Soft measurement method and device based on depth residual PLSR

Abstract

The invention discloses a deep residual PLSR-based soft measurement method and device, which deeply represents nonlinear features beneficial to quality data soft measurement in learning residual information on the basis of a PLSR model, thereby improving the performance of a soft measurement model. According to the soft measurement method disclosed by the invention, the linear relationship and the nonlinear relationship between the input and the output are simultaneously considered through the deep residual PLSR model, and the residual which cannot be fitted by the previous PLSR model is further fitted through deep nonlinear transformation, so that the prediction precision of the soft measurement model on the quality data can be effectively improved. In addition, the invention further discloses a plurality of improved technical schemes related to parameter optimization, orthogonal signal correction processing, filtering processing and weighting processing, so that the soft measurement performance of the soft measurement method on quality data is further improved. Finally, the invention designs and builds a corresponding soft measurement device, and aims to implement soft measurement according to the disclosed soft measurement method.

Description

Technical Field

[0001] This invention belongs to the field of data-driven soft measurement technology, and specifically relates to a soft measurement method and device based on depth residual PLSR. Background Technology

[0002] In modern, increasingly complex and large-scale industrial production processes, timely measurement of key product quality variables is crucial for process monitoring, control, and optimization. Unlike other easily measurable process variables, quality variables are typically measured offline using analyzers or laboratory testing, rather than online. This is primarily because online measuring instruments are not only expensive and have low reliability, but also operate in harsh environments. Therefore, soft measurement technology has found its niche. It uses mathematical models built from easily measurable process variables to estimate product quality variables in real time. In recent years, data-driven soft sensors have received widespread attention in the industrial field. They can be applied without prior knowledge of the process background, effectively solving the challenge of measuring product quality in large-scale, automated, and complex modern industrial environments.

[0003] Chinese invention patent No. 202211611844.0 discloses a deep learning-based soft measurement method for bottomhole flowing pressure in oil wells. This method combines dynamic and static data for in-depth data analysis, enabling real-time continuous soft measurement of bottomhole flowing pressure. Chinese invention patent No. 202311550643.9 discloses a soft measurement modeling method integrating feature extraction and adaptive mapping. This method learns feature representations through correlation analysis and incorporates adaptive mapping optimization to improve the generalization ability of the soft measurement model. Chinese invention patent No. 202410564231.9 discloses a soft measurement modeling method for anaerobic fermentation processes based on RF-IHHO-LSTM. This method combines the technical advantages of three algorithms to ensure the accuracy of the soft measurement model.

[0004] In recent years, deep learning technology has transformed the way models learn, inspiring numerous innovations. One such innovation is the evolution of models from shallow to deep, allowing for the extraction of more abstract features for data feature representation learning. A 2023 paper published in *IEEE Transactions on Neural Networks and Learning Systems*, titled "Deep PLS: A lightweight deep learning model for interpretable and efficient data analytics," disclosed a multi-level cascaded PLS model for interpretable data analysis. However, the model presented in this paper remains a linear model and cannot perform non-linear feature representation learning. Furthermore, a 2023 paper published in *IEEE Transactions on Industrial Informatics*, titled "A deepsupervised learning framework based on kernel partial least squares for industrial soft sensing," disclosed a deep learning model based on kernel partial least squares to address soft sensing problems.

[0005] However, these existing soft measurement methods only focus on deep representation learning of latent feature components, without considering the full utilization of residual information in Partial Least Squares Regression (PLSR) modeling. For data feature representation learning, the residual information of PLSR models still contains some useful nonlinear feature components. If this residual information can be fully utilized, richer features can be obtained for soft measurement tasks, thereby improving soft measurement performance. Summary of the Invention

[0006] The main technical problem to be solved by this invention is: how to use deep representation learning to learn the nonlinear features in the residual information that are beneficial to improving the accuracy of soft measurement based on the PLSR model, thereby improving the generalization ability of the soft measurement model and effectively solving the problem of soft measurement.

[0007] The technical solution adopted by the present invention to solve the above problems is: a soft measurement method and device based on depth residual PLSR; wherein, the implementation process of the soft measurement method based on depth residual PLSR disclosed in the present invention includes the following steps 1 to 3.

[0008] Step 1: Obtain from the industrial process database The process data and their corresponding quality data, after being standardized, form an input matrix. and output vector .

[0009] Step 2, using and Training a by A deep residual PLSR model composed of layered PLSR models.

[0010] Step 3: Based on the sampling time interval, periodically execute steps 3-1 to 3-2 to perform soft measurement on the quality data.

[0011] Step 3-1: Obtain the latest sampling time A set of process data, after being standardized, forms a row vector. .

[0012] Step 3-2, with As the input vector, the terminal estimate is calculated using the deep residual PLSR model trained in step 2. Then, on The soft measurement values ​​of the quality data are obtained by performing de-standardization processing.

[0013] The core technical feature of this invention lies in: for any given input vector The deep residual PLSR model sequentially calls the first Transformation matrix of layer PLSR model Load matrix and regression vector and through , and Calculate the corresponding score vector Input residual vector and output valuation Then, set the terminal valuation. ;in, For the first The input to the layer PLSR model, and when hour, ,when hour, Two adjustable parameters and The range of values ​​for are: and .

[0014] Furthermore, the deep residual PLSR model was trained using cross-validation, sequentially training layers 1, 2, and up to layer C; among which, the training of the layer C... The process for training the layer PLSR model is as follows: First, the model will be used to train the layer PLSR model. The input matrix and output vector of the layer PLSR model are denoted as follows: and Afterwards, and The dataset is divided into several sub-blocks, with one sub-block serving as the validation dataset and the remaining sub-blocks as the training dataset. Next, the score vector dimension is constructed using the training dataset. They are respectively equal to The PLSR model is then used to calculate the corresponding mean squared error using the validation dataset. Finally, the dimension of the score vector corresponding to the minimum mean square error is denoted as... , and then and To construct a score vector for the training dataset with a dimension equal to [missing value], [missing value]. The PLSR model is obtained, and the corresponding transformation matrix is ​​preserved. Load matrix and regression vector Among them, when hour, and They are respectively equal to and ,when At that time, first through and Calculate the input residual matrix respectively and output valuation vector Then set and ; for The total number of column vectors in the middle.

[0015] From the deep residual PLSR model Calculate the corresponding terminal valuation The process shows that the first layer, the second layer, and so on were called sequentially. The transformation matrix, loading matrix, and regression vector of the layer PLSR model; moreover, the input residual vector generated by the previous layer PLSR model undergoes a nonlinear transformation. Only after this is it used as input for the next layer of the PLSR model; therefore, the deep residual PLSR model gradually reduces the estimation error of the output data through multiple layers of PLSR models.

[0016] Furthermore, except for the first-layer PLSR model which performs linear regression on the input and output, starting from the second-layer PLSR model, nonlinear transformations are employed. The corresponding PLSR model is actually a nonlinear regression of the input residuals and output residuals. From this perspective, the deep residual PLSR model considers both the linear and nonlinear relationships between the input and output. Furthermore, it can further fit the residuals that the first-level PLSR model cannot fit through deep nonlinear transformations, which can effectively improve the prediction accuracy of the soft measurement model for quality data and bring significant and beneficial technical effects.

[0017] Due to the Adjustable parameters of the layer PLSR model and This will affect the performance of soft measurement. As an improvement of this invention, it is preferable to use an intelligent optimization algorithm for the first... Layer PLSR model determination and The optimal value; where the intelligent optimization algorithm represents the value for each individual in the iteration process. and All of these need to be based on Calculate the corresponding input matrix Then, the corresponding PLSR model is trained using the cross-validation method, and then... Calculate the corresponding output estimation vector Thus through Calculate the objective function value for this individual. The intelligent optimization algorithm provides a result after the iteration process is completed, which makes... The minimal individual, which represents and That is, the optimal value; and They represent and The Middle Each element.

[0018] Intelligent optimization algorithms, as a classic tool for solving optimization problems, can effectively solve the problem of minimizing or maximizing complex objective functions by providing a method for calculating the objective function value corresponding to each individual. The intelligent optimization algorithms applicable to this improved technical solution include, but are not limited to, genetic algorithms, particle swarm optimization, differential evolution, ant colony optimization, tabu search, and their various improved algorithms.

[0019] Determined through intelligent optimization algorithms and After finding the optimal value, it can be ensured that the PLSR models of layers 2 to C can be trained while minimizing the output estimation error, and the corresponding transformation matrix, load matrix and regression vector are retained, thereby making the soft measurement performance of the deep residual PLSR model optimal.

[0020] As another improvement to the present invention, in step 2, it is preferable to first pass through right After performing orthogonal signal correction processing, the deep residual PLSR model is then trained; correspondingly, in step 3-2, it is preferable to first train the deep residual PLSR model. right Orthogonal signal correction processing is performed, and then the terminal estimate is calculated using a deep residual PLSR model; among which, for An identity matrix of dimension 1, orthogonal transformation matrix It is determined according to steps A1 to A4.

[0021] Step A1: Set the orthogonal score vector equal Any column vector in the vector;

[0022] Step A2, according to right After implementing the update, then according to Calculate the regression coefficient vector ; where the superscript T denotes the transpose of a matrix or vector.

[0023] Step A3, according to Again After implementing the update process, determine Has convergence been achieved? If not, return to step A2; if yes, proceed to step A4.

[0024] Step A4, according to Calculate the load vector Then, according to Calculate the orthogonal transformation matrix .

[0025] Orthogonal signal correction processing can be used to eliminate [these signals]. and The negative impact of information components unrelated to the output on regression modeling can be mitigated, thereby improving the measurement performance of the deep residual PLSR model.

[0026] because and The input is the first-level PLSR model. Therefore, the above-mentioned improved technical solution involving orthogonal signal correction processing actually only performs orthogonal signal correction processing on the input of the first-level PLSR model. Given the beneficial effects that orthogonal signal correction processing can bring, orthogonal signal correction processing is not limited to the first-level PLSR model.

[0027] As another improvement of the present invention, when training the PLSR model of layer 1, layer 2, up to layer C in sequence, it is preferable to follow the procedure... After performing orthogonal signal correction processing on the input matrices of each layer of the PLSR model, the corresponding PLSR model is trained using cross-validation; correspondingly, the deep residual PLSR model is used as the basis for... In the process of calculating the terminal valuation, the preferred method is to first follow... After performing orthogonal signal correction processing on the inputs of each layer of the PLSR model, the transformation matrix, loading matrix, and regression vector are then called for corresponding calculations; among them, those belonging to the first layer... Orthogonal transformation matrix of layer PLSR model It is determined according to steps B1 to B4.

[0028] Step B1: Set the orthogonal score vector equal Any column vector in the vector;

[0029] Step B2, according to right After implementing the update, then according to Calculate the regression coefficient vector .

[0030] Step B3, according to Again After implementing the update process, determine Has convergence been achieved? If not, return to step B2; if yes, proceed to step B4.

[0031] Step B4, according to Calculate the load vector Then, according to Calculate the orthogonal transformation matrix .

[0032] By performing orthogonal signal correction processing on the input of each layer of the PLSR model, the negative interference of information components that are orthogonal to and unrelated to the output can be eliminated more comprehensively, thereby further improving the soft measurement performance.

[0033] Since the first layer PLSR model is based on the input vector Perform a linear transformation, that is: To obtain the corresponding output valuation , Not all data is favorable for linear regression. Therefore, in order to further improve the performance of soft measurement, it is possible to... After filtering and processing each data point, the corresponding calculations are performed using the deep residual PLSR model.

[0034] As another improvement of the present invention, the deep residual PLSR model further includes a layer before the first layer PLSR model. right The filtering process is performed, and after filtering, the first-layer PLSR model is... , and Perform the corresponding calculations; among them, This represents the input vector after filtering. This represents element-wise multiplication at the same position, in binary vectors. It is determined using a genetic algorithm. During the iterative process of the genetic algorithm, each individual is... A binary vector of dimension, representing any individual The process of calculating the corresponding objective function value is shown in steps C1 to C3.

[0035] Step C1, through right Each row vector in the input matrix is ​​filtered to obtain the filtered input matrix. ;in, and They represent and The first in Row vector, number .

[0036] Step C2, and Using these as the input matrix and output vector of the PLSR model respectively, and training the corresponding PLSR model using cross-validation, the corresponding transformation matrix is ​​obtained. and .

[0037] Step C3, according to Calculate the output valuation vector Then, according to Calculate the objective function value for this individual. ;in, and They represent and The first in Each element.

[0038] After the iterative process is complete, the genetic algorithm provides the result that satisfies the objective function value. The minimized individual, whose binary vector is the binary vector for which filtering is performed. And according to steps C1 to C2, the results are the same as those obtained from the previous steps. The corresponding PLSR model is the first layer PLSR model in the deep residual PLSR model.

[0039] Since the elements in a binary vector can only take the values ​​0 or 1, by multiplying the elements at the same position, the elements in certain columns of the filtered input vector or data matrix can be made to be 0. This makes the columns with 0 have no effect on the regression estimation output, thus eliminating the influence of the data in that column on the output estimation. Therefore, this improved technical solution can further enhance the soft measurement performance of the deep residual KPLSR model.

[0040] In addition to filtering the input to the first-layer PLSR model, the input can also be weighted. Therefore, as another improvement of the present invention, the deep residual PLSR model further includes a filter before the first-layer PLSR model. For any given input vector The process of implementing weighted processing, weight vector The nearest neighbor component analysis algorithm is preferred for determination; wherein, the optimization objective of the nearest neighbor component analysis algorithm is to make Minimum; express The Middle elements, regularization coefficient The range of values ​​is probability regression value The calculation process is as follows: First, calculate the probability according to the following formula ① Then calculate according to formula ② below. : ① ②

[0041] in, express The first in Line 1 Column elements, express The first in Line 1 Column elements, and They represent The first in and the One element, Indicates calculation The absolute value, number .

[0042] In the weight vector Once determined, the training of the first-layer PLSR model must first follow the instructions. right Each row vector in the middle is weighted and then... and The first-layer PLSR model is trained using cross-validation, with the input matrix and output vector serving as the input and output vectors.

[0043] In this technical field, nearest neighbor component analysis (NNB) algorithms, as a feature selection or weighting tool, have been widely recognized for improving the classification accuracy of classification models and the regression estimation accuracy of regression models; weighting is achieved through... Distinguishing between the magnitude changes of medium-weighted elements The magnitude of the influence of each data point on the regression estimation output is determined, thereby improving the soft measurement performance of the deep residual PLSR model.

[0044] In addition to weighting the input before the first-level PLSR model, the nearest neighbor component analysis algorithm can also be used to determine the corresponding weighting vectors before the second to C-level PLSR models, thereby improving the weighting efficiency. Weighting is applied; correspondingly, the input matrix is ​​processed... Based on weighted processing, the PLSR model is trained using cross-validation; where, .

[0045] As a further improvement to the present invention and its various improved technical solutions described above, it is preferable to determine the total number of layers of the PLSR model according to steps D1 to D2 as shown below. .

[0046] Step D1: Train the PLSR model with layers 1, 2, and 3 using cross-validation, and then respectively... Calculate the mean square error of the depth residual PLSR model ;in, for The Middle One element, Indicates passing through the first The output estimate calculated by the layer PLSR model.

[0047] Step D2, Settings Then, cross-validation was used to train the first... Layered PLSR model, and then calculate the mean square error accordingly. Then determine whether the conditions are met. If so, then set Then, continue training the next layer of the PLSR model until the conditions are no longer met. If not, then the total number of layers in the PLSR model. .

[0048] Therefore, this improved technical solution involves gradually increasing the number of layers in the PLSR model and determining whether the mean squared error significantly decreases, i.e., whether the judgment condition is met. If adding a new PLSR model layer does not significantly reduce the mean square error, then it is unnecessary to add this new PLSR model layer. In this case, the optimal deep residual PLSR model is composed of the previous PLSR models, which ensures that the soft measurement of quality data is carried out with the PLSR model with the optimal total number of layers.

[0049] The present invention also provides a soft measurement device based on deep residual PLSR, including a data acquisition module, a central controller, a display unit, and a storage medium; the functions of each component are as follows.

[0050] The storage medium contains the mean and standard deviation for implementing standardization, the transformation matrix, loading matrix and regression vector required to calculate the terminal prediction, and the program for executing the cross-validation training of the PLSR model.

[0051] The central controller periodically controls the data acquisition module to retrieve the latest sampling time from the industrial process database according to the set sampling time interval. A set of process data is used, and the mean and standard deviation stored in the storage medium are called to perform standardization processing to obtain a column vector. Then, the transformation matrix, load matrix, and regression vector stored in the storage medium are called to calculate the corresponding terminal prediction value. After de-standardization, the soft measurement value of the quality data is obtained, and then the soft measurement value is sent to the display unit for real-time display.

[0052] The central controller can also, based on initialization commands, control the data acquisition module to retrieve data from the industrial process database in real time. After calculating the mean and standard deviation of the process data and its corresponding quality data, standardization is then performed to form the input matrix. and output vector The mean and standard deviation are then sent to the storage medium for re-storage; next, the program stored in the storage medium is executed to train the PLSR model at layers 1, 2, up to C in sequence; finally, the... The transformation matrix, load matrix, and regression vector of the layer PLSR model are sent to the storage medium for re-storage.

[0053] This soft measurement device and the soft measurement method disclosed in this invention are based on the same inventive concept, specifically in that: the process of the central controller calculating the terminal prediction value is as follows: first, setting up sequentially... and through , and Calculate the corresponding score vector Input residual vector and output valuation Then, set the terminal valuation. .

[0054] The design flow for programs stored in storage media is as follows: First, ... and The dataset is divided into several sub-blocks, with one sub-block serving as the validation dataset and the remaining sub-blocks as the training dataset. Next, the score vector dimension is constructed using the training dataset. They are respectively equal to The PLSR model is then used to calculate the corresponding mean squared error using the validation dataset. Finally, the dimension of the score vector corresponding to the minimum mean square error is denoted as... , and then and To construct a score vector for the training dataset with a dimension equal to [missing value], [missing value]. The PLSR model is obtained, and the corresponding transformation matrix is ​​preserved. Load matrix and regression vector .

[0055] Furthermore, the present invention also provides an improved soft measurement device, wherein the storage medium of the device further stores the execution implementation using an intelligent optimization algorithm for the first... Layer PLSR model determination and The optimal value is determined by the program executed first when the central controller receives the initialization command. and After determining the optimal value, the program for training the PLSR model using cross-validation is executed; the intelligent optimization algorithm involved in the program represents the value for each individual in the iteration process. and All of these must be based on Calculate the corresponding input matrix Then, the corresponding PLSR model is established using the cross-validation method, and then... Calculate the corresponding output estimation vector Thus through Calculate the objective function value for this individual. The intelligent optimization algorithm provides a result after the iteration process is completed, which makes... The minimal individual, which represents and That is the optimal value. Attached Figure Description

[0056] Figure 1 This is a flow chart of the hydrocracking process.

[0057] Figure 2 This is a schematic diagram of the composition of the depth residual PLSR model involved in this invention.

[0058] Figure 3 This is a schematic diagram illustrating the soft measurement effect of the first embodiment of the present invention.

[0059] Figure 4 This is a comparison chart of the soft measurement performance indicators of the first embodiment of the present invention and three comparative methods.

[0060] Figure 5 Box plot comparison of soft measurement errors of quality data in the first to fourth embodiments of the present invention.

[0061] Figure 6 This is a schematic diagram of the composition and structure of the soft measurement device disclosed in this invention. Detailed Implementation

[0062] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0063] Figure 1 The hydrocracking process illustrated is a crucial step in petroleum refining. Feedstock oil and heated hydrogen undergo a chemical reaction in a hydrocracking reactor. The resulting products, obtained through multi-stage separation towers, a desulfurization hydrogen separator, and a fractionation tower, are then differentiated into oil and gas products with varying boiling points. Among the monitoring of these oil and gas products, the initial distillation point of aviation kerosene is critical to ensuring its quality, thus requiring real-time measurement and monitoring. However, the hardware sensors used to measure the initial distillation point of aviation kerosene operate in harsh environments, typically requiring periodic maintenance to ensure reliability. Therefore, soft measurement technology is essential to meet the requirement of real-time measurement of the initial distillation point of aviation kerosene.

[0064] Figure 1 The circled F and T represent flow rate and temperature, respectively. Figure 1The hydrocracking process shown can measure data of 43 variables such as temperature, pressure, and flow rate in real time. Therefore, the corresponding mass data, namely the initial distillation point of aviation kerosene, can be soft-measured based on the process data corresponding to these 43 easily measurable variables. The first embodiment of the present invention applies the disclosed soft-measurement method to solve the problem of soft-measurement of the mass data of the initial distillation point of aviation kerosene. The corresponding process includes the following steps 1 to 3.

[0065] Step 1: Obtain from the industrial process database =966 sets of process data and their corresponding quality data, after standardization, form an input matrix. and output vector .

[0066] Step 2, using and Training a Figure 2 The diagram shows the composition of the diagram. A deep residual PLSR model composed of layered PLSR models; wherein, for any given input vector , Figure 2 The depth residual PLSR model shown calls the first one in sequence. Transformation matrix of layer PLSR model Load matrix and regression vector and through , and Calculate the corresponding score vector Input residual vector and output valuation Then, set the terminal valuation. Among them, adjustable parameters are set. , , , .

[0067] In the first embodiment, cross-validation is used to train the PLSR model at layers 1, 2, and 3 sequentially; wherein, training the layers 1 through 3... The process for training the layer PLSR model is as follows: First, the model will be used to train the layer PLSR model. The input matrix and output vector of the layer PLSR model are denoted as follows: and Afterwards, and The dataset is divided into 5 sub-blocks, with one sub-block serving as the validation dataset and the remaining sub-blocks as the training dataset. Next, the score vector dimension is constructed using the training dataset. They are respectively equal to The PLSR model is then used to calculate the corresponding mean squared error using the validation dataset. Finally, the dimension of the score vector corresponding to the minimum mean square error is denoted as... , and then and To construct a score vector for the training dataset with a dimension equal to [missing value], [missing value]. The PLSR model is obtained, and the corresponding transformation matrix is ​​preserved. Load matrix and regression vector .

[0068] The dimension of the score vector is well known to those skilled in the art. The PLSR model is generally implemented through a nonlinear iterative partial least squares algorithm, specifically including steps 2-1 to 2-5 as shown below.

[0069] Step 2-1: Initialize the transformation vector .

[0070] Step 2-2, according to Calculate the score vector Then, according to right Implement the update process.

[0071] Steps 2-3, according to Calculate the correlation coefficient And then according to Updated Then, make a judgment Has convergence been achieved? If not, return to step 2-2; if yes, proceed to step 2-4.

[0072] Steps 2-4, according to and Calculate the load vectors separately and regression coefficients Then, according to and For the input matrix respectively and output vector After performing deflationary processing, return to step 2-1 to continue obtaining the next transformation vector, loading vector, and regression coefficients, until the desired result is obtained. A transformation vector, a load vector, and regression coefficients.

[0073] Steps 2-5, obtain Each transformation vector, load vector, and regression coefficient is combined to form the corresponding transformation matrix. Load matrix Regression vector Then, according to right Implement the update.

[0074] Step 3: Based on the sampling time interval, periodically execute steps 3-1 to 3-2 to perform soft measurement on the quality data.

[0075] Step 3-1: Obtain the latest sampling time A set of process data, after being standardized, forms a row vector. .

[0076] Step 3-2, with As the input vector, the terminal estimate is calculated using the deep residual PLSR model trained in step 2. Then, on The soft measurement values ​​of the quality data are obtained by performing de-standardization processing.

[0077] Step 3 involves performing soft measurement on the quality data at the 483 most recent sampling times, and then plotting the curves showing the changes in the soft measurement values ​​and actual measurement values ​​at these 483 most recent sampling times. Figure 3 From; Figure 3 It can be seen that the soft measurement values ​​given by the soft measurement method disclosed in this invention have a very high degree of overlap with the actual measurement values, which fully demonstrates the effectiveness of the soft measurement method disclosed in this invention.

[0078] Furthermore, to illustrate the significant improvement in soft measurement performance that the method of this invention can bring through the deep residual PLSR model, the soft measurement method disclosed in this invention is compared with three other soft measurement methods based on PLSR, neural network PLSR, and deep neural networks. The root mean square error (RMSE) of the quality data at the 483 most recent sampling times is used as the metric to evaluate the soft measurement performance. The corresponding soft measurement performance results are compared to, for example... Figure 4 As shown.

[0079] from Figure 4 It can be seen that the RMSE index corresponding to the method of the present invention is significantly smaller than that of the other three comparison methods, which fully demonstrates that the present invention can bring significant beneficial technical effects by using the deep residual PLSR model.

[0080] The second embodiment of the present invention is an improvement on the optimal value of the adjustable parameter, which is achieved by using an intelligent optimization algorithm to obtain the optimal value for the first parameter. Layer PLSR model determination and The optimal value; where the intelligent optimization algorithm represents the value for each individual in the iteration process. and All of these need to be based on Calculate the corresponding input matrix Then, the corresponding PLSR model is trained using the cross-validation method, and then... Calculate the corresponding output estimation vector Thus through Calculate the objective function value for this individual. The intelligent optimization algorithm provides a result after the iteration process is completed, which makes... The minimal individual, which represents and That is, the optimal value; and They represent and The Middle Each element.

[0081] The third embodiment of the present invention involves performing orthogonal correction signal processing on the input of each layer of the PLSR model. Specifically, when training the PLSR models of layer 1, layer 2, and up to layer C sequentially, it is preferable to follow the... After performing orthogonal signal correction processing on the input matrices of each layer of the PLSR model, the corresponding PLSR model is trained using cross-validation; correspondingly, the deep residual PLSR model is used as the basis for... In the process of calculating the terminal valuation, the preferred method is to first follow... After performing orthogonal signal correction processing on the inputs of each layer of the PLSR model, the transformation matrix, loading matrix, and regression vector are then called for corresponding calculations; among them, those belonging to the first layer... Orthogonal transformation matrix of layer PLSR model It is determined according to the aforementioned steps B1 to B4.

[0082] The fourth embodiment of the present invention performs filtering processing on the input vector of the deep residual PLSR model, that is, before the first layer of the PLSR model, it also includes a filtering process. right The filtering process is performed, and after filtering, the first-layer PLSR model is... , and Perform the corresponding calculations; among which, binary vectors It is determined using a genetic algorithm. During the iterative process of the genetic algorithm, each individual is... A binary vector of dimension, representing any individual The process of calculating the corresponding objective function value is as described in steps C1 to C3 above.

[0083] The fifth embodiment of the present invention further includes a method before the first-layer PLSR model. For any given input vector The process of implementing weighted processing; weight vector It is determined using the nearest neighbor component analysis algorithm, and in the weight vector Once determined, the training of the first-layer PLSR model must first follow the instructions. right Each row vector in the middle is weighted and then... and The first-layer PLSR model is trained using cross-validation, with the input matrix and output vector serving as the input and output vectors.

[0084] The second to fifth embodiments are all based on the first embodiment, but with different improvements. To verify the beneficial effects of these improved technical solutions, the soft measurement errors of the quality data at the same 483 latest sampling times in the first to fifth embodiments are compared and displayed in the form of box plots. Figure 5 middle; Figure 5 The box plot shows the median, minimum, and maximum values ​​of soft measurement errors, as well as possible gross error data points.

[0085] from Figure 5 As can be seen, compared with the first embodiment, the soft measurement error corresponding to the second to fifth embodiments can be further reduced, which fully demonstrates that the improved technical solution can bring about a certain degree of improvement in soft measurement performance.

[0086] The sixth embodiment of the present invention is based on any one of the preceding five embodiments, and determines the total number of layers of the PLSR model according to steps D1 to D2 as shown below. .

[0087] Step D1: Train the PLSR model with layers 1, 2, and 3 using cross-validation, and then... Calculate the mean square error of the depth residual PLSR model .

[0088] Step D2, Settings Then, cross-validation was used to train the first... Layered PLSR model, and then calculate the mean square error accordingly. Then determine whether the conditions are met. If so, then set Then, continue training the next layer of the PLSR model until the conditions are met. If not, then the total number of layers in the PLSR model. .

[0089] The seventh embodiment of the present invention is to construct and design a soft measurement device for implementing the soft measurement method disclosed in the present invention; specifically, the soft measurement device is composed as follows: Figure 6 As shown, it includes a data acquisition module, a central controller, a display unit, and a storage medium; wherein, the storage medium stores the mean and standard deviation for implementing standardization, the transformation matrix, loading matrix and regression vector required for calculating the terminal prediction value, and the program for executing the cross-validation training of the PLSR model.

[0090] The central controller periodically controls the data acquisition module to retrieve the latest sampling time from the industrial process database according to the set sampling time interval. A set of process data is used, and the mean and standard deviation stored in the storage medium are called to perform standardization processing to obtain a column vector. Then, the transformation matrix, load matrix, and regression vector stored in the storage medium are called, and the corresponding terminal prediction value is calculated according to the following process. After the prediction value is de-standardized, the soft measurement value of the quality data is obtained, and then the soft measurement value is sent to the display unit for real-time display.

[0091] First, set them in sequence. and through , and Calculate the corresponding score vector Input residual vector and output valuation .

[0092] Then, set the terminal valuation. .

[0093] The central controller can also, based on initialization commands, control the data acquisition module to retrieve data from the industrial process database in real time. After calculating the mean and standard deviation of the process data and its corresponding quality data, standardization is then performed to form the input matrix. and output vector The mean and standard deviation are then sent to the storage medium for re-storage; next, the program stored in the storage medium is executed to train the PLSR model at layers 1, 2, up to C in sequence; finally, the... The transformation matrix, load matrix, and regression vector of the layer PLSR model are sent to the storage medium for re-storage.

[0094] The design flow for programs stored in storage media is as follows: and The dataset is divided into several sub-blocks, with one sub-block serving as the validation dataset and the remaining sub-blocks as the training dataset. Next, the score vector dimension is constructed using the training dataset. They are respectively equal to The PLSR model is then used to calculate the corresponding mean squared error using the validation dataset. Finally, the dimension of the score vector corresponding to the minimum mean square error is denoted as... , and then and To construct a score vector for the training dataset with a dimension equal to [missing value], [missing value]. The PLSR model is obtained, and the corresponding transformation matrix is ​​preserved. Load matrix and regression vector .

[0095] Of course, the storage medium also stores the execution implementation using intelligent optimization algorithms for any second... Layer PLSR model determination and The optimal value is determined by the program executed first when the central controller receives the initialization command. and After determining the optimal value, the program for training the PLSR model using cross-validation is executed; the intelligent optimization algorithm involved in the program represents the value for each individual in the iteration process. and All of these must be based on Calculate the corresponding input matrix Then, the corresponding PLSR model is established using the cross-validation method, and then... Calculate the corresponding output estimation vector Thus through Calculate the objective function value for this individual. The intelligent optimization algorithm provides a result after the iteration process is completed, which makes... The minimal individual, which represents and That is the optimal value.

[0096] The above explanations and descriptions of the technical solutions disclosed in this invention in conjunction with the accompanying drawings are illustrative rather than restrictive. The technical solutions disclosed in this invention are not limited to the above-described embodiments. Modifications made without departing from the spirit of the disclosed technical solutions and the scope of protection of the claims shall fall within the protection scope of this invention.

Claims

1. A soft measurement method based on depth residual PLSR, comprising the following steps 1 to 3: Step 1: Obtain from the industrial process database The process data and their corresponding quality data, after being standardized, form an input matrix. and output vector ; Step 2, using and Training a by A deep residual PLSR model composed of layered PLSR models; Step 3: According to the sampling time interval, periodically execute steps 3-1 to 3-2 to perform soft measurement on the quality data; Step 3-1: Obtain the latest sampling time A set of process data, after being standardized, forms a row vector. ; Step 3-2, with As the input vector, the terminal estimate is calculated using the deep residual PLSR model trained in step 2. Then, on The soft measurement values ​​of the quality data are obtained by performing de-standardization processing; Its features are: For any given input vector The deep residual PLSR model sequentially calls the first Transformation matrix of layer PLSR model Load matrix and regression vector and through , and Calculate the corresponding score vector Input residual vector and output valuation Then, set the terminal valuation. Among them, when At that time, the first Input of the layer PLSR model ,when hour, Two adjustable parameters and The range of values ​​for are: and ; The deep residual PLSR model was trained using cross-validation, sequentially training layers 1, 2, and up to layer C. The process for training the layer PLSR model is as follows: First, the model will be used to train the layer PLSR model. The input matrix and output vector of the layer PLSR model are denoted as follows: and Afterwards, and The dataset is divided into several sub-blocks, with one sub-block serving as the validation dataset and the remaining sub-blocks as the training dataset. Next, the score vector dimension is constructed using the training dataset. They are respectively equal to The PLSR model is then used to calculate the corresponding mean squared error using the validation dataset. Finally, the dimension of the score vector corresponding to the minimum mean square error is denoted as... , and then and To construct a score vector for the training dataset with a dimension equal to [missing value], [missing value]. The PLSR model is obtained, and the corresponding transformation matrix is ​​preserved. Load matrix and regression vector Among them, when hour, and They are respectively equal to and ,when At that time, first through and Calculate the input residual matrix respectively and output valuation vector Then set and ; for The total number of column vectors in the middle.

2. The soft measurement method according to claim 1, characterized in that, Using intelligent optimization algorithms for the first Layer PLSR model determination and The optimal value; where the intelligent optimization algorithm represents the value for each individual in the iteration process. and All of these need to be based on Calculate the corresponding input matrix Then, the corresponding PLSR model is trained using the cross-validation method, and then... Calculate the corresponding output estimation vector Thus through Calculate the objective function value for this individual. The intelligent optimization algorithm provides a result after the iteration process is completed, which makes... The minimal individual, which represents and That is, the optimal value; and They represent and The Middle Each element.

3. The soft measurement method according to claim 1, characterized in that, in In step 2, firstly through right After performing orthogonal signal correction processing, the deep residual PLSR model is then trained; correspondingly, in step 3-2, the model is first trained by... right Orthogonal signal correction processing is performed, and then the terminal estimate is calculated using a deep residual PLSR model; among which, The identity matrix and the orthogonal transformation matrix It is determined according to the following steps A1 to A4: Step A1: Set the orthogonal score vector equal Any column vector in the vector; Step A2, according to right After implementing the update, then according to Calculate the regression coefficient vector ; where the superscript T denotes the transpose of a matrix or vector; Step A3, according to Again After implementing the update process, determine Has convergence been achieved? If not, return to step A2; if yes, proceed to step A4. Step A4, according to Calculate the load vector Then, according to Calculate the orthogonal transformation matrix .

4. A soft measurement method according to claim 1, characterized in that, in When training the PLSR model at layers 1, 2, up to C in sequence, it is preferable to follow the order of training. After performing orthogonal signal correction processing on the input matrices of each layer of the PLSR model, the corresponding PLSR model is trained using cross-validation; correspondingly, the deep residual PLSR model is used as the basis for... In the process of calculating the terminal valuation, first follow After performing orthogonal signal correction processing on the inputs of each layer of the PLSR model, the transformation matrix, loading matrix, and regression vector are then called for corresponding calculations; among them, those belonging to the first layer... Orthogonal transformation matrix of layer PLSR model It is determined according to the following steps B1 to B4: Step B1: Set the orthogonal score vector equal Any column vector in the vector; Step B2, according to right After implementing the update, then according to Calculate the regression coefficient vector ; Step B3, according to Again After implementing the update process, determine Has convergence been achieved? If not, return to step B2; if yes, proceed to step B4. Step B4, according to Calculate the load vector Then, according to Calculate the orthogonal transformation matrix .

5. The soft measurement method according to claim 1, characterized in that, The deep residual PLSR model includes a layer before the first-layer PLSR model, which is obtained through... right The filtering process is performed, and after filtering, the first-layer PLSR model is... , and Perform the corresponding calculations; among them, This represents the input vector after filtering. This represents element-wise multiplication at the same position, in binary vectors. It is determined using a genetic algorithm. During the iterative process of the genetic algorithm, each individual is... A binary vector of dimension, representing any individual The process of calculating the corresponding objective function value is shown in steps C1 to C3: Step C1, through right Each row vector in the input matrix is ​​filtered to obtain the filtered input matrix. ;in, and They represent and The first in Row vector, number ; Step C2, and Using these as the input matrix and output vector of the PLSR model respectively, and training the corresponding PLSR model using cross-validation, the corresponding transformation matrix is ​​obtained. and ; Step C3, according to Calculate the output valuation vector Then, according to Calculate the objective function value for this individual. ;in, and They represent and The first in One element; After the iterative process is complete, the genetic algorithm provides the result that satisfies the objective function value. The minimized individual, whose binary vector is the binary vector for which filtering is performed. The results obtained according to steps C1 to C2 are consistent with... The corresponding PLSR model is the first layer PLSR model in the deep residual PLSR model.

6. The soft measurement method according to claim 1, characterized in that, The deep residual PLSR model also includes a layer before the first-layer PLSR model. For any given input vector The process of implementing weighted processing, weight vector It is determined using the nearest neighbor component analysis algorithm; where the optimization objective of the nearest neighbor component analysis algorithm is to make Minimum; express The Middle elements, regularization coefficient The range of values ​​for is probability regression value The calculation process is as follows: First, calculate the probability according to the following formula ① Then calculate according to formula ② below. : ① ② in, express The first in Line 1 Column elements, express The first in Line 1 Column elements, and They represent The first in and the Each element, numbered ; In the weight vector Once determined, the training of the first-layer PLSR model must first be performed according to... right Each row vector in the middle is weighted and then... and The first-layer PLSR model is trained using cross-validation, with the input matrix and output vector serving as the input and output vectors.

7. A soft measurement method according to claim 6, characterized in that, The nearest neighbor component analysis algorithm is used to determine the corresponding weighting vectors before the PLSR model at layers 2 to C, thereby improving the performance of the model. Weighting is applied; simultaneously, the input matrix is ​​processed... Based on weighted processing, the corresponding PLSR model is trained using cross-validation; where, .

8. A soft measurement method according to any one of claims 1 to 7, characterized in that, Determine the total number of layers in the PLSR model according to steps D1 to D2 as shown below. : Step D1: Train the PLSR model with layers 1, 2, and 3 using cross-validation, and then... Calculate the mean square error of the depth residual PLSR model ;in, for The Middle One element, Indicates passing through the first Output estimate calculated by the layer PLSR model; Step D2, Settings Then, cross-validation was used to train the first... Layered PLSR model, and then calculate the mean square error accordingly. Then determine whether the conditions are met. If so, then set Then, continue training the next layer of the PLSR model until the conditions are no longer met. If not, then the total number of layers in the PLSR model. .

9. A soft measurement device based on deep residual PLSR, comprising a data acquisition module, a central controller, a display unit, and a storage medium; wherein, The storage medium contains the mean and standard deviation for implementing standardization, the transformation matrix, loading matrix and regression vector required to calculate the terminal prediction, and the program for training the PLSR model using cross-validation. The central controller periodically controls the data acquisition module to retrieve the latest sampling time from the industrial process database according to the set sampling time interval. A set of process data is used, and the mean and standard deviation stored in the storage medium are called to perform standardization processing to obtain a column vector. Then, the transformation matrix, load matrix and regression vector stored in the storage medium are called to calculate the corresponding terminal prediction value. After the prediction value is calculated, it is de-standardized to obtain the soft measurement value of the quality data. The soft measurement value is then sent to the display unit for real-time display. The central controller can also, based on initialization commands, instantly control the data acquisition module to retrieve data from the industrial process database. After calculating the mean and standard deviation of the process data and its corresponding quality data, standardization is then performed to form the input matrix. and output vector The mean and standard deviation are then sent to the storage medium for re-storage. Then, the program stored in the storage medium is executed to train the PLSR model of layer 1, layer 2, up to layer C in sequence; finally, the layer C is... The transformation matrix, load matrix, and regression vector of the layered PLSR model are sent to the storage medium for re-storage; Its characteristic is that the process by which the central controller calculates the predicted values ​​of the terminal is as follows: First, it sets up the following sequentially. and through , and Calculate the corresponding score vector Input residual vector and output valuation Then, set the terminal valuation. ; The design flow for programs stored in storage media is as follows: First, ... and The dataset is divided into several sub-blocks, with one sub-block serving as the validation dataset and the remaining sub-blocks as the training dataset. Next, the score vector dimension is constructed using the training dataset. They are respectively equal to The PLSR model is then used to calculate the corresponding mean squared error using the validation dataset. Finally, the dimension of the score vector corresponding to the minimum mean square error is denoted as... , and then and To construct a score vector for the training dataset with a dimension equal to [missing value], [missing value]. The PLSR model is obtained, and the corresponding transformation matrix is ​​preserved. Load matrix and regression vector .

10. A soft measuring device according to claim 9, characterized in that, The storage medium also stores the execution implementation using an intelligent optimization algorithm for the first... Layer PLSR model determination and The optimal value is determined by the program executed first when the central controller receives the initialization command. and After determining the optimal value, the program for training the PLSR model using cross-validation is executed; the intelligent optimization algorithm involved in the program represents the value for each individual in the iteration process. and All of these must be based on Calculate the corresponding input matrix Then, the corresponding PLSR model is established using the cross-validation method, and then... Calculate the corresponding output estimation vector Thus through Calculate the objective function value for this individual. The intelligent optimization algorithm provides a result after the iteration process is completed, which makes... The minimal individual, which represents and That is the optimal value.

Images