CN-121997177-A - Intelligent diagnosis method, system and storage medium for faults in krypton-xenon extraction process

CN121997177ACN 121997177 ACN121997177 ACN 121997177ACN-121997177-A

Abstract

The invention provides an intelligent diagnosis method, a system and a storage medium for faults in a krypton-xenon extraction process, which are characterized in that multivariate time sequence data in the krypton-xenon extraction process are obtained, a multi-scale residual dense cavity convolution network is input to extract a high-dimensional depth feature sequence, the network uses cavity convolutions with different expansion rates in different residual blocks and fuses the multi-scale time sequence features through dense connection, a channel attention and time self-attention mechanism is utilized to weight the feature sequence to obtain time weighted features, meanwhile, a feature dimension attention score is calculated based on modeling of covariance among feature channels to obtain variable associated weighted features, the time weighted features and the variable associated weighted features are fused in a first outer product mode, an original depth feature and the time weighted features are fused in a second outer product mode, fault expression vectors are generated through splicing of the two fusion results, and the fault expression vectors are input into a classifier to output fault type diagnosis results.

Inventors

XU HONGBO
CHEN QINGSHAN
ZHONG JIANMING
GUO WEI

Assignees

北京舞水科技有限公司
河南科益气体股份有限公司

Dates

Publication Date: 20260508
Application Date: 20260129

Claims (10)

1. The intelligent diagnosis method for the faults in the krypton-xenon extraction process is characterized by comprising the following steps of: The method comprises the steps of obtaining time sequence data formed by a plurality of process variables in a krypton-xenon extraction process, inputting the time sequence data into a multi-scale residual dense cavity convolution network, extracting a high-dimensional depth feature sequence containing multi-scale time sequence dependence, setting cavity convolutions with different expansion rates in different residual blocks, and densely connecting output features of all scale convolutions by the network; weighting the high-dimensional depth feature sequence by combining a channel attention and a time self-attention mechanism, detecting importance of different feature channels and dependency of different time steps, and obtaining a time weighting feature; Based on covariance among feature channels of the high-dimensional depth feature sequence, a feature association model is established, attention scores in feature dimensions are calculated, and the high-dimensional depth feature sequence is weighted to obtain variable association weighted features; And performing first outer product fusion on the time weighted feature and the variable associated weighted feature, performing second outer product fusion on the high-dimensional depth feature sequence and the time weighted feature, splicing the two fusion results to generate a fault expression vector, inputting the fault expression vector into a preset classifier, and outputting a fault type diagnosis result in the krypton-xenon extraction process.
2. The method of claim 1, wherein the process variables comprise key operating parameters and product quality metrics of the rectifying column.
3. The method of claim 1, wherein the combining the channel attention and time self-attention mechanism, weighting the high-dimensional depth feature sequence, detecting importance of different feature channels and dependency of different time steps, comprises: The channel attention mechanism generates the weight of each characteristic channel by carrying out global information compression and transformation on the high-dimensional depth characteristic sequence, and weights the original high-dimensional depth characteristic sequence; The time self-attention mechanism obtains attention scores by linearly mapping the feature sequences into queries, keys and values, calculating the similarity between the queries and the keys, and weighting the values according to the scores so as to detect the dependency relationship between time steps.
4. The method of claim 1, wherein the establishing a feature correlation model based on covariance between feature channels of the high-dimensional depth feature sequence, and calculating an attention score in a feature dimension, and weighting the high-dimensional depth feature sequence to obtain a variable correlation weighted feature, comprises: Based on the high-dimensional depth feature sequence Calculating covariance matrix between characteristic channels Wherein T is the time step and C is the number of characteristic channels; generating an attention score vector in a feature dimension by applying a Softmax function to the rows of the covariance matrix The calculation formula is as follows: Wherein the method comprises the steps of Is a full 1 vector; using the attention score vector Channel weighting is carried out on the high-dimensional depth feature sequence H, and variable association weighting features are obtained Wherein +.is the multiplication element by element, Representation of Is a transpose of (a).
5. The method according to any of claims 1-4, wherein the pre-set classifier is a Softmax classifier for mapping fault representation vectors onto N pre-defined fault classes.
6. The intelligent diagnosis system for faults in the krypton-xenon extraction process is characterized by comprising the following modules: The system comprises a connection module, a multi-scale residual error dense cavity convolution network, a multi-scale time sequence dependency extraction module and a storage module, wherein the connection module is used for acquiring time sequence data formed by a plurality of process variables in a krypton-xenon extraction process; The detection module is used for combining a channel attention and a time self-attention mechanism, weighting the high-dimensional depth feature sequence, detecting the importance of different feature channels and the dependency of different time steps, and obtaining a time weighting feature; The weighting module is used for establishing a feature association model based on covariance among feature channels of the high-dimensional depth feature sequence, calculating attention scores on feature dimensions, and weighting the high-dimensional depth feature sequence to obtain variable association weighting features; The output module is used for carrying out first outer product fusion on the time weighted feature and the variable associated weighted feature, carrying out second outer product fusion on the high-dimensional depth feature sequence and the time weighted feature, splicing the two fusion results to generate a fault expression vector, inputting the fault expression vector into a preset classifier, and outputting a fault type diagnosis result in the krypton-xenon extraction process.
7. The system of claim 6, wherein the process variables include key operating parameters and product quality metrics of the rectifying column.
8. The system of claim 6, wherein the combining channel attention and time self-attention mechanisms, weighting the high-dimensional depth feature sequences, detecting importance of different feature channels and dependency of different time steps, comprises: The channel attention mechanism generates the weight of each characteristic channel by carrying out global information compression and transformation on the high-dimensional depth characteristic sequence, and weights the original high-dimensional depth characteristic sequence; The time self-attention mechanism obtains attention scores by linearly mapping the feature sequences into queries, keys and values, calculating the similarity between the queries and the keys, and weighting the values according to the scores so as to detect the dependency relationship between time steps.
9. The system of claim 6, wherein the establishing a feature correlation model based on covariance between feature channels of the high-dimensional depth feature sequence and calculating an attention score in a feature dimension weights the high-dimensional depth feature sequence to obtain a variable correlation weighted feature comprises: Based on the high-dimensional depth feature sequence Calculating covariance matrix between characteristic channels Wherein T is the time step and C is the number of characteristic channels; generating an attention score vector in a feature dimension by applying a Softmax function to the rows of the covariance matrix The calculation formula is as follows: Wherein the method comprises the steps of Is a full 1 vector; using the attention score vector Channel weighting is carried out on the high-dimensional depth feature sequence H, and variable association weighting features are obtained Wherein +.is the multiplication element by element, Representation of Is a transpose of (a).
10. A computer readable storage medium having stored thereon a computer program, which when executed by a processor, implements the method according to any of claims 1-5.

Description

Intelligent diagnosis method, system and storage medium for faults in krypton-xenon extraction process Technical Field The application belongs to the field of intelligent diagnosis, and particularly relates to an intelligent diagnosis method, an intelligent diagnosis system and a storage medium for faults in a krypton-xenon extraction process. Background The extraction of krypton and xenon adopts a cryogenic rectification method, which relates to a plurality of rectification towers, heat exchangers and key equipment of compressors, and has a plurality of operation parameters and mutual influence. In actual production, krypton-xenon extraction is easy to generate various faults such as leakage, blockage, abnormal temperature, out-of-control pressure and the like due to the fluctuation of raw material components, equipment aging, external environment interference or artificial misoperation factors. The existing fault diagnosis method is based on a mechanism model and an expert system, but the former is difficult to build a mathematical model, and the latter is difficult to build a knowledge base and inconvenient to update. And extracting spatial characteristics of the data by using a convolutional neural network, or detecting the dependency relationship of the time sequence by using the convolutional neural network and the variant. The process data simultaneously contains short-term fluctuation generated by the periodical operation of the equipment and long-term trend caused by slow process drift, the single-scale feature extraction network is difficult to detect the multi-scale time sequence dependency, and fault symptom information is easy to lose. Attention mechanisms are mostly focused on weighting the explicit importance of time steps or characteristic channels, but neglecting the correlation of potential variations between different process variables in a fault state. Failure to model the correlation patterns between the variables may limit the recognition ability of the diagnostic model, rather than anomalies in a single variable, but rather the result of multiple variables synergistically deviating from normal operation. When different feature views are fused, the splicing or linear combination mode cannot fully mine high-order interaction among various feature information, so that the generated fault expression vector is insufficient in distinguishing degree, and the accuracy of fault classification is affected. Disclosure of Invention The invention provides an intelligent diagnosis method for faults in a krypton-xenon extraction process, which is used for solving the problem that the prior art ignores the relevance of potential changes among different process variables in a fault state and cannot fully mine high-order interaction among various characteristic information, and comprises the following steps: The method comprises the steps of obtaining time sequence data formed by a plurality of process variables in a krypton-xenon extraction process, inputting the time sequence data into a multi-scale residual dense cavity convolution network, extracting a high-dimensional depth feature sequence containing multi-scale time sequence dependence, setting cavity convolutions with different expansion rates in different residual blocks, and densely connecting output features of all scale convolutions by the network; weighting the high-dimensional depth feature sequence by combining a channel attention and a time self-attention mechanism, detecting importance of different feature channels and dependency of different time steps, and obtaining a time weighting feature; Based on covariance among feature channels of the high-dimensional depth feature sequence, a feature association model is established, attention scores in feature dimensions are calculated, and the high-dimensional depth feature sequence is weighted to obtain variable association weighted features; And performing first outer product fusion on the time weighted feature and the variable associated weighted feature, performing second outer product fusion on the high-dimensional depth feature sequence and the time weighted feature, splicing the two fusion results to generate a fault expression vector, inputting the fault expression vector into a preset classifier, and outputting a fault type diagnosis result in the krypton-xenon extraction process. In addition, the invention also relates to an intelligent diagnosis system for faults in the krypton-xenon extraction process, which comprises the following modules: The system comprises a connection module, a multi-scale residual error dense cavity convolution network, a multi-scale time sequence dependency extraction module and a storage module, wherein the connection module is used for acquiring time sequence data formed by a plurality of process variables in a krypton-xenon extraction process; The detection module is used for combining a channel attention and a time self-attention mechanism, weighting