Method for classifying states of oxygen in liquid lead bismuth

Abstract

The invention discloses a method for classifying the state of oxygen in liquid lead and bismuth, and relates to the technical field of intelligent classification and recognition, and a model for classifying the state of oxygen in liquid lead and bismuth comprises a feature embedded network, a position coding layer, an attention network, a dynamic feature fusion gating network and an output layer. Wherein the attention network is formed by stacking a plurality of Transform networks, and each Transform network comprises a feature inner self-attention layer, an inter-feature cross attention layer and a feedforward network layer which are connected in sequence; the method for classifying the oxygen state in the liquid lead bismuth comprises the following steps of: performing wavelet transformation on an oxygen potential time subsequence and a temperature time subsequence to obtain an approximate coefficient mean value characteristic; enabling the approximate coefficient mean value features to sequentially pass through a feature embedding network, a position coding layer, an attention network, a dynamic feature fusion gating network and an output layer in a liquid lead and bismuth oxygen state classification model to obtain a liquid lead and bismuth oxygen state classification prediction result; the state transition region and the complex state boundary are classified, and the classification precision is improved.

Description

Technical Field

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

[0002] 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, the problem of material corrosion is serious. Maintaining a stable oxygen state is the key to its long-term safe operation. Therefore, identifying and monitoring the oxygen state in liquid lead-bismuth is the key to the intelligent operation and maintenance of the nuclear energy system; In the prior art, the oxygen state in liquid lead-bismuth is usually divided into three states: saturated, unsaturated, and sensor abnormal by using hard labels. The traditional hard-label classification method mainly relies on discrete labels for state recognition. However, this method is difficult to capture the ambiguity during state transitions. Especially when the oxygen state in liquid lead-bismuth is in the transition stage, the classification accuracy is often reduced due to the imbalance of data categories (saturated, unsaturated, sensor abnormal), and it cannot fully reflect the actual operation situation, resulting in low classification accuracy. Summary of the Invention

[0003] In order to overcome the defects of low classification accuracy of the oxygen state in liquid lead-bismuth and difficulty in capturing the ambiguity during the transition of the oxygen state in liquid lead-bismuth in the above-mentioned prior art, the present invention proposes a method for classifying the oxygen state in liquid lead-bismuth.

[0004] To achieve the above object, the present invention adopts the following technical solutions, including: A method for classifying the oxygen state in liquid lead-bismuth, characterized in that it uses an oxygen state classification model in liquid lead-bismuth, and the model includes 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 multiple stacked Transformer networks. Between two adjacent Transformer networks, the output of the previous Transformer network is used as the input of the next Transformer network. Each Transformer network includes a feature intra-self-attention layer, a feature inter-cross-attention layer, and a feed-forward network layer connected in sequence. The method for classifying the oxygen state in liquid lead-bismuth includes: Obtain the oxygen potential time subsequence and the temperature time subsequence, and obtain the approximate coefficient mean feature through wavelet transform; Perform non-linear projection on the approximate coefficient mean feature through the feature embedding network to obtain the oxygen potential projection feature and the temperature projection feature; Encode the oxygen potential projection feature and the temperature projection feature through the position encoding layer to obtain the encoded projection feature E; The encoded projection feature E is subjected to feature extraction through an attention network to obtain a non-linearly transformed enhanced feature; the non-linearly transformed enhanced feature consists of an oxygen potential non-linearly transformed enhanced feature F v and a temperature non-linearly transformed enhanced feature F t and is composed of; The non-linearly transformed enhanced feature obtained by the attention network is input into a dynamic feature fusion gating network for feature fusion to obtain a dynamically weighted fusion feature; The dynamically weighted fusion feature is input into an output layer to obtain a classification prediction result of the oxygen state in liquid lead-bismuth.

[0005] Preferably, before obtaining the oxygen potential time subsequence and the temperature time subsequence, it further includes: Obtaining the oxygen potential time series data and the temperature time series data in the lead-bismuth alloy and adding noise perturbations to obtain the enhanced oxygen potential time series data and the temperature time series data; The enhanced oxygen potential time series data and the temperature time series data are segmented into an oxygen potential time subsequence and a temperature time subsequence with time continuity; Based on the oxygen potential time subsequence and the temperature time subsequence, the hard label describing the oxygen state classification is converted into an oxygen state classification soft label with ambiguity and continuity, and the oxygen state classification result is marked to obtain an oxygen state classification data set; the oxygen state classification soft label is the probability distribution of several oxygen state categories.

[0006] Preferably, obtaining the oxygen potential time subsequence and the temperature time subsequence, and obtaining the approximate coefficient mean feature through wavelet transform includes: First, wavelet decomposition is performed on the oxygen potential time subsequence and the temperature time subsequence by using wavelet transform to obtain approximate coefficients, and the calculation formula is: ; where a is the decomposition layer number; m is the position index in each decomposition layer; is the approximate coefficient at the mth position in the ath decomposition layer; k is the position index of the filter coefficient, and K is the total number of filter coefficient positions; h k is the filter coefficient at the kth position; is the approximate coefficient at the th position in the (a - 1)th decomposition layer; Then, based on the approximate coefficients, the approximate coefficient mean feature is obtained, and the calculation formula is: ; where M is the number of approximate coefficients, , and U is the length of the oxygen potential time subsequence or the temperature time subsequence.

[0007] Preferably, the encoded projection feature E is subjected to feature extraction through an attention network to obtain a non-linear transformation enhanced feature, including: First, the encoded projection feature is subjected to feature extraction through an intra-feature self-attention layer to obtain a self-attention feature E intra , and the self-attention updated feature E intra consists of an oxygen potential self-attention updated feature and a temperature self-attention updated feature ; Next, the self-attention updated feature is passed through an inter-feature cross-attention layer to obtain a cross-attention updated feature E inter ; Then, the cross-attention updated feature E inter is input into a feed-forward network for non-linear feature enhancement to obtain a non-linear transformation enhanced feature; among them, the non-linear transformation enhanced feature consists of an oxygen potential non-linear transformation enhanced feature F v and a temperature non-linear transformation enhanced feature F t ; Passing the self-attention updated feature through the inter-feature cross-attention layer to obtain the cross-attention updated feature E inter includes: First, based on the oxygen potential self-attention updated feature and the temperature self-attention updated feature , calculate the cross-attention from oxygen potential to temperature ; Secondly, based on the oxygen potential self-attention updated feature and the temperature self-attention updated feature , calculate the cross-attention from temperature to oxygen potential ; Finally, based on the cross-attention from oxygen potential to temperature and the cross-attention from temperature to oxygen potential , obtain the cross-attention updated feature E inter ; Among them, the intra-feature self-attention layer is used to separately extract different features; the inter-feature cross-attention layer is used to mix and extract different features; the feed-forward network layer is used to enhance the features; Preferably, the non-linear transformation enhanced feature obtained by the attention network is input into a dynamic feature fusion gating network for feature fusion to obtain a dynamically weighted fusion feature, including: First, the oxygen potential non-linear transformation enhanced feature F v and the temperature non-linear transformation enhanced feature F t obtained by the last Transformer network are input into the dynamic feature fusion gating network for feature fusion to obtain a dynamic weight matrix , the calculation formula is: ; Then, based on the dynamic weight matrix, the dynamic weighted fusion feature f is obtained, and the calculation formula is: ; Among them, is the mapping probability function; ⊙ represents element-wise multiplication; represents the fused feature F of the non-linear transformation of the oxygen potential v and the fused feature F of the non-linear transformation of the temperature t perform eigenvector splicing; g is the dynamic weight matrix generated by the Sigmoid function.

[0008] Preferably, the prediction result of the oxygen state classification in liquid lead-bismuth has the calculation formula: ; Among them, is the 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(.) performs layer normalization; Softmax(.) is the normalization exponential function.

[0009] Preferably, the obtained oxygen state classification data set is used to train the oxygen state classification model in liquid lead-bismuth using the loss function L, and 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 balance 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 the oxygen state classification in the i-th sample for the j-th category of liquid lead-bismuth, , is the label smoothing factor; is the predicted probability of the j-th category of the oxygen state classification in the i-th sample of liquid lead-bismuth; is the true probability of the j-th category of the oxygen state classification in the i-th sample of liquid lead-bismuth.

[0010] Preferably, the cross-attention from oxygen potential to temperature has the calculation formula: ; Among them, respectively represent the query, key, and value of the cross-attention mechanism from oxygen potential to temperature; is the learnable weight matrix for the cross-attention from oxygen potential to temperature, and is used to generate queries, keys, and values 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: ; Among them, respectively represent the queries, keys, and values of the cross-attention mechanism from temperature to oxygen potential; is the learnable weight matrix for the cross-attention from temperature to oxygen potential, and is used to generate queries, keys, and values respectively; is the cross-attention bias from temperature to oxygen potential.

[0011] Preferably, the cross-attention updated feature E inter The calculation formula is: ; Among them, LayerNorm(.) is the layer normalization function; Dropout(.) is the random regularization probability.

[0012] Preferably, the calculation formulas for the oxygen potential projection feature and the temperature projection feature are respectively: ; ; Among them, e v is the oxygen potential projection feature; e t is the temperature projection feature; v and t are respectively the oxygen potential approximate coefficient mean feature and the temperature approximate coefficient mean feature; , , and are respectively the first layer weight matrix and the second layer weight matrix when projecting the oxygen potential approximate coefficient mean feature and the temperature approximate coefficient mean feature; , , and are respectively the first layer bias and the second layer bias when projecting the oxygen potential approximate coefficient mean feature and the temperature approximate coefficient mean feature; (.)is the Gaussian error activation function.

[0013] The advantages of the present invention are: (1) The present invention constructs a classification model for the oxygen state in liquid lead-bismuth. 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 multiple 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 intra-self-attention layer, a feature inter-cross-attention layer, and a feed-forward network layer connected in sequence. A feature inter-cross-attention layer is added to the basic Transformer network, and a dynamic feature fusion gating network is added after the Transformer network. Based on the self-attention updated features, the cross-attention updated features are obtained through the feature inter-cross-attention layer. The cross-attention updated features are input into the feed-forward network for non-linear feature enhancement, and the non-linearly transformed enhanced features F of the oxygen potential and the non-linearly transformed enhanced features F of the temperature obtained in the attention network v are adaptively fused to obtain a dynamic weight matrix. Based on the dynamic weight matrix, a dynamically weighted fusion feature is obtained. Finally, the classification prediction result of the oxygen state in liquid lead-bismuth is obtained through the dynamically weighted fusion feature. The probability distribution is used to describe the over-transition stage of the oxygen state in liquid lead-bismuth, so as to realize the classification of the state transition region and the complex state boundary, and at the same time improve the classification accuracy. t

[0014] (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 relationships between the features, thereby enhancing the ability to capture complex state boundaries and local features.

[0015] (3) The present invention is trained through a weighted combination of cross-entropy and KL divergence loss, which can not only ensure the accuracy of the model in probability prediction but also improve the distribution alignment and reduce the risk of overfitting.

[0016] (4) The present invention uses the oxygen potential and temperature in the lead-bismuth alloy to evaluate the oxygen state in liquid lead-bismuth. By processing the oxygen potential time series and the temperature time series through wavelet transform, the approximate coefficient mean features that effectively capture the low-frequency trend features of the signals are obtained, overcoming the defect of the singularity of traditional time-domain / frequency-domain analysis, and is especially suitable for the complex oxygen state dynamic change scenario of the lead-bismuth eutectic system.

[0017] (5) The present invention uses a feature embedding network to perform non - linear projection on approximate coefficients, mapping the original features to a high - dimensional hidden space, solving the problem of insufficient feature separability of traditional linear projections (such as PCA) in the oxygen state classification task, and significantly improving the joint representation ability of oxygen potential features and temperature features.

[0018] (6) The present invention uses a position encoding layer to introduce temporal position information, making up for the problem of temporal information loss caused by permutation invariance in traditional Transformers when dealing with time series, enabling the model to accurately identify the temporal dependence of oxygen state changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is the model architecture diagram of the present invention; Figure 2 is the technical flow chart of the present invention; Figure 3 is the feature extraction result diagram of the present invention; Figure 4 is the comparison effect diagram between the model of the present invention and the existing model. DETAILED DESCRIPTION OF THE INVENTION

[0020] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0021] 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, the problem of material corrosion is serious. Therefore, maintaining a stable oxygen state is the key to its long - term safe operation. Since it is difficult to accurately depict the transitional ambiguity of the oxygen state in lead - bismuth alloy and it is impossible to adaptively focus on the key features extracted, the present invention proposes a method for classifying the oxygen state in liquid lead - bismuth. Using an oxygen state classification model in liquid lead - bismuth, the oxygen state classification model in liquid lead - bismuth 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 Lx stacked Transformer networks. Between two adjacent Transformer networks, the output of the previous Transformer network is used as the input of the next Transformer network; each Transformer network includes a feature intra - self - attention layer, a feature inter - cross - attention layer, and a feed - forward network layer connected in sequence. The input of the oxygen state classification model in liquid lead - bismuth is the mean feature of oxygen potential approximate coefficients and the mean feature of temperature approximate coefficients, and the output is the oxygen state classification prediction result in liquid lead - bismuth.

[0022] As Figures 1-4 shown, a method for classifying the oxygen state in liquid lead-bismuth includes: S1: Obtain the oxygen potential time series data and temperature time series data in the lead-bismuth alloy, and use an enhancement method based on smooth perturbation to obtain the enhanced oxygen potential time series data and temperature time series data.

[0023] Since the oxygen state is usually described by hard labels for classification, including saturated state, unsaturated state, and sensor anomaly. Due to the small amount of saturated state data and sensor anomaly data, the accuracy of the classification model decreases. To improve the robustness of the model to the fluctuations of oxygen potential and temperature in the lead-bismuth alloy under different scenarios, the present invention adopts a time series data enhancement method based on smooth perturbation, and the perturbation signal has the following expression: ; where A represents the perturbation amplitude, f represents the perturbation frequency, t is the current time step, and N is the length of the oxygen potential time series data set or the temperature time series data set in the lead-bismuth alloy.

[0024] Apply the perturbation signal to the oxygen potential time series data and temperature time series data in the lead-bismuth alloy respectively to obtain the enhanced oxygen potential time series data and temperature time series data, so as to simulate the possible dynamic changes in the actual operating environment, and solve the problem of the decrease in classification accuracy caused by the class imbalance of the oxygen potential time series data and temperature time series data in the lead-bismuth alloy.

[0025] S2: Use a sliding window to divide the enhanced oxygen potential time series data and temperature time series data into oxygen potential time subsequences and temperature time subsequences with time continuity.

[0026] The length of the sliding window is not limited, and the step size is usually 1.

[0027] This method can capture the dynamic change characteristics of the oxygen potential time series data and temperature time series data, and at the same time lay a foundation for subsequent feature extraction. Use the sliding window combined with wavelet transform to extract the time-frequency features of the oxygen potential time series data and temperature time series data, and convert the hard label into a soft label to represent the ambiguity of the transition stage.

[0028] S3: Based on the oxygen potential time subsequences and temperature time subsequences, convert the hard labels describing the oxygen state classification into soft labels of oxygen state classification with ambiguity and continuity, mark the oxygen state classification results, and obtain the oxygen state classification data set.

[0029] To improve the classification ability of the model during the oxygen state transition stage (such as from saturated to unsaturated state), the hard labels (saturated, unsaturated, sensor anomaly) describing the oxygen state are converted from single discrete values to soft labels with ambiguity and continuity. By normalizing the frequency of each category, it is ensured that the labels of each sample satisfy the probability distribution characteristics.

[0030] Soft label classification means that during the training process of the model, instead of using traditional discrete hard labels, probability distributions are used to describe the ambiguity and uncertainty of the labels, thereby realizing a classification method for the state transition region and complex state boundaries, which is mainly applied to the field of oxygen concentration state classification in lead-bismuth reactors. Among them, soft label classification has the advantages of being able to capture the ambiguity in the state transition stage, providing a fine-grained expression of uncertainty; reducing the overfitting risk brought by hard labels through probability distribution training, and enhancing the generalization ability of the model.

[0031] In the present invention, the hard label categories of the oxygen state are divided into three categories, and the oxygen state classification result y is expressed as: ; where 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 oxygen state category frequency in the hard label.

[0032] S4: Based on the oxygen potential time subsequence and the temperature time subsequence, the approximate coefficient mean features are obtained through wavelet transform, and the approximate coefficient mean features include the oxygen potential approximate coefficient mean feature V_μ approx and the temperature approximate coefficient mean feature T_μ approx ; Before calculating the approximate coefficient mean features, first, the oxygen potential time subsequence and the temperature time subsequence are wavelet decomposed using wavelet transform to obtain the approximate coefficients, The calculation formula is: ; where a is the decomposition layer number, a≥1, and the value range in this embodiment is [1, 2, 3], and the initial layer A0 is the original signal; m is the position index in each decomposition layer; is the approximate coefficient at the mth position in the ath decomposition layer; k is the position index of the filter coefficient, and the value range is [0, 1,..., 15]; h k is the filter coefficient at the kth position; is the approximate coefficient at the th position in the (a - 1)th decomposition layer; The approximate coefficients of each decomposition layer are obtained through the convolution of the coefficients of the previous decomposition layer and the filter coefficients and downsampling (step size is 2).

[0033] Approximate coefficient mean feature The calculation formula is as follows: ; where M is the number of approximate coefficients, , and U is the length of the oxygen potential time subsequence or the temperature time subsequence.

[0034] Wavelet transform can analyze the local dynamic characteristics of the oxygen potential time subsequence and the temperature time subsequence at different times, and realize the feature capture of the signal in the local time domain and frequency domain.

[0035] In this embodiment, Symlets8 (sym8) wavelet is used as the wavelet basis function to extract and decompose the time-frequency features of the oxygen potential time subsequence and the temperature time subsequence respectively, and obtain the oxygen potential approximate coefficient mean feature and the temperature approximate coefficient mean feature.

[0036] S5: Project the approximate coefficient mean feature through the feature embedding network to obtain the projection feature e. The projection feature includes the oxygen potential projection feature e v and the temperature projection feature e t . The calculation formulas for the oxygen potential projection feature e v and the temperature projection feature e t are respectively: ; ; where, and t are the abbreviations of the oxygen potential approximate coefficient mean feature V_μ approx and the temperature approximate coefficient mean feature T_μ approx respectively; and both have dimensions of ; represents a real number matrix of size B×d model ; and both have dimensions of ; and both have dimensions of ; , , and are all learnable weight matrices used for non-linearly projecting the approximate coefficient mean features of the oxygen potential time subsequence and the temperature time subsequence; , , and are respectively the first-layer weight and the second-layer weight for projecting the oxygen potential approximate coefficient mean feature and the temperature approximate coefficient mean feature; , , and are the first - layer bias and the second - layer bias respectively when projecting the mean characteristics of the oxygen potential approximation coefficient and the mean characteristics of the temperature approximation coefficient; and both have dimensions of , and both have dimensions of ; GELU(.) is the Gaussian error activation function; B represents the batch size, which is the number of samples input into the model at one time. In classification tasks, the Batch Size is usually between 32 - 512; is the hidden - layer dimension, such as 768.

[0037] S6: Input the oxygen - potential projection feature e v and the temperature - projection feature e t into the position - encoding layer for position encoding to obtain the encoded projection feature E, and the expression is: ; where ; P is the learnable position - encoding matrix; S7: Extract features from the encoded projection feature through the attention network to obtain non - linear transformation enhanced features; The intra - feature self - attention layer is used to separately extract different features; The inter - feature cross - attention layer is used to mix - extract different features; The feed - forward network layer is used to enhance the features; including: S71: Extract features from the encoded projection feature through the intra - feature self - attention layer to obtain self - attention features, including: S711: Obtain the encoded self - attention Attn based on the encoded projection feature E, and the calculation formula is: ; ; ; ; where Q, K, and V are the query (Query), key (Key), and value (Value) of the self - attention mechanism respectively; , and are the self - attention learnable weight matrices for generating the query, key, and value respectively; , and are the self - attention biases for generating the query, key, and value respectively; Softmax(.) is the normalization exponential function; T is the transpose of the matrix; The (head) is the head of the self-attention mechanism.

[0038] S712: Based on the encoded self-attention, obtain the self-attention updated feature E intra , and the calculation formula is: ; where LayerNorm(.) is the layer normalization function and Dropout(.) is the random regularization probability.

[0039] The self-attention updated feature consists of the oxygen potential self-attention updated feature and the temperature self-attention updated feature .

[0040] S72: Pass the self-attention updated feature E intra through the cross-attention layer between features to obtain the cross-attention updated feature E inter , including: S721: Calculate the cross-attention from oxygen potential to temperature , and the calculation formula is: ; where respectively represent the query, key, and value of the cross-attention mechanism from oxygen potential to temperature; is the learnable weight matrix for the cross-attention from oxygen potential to temperature, which is used to generate the query (Query), key (Key), and value (Value) respectively; is the bias of the cross-attention from oxygen potential to temperature; Softmax(.) is the normalization exponential function; , , and all have dimensions of .

[0041] S722: Calculate the cross-attention from temperature to oxygen potential , and the calculation formula is: ; where respectively represent the query, key, and value of the cross-attention mechanism from temperature to oxygen potential; is the learnable weight matrix for the cross-attention from temperature to oxygen potential, which is used to generate the query (Query), key (Key), and value (Value) respectively; is the bias of the cross-attention from temperature to oxygen potential; Softmax(.) is the normalization exponential function; , , and All dimensions are .

[0042] S723: Obtain the cross-attention updated feature E based on the cross-attention from oxygen potential to temperature and the cross-attention from temperature to oxygen potential. The calculation formula is: inter , and the calculation formula is: .

[0043] S73: Input the cross-attention updated feature E inter into the feed-forward network for non-linear feature enhancement to obtain the non-linearly transformed enhanced feature F. The calculation formula is: ; ; where, W1 is the weight matrix for expanding the dimension, ; W2 is the weight matrix for compressing the dimension, maintaining the input-output consistency and enhancing the non-linear ability of the model. ; b1 and b2 are the biases for the expanding dimension and the compressing dimension in the feed-forward network respectively; H is the non-linearly transformed enhanced matrix, ; LayerNorm(.) is for layer normalization.

[0044] where, the input of the first Transformer network is the output of the position encoding layer; In this embodiment, the attention network has Lx Transformer network layers.

[0045] where, the non-linearly transformed enhanced feature is composed of the oxygen potential non-linearly transformed enhanced feature F v and the temperature non-linearly transformed enhanced feature F t .

[0046] S8: Input the non-linearly transformed enhanced feature obtained by the attention network into the dynamic feature fusion gating network for feature fusion to obtain the dynamically weighted fusion feature f; where, the input of the dynamic feature fusion gating network is the output of the last Transformer network; including: First, input the oxygen potential non-linearly transformed enhanced feature F v and the temperature non-linearly transformed enhanced feature F t obtained by the Transformer network into the dynamic feature fusion gating network for feature fusion to obtain the dynamic weight matrix g. The calculation formula is: ; Then, based on the dynamic weight matrix, obtain the dynamically weighted fusion feature f. The calculation formula is: ; is the mapping probability function, ⊙ represents element-wise multiplication, represents the fused feature F of the non-linear transformation of the oxygen potential v and the fused feature F of the non-linear transformation of the temperature t perform eigenvector concatenation to form a joint feature containing dual feature information; g is the dynamic weight matrix generated by the Sigmoid function and is used to determine the fusion ratio of the final output feature; the dimensions of both g and f are .

[0047] S9: Input the dynamically weighted fused feature f into the output layer to obtain the classification prediction result of the oxygen state in liquid lead-bismuth , and the calculation formula is: ; wherein, is the weight matrix of the classification head, which maps the high-dimensional dynamically weighted fused feature to the category space; is the bias vector of the classification head, which adjusts the offset of the classification boundary; z is the unnormalized output of the classification head and is used to calculate the loss (such as cross-entropy loss); LayerNorm(.) performs layer normalization; Softmax(.) is the normalized exponential function; z ; .

[0048] In order to train the model more efficiently and improve its classification ability during the oxygen state transition phase, especially during the transition from the saturated state to the non-saturated state, the class label is no longer a single discrete value but shows a certain degree of ambiguity and continuity. A hybrid loss function based on cross-entropy loss and Kullback-Leibler divergence loss is designed. This loss function can not only handle traditional hard-label classification problems but also capture the fuzzy transition characteristics between classes through the soft-label mechanism (soft labels).

[0049] The total number of training samples is N, and the number of classes is C. For each sample, the soft-label probability distribution of the oxygen state classification in the real liquid lead-bismuth is , and the predicted soft-label probability distribution of the oxygen state classification in the liquid lead-bismuth is .

[0050] The loss function adopted in this study is the weighted sum of the cross-entropy loss with label smoothing and the KL (Kullback-Leibler) divergence loss , and the calculation formula is: ; Among them, α is a hyperparameter used to balance the two parts of the loss.

[0051] To reduce the overfitting of the model to the labels in the training data, we introduce the label smoothing technique, that is, using the label smoothing factor as a hyperparameter to generate a smoothed target distribution, where = 0.1. This not only enhances the robustness of the model but also effectively reduces the adverse effects brought by mislabeling. The smoothed target probability for the j-th class of the sample is: ; where C is the total number of classes, refers to the probability value of the j-th class in the sample; the label smoothing technique enhances the generalization ability of the model by shifting the target label to a uniform distribution, thus retaining a certain degree of fault tolerance for each class during training; is the label smoothing factor. Based on the above smoothed target distribution, the cross-entropy loss L CE is defined as: ; where 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 j-th class in the i-th sample; is the predicted probability value of the j-th class in the i-th sample.

[0052] To further utilize the probability distribution information of the soft labels, the KL divergence loss metric is introduced to measure the difference between the model prediction result and the soft label probability distribution: ; is the true probability of the j-th class of the oxygen state classification in the liquid lead-bismuth of the i-th sample; the KL divergence penalty term forces the model's prediction to align with the original soft label probability distribution, enhancing the model's sensitivity to the state transition stage and enabling it to provide a smoother decision in the category boundary region This design allows the model to simultaneously optimize the following objectives: 1. Discriminative learning: Enhance the separability of category boundaries through the label smoothing cross-entropy loss.

[0053] 2. Distribution alignment: Maintain the consistency between the prediction distribution and the original label through the KL divergence loss.

[0054] To further prove the effectiveness of the oxygen state classification model (GIDST) designed in the present invention for liquid lead-bismuth, the present invention conducted a comparative experiment. The experimental results are shown in Table 1. The experimental results show that the accuracy rate of the model of the present invention reaches 95.4% compared with the comparative model. The accuracy rate, macro-average precision rate, macro-average recall rate, and macro-average F1-score indicators are all superior to the comparative model. The overall model performance is more superior than that of common FCN and MLP models.

[0055] Table 1 Comparison results between the model of the present invention and the existing model ; Of course, for those skilled in the art, the present invention is not limited to the details of the above exemplary embodiments, but also includes the same or similar structures that can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be embraced within the present invention. Any reference signs in the claims should not be construed as limiting the claimed rights.

[0056] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for 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 embodiments that can be understood by those skilled in the art.

[0057] The technologies, shapes, and structures not described in detail in the present invention are all well-known technologies.

Claims

1. A method for classifying the oxygen state in liquid lead-bismuth, characterized in that, Using an oxygen state classification model in liquid lead-bismuth, the model includes 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 multiple 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 intra-self-attention layer, a feature inter-cross-attention layer, and a feed-forward network layer connected in sequence. The method for classifying the oxygen state in liquid lead-bismuth includes: Obtain the oxygen potential time subsequence and the temperature time subsequence, and obtain the approximate coefficient mean feature through wavelet transform; Non-linearly project the approximate coefficient mean feature through the feature embedding network to obtain the oxygen potential projection feature and the temperature projection feature; Encode the oxygen potential projection feature and the temperature projection feature through the position encoding layer to obtain the encoded projection feature; Extract features from the encoded projection features through the attention network to obtain non-linear transformation enhanced features; the non-linear transformation enhanced features consist of the non-linear transformation enhanced feature F of oxygen potential v and the non-linear transformation enhanced feature F of temperature t and are composed of Input the non-linearly transformed enhanced feature obtained by the attention network into the dynamic feature fusion gating network for feature fusion to obtain the dynamically weighted fusion feature; Input the dynamically weighted fusion feature into the output layer to obtain the oxygen state classification prediction result in liquid lead-bismuth.

2. The method for classifying the oxygen state in liquid lead-bismuth according to claim 1, characterized in that, Before obtaining the oxygen potential time subsequence and the temperature time subsequence, it further includes: Obtain the oxygen potential time series data and the temperature time series data in the lead-bismuth alloy and add noise perturbation to obtain the enhanced oxygen potential time series data and temperature time series data; Segment the enhanced oxygen potential time series data and temperature time series data into oxygen potential time subsequences and temperature time subsequences with time continuity; Based on the oxygen potential time subsequence and the temperature time subsequence, convert the hard label describing the oxygen state classification into an oxygen state classification soft label with ambiguity and continuity, mark the oxygen state classification result, and obtain the oxygen state classification data set; the oxygen state classification soft label is the probability distribution of several oxygen state categories.

3. A method for classifying the oxygen state in liquid lead-bismuth according to claim 1, characterized in that, Obtain the oxygen potential time subsequence and the temperature time subsequence, and obtain the approximate coefficient mean feature through wavelet transform, including: First, use wavelet transform to perform wavelet decomposition on the oxygen potential time subsequence and the temperature time subsequence to obtain the approximate coefficient. The calculation formula is: ； Among them, a is the decomposition layer number; m is the position index in each decomposition layer; is the approximation coefficient at the m-th position in the a-th decomposition layer; k is the position index of the filter coefficient, and K is the total number of filter coefficient position indices; h k is the filter coefficient at the k-th position; is the approximation coefficient at the -th position in the (a - 1)-th decomposition layer; Then, based on the approximation coefficients, the mean feature of the approximation coefficients is obtained , and the calculation formula is as follows: ； where M is the number of approximation coefficients, , and U is the length of the oxygen potential time subsequence or the temperature time subsequence.

4. A method for classifying the oxygen state in liquid lead-bismuth as described in claim 1, characterized in that, Extract the feature of the encoded projection feature E through the attention network to obtain the non-linearly transformed enhanced feature, including: First, the encoded projection features are subjected to feature extraction through the intra-feature self-attention layer to obtain the self-attention feature E intra , the self-attention updated feature E intra consists of the oxygen potential self-attention updated feature and the temperature self-attention updated feature ; Next, the self-attention updated features are passed through the cross-attention layer between features to obtain the cross-attention updated features E inter ; Then update the cross-attention feature E inter and input it into the feed-forward network for non-linear feature enhancement to obtain the non-linearly transformed enhanced feature; among them, the non-linearly transformed enhanced feature consists of the non-linearly transformed enhanced feature F of the oxygen potential v and the non-linearly transformed enhanced feature F of the temperature t ; The self-attention updated features are passed through the feature cross-attention layer to obtain the cross-attention updated feature E inter including: First, update the features based on the oxygen potential self-attention and the temperature self-attention to update the features , and calculate the cross-attention from the oxygen potential to the temperature ; Secondly, update the features based on the oxygen potential self-attention and the temperature self-attention to update the features , and calculate the cross-attention from temperature to oxygen potential ; Finally, based on the cross-attention from oxygen potential to temperature and the cross-attention from temperature to oxygen potential , the cross-attention updated feature E is obtained inter ; Among them, the feature intra-self-attention layer is used to separately extract different features; the feature inter-cross-attention layer is used to mix and extract different features; the feed-forward network layer is used to enhance the features.

5. A method for classifying the oxygen state in liquid lead-bismuth as described in claim 1, characterized in that, Input the non-linearly transformed enhanced feature obtained by the attention network into the dynamic feature fusion gating network for feature fusion to obtain the dynamically weighted fusion feature, including: First, the non-linearly transformed and enhanced oxygen potential feature F obtained from the last Transformer network v and the non-linearly transformed and enhanced temperature feature F t are input into the dynamic feature fusion gating network for feature fusion to obtain the dynamic weight matrix , and the calculation formula is as follows: ； Then, based on the dynamic weight matrix, obtain the dynamically weighted fusion feature f. The calculation formula is: ； Among them, is the mapping probability function; ⊙ represents element-wise multiplication; represents the fused feature F of the non-linear transformation of the oxygen potential v F t performs eigenvector concatenation; g is the dynamic weight matrix generated by the Sigmoid function.

6. The method for classifying the oxygen state in liquid lead-bismuth according to claim 1, wherein Prediction Results of Oxygen State Classification in Liquid Lead-Bismuth The calculation formula is as follows: ； Among them, is the dynamically weighted fusion feature; is the weight matrix of the classification head; is the bias vector of the classification head; z is the unnormalized output of the classification head; LayerNorm(.) is for layer normalization; Softmax(.) is the normalization exponential function.

7. The method for classifying the oxygen state in liquid lead-bismuth according to claim 2, wherein, Use the obtained oxygen state classification data set to train the oxygen state classification model in liquid lead-bismuth using the loss function L. The calculation formula of the loss function L is: ； where α is a hyperparameter; L CE is the cross-entropy loss; L KL is the KL divergence loss; δ is the balance 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 the oxygen state classification in the j-th category of liquid lead-bismuth in the i-th sample, , is the label smoothing factor; is the predicted probability of the j-th category of the oxygen state classification in liquid lead-bismuth in the i-th sample; is the true probability of the j-th category of the oxygen state classification in liquid lead-bismuth in the i-th sample.

8. The method for classifying the oxygen states in liquid lead-bismuth according to claim 4, wherein, Cross-attention of oxygen potential to temperature The calculation formula is as follows: ； Among them, respectively represent the query, key, and value of the cross-attention mechanism from oxygen potential to temperature; is the learnable weight matrix for cross-attention from oxygen potential to temperature, which is used to generate query, key, and value respectively; is the bias of the cross-attention 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 as follows: ； Among them, respectively represent the query, key, and value of the cross-attention mechanism from temperature to oxygen potential; is the learnable weight matrix for cross-attention from temperature to oxygen potential, which is used to generate the query, key, and value respectively; is the cross-attention bias from temperature to oxygen potential.

9. A method for classifying the oxygen state in liquid lead-bismuth according to claim 4, characterized in that, Cross-attention updates feature E inter The calculation formula is as follows: ； Among them, LayerNorm(.) is the layer normalization function; Dropout(.) is the random regularization probability.

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

Images