential function; is the head of the self-attention mechanism; is the hidden layer dimension;

[0048] Cross-attention from temperature to oxygen potential The calculation formula is:

[0049] ;

[0050] in, represent the query, key, and value of the cross-attention mechanism from temperature to oxygen potential, respectively; is the temperature to oxygen potential cross-attention learnable weight matrix, used to generate query, key and value respectively; is the cross attention bias from temperature to oxygen potential.

[0051] Preferably, the cross attention update feature E inter The calculation formula is:

[0052] ;

[0053] Among them, LayerNorm(.) is the layer normalization function; Dropout(.) is the random regularization probability.

[0054] Preferably, the calculation formulas for the oxygen potential projection feature and the temperature projection feature are:

[0055] ;

[0056] ;

[0057] Among them, e v is the oxygen potential projection feature; e t is the temperature projection feature; v and t are the oxygen potential approximation coefficient mean feature and the temperature approximation coefficient mean feature respectively; 、 、 and They are the first-layer weight matrix and the second-layer weight matrix when projecting the mean characteristics of the oxygen potential approximation coefficient and the mean characteristics of the temperature approximation coefficient, respectively; 、 、 and are the first layer bias and the second layer bias when projecting the mean feature of the oxygen potential approximation coefficient and the mean feature of the temperature approximation coefficient, respectively; (.) is the Gaussian error activation function.

[0058] The advantages of the present invention are:

[0059] (1) The present invention constructs a classification model of oxygen state in liquid lead and bismuth, which includes a feature embedding network, a position encoding layer, an attention network, a dynamic feature fusion gating network and an output layer. The attention network is composed of a plurality of stacked Transformer networks. Between two adjacent Transformer networks, the output of the previous Transformer network serves as the input of the next Transformer network. Each Transformer network includes a feature self-attention layer, an inter-feature cross-attention layer and a feedforward network layer connected in sequence. An inter-feature cross-attention layer is added to the basic Transformer network, and a dynamic feature fusion gating network is added after the Transformer network. The cross-attention update feature is obtained through the inter-feature cross-attention layer based on the self-attention update feature, and the cross-attention update feature is converted into the cross-attention update feature. Input to the feedforward network for nonlinear feature enhancement, and the oxygen potential obtained in the attention network is nonlinearly transformed to enhance the feature F v and temperature nonlinear transformation enhancement feature F tAdaptive fusion is performed to obtain a dynamic weight matrix. Based on the dynamic weight matrix, dynamic weighted fusion features are obtained. Finally, the classification prediction results of the oxygen state in liquid lead and bismuth are obtained through the dynamic weighted fusion features. The probability distribution is used to describe the transition stage of the oxygen state in liquid lead and bismuth, thereby realizing the classification of state transition areas and complex state boundaries, and at the same time improving the classification accuracy.

[0060] (2) The present invention combines cross-attention with a learnable gating mechanism, enabling the model to adaptively focus on the time-frequency features of the oxygen potential time series data and the temperature time series data in liquid lead bismuth and learn the complex interactive relationship between the features, thereby improving the ability to capture complex state boundaries and local features.

[0061] (3) The present invention uses a weighted combination of cross entropy and KL divergence loss for training, which can not only ensure the accuracy of the model in probability prediction, but also improve distribution alignment and reduce the risk of overfitting.

[0062] (4) The present invention uses the oxygen potential and temperature in the lead-bismuth alloy to evaluate the oxygen state in liquid lead-bismuth. The oxygen potential time series and temperature time series are processed by wavelet transform, which effectively captures the approximate coefficient mean characteristics of the low-frequency trend characteristics of the signal, overcomes the defects of the singleness of traditional time domain / frequency domain analysis, and is particularly suitable for the complex dynamic change of oxygen state in the lead-bismuth eutectic system.

[0063] (5) The present invention uses a feature embedding network to perform nonlinear projection on the approximate coefficients and map the original features into a high-dimensional latent space, which solves the problem of insufficient feature separability of traditional linear projections (such as PCA) in oxygen state classification tasks and significantly improves the joint characterization capability of oxygen potential features and temperature features.

[0064] (6) The present invention uses the position encoding layer to introduce temporal position information, which makes up for the problem of temporal information loss caused by permutation invariance when the traditional Transformer processes time series, so that the model can accurately identify the temporal dependence of oxygen state changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a model architecture diagram of the present invention;

[0066] Figure 2 This is a technical flow chart of the present invention;

[0067] Figure 3 This is the feature extraction result diagram of the present invention;

[0068] Figure 4 This is a comparison diagram of the model of the present invention and the existing model. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0070] As a fourth-generation advanced nuclear energy system, the lead-bismuth reactor (LBE) has attracted much attention due to its high safety, stability, and cooling performance. However, it suffers from serious material corrosion problems. Therefore, maintaining a stable oxygen state is key to its long-term safe operation. Due to the difficulty in accurately characterizing the transition fuzziness of the oxygen state in lead-bismuth alloys and the inability to adaptively focus on the extracted key features, the present invention proposes a method for classifying the oxygen state in liquid lead-bismuth. The liquid lead-bismuth oxygen state classification model is utilized. The liquid lead-bismuth oxygen state classification model includes a feature embedding network, a position encoding layer, an attention network, a dynamic feature fusion gating network, and an output layer. The attention network is composed of a stack of Lx Transformer networks. Between two adjacent Transformer networks, the output of the previous Transformer network serves as the input of the next Transformer network. Each Transformer network includes a sequentially connected feature intra-feature self-attention layer, an inter-feature cross-attention layer, and a feedforward network layer. The input of the liquid lead-bismuth oxygen state classification model is the oxygen potential approximation coefficient mean feature and the temperature approximation coefficient mean feature, and the output is the liquid lead-bismuth oxygen state classification prediction result.

[0071] like Figure 1-4 As shown, a method for classifying oxygen states in liquid lead and bismuth comprises:

[0072] S1: Obtain oxygen potential time series data and temperature time series data in lead-bismuth alloy, and use an enhancement method based on smooth perturbation to obtain enhanced oxygen potential time series data and temperature time series data.

[0073] Since oxygen states are usually described and classified using hard labels, including saturated state, unsaturated state and sensor abnormality, the accuracy of the classification model is reduced due to the lack of saturated state data and sensor abnormality data. In order to improve the robustness of the model to oxygen potential and temperature fluctuations in lead-bismuth alloy under different scenarios, the present invention adopts a time series data enhancement method based on smooth perturbation. The perturbation signal The expression is as follows:

[0074] ;

[0075] Where A represents the perturbation amplitude, f represents the perturbation frequency, t represents the current time step, and N represents the length of the oxygen potential time series dataset or the temperature time series dataset in lead-bismuth alloy.

[0076] The disturbance signal They were applied to the oxygen potential time series data and temperature time series data in lead-bismuth alloy respectively to obtain enhanced oxygen potential time series data and temperature time series data to simulate the possible dynamic changes in the actual operating environment, solving the problem of decreased classification accuracy caused by category imbalance of oxygen potential time series data and temperature time series data in lead-bismuth alloy.

[0077] S2: Use a sliding window to split the enhanced oxygen potential time series data and temperature time series data into oxygen potential time subseries and temperature time subseries with time continuity.

[0078] The length of the sliding window is not limited, and the step size is usually 1.

[0079] This method can capture the dynamic characteristics of oxygen potential and temperature time series data, and lay the foundation for subsequent feature extraction. The time-frequency characteristics of the oxygen potential and temperature time series data are extracted using a sliding window combined with wavelet transform, and the hard labels are converted into soft labels to represent the ambiguity of the transition phase.

[0080] S3: Based on the oxygen potential time subsequence and temperature time subsequence, the hard labels describing the oxygen state classification are converted into oxygen state classification soft labels with fuzziness and continuity, and the oxygen state classification results are marked to obtain the oxygen state classification dataset.

[0081] To improve the model's classification capabilities during oxygen state transitions (e.g., from saturated to unsaturated), the hard labels describing oxygen states (saturated, unsaturated, sensor abnormality) were converted from single discrete values to soft labels with both fuzziness and continuity. The frequency of each category was normalized to ensure that the label for each sample adhered to a probability distribution.

[0082] Soft-label classification refers to a method that uses probability distributions to describe the ambiguity and uncertainty of labels rather than traditional discrete hard labels during model training. This allows for classification of state transition regions and complex state boundaries. This method is primarily used in the classification of oxygen concentration states in lead-bismuth reactors. Soft-label classification can capture the ambiguity of state transitions and provide a fine-grained representation of uncertainty. Probability distribution training reduces the risk of overfitting associated with hard labels and enhances model generalization.

[0083] In the present invention, the oxygen status hard label categories are divided into three categories, and the oxygen status classification result y is expressed as:

[0084] ;

[0085] Among them, the oxygen state classification result y is the probability distribution of the soft label; j is the oxygen state category number in the hard label, is the frequency of oxygen status categories in the hard label.

[0086] S4: Based on the oxygen potential time subsequence and the temperature time subsequence, obtain the approximate coefficient mean feature through wavelet transform, wherein the approximate coefficient mean feature includes the oxygen potential approximate coefficient mean feature V_μ approx and temperature approximation coefficient mean characteristic T_μ approx ;

[0087] Before calculating the approximate coefficient mean characteristics, the oxygen potential time subseries and temperature time subseries are first decomposed by wavelet transform to obtain the approximate coefficients.

[0088] The calculation formula is:

[0089] ;

[0090] Wherein, a is the decomposition layer number, a≥1, and in this embodiment, the value range is [1,2,3]. The initial layer A0 is the original signal; m is the position index in each decomposition layer; is the approximate coefficient of the mth position in the ath decomposition layer; k is the position index of the filter coefficient, and its value range is [0,1,...,15]; h k is the filter coefficient at the kth position; is the first in the a-1th decomposition layer The approximate coefficients of the positions;

[0091] The approximate coefficients of each decomposition layer are obtained by convolution and downsampling (with a step size of 2) of the coefficients of the previous decomposition layer with the filter coefficients.

[0092] Approximate coefficient mean characteristics The calculation formula is:

[0093] ;

[0094] Where M is the number of approximation coefficients, , U is the length of the oxygen potential time subsequence or temperature time subsequence.

[0095] Wavelet transform can analyze the local dynamic characteristics of oxygen potential time subseries and temperature time subseries at different times, and realize the feature capture of signals in local time domain and frequency domain.

[0096] In this embodiment, the Symlets8 (sym8) wavelet is used as the wavelet basis function, and the time-frequency features of the oxygen potential time subsequence and the temperature time subsequence are respectively extracted and decomposed to obtain the oxygen potential approximate coefficient mean feature and the temperature approximate coefficient mean feature.

[0097] S5: The approximate coefficient mean feature is nonlinearly projected through the feature embedding network to obtain the projection feature e, which includes the oxygen potential projection feature e v and temperature projection characteristics e t , oxygen potential projection characteristic e v and temperature projection characteristics e t The calculation formulas are:

[0098] ;

[0099] ;

[0100] in, and t are the mean characteristic V_μ of the oxygen potential approximation coefficient approx and temperature approximation coefficient mean characteristic T_μ approx abbreviation of; and The dimensions of ; Represents a size of B×d model A real matrix of ; and The dimensions of ; and The dimensions of ; 、 、 and Both are learnable weight matrices used to perform nonlinear projection on the approximate coefficient mean features of the oxygen potential time subseries and the temperature time subseries; 、 、 and They are the first-layer weight and the second-layer weight when projecting the mean feature of the oxygen potential approximation coefficient and the mean feature of the temperature approximation coefficient respectively; 、 、 and are the first layer bias and the second layer bias when projecting the mean feature of the oxygen potential approximation coefficient and the mean feature of the temperature approximation coefficient, respectively; and The dimensions of , and The dimensions of ; GELU(.) is the Gaussian error activation function; B represents batch size, which is the number of samples input into the model at one time. In classification tasks, batch size is usually between 32 and 512; is the hidden layer dimension, such as 768.

[0101] S6: Project the oxygen potential feature e v and temperature projection characteristics e t Input to the position encoding layer for position encoding, and obtain the encoded projection feature E, which is expressed as:

[0102] ;

[0103] in, ; P is the learnable position encoding matrix;

[0104] S7: Extract the encoded projected features through the attention network to obtain nonlinear transformation enhanced features; the self-attention layer within the feature is used to extract different features separately; the cross-attention layer between features is used to extract different features in a mixed manner; the feedforward network layer is used to enhance the features; including:

[0105] S71: Extract the encoded projection features through the self-attention layer within the feature to obtain self-attention features, including:

[0106] S711: Obtain the encoded self-attention Attn based on the encoded projection feature E. The calculation formula is:

[0107] ;

[0108] ;

[0109] ;

[0110] ;

[0111] Among them, Q, K, and V are the query, key, and value of the self-attention mechanism respectively; 、 and are the self-attention learnable weight matrices used to generate queries, keys, and values, respectively; 、 and are the self-attention biases used to generate queries, keys, and values, respectively; Softmax(.) is the normalized exponential function; T is the transpose of the matrix; (head) is the head of the self-attention mechanism.

[0112] S712: Based on the encoded self-attention, obtain the self-attention update feature E intra , the calculation formula is:

[0113] ;

[0114] Among them, LayerNorm(.) is the layer normalization function, and Dropout(.) is the random regularization probability.

[0115] The self-attention update feature Update features by oxygen potential self-attention and temperature self-attention update features composition.

[0116] S72: Update the self-attention feature E intra Through the cross-attention layer between features, obtain the cross-attention update feature E inter ,include:

[0117] S721: Calculating the cross-attention of oxygen potential to temperature , the calculation formula is:

[0118] ;

[0119] in, Cross-attention mechanism query, key, and value representing oxygen potential to temperature, respectively; is the learnable weight matrix of the oxygen potential to temperature cross attention, which is used to generate query, key, and value respectively; is the cross-attention bias from oxygen potential to temperature; Softmax(.) is the normalized exponential function; 、 、 and The dimensions of .

[0120] S722: Calculating the cross-attention from temperature to oxygen potential , the calculation formula is:

[0121] ;

[0122] in, represent the cross-attention mechanism query, key, and value from temperature to oxygen potential, respectively; is the learnable weight matrix of the temperature-to-oxygen potential cross-attention, which is used to generate query, key, and value respectively; is the cross-attention bias from temperature to oxygen potential; Softmax(.) is the normalized exponential function; 、 、 and The dimensions of .

[0123] S723: Based on the cross attention from oxygen potential to temperature and the cross attention from temperature to oxygen potential, obtain the cross attention update feature E inter , the calculation formula is:

[0124] .

[0125] S73: Update the cross attention feature E inter Input into the feedforward network for nonlinear feature enhancement to obtain the nonlinear transformation enhanced feature F, which is calculated as follows:

[0126] ;

[0127] ;

[0128] Among them, W1 is the weight matrix of the extended dimension, ; W2 is the weight matrix of the compressed dimension, which maintains the consistency of input and output and enhances the nonlinear ability of the model. ; b1 and b2 are the biases of the expanded dimension and compressed dimension in the feedforward network respectively; H is the nonlinear transformation enhancement matrix, ; LayerNorm(.) is for layer normalization.

[0129] Among them, the input of the first Transformer network is the output of the position encoding layer;

[0130] In this embodiment, the attention network has Lx Transformer network layers.

[0131] Among them, nonlinear transformation enhances features Enhanced characteristic F by nonlinear transformation of oxygen potential v and temperature nonlinear transformation enhancement feature F t composition.

[0132] S8: Input the nonlinear transformation enhancement features obtained by the attention network into the dynamic feature fusion gating network for feature fusion to obtain the dynamic weighted fusion feature f; where the input of the dynamic feature fusion gating network is the output of the last Transformer network; including:

[0133] First, the oxygen potential nonlinear transformation enhancement feature F obtained by the Transformer network is vand temperature nonlinear transformation enhancement feature F t Input into the dynamic feature fusion gating network for feature fusion to obtain the dynamic weight matrix g, which is calculated as follows:

[0134] ;

[0135] Then, based on the dynamic weight matrix, the dynamic weighted fusion feature f is obtained. The calculation formula is:

[0136] ;

[0137] is the mapping probability function, ⊙ represents element-by-element multiplication, Represents the oxygen potential nonlinear transformation fusion feature F v and temperature nonlinear transformation fusion feature F t The feature vectors are concatenated to form a joint feature containing dual feature information; g is a dynamic weight matrix generated by the Sigmoid function, which is used to determine the fusion ratio of the final output feature; the dimensions of g and f are .

[0138] S9: Input the dynamic weighted fusion feature f into the output layer to obtain the classification prediction results of the oxygen state in liquid lead and bismuth , the calculation formula is:

[0139] ;

[0140] in, is the weight matrix of the classification head, which maps the high-dimensional dynamic weighted fusion features to the category space; The bias vector of the classification head adjusts the offset of the classification boundary; z is the unnormalized output of the classification head, which is used to calculate the loss (such as cross entropy loss); LayerNorm(.) is layer normalization; Softmax(.) is the normalized exponential function; z ; .

[0141] In order to train the model more efficiently and improve its classification ability in the oxygen state transition stage, especially in the transition from saturated state to unsaturated state, the category label is no longer a single discrete value, but shows a certain fuzziness and continuity. and Kullback-Leibler divergence loss This loss function can not only handle the traditional hard label classification problem, but also capture the fuzzy transition characteristics between categories through the soft label mechanism (softlabels).

[0142] The total number of training samples is N, the number of categories is C, and for each sample, the soft label probability distribution of oxygen state classification in real liquid lead and bismuth is , the predicted probability distribution of the oxygen state classification soft label in liquid lead bismuth is .

[0143] The loss function used in this study is the cross entropy loss with label smoothing and KL (Kullback-Leibler) divergence loss The weighted sum of is calculated as:

[0144] ;

[0145] Among them, α is a hyperparameter used to balance the two parts of the loss.

[0146] In order to reduce the overfitting of the model to the labels in the training data, we introduced the label smoothing technique, that is, using the label smoothing factor As a hyperparameter to generate a smooth target distribution, where =0.1. This not only enhances the robustness of the model, but also effectively reduces the adverse effects of mislabeling. The smoothed target probability of the jth category of the sample for:

[0147] ;

[0148] Where C is the total number of categories, refers to the probability value of the jth category in the sample; label smoothing technology enhances the generalization ability of the model by shifting the target label to a uniform distribution, thereby retaining a certain degree of error tolerance for each category during training; is the label smoothing factor. Based on the above smoothed target distribution, the cross entropy loss L CE Defined as:

[0149] ;

[0150] Among them, δ is a very small constant (δ=10e-7) used to avoid the zero value problem in logarithmic operations; is the smoothed target probability of the jth category in the i-th sample; is the predicted probability value of the jth category in the i-th sample.

[0151] In order to further utilize the probability distribution information of soft labels, the KL scattering loss is introduced Metric to measure the difference between the model prediction result and the soft label probability distribution:

[0152] ;

[0153] The true probability of classifying the jth category for the oxygen state in liquid lead bismuth in the i-th sample; the KL diffusion penalty term forces the model's prediction to be aligned with the original soft label probability distribution, enhancing the model's sensitivity to the state transition stage, enabling it to provide smoother decisions in the category boundary area

[0154] This design allows the model to simultaneously optimize the following objectives:

[0155] 1. Discriminative learning: Enhance the separability of category boundaries through label smoothing cross entropy loss.

[0156] 2. Distribution alignment: Keep the predicted distribution consistent with the original label through KL divergence loss.

[0157] To further demonstrate the effectiveness of the liquid lead-bismuth oxygen state classification model (GIDST) designed by the present invention, a comparative experiment was conducted. The experimental results are shown in Table 1. The experimental results show that the accuracy of the present model reached 95.4% compared with the comparative model. The accuracy, macro-average precision, macro-average recall, and macro-average F1 score indicators were all better than the comparative model. The overall model performance was superior to that of the common FCN and MLP models.

[0158] Table 1 Comparison results between the model of the present invention and the existing model

[0159] ;

[0160] Of course, it will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, but also encompasses the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and it is intended that all variations that fall within the meaning and range of equivalents of the claims be encompassed within the present invention. Any reference signs in the claims should not be construed as limiting the claim to which they relate.

[0161] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0162] The technology, shape, and structure not described in detail in the present invention are all well-known technologies.

Claims

1. A method for classifying oxygen states in liquid lead and bismuth, characterized in that: A model for classifying oxygen states in liquid lead and bismuth is used. The model includes a feature embedding network, a position encoding layer, an attention network, a dynamic feature fusion gating network, and an output layer. The attention network is composed of a stack of multiple Transformer networks. Between two adjacent Transformer networks, the output of the previous Transformer network serves as the input of the next Transformer network. Each Transformer network includes a sequentially connected intra-feature self-attention layer, an inter-feature cross-attention layer, and a feedforward network layer. The method for classifying oxygen states in liquid lead and bismuth includes: Obtain oxygen potential time subseries and temperature time subseries, and obtain approximate coefficient mean characteristics through wavelet transform; The approximate coefficient mean feature is nonlinearly projected through the feature embedding network to obtain the oxygen potential projection feature and temperature projection feature; The oxygen potential projection features and the temperature projection features are encoded through the position encoding layer to obtain the encoded projection features; The encoded projection features are extracted through the attention network to obtain the nonlinear transformation enhancement features; the nonlinear transformation enhancement features are obtained by the oxygen potential nonlinear transformation enhancement features F v and temperature nonlinear transformation enhancement feature F t composition; The nonlinear transformation enhancement features obtained by the attention network are input into the dynamic feature fusion gating network for feature fusion to obtain dynamic weighted fusion features; The dynamically weighted fusion features are input into the output layer to obtain the classification prediction results of oxygen state in liquid lead and bismuth.

2. The method for classifying oxygen states in liquid lead and bismuth according to claim 1, wherein: Before obtaining the oxygen potential time subseries and temperature time subseries, it also includes: Obtaining oxygen potential time series data and temperature time series data in a lead-bismuth alloy and adding noise disturbance to obtain enhanced oxygen potential time series data and temperature time series data; dividing the enhanced oxygen potential time series data and temperature time series data into oxygen potential time subseries and temperature time subseries with time continuity; Based on the oxygen potential time subsequence and temperature time subsequence, the hard labels describing the oxygen state classification are converted into oxygen state classification soft labels with fuzziness and continuity. The oxygen state classification results are marked to obtain an oxygen state classification dataset; the oxygen state classification soft labels are the probability distribution of several oxygen state categories.

3. The method for classifying oxygen states in liquid lead and bismuth according to claim 1, wherein: Obtain oxygen potential time subseries and temperature time subseries, and obtain approximate coefficient mean characteristics through wavelet transform, including: First, wavelet transform is used to decompose the oxygen potential time subseries and temperature time subseries to obtain the approximate coefficients. The calculation formula is: ； Where a is the decomposition layer number; m is the position index in each decomposition layer; is the approximate coefficient of the mth position in the ath decomposition layer; k is the position index of the filter coefficient, K is the total index number of the filter coefficient position; h k is the filter coefficient at the kth position; is the first in the a-1th decomposition layer The approximate coefficients of the positions; Then, based on the approximate coefficient, we get the approximate coefficient mean feature. , the calculation formula is: ； Where M is the number of approximation coefficients, , U is the length of the oxygen potential time subsequence or temperature time subsequence.

4. The method for classifying oxygen states in liquid lead and bismuth according to claim 1, wherein: The encoded projection feature E is extracted through the attention network to obtain nonlinear transformation enhancement features including: First, the encoded projection feature is extracted through the self-attention layer within the feature to obtain the self-attention update feature E intra , the self-attention update feature E intra Update features by oxygen potential self-attention and temperature self-attention update features composition; Then the self-attention update feature is passed through the cross-attention layer between features to obtain the cross-attention update feature E inter ; Then the cross attention is used to update the feature E inter Input to the feedforward network for nonlinear feature enhancement to obtain nonlinear transformation enhancement features; among them, nonlinear transformation enhancement features Enhanced characteristic F by nonlinear transformation of oxygen potential v and temperature nonlinear transformation enhancement feature F t composition; Pass the self-attention update feature through the cross-attention layer between features to obtain the cross-attention update feature E inter include: First, the feature is updated based on oxygen potential self-attention and temperature self-attention update features , calculate the cross attention of oxygen potential to temperature ; Secondly, based on the oxygen potential self-attention update feature and temperature self-attention update features , calculate the cross attention from temperature to oxygen potential ; Finally, the cross-attention based on oxygen potential to temperature and temperature to oxygen potential , get the cross attention update feature E inter ; Among them, the intra-feature self-attention layer is used to extract different features separately; the inter-feature cross-attention layer is used to extract different features in a mixed manner; and the feedforward network layer is used to enhance the features.

5. The method for classifying oxygen states in liquid lead and bismuth according to claim 1, wherein: The nonlinear transformation enhancement features obtained by the attention network are input into the dynamic feature fusion gating network for feature fusion. The dynamic weighted fusion features obtained include: First, the oxygen potential nonlinear transformation enhancement feature F obtained by the last Transformer network is v and temperature nonlinear transformation enhancement feature F t Input into the dynamic feature fusion gating network for feature fusion to obtain the dynamic weight matrix , the calculation formula is: ； Then, based on the dynamic weight matrix, the dynamic weighted fusion feature f is obtained. The calculation formula is: ； in, is the mapping probability function; ⊙ represents element-by-element multiplication; Represents the oxygen potential nonlinear transformation fusion feature F v and temperature nonlinear transformation fusion feature F t Perform feature vector splicing; g is the dynamic weight matrix generated by the Sigmoid function.

6. The method for classifying oxygen states in liquid lead and bismuth according to claim 1, wherein: Classification prediction results of oxygen states in liquid lead and bismuth The calculation formula is: ； in, It is a dynamic weighted fusion feature; is the weight matrix of the classification head; The bias vector of the classification head; z is the unnormalized output of the classification head; LayerNorm(.) is layer normalization; Softmax(.) is the normalized exponential function.

7. The method for classifying oxygen states in liquid lead and bismuth according to claim 2, wherein: The oxygen state classification data set is used to train the oxygen state classification model in liquid lead and bismuth using the loss function L. The calculation formula of the loss function L is: ； Among them, α is a hyperparameter; L CE is the cross entropy loss; L KL is the KL divergence loss; δ is the equilibrium constant; N is the total number of samples; i is the sample number; C is the total number of categories; j is the category number; is the smoothed target probability of oxygen state classification in liquid lead bismuth of the jth category in the i-th sample, , is the label smoothing factor; is the predicted probability of classifying the jth category of oxygen state in liquid lead bismuth in the i-th sample; is the true probability of classifying the j-th category for the oxygen state in liquid lead bismuth in the i-th sample.

8. The method for classifying oxygen states in liquid lead and bismuth according to claim 4, wherein: Cross-Attention from Oxygen Potential to Temperature The calculation formula is: ； in, represent the query, key, and value of the cross-attention mechanism from oxygen potential to temperature, respectively; is the oxygen potential to temperature cross-attention learnable weight matrix, used to generate query, key and value respectively; is the cross-attention bias from oxygen potential to temperature; Softmax(.) is the normalized exponential function; is the head of the self-attention mechanism; is the hidden layer dimension; Cross-attention from temperature to oxygen potential The calculation formula is: ； in, represent the query, key, and value of the cross-attention mechanism from temperature to oxygen potential, respectively; is the temperature to oxygen potential cross-attention learnable weight matrix, used to generate query, key and value respectively; is the cross attention bias from temperature to oxygen potential.

9. The method for classifying oxygen states in liquid lead and bismuth according to claim 4, wherein: Cross-attention update feature E inter The calculation formula is: ； Among them, LayerNorm(.) is the layer normalization function; Dropout(.) is the random regularization probability.

10. The method for classifying oxygen states in liquid lead and bismuth according to claim 1, wherein: The calculation formulas for oxygen potential projection characteristics and temperature projection characteristics are: ； ； Among them, e v is the oxygen potential projection feature; e t is the temperature projection feature; v and t are the oxygen potential approximation coefficient mean feature and the temperature approximation coefficient mean feature respectively; 、 、 and They are the first-layer weight matrix and the second-layer weight matrix when projecting the mean characteristics of the oxygen potential approximation coefficient and the mean characteristics of the temperature approximation coefficient, respectively; 、 、 and are the first layer bias and the second layer bias when projecting the mean feature of the oxygen potential approximation coefficient and the mean feature of the temperature approximation coefficient, respectively; (.) is the Gaussian error activation function.

Images