Search

CN-122020063-A - Neutron tube life prediction evaluation method based on machine learning

CN122020063ACN 122020063 ACN122020063 ACN 122020063ACN-122020063-A

Abstract

The invention discloses a neutron tube life prediction assessment method based on machine learning, and relates to the technical field of neutron tube life prediction. The method comprises the steps of collecting historical working condition data of the neutron tube, carrying out integrated clustering on the historical working condition data by using a plurality of clustering algorithms to obtain working condition clusters, collecting sequenced operating data of the neutron tube, carrying out feature selection to determine a life characterization vector template, extracting historical life characterization vectors of the neutron tube from the working condition clusters based on the life characterization vector template, constructing a cross-working condition life prediction model, training the cross-working condition life prediction model by using the historical life characterization vectors to obtain a trained cross-working condition life prediction model, collecting real-time life characterization vectors of the neutron tube, inputting the real-time life characterization vectors into the trained cross-working condition life prediction model, and outputting to obtain a life prediction evaluation result. According to the invention, the predicted life prediction evaluation result is generated through machine learning, so that the life prediction of the neutron tube is realized.

Inventors

  • LI KANG
  • LIU YANG
  • WANG ZIHAO

Assignees

  • 西安冠能中子探测技术有限公司

Dates

Publication Date
20260512
Application Date
20260310

Claims (10)

  1. 1. The neutron tube life prediction evaluation method based on machine learning is characterized by comprising the following steps of: step S1, collecting historical working condition data of a neutron tube, and carrying out integrated clustering on the historical working condition data by using a plurality of clustering algorithms to obtain a working condition cluster; s2, collecting sequential operation data of the neutron tube, performing feature selection on the sequential operation data, and determining a life characterization vector template; S3, extracting a historical service life representation vector of a neutron tube from a working condition cluster based on a service life representation vector template, constructing a cross-working condition service life prediction model based on a long-short-term memory network, and training the cross-working condition service life prediction model by using the historical service life representation vector to obtain a trained cross-working condition service life prediction model; And S4, collecting a real-time life representation vector of the neutron tube, inputting the real-time life representation vector into a trained cross-working condition life prediction model, and outputting to obtain a life prediction evaluation result.
  2. 2. The machine learning-based neutron tube life prediction evaluation method of claim 1, wherein the integrated clustering of the historical working condition data using a plurality of clustering algorithms to obtain working condition clusters comprises: collecting historical working condition data of a neutron tube, selecting representative points from the historical working condition data through a density peak clustering algorithm, and firstly calculating local density of each data point And further calculate the superior point distance for each data point Then, combining the local density and the upper-level point distance to obtain a comprehensive index By setting the parameter c, the data point of c% before the gamma value is selected as the representative point, and a representative set R= { is formed , ,..., }, Wherein Is a representative point; after the representative points are obtained, a sparse affinity sub-matrix between the data points and the representative points is constructed by adopting a nearest neighbor principle, and elements of the sparse affinity sub-matrix The calculation is as follows: ; Wherein, the Representing data points And representative point Similarity between; Is a data point in the original dataset; Is a representative point; Is that A set of K nearest representative points; sigma is a gaussian kernel parameter; Dividing the sparse affinity sub-matrix by a bipartite graph dividing method to obtain a plurality of base clustering results, integrating the base clustering results by consistency fusion, and finally outputting a dividing result of the working condition cluster; and unifying the degenerated expression caliber under different working conditions through the service life semantic tags.
  3. 3. The machine learning-based neutron tube life prediction assessment method of claim 2, wherein the operating mode fingerprint comprises: The working condition fingerprint is constructed by extracting the characteristic combination of each working condition cluster; Specifically, for each working condition cluster, feature vectors of all historical samples of the working condition clusters are calculated, and then representative fingerprint feature vectors and working condition fingerprints are obtained through principal component analysis or weighted average The preset rule for the working condition cluster k is as follows: ; Wherein, the Is a working condition fingerprint; is a fingerprint feature vector, and m is the total number of fingerprint feature vectors.
  4. 4. The machine learning-based neutron tube life prediction assessment method of claim 3, wherein the feature selection of the sequenced operational data, determining a life characterization vector template, comprises: The method comprises the steps of collecting sequential operation data of all working conditions, marking degradation states of each data segment according to life semantic labels, and determining a life representation vector template based on a feature extraction framework, wherein the feature extraction framework comprises the following calculation flows: the input layer receives the original time sequence data after pre-cleaning and normalization processing, the original time sequence data flows into the time sequence feature extraction layer, the first layer LSTM receives the input original time sequence data, and the hidden state generated by the first layer LSTM in the time step t is generated Inputting the second LSTM network to abstract at higher level to generate hidden state ; Enters the attention convergence layer, introduces an attention mechanism, and { a hidden state sequence outputted by the LSTM of the second layer Self-adaptive weighting is carried out to obtain the attention score of the template Attention score of the template Normalizing to obtain the attention weight of the template ; Weighting and summing the hidden states of all time steps to obtain a template context vector ; And at the projection and standardization output layer, performing linear transformation through a fully connected projection layer, performing layer normalization processing, and finally converting the template context vector c into a service life representation vector template z.
  5. 5. The machine learning-based neutron tube life prediction assessment method of claim 4, wherein constructing a cross-working condition life prediction model based on a long-term and short-term memory network comprises: the cross-working condition life prediction model takes a life characterization vector obtained based on a life characterization vector template as input, adopts a long-period memory network to capture a time sequence dependency relationship, and is combined with a self-attention mechanism to focus on a key degradation stage; Updating hidden state of long-term and short-term memory network layer at each time step t ; The self-attention mechanism is introduced to carry out weighted aggregation on the hidden state sequence output by the LSTM, and the original attention score is obtained through calculation Based on the original attention score Calculating attention weights And based on the attention weight Calculating a weighted context vector o; finally, life prediction value Output through a full connectivity layer: ; Wherein, the And Is the weight and bias parameters of the prediction layer; Is a life prediction value.
  6. 6. The machine learning-based neutron tube life prediction assessment method of claim 5, wherein training the cross-condition life prediction model using the historical life characterization vector to obtain a trained cross-condition life prediction model comprises: the cross-working condition life prediction model training is divided into two stages; In the source domain pre-training stage, firstly, based on a life representation vector template, a working condition cluster with a sufficient coverage of life samples is selected from working condition clusters to serve as a source working condition, a corresponding historical life representation vector is generated, and a source domain data set is formed In source domain data sets As training data, the cross-working condition life prediction model is trained for the first time, and a mean square error loss function is adopted during training ; In the target domain migration learning stage, aiming at the target working condition of data scarcity, a small amount of historical data is collected, and a target domain life characterization vector is generated according to a life characterization vector template to form a target domain data set Performing migration learning fine tuning on the basis of the cross-working condition life prediction model obtained by training in the first stage, and introducing the maximum mean difference loss in the fine tuning process ; Total loss function Combining mean square error loss And maximum mean difference loss The method is characterized by comprising the following steps: ; Wherein, the Is a super parameter.
  7. 7. The machine learning-based neutron tube life prediction assessment method of claim 6, wherein inputting the real-time life characterization vector into the trained cross-working condition life prediction model and outputting the obtained life prediction assessment result comprises: Collecting real-time life characterization data of target working conditions, wherein the real-time life characterization data are derived from real-time observation in the neutron tube operation process, and converting the real-time life characterization data into a real-time life characterization vector by applying a life characterization vector template ; Characterizing vectors for real-time life Input trained cross-working condition life prediction model Obtaining a life prediction value ; Finally, a life prediction evaluation result is output, wherein the life prediction evaluation result comprises a life prediction value, a trusted interval and a risk level.
  8. 8. The machine learning-based neutron tube life prediction assessment method of claim 7, wherein the trusted interval comprises: Generating a trusted interval by adopting a Monte Carlo Dropout method, starting a Dropout layer according to a selected Dropout rate p and carrying out forward propagation for N times during model reasoning to obtain N life prediction values Then calculate life prediction mean And lifetime prediction variance ; Mean value based on life prediction And lifetime prediction variance Calculating to obtain a trusted interval Wherein Is the standard deviation of life prediction, and 1.96 is the 97.5% quantile of the standard normal distribution.
  9. 9. The machine learning-based neutron tube life prediction assessment method of claim 8, wherein the risk level comprises: based on trusted interval and cosine similarity Dividing the risk level Q according to the prediction standard deviation And maximum similarity Dividing: ; Wherein, the And Is the minimum and maximum thresholds of the prediction standard deviation, And Is the minimum and maximum thresholds for cosine similarity.
  10. 10. The machine learning-based neutron tube life prediction assessment method of claim 9, wherein the cosine similarity comprises: Working condition drift is detected according to working condition fingerprint rules, and working condition fingerprints of real-time running states are calculated first ; Then, working condition fingerprints of real-time running states are calculated Working condition fingerprint of each working condition cluster Cosine similarity of (2) The calculation formula is as follows: ; Wherein, the The dot product is represented by a graph of the dot product, Representing the euclidean norm of the vector, Is the working condition fingerprint of the kth working condition cluster; If the maximum similarity is the same as the first threshold value Below a preset threshold Judging that working condition drift occurs, and outputting an adjusted conservative predicted value by the cross-working condition life prediction model 。

Description

Neutron tube life prediction evaluation method based on machine learning Technical Field The invention relates to the technical field of neutron tube life prediction, in particular to a neutron tube life prediction assessment method based on machine learning. Background The neutron tube is used as core key equipment in the fields of nuclear detection, material analysis and the like, and the running state of the neutron tube directly influences the stability and reliability of a related system. In practical application, the neutron tube is often operated under various complex working conditions, such as different working modes of continuous output type, narrow pulse high repetition type and the like, and the performance degradation rules of the equipment under different working conditions are obviously different, so that the prediction of the residual life of the neutron tube under a cross-working condition scene is accurately realized, and the method has become a core technical requirement for guaranteeing the safety of the equipment and reducing the operation and maintenance cost. The existing neutron tube life prediction and evaluation method often adopts a long-period and short-period memory network prediction method based on single working condition data. The method comprises the core processes of firstly collecting sequential operation data of the neutron tube under a certain fixed working condition, including output intensity, fluctuation amplitude and the like, then screening life representation vectors which are related to strong life degradation, finally training a long-period memory network model based on the life representation vectors under the single working condition, and predicting the residual life of the neutron tube under the same working condition by using the trained long-period memory network model. However, the long-short-term memory network prediction method based on the single working condition data only depends on the data under the single working condition to complete the long-short-term memory network model training, and the differentiation characteristic of the degradation rule of the neutron tube under different working conditions cannot be adapted, so that the degradation characteristic distribution learned by the long-short-term memory network model is obviously deviated from the actual characteristic distribution under the new working condition, and finally, the residual life prediction precision of the neutron tube under the cross-working condition scene is difficult to meet the requirements of engineering application on reliability and accuracy. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a neutron tube life prediction evaluation method based on machine learning, which solves the problems in the background art. In order to achieve the purpose, the invention is realized by the following technical scheme that the neutron tube life prediction evaluation method based on machine learning comprises the following steps: step S1, collecting historical working condition data of a neutron tube, and carrying out integrated clustering on the historical working condition data by using a plurality of clustering algorithms to obtain a working condition cluster; s2, collecting sequential operation data of the neutron tube, performing feature selection on the sequential operation data, and determining a life characterization vector template; S3, extracting a historical service life representation vector of a neutron tube from a working condition cluster based on a service life representation vector template, constructing a cross-working condition service life prediction model based on a long-short-term memory network, and training the cross-working condition service life prediction model by using the historical service life representation vector to obtain a trained cross-working condition service life prediction model; And S4, collecting a real-time life representation vector of the neutron tube, inputting the real-time life representation vector into a trained cross-working condition life prediction model, and outputting to obtain a life prediction evaluation result. Preferably, the method for performing integrated clustering on the historical working condition data by using a plurality of clustering algorithms, and obtaining the working condition cluster comprises: collecting historical working condition data of a neutron tube, selecting representative points from the historical working condition data through a density peak clustering algorithm, and firstly calculating local density of each data point And further calculate the superior point distance for each data pointThen, combining the local density and the upper-level point distance to obtain a comprehensive indexBy setting the parameter c, the data point of c% before the gamma value is selected as the representative point, and a representative set R= { is formed,,...,}, WhereinIs a represent