Search

CN-122020048-A - Quality prediction method and device based on deep attention network

CN122020048ACN 122020048 ACN122020048 ACN 122020048ACN-122020048-A

Abstract

The invention discloses a quality prediction method and a quality prediction device based on a deep attention network, which solve the problems of how to expand the information quantity of input data, and can deeply represent deep features beneficial to learning and quality prediction tasks on the basis of effectively fusing expanded information so as to improve the online prediction precision of industrial process quality indexes. The invention designs a novel deep neural network structure of a deep attention network formed by connecting multiple layers of attention units in series, and aims to expand input information through a neighbor relation and realize effective fusion of expanded information through an attention mechanism. Compared with the traditional method, the method can extract deep features which are more beneficial to predicting quality data. On the basis, the invention discloses various technical schemes related to network structure improvement and the like so as to further improve the accuracy of quality prediction. Based on the same inventive concept, the invention further provides a quality prediction device, which executes implementation steps for realizing the quality prediction method.

Inventors

  • TONG CHUDONG
  • DAI HUATONG

Assignees

  • 宁波职业技术大学

Dates

Publication Date
20260512
Application Date
20260202

Claims (9)

  1. 1. A quality prediction method based on a deep attention network comprises the following steps of 1 to 3: Step 1, acquiring from a history database of an industrial process The group process data and the corresponding quality data are normalized to form a process data matrix respectively And quality data vector ; Step 2, building a slave Deep attention network composed of serially connected layer attention units and its control method As input, to each column vector of Corresponding quality data of the deep attention network is used as output, and the parameters of the deep attention network are sequentially pre-trained and fine-tuned; step3, periodically executing the steps 3-1 to 3-2 according to the sampling time interval, and predicting the quality data; step 3-1, obtaining the latest sampling moment Is normalized to form a column vector ; Step 3-2 to As input, after calculating the terminal estimation by using the deep attention network in step 2, performing inverse normalization processing on the terminal estimation to obtain a quality prediction value ; It is characterized in that arbitrary first The layer attention units each include an implementation transform Is a query encoding layer of (1), an implementation transform Key coding layer, an implementation transform Value encoding layer of (a), an implementation transform And Is a SoftMax layer of (C), and an implementation transform Wherein, 、 、 、 、 And Respectively represent warp lengths Layer attention unit transformed A number of query vectors, a first Key vector number 1 A value vector of A weighted probability value, a feature vector and an output estimate, The tanh function is represented by a table, Represent the first Layer correlation attention unit The number of input vectors is chosen to be, Equal to Is used for the number of dimensions of (c), Numbering for integers not less than 5 and not more than 10 First, a third step The parameters of the layer dependency attention unit comprise 4 weight matrices 、 、 And 4 Bias vectors 、 、 And ; For any given input The deep attention network calculates terminal estimates according to the following steps A1 to A4 : Step A1, slave Find out in (a) With minimum distance therebetween Individual column vectors and are marked as Re-recording this The individual column vectors are in After column numbering in (3), initializing ; Step A2, by setting Obtain the first Layer correlation attention units 0 to After inputting the vector, utilize the first Layer attention unit transformation to obtain feature vector And output estimation ; A3, according to the recorded column numbers, the feature matrix is obtained Column vectors of the same column are re-labeled as Then, reset Wherein, the method comprises the steps of, Is composed of The column vectors of the three-dimensional vector are used as the input of a deep attention network and pass through the first The feature vectors obtained after the layer attention unit transformation are formed; Step A4, judging Whether or not to be smaller than If yes, then set up Returning to the step A2, if not, obtaining Individual feature vectors And Individual output estimates And is provided with 。
  2. 2. A quality prediction method according to claim 1, characterized in that for any given input After the step A1 is executed, the formula is sequentially passed through And Calculating coefficient vectors respectively Updating Then according to the steps A2 to A4, the method is calculated Wherein, the neighbor matrix From the following components And (5) combining.
  3. 3. A quality prediction method according to claim 1, further comprising, between step 1 and step 2, step 1-1 shown below, and further, for any given input After the step A1 is performed, the method comprises the following steps of Inputting the vector information to the self-encoder trained in the step 1-1 to obtain corresponding reconstruction vector Through again Updating Then, according to the steps A2 to A4, the method is calculated ; Step 1-1 to The column vectors in (a) are used as input and output at the same time, and a self-encoder capable of realizing input reconstruction is built and trained.
  4. 4. A quality prediction method according to any one of claims 1 to 3, characterized in that the layers of attention units in the deep attention network further comprise an implementation transformation Is provided with a gate control layer of the (c), the gating layer is further according to For characteristic vector After weighting processing, the output layer performs transformation to obtain an output estimated value, wherein, And (3) with Respectively represent the first The weight matrix and bias vector of the gating layer in the layer attention unit, Representation of And gating vector The elements of the same position in (a) are multiplied, Representing a sigmoid function.
  5. 5. A quality prediction method according to any one of claims 1 to 3, wherein said deep attention network further comprises a transform implementation And said step A4 is performed by setting up Computing terminal estimates Wherein, the method comprises the steps of, Indicating that layer 1 to layer 1 Feature vectors obtained by layer attention unit transformation Is combined into a column vector which is then used to generate a vector, And (3) with The weight matrix and bias vector of the regression unit are represented respectively, Representing regression vectors transformed by regression units The first of (3) The elements.
  6. 6. A quality prediction method according to any one of claims 1 to 3, wherein said step 2 is performed immediately after completion of said step Calculating regression coefficient vectors And correspondingly, the step A4 is carried out according to the following steps Computing terminal estimates Wherein an estimation matrix is output Middle (f) The elements of the row being equal to By training the first column vector in the deep attention network The layer attention unit transforms the resulting output estimate, with the reference T representing the transpose of the matrix or vector.
  7. 7. The method of claim 4, wherein the gating layer is according to A transformation is performed, wherein, Representing the presentation to be And And are combined into a column vector.
  8. 8. A quality prediction device based on a deep attention network comprises a data acquisition module, a central control module, a display module and a storage module, wherein the storage module stores minimum values and maximum values required by normalization processing and inverse normalization processing, parameters of the deep attention network and a process data matrix Feature matrix corresponding to the feature matrix And an executive for pre-training and fine tuning parameters of the deep attention network; the central control module periodically controls the data acquisition module to acquire the latest sampling time from the database of the industrial process according to the set sampling time interval And invoking minimum and maximum values stored in the memory module to perform standardized processing on the set of process data to obtain column vectors Thereafter, by Input as a deep attention network Calling the stored parameters of the deep attention network, calculating to obtain corresponding terminal estimated values, and performing inverse normalization processing on the estimated values to obtain quality predicted values And will immediately Sending the display information to a display module for real-time display; The central control module can also control the data acquisition module to acquire from the database of the industrial process in real time according to the initialization instruction The group process data and the corresponding quality data are normalized to form a process data matrix after corresponding minimum and maximum values are calculated And quality data vector And sending the minimum value and the maximum value to the storage module for re-storage, and then, according to the execution program stored in the storage module, performing the data processing After the parameters of the deep attention network formed by the serial connection of the layer attention units are pre-trained and fine-tuned in sequence, the parameters of the deep attention network are obtained, Feature matrix corresponding to the feature matrix Sending the data to a storage module for re-storage; the central controller calculates corresponding terminal estimated value by executing steps A1 to A4 By using arbitrary first Layer attention unit transformation to obtain feature vector And output estimation The process of (1) is that firstly, respectively pass through 、 And Calculate the first Individual query vectors First, the Personal key vector And (d) Personal value vector Through again Respectively calculate the first Weighted probability values Then sequentially pass through And Computing corresponding feature vectors And output estimation Wherein, the method comprises the steps of, 。
  9. 9. The apparatus of claim 8, wherein the central controller sequentially passes through the formula after performing step A1 And Calculating coefficient vectors respectively Updating Then according to the steps A2 to A4, the method is calculated Wherein, the neighbor matrix From the following components Composition is prepared.

Description

Quality prediction method and device based on deep attention network Technical Field The invention belongs to the technical field of data-driven soft measurement, and particularly relates to a quality prediction method and device based on a deep attention network. Background The industry is often equipped with a large number of sensors, primarily for real-time monitoring of the normal operation of the process. In recent years, researchers have come to pay attention to and utilize a large amount of data stored by measurement in industrial processes, and based on these data, a model is built with easily measurable auxiliary variables as inputs and difficult-to-measure key indexes as outputs, and such a technique is called a data-driven soft measurement technique. In order to quickly and accurately estimate or predict important variables, the soft measurement technique is used as a virtual measurement technique, which calculates the estimated value of a key variable which is difficult to measure based on an auxiliary variable which is easy to measure, and provides a feasible and economical alternative to an expensive or impractical physical sensor. In comparison with conventional hardware sensors, soft measurement techniques are a combination of data processing, mathematical models and software techniques, and have received a great deal of attention. A paper titled "a soft measurement algorithm based on improved gradient clipping and gating cycle unit" published in the university of Qilu industry university journal in 2025 proposes a gating cycle unit of a normal differential equation as a basic model for processing nonlinear data, and a dynamic gradient clipping method is assisted, so that the quality prediction precision of the model to key indexes is improved. The Chinese patent application No. 202411796939.3 discloses an engineering quality prediction method based on process variables, and the quality index is predicted on line through a multi-layer graph neural network consisting of a multi-layer convolution network and a graph meaning network. In addition, the Chinese patent application number/patent number 202510012580.4 discloses an industrial process soft measurement method of a plurality of attention-gating networks, and the quality prediction accuracy of the model is improved by combining the technical advantages of isomorphic self-encoder networks and attention mechanisms. The Chinese patent application number 202510781128.4 discloses an industrial process soft measurement method based on graph attention, and the graph attention mechanism is utilized to improve the quality prediction performance. In the prior art and scientific literature, deep learning network models including self-encoders, convolutional neural networks, attention mechanisms, etc. have been successfully applied to solve different soft measurement problems. In the technical field, the main technical advantage of the deep learning model for solving the soft measurement problem is that deep abstract features can be learned through layer-by-layer feature extraction representation, so that the prediction accuracy of quality indexes is improved. However, since the deep learning model is built as a black box model between input and output, the quality of input data and the comprehensiveness of information thereof can directly affect the accuracy of quality prediction. It can be said that how to ensure the sufficiency of the information amount of the input data is an important way to improve the accuracy of quality prediction. Disclosure of Invention The invention aims to solve the main technical problems of how to expand the information quantity of input data, and can deeply represent deep features beneficial to learning and quality prediction tasks on the basis of effectively fusing the expanded information, thereby improving the on-line prediction precision of quality indexes of industrial processes. The technical scheme adopted by the invention for solving the problems is that the quality prediction method and the device based on the deep attention network are adopted, wherein the implementation process of the quality prediction method based on the deep attention network disclosed by the invention comprises the following steps 1 to 3. Step 1, acquiring from a history database of an industrial processThe group process data and the corresponding quality data are normalized to form a process data matrix respectivelyAnd quality data vector。 Step 2, building a slaveDeep attention network composed of serially connected layer attention units and its control methodAs input, to each column vector ofCorresponding quality data of the deep attention network is used as output, and the pre-training and the fine tuning are sequentially carried out on parameters of the deep attention network. And 3, periodically executing the following steps 3-1 to 3-2 according to the sampling time interval, and predicting the quality data. Step 3-1, obtaining the