An intelligent recommendation method and system for industrial process monitoring models

Abstract

The present invention discloses an intelligent recommendation method and system for industrial process monitoring models, which includes: firstly, facing the problem of lack of intuitive features in industrial time series data, a dataset feature engineering method is used to extract the dataset features of each case from four perspectives: general statistical features, distribution shape features, task-related features, and exogenous knowledge embedding; then, a mapping relationship learning strategy based on width learning is designed to map the dataset features of each case to its corresponding model score vector, thereby giving the width learning network the ability to predict model performance; finally, driven by the multiple incremental requirements of industrial scenarios, a triple incremental strategy is proposed, and the update strategy of the width learning network is designed for the three situations of monitoring case increment, candidate model increment, and recommended node increment, thereby continuously expanding and enhancing the recommendation ability of this method. The present invention can recommend the detection model with the best performance for industrial processes.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial model recommendation, and in particular to an industrial process monitoring model intelligent recommendation method and system. Background Art

[0002] With the advancement of technologies such as intelligent manufacturing and automation, modern industrial processes are moving towards large-scale and intelligent development. However, increasingly complex industrial processes have brought great challenges to the operation and maintenance of equipment. Therefore, in-depth research on industrial monitoring applications (such as fault detection, fault diagnosis, soft measurement, etc.) is of great significance, which helps to avoid accidents and improve production quality. In recent years, with the development of industrial sensors and machine learning algorithms, data-based process monitoring algorithms have attracted widespread attention in the industrial field. Taking fault detection as an example, one of the most popular multivariate statistical process monitoring methods is the principal component analysis method, which explores the autocorrelation of variables by projecting observed data into a latent space. To deal with the non-stationary characteristics of industrial data, researchers have proposed methods such as stationary subspace analysis. Considering that industrial processes exhibit temporal correlation, dynamic latent variable algorithms including slow feature analysis have been proposed. At the same time, monitoring methods based on deep learning are popular due to their effectiveness in nonlinear information processing, including autoencoders, convolutional neural networks, recurrent neural networks, Transformers, etc. In summary, inspired by the characteristics of industrial processes, researchers have proposed and continuously developed a series of methods.

[0003] Although these algorithms have achieved good results in specific industrial applications, no single algorithm can currently handle all scenarios. For example, principal component analysis may not be able to handle non-Gaussian or multimodal process data, slow feature analysis methods may perform poorly on multi-order time-correlated data, and neural network methods may memorize patterns in training data but fail to generalize to unseen samples when training data is limited. In fact, a study proposed the no free lunch theorem, which proves that if learning algorithm A is better than learning algorithm B on some problems, then B must be better than A on other problems. The theorem shows that no algorithm can perform well in all scenarios. Therefore, faced with diverse scenarios and the rapid growth of industrial time series monitoring algorithms, there is no universal winner, but experts suitable for specific scenarios can be found. Therefore, a key issue is how to automatically identify a suitable monitoring model given an industrial case and its corresponding time series data.

[0004] To solve this problem, an intuitive solution is to train and evaluate all available models one by one and select the one with the best performance. However, training and evaluating all candidate models one by one is time-consuming, especially for deep learning models with a large number of hyperparameters. When the number of devices to be monitored is large, the time spent is unacceptable. More importantly, the evaluation process may be difficult to carry out if only normal or unlabeled data is available. However, choosing a suitable model for a given case is similar to recommending products to users who may be interested, which can be achieved through recommendation systems. In the past few decades, recommendation systems have achieved great success and can provide the most relevant recommendations for specific users. In general, existing recommendation systems are mainly concentrated in the commercial field and can be divided into content-based filtering, collaborative filtering, and hybrid filtering, all of which rely on intuitive user profile information such as gender, age, purchase history, product evaluation, etc. However, when designing model recommendation methods for industrial monitoring, time series process data (similar to users) are usually composed of high-dimensional sensor signals, which lack intuitive features associated with relevant purchase behaviors in commercial scenarios. In addition, industrial monitoring scenarios are dynamically changing, including new candidate algorithms that emerge with technological advances and new monitoring cases that emerge with the deployment of new equipment. Industrial model recommendation methods need to have incremental capabilities. Therefore, inspired by the concept of recommendation systems, this paper proposes to construct an industrial model recommendation method that overcomes the challenges of industrial signals lacking intuitive features and dynamic changes in monitoring scenarios. This method can treat the monitored object as a user and automatically recommend appropriate monitoring models as products for them. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent recommendation method and system for industrial process monitoring models in view of the gap in the research field of industrial model recommendation in the prior art. The present invention proposes a new task paradigm for industrial model recommendation for the first time, and constructs a recommendation method based on a wide learning network, which can map the data set features obtained through data set feature engineering to the scores of the data set on each model, thereby providing preliminary recommendation results for model recommendation, and greatly avoiding the high cost of one-by-one testing and evaluation; the present invention can be applied to various industrial fields including but not limited to energy, chemical industry, injection molding, etc.

[0006] The objective of the present invention is achieved through the following technical solutions: In a first aspect, an embodiment of the present invention provides an intelligent recommendation method for an industrial process monitoring model, comprising the following steps:

[0007] (1) Obtain time series data sets of N industrial objects, process all time series data sets using data set feature engineering methods to obtain data set feature vectors corresponding to each time series data set, and concatenate all data set feature vectors together to obtain a data set feature matrix;

[0008] (2) For each time series data set, use the training set data in the time series data set to train the preset M different models, and test them on the test set data to obtain a data set score vector composed of the test results of all models corresponding to the time series data set, and then concatenate all the data set score vectors to obtain a data set score matrix;

[0009] (3) Using the dataset feature matrix and the dataset score matrix, a width learning network for establishing a mapping relationship between the feature matrix and the score matrix is ​​trained to obtain a trained width learning network; and a triple incremental strategy is used to incrementally update and optimize the trained width learning network to obtain an updated width learning network;

[0010] (4) When applied online, the target dataset is processed using a dataset feature engineering method to obtain a dataset feature vector corresponding to the target dataset, which is then input into a trained or updated width learning network to obtain a dataset score vector corresponding to all models. The model with the highest score is selected as the recommended model.

[0011] Furthermore, the step (1) specifically includes:

[0012] Get a time series dataset of N industrial objects and record each time series dataset as , where m represents the total number of variables in the time series dataset, Represents the total number of samples in the time series dataset; all time series datasets are processed using the dataset feature engineering method to convert each time series dataset into a one-dimensional feature vector, specifically extracting the general statistical features, distribution shape features, task-related features, and exogenous knowledge embedding features of each time series dataset, and using the one-dimensional feature vector as the dataset feature vector corresponding to the time series dataset ,in represents the feature vector of the i-th data set, K represents the number of data set features; the data set feature vectors of N time series data sets are concatenated together to obtain the data set feature matrix .

[0013] Furthermore, the general statistical features include the total number of samples in the data set , total number of data set variables m, mean value of data set , median , Standard Deviation , maximum value , minimum , four sample quantiles , interquartile range , normalized mean , normalized median , sample range, sample Gini coefficient, absolute median deviation , absolute mean deviation , Quantile coefficient dispersion , Variance Coefficient , in and The number of outliers outside and The proportion of abnormal points outside and The number of outliers outside and The number of abnormal points outside the range of 1% to 99% and the number of abnormal points outside the range of 1% to 99% and the number of abnormal points outside the range of 1% to 99% and The number of abnormal points other than The proportion of abnormal points other than the above, the mean of the sparsity of each variable, the maximum value of the sparsity of each variable, the minimum value of the sparsity of each variable, the standard deviation of the sparsity of each variable, the skewness of the sparsity of each variable, the kurtosis of the sparsity of each variable, the mean of the entropy of each variable, the maximum value of the entropy of each variable, the minimum value of the entropy of each variable, the standard deviation of the entropy of each variable, the skewness of the entropy of each variable, the kurtosis of the entropy of each variable, the mean of the covariance between variables, the maximum value of the covariance between variables, the minimum value of the covariance between variables, the standard deviation of the covariance between variables, the The skewness of variance, the kurtosis of covariance between variables, the mean of correlation between variables, the maximum value of correlation between variables, the minimum value of correlation between variables, the standard deviation of correlation between variables, the skewness of correlation between variables, the kurtosis of correlation between variables, the mean of P value of ANOVA for each variable, the maximum value of P value of ANOVA for each variable, the minimum value of P value of ANOVA for each variable, the standard deviation of P value of ANOVA for each variable, the skewness of P value of ANOVA for each variable, the kurtosis of P value of ANOVA for each variable; where IQR means interquartile range;

[0014] The distribution shape characteristics include normality test, mean of skewness of each variable, maximum value of skewness of each variable, minimum value of skewness of each variable, standard deviation of skewness of each variable, skewness of skewness of each variable, kurtosis of skewness of each variable, mean of kurtosis of each variable, maximum value of kurtosis of each variable, minimum value of kurtosis of each variable, standard deviation of kurtosis of each variable, skewness of kurtosis of each variable, kurtosis of kurtosis of each variable, mean of fifth-order moment of each variable, maximum value of fifth-order moment of each variable, minimum value of fifth-order moment of each variable, standard deviation of fifth-order moment of each variable, skewness of fifth-order moment of each variable, kurtosis of fifth-order moment of each variable, mean of sixth-order moment of each variable, maximum value of sixth-order moment of each variable, minimum value of sixth-order moment of each variable, standard deviation of sixth-order moment of each variable, skewness of sixth-order moment of each variable, kurtosis of sixth-order moment of each variable degree, the mean of the seventh-order moment of each variable, the maximum value of the seventh-order moment of each variable, the minimum value of the seventh-order moment of each variable, the standard deviation of the seventh-order moment of each variable, the skewness of the seventh-order moment of each variable, the kurtosis of the seventh-order moment of each variable, the mean of the eighth-order moment of each variable, the maximum value of the eighth-order moment of each variable, the minimum value of the eighth-order moment of each variable, the standard deviation of the eighth-order moment of each variable, the skewness of the eighth-order moment of each variable, the kurtosis of the eighth-order moment of each variable, the mean of the ninth-order moment of each variable, the maximum value of the ninth-order moment of each variable, the minimum value of the ninth-order moment of each variable, the standard deviation of the ninth-order moment of each variable, the skewness of the ninth-order moment of each variable, the kurtosis of the ninth-order moment of each variable, the mean of the tenth-order moment of each variable, the maximum value of the tenth-order moment of each variable, the minimum value of the tenth-order moment of each variable, the standard deviation of the tenth-order moment of each variable, the skewness of the tenth-order moment of each variable, the kurtosis of the tenth-order moment of each variable;

[0015] The task-related characteristics include the standard deviation of the first principal component, the skewness of the first principal component, the kurtosis of the first principal component, the proportion of explained variance of the first three principal components, the proportion of singular values ​​of the first three typical correlation components, the ADF test of the first three stationary components, the proportion of singular values ​​of the first three stationary components, the filling error of linear regression, the filling error of nonlinear regression, and the linearity based on the proportion of filling error;

[0016] The exogenous knowledge embedding feature is specifically obtained by the following method: the industrial process and equipment to which the time series data set belongs are input into a large language model, and the answers of the large language model to five questions are encoded 0-1 as exogenous knowledge embedding features; the five questions are: whether it is a multi-stage industrial process, whether there are fluctuations in raw materials, whether there are changes in set points, whether it is a closed-loop system, and whether complex chemical reactions occur.

[0017] Furthermore, the method of using the dataset feature matrix and the dataset score matrix to train a width learning network for establishing a mapping relationship between the feature matrix and the score matrix to obtain a trained width learning network specifically includes:

[0018] (3.1) The feature matrix of the dataset is mapped to feature nodes through the mapping matrix, which is expressed as:

[0019]

[0020] in, represents the rth feature node, represents the activation function, represents the feature matrix of the dataset, represents a vector whose elements are all 1. Represents a randomly initialized orthogonal matrix corresponding to the feature node;

[0021] (3.2) Connect the R groups of feature nodes obtained in step (3.1) to obtain the connected feature nodes ;

[0022] (3.3) Using the spliced ​​feature nodes Compute the enhanced node, expressed as:

[0023]

[0024] in, represents an enhanced node, represents the activation function, represents a randomly initialized orthogonal matrix corresponding to the enhanced node, and E represents the output dimension of the enhanced node;

[0025] (3.4) Concatenate the feature nodes and the enhanced nodes, denoted as , and use it as the input data of the width learning network. Then the objective function of the width learning network is:

[0026]

[0027] in, is the linear mapping to be solved, represents the dataset score matrix, represents the score vector of the i-th data set, represents the penalty coefficient;

[0028] (3.5) Linear mapping to be solved in the objective function of the width learning network , first use the ridge regression method to solve the matrix The pseudo-inverse of The pseudo-inverse of the dataset score matrix is ​​solved and expressed as:

[0029]

[0030]

[0031] in, Representation Matrix The pseudo-reversal, Representation Matrix The transpose of Represents the identity matrix.

[0032] Furthermore, the triple incremental strategy includes an update strategy of monitoring case increment, candidate model increment and recommendation node increment.

[0033] Furthermore, the update strategy of the monitoring case increment specifically includes:

[0034] When a new time series data set is collected, first obtain the feature matrix of the newly added data set And the dataset score matrix , then the feature nodes and enhanced nodes corresponding to the newly added time series dataset are expressed as:

[0035]

[0036]

[0037] in, Represents the concatenated feature nodes corresponding to the newly added time series dataset. Indicates the number of newly added time series datasets, Indicates the enhanced node corresponding to the newly added time series dataset;

[0038] Note the newly added input data , then the updated complete input data is:

[0039]

[0040] in, represents the updated complete input data, then its pseudo-inverse is expressed as:

[0041]

[0042]

[0043]

[0044] in, for The pseudo-reversal, and It is an intermediate calculation variable and has no practical significance;

[0045] Then the updated linear mapping to be solved is:

[0046]

[0047] in, Represents the updated linear mapping to be solved.

[0048] Furthermore, the update strategy of the candidate model increment specifically includes:

[0049] As the number of candidate models increases, the dataset score matrix changes from Expanded to ,in represents the number of newly added models, then the linear mapping to be solved is updated as:

[0050]

[0051] in, Indicates newly added The dataset score matrix corresponding to the candidate models.

[0052] Furthermore, the update strategy of the recommended node increment specifically includes:

[0053] Before the feature nodes and enhancement nodes are incremented, their respective numbers are and , assuming that the number of newly added feature nodes is , the number of newly added enhanced nodes is , then the newly added feature nodes and enhanced nodes are expressed as:

[0054]

[0055]

[0056] in, Represents the newly added feature node after splicing. Indicates a newly added enhanced node. Indicates the number of nodes in a newly added set of feature nodes, Indicates The newly added feature nodes are linearly mapped. Indicates a newly added enhanced node, then the updated input data It is expressed as:

[0057]

[0058] in, Represents the newly added input data; based on the above formula, solve the matrix The pseudo-inverse of is calculated as:

[0059]

[0060]

[0061]

[0062] in, Representation Matrix The pseudo-inverse of The pseudo-inverse of and the dataset score matrix obtain the solution of the updated linear mapping in the case of incremental recommendation nodes, expressed as:

[0063]

[0064] in, Represents the updated linear mapping.

[0065] A second aspect of an embodiment of the present invention provides a recommendation system for implementing the above-mentioned industrial process monitoring model intelligent recommendation method, including:

[0066] A case library, consisting of various failure cases of industrial equipment involved in the industrial process, for providing failure cases for the recommendation system;

[0067] A candidate model library, consisting of industrial process monitoring models of different types and with different hyperparameter settings, for providing different preset models for the recommendation system;

[0068] A scoring rule library, including a plurality of scoring rules, wherein the scoring rules are used to evaluate the monitoring performance of models of different types and different hyperparameters in the model library;

[0069] The data set feature matrix acquisition module is used to call the fault case library to obtain the time series data sets of N industrial objects, and use the data set feature engineering method to process all the time series data sets to obtain the data set feature vectors corresponding to each time series data set, and splice all the data set feature vectors together to obtain the data set feature matrix;

[0070] The data set score matrix acquisition module is used to train different models in the candidate model library using the training set data in the fault case library, and use the test set data in the fault case library to perform performance evaluation on different models according to the scoring rules in the scoring rule library, and obtain the data set score vector composed of the performance evaluation results of all models, and splice all the data set score vectors to obtain the data set score matrix;

[0071] A width learning network training and updating module is used to train a width learning network for establishing a mapping relationship between a feature matrix and a score matrix using a data set feature matrix and a data set score matrix, so as to obtain a trained width learning network; and to incrementally update and optimize the trained width learning network using a triple incremental strategy, so as to obtain an updated width learning network; and

[0072] The online application module is used to process the target dataset using the dataset feature engineering method to obtain the dataset feature vector corresponding to the target dataset, and input it into the trained or updated width learning network to obtain the dataset score vectors corresponding to all models, and select the model with the highest score as the recommended model.

[0073] Furthermore, the fault cases cover a variety of fault types, including step faults, intermittent faults, initial faults, slow-changing faults, drift faults, and stuck faults; each fault case contains corresponding training data, test data, and fault occurrence time, wherein the training data is used for model training of the model library and construction of data set feature vectors, and the test data is used to evaluate the performance of each model in the model library according to the fault occurrence time and the scoring rule library;

[0074] The candidate model library includes principal component analysis model, slow feature analysis model, stationary subspace analysis model, canonical correlation analysis model, autoencoder model, variational autoencoder model, convolutional autoencoder model, recurrent neural network model, and Transformer model;

[0075] The specific scoring rule is: use the F1 score of the model as the evaluation result, and its calculation formula is:

[0076]

[0077] Among them, Precision is the model accuracy, which indicates the proportion of real fault samples among the samples predicted to be faults; Recall is the model recall rate, which indicates the proportion of samples predicted to be faults among the real fault samples.

[0078] Compared with the prior art, the present invention has the following beneficial effects:

[0079] (1) The present invention designs a dataset feature engineering method that can uniformly extract dataset-level features from datasets of different variables and samples, and on this basis designs a width learning network to learn the mapping relationship from dataset features to model scores, thereby facilitating the rapid prediction of the model with the best performance based on the dataset features, greatly reducing the cost of individual experiments and evaluations.

[0080] (2) In addition, the present invention also makes update designs according to three incremental situations that may be encountered in reality, so as to facilitate the rapid update and migration of the model in various incremental situations.

[0081] (3) The present invention can discover statistical characteristics, task characteristics, mechanism knowledge, etc. from the abstract time series of the target object, build a bridge from the industrial process to the monitoring model, automatically recommend a suitable model for the target object, get rid of the costly one-by-one experiments and evaluations, and provide a modeling paradigm for the industrial model recommendation of time series monitoring tasks.

[0082] (4) The present invention proposes for the first time a new task paradigm for industrial model recommendation. The intelligent recommendation system for industrial process monitoring models constructed by the present invention provides a reference for case library construction, model library construction, and model evaluation rules. It also provides a reference benchmark and practical assistance for feature engineering of large-scale data sets, recommendation method training, and model online recommendation. The present invention is universal and can be extended to model recommendation for other industrial tasks, including but not limited to fault diagnosis, soft measurement, data filling, etc., by adjusting the design of task-related features. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 A flowchart of the industrial process monitoring model intelligent recommendation method of the present invention;

[0084] Figure 2 It is a schematic diagram of monitoring case increment of the present invention;

[0085] Figure 3 It is a schematic diagram of the candidate model increment of the present invention;

[0086] Figure 4 It is a schematic diagram of the recommended node increment of the present invention;

[0087] Figure 5 It is a flow chart of the architecture of the intelligent recommendation system for industrial process monitoring model of the present invention;

[0088] Figure 6 A heat map of some weights in the intelligent recommendation method for industrial process monitoring models of the present invention;

[0089] Figure 7 This is the influence curve of the recommended node increment on the recommendation performance in the industrial process monitoring model intelligent recommendation method of the present invention. DETAILED DESCRIPTION

[0090] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0091] The terms used in the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0092] It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0093] The present invention is described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the features of the following embodiments and implementations can be combined with each other.

[0094] The overall framework process of the industrial process monitoring model intelligent recommendation method of the present invention is as follows: Figure 1 As shown in the figure, its working principle is as follows: first, obtain time series data sets of multiple industrial objects, and perform data set feature engineering on these data sets; then, splice the obtained data set features into a data set feature matrix, and test all models in the model library on these data sets one by one to obtain a score matrix; then, train the width learning network to learn the mapping relationship between the data set feature matrix and the score matrix; finally, during the online test, the feature vector of the target data set after feature engineering is input into the trained width learning network to obtain a score vector, and the model with the highest score is the model recommendation result.

[0095] like Figure 1 As shown, the model intelligent recommendation method specifically includes the following steps:

[0096] (1) Obtain time series data sets of N industrial objects, process all time series data sets using data set feature engineering methods to obtain data set feature vectors corresponding to each time series data set, and concatenate all data set feature vectors together to obtain a data set feature matrix.

[0097] It should be understood that industrial objects refer to specific equipment involved in industrial processes, such as pre-pumps, air preheaters, and blowers in coal-fired power generation processes.

[0098] Specifically, obtain a time series data set of N industrial objects and record each time series data set as , where m represents the total number of variables in the time series dataset, Represents the total number of samples in the time series dataset; all time series datasets are processed using the dataset feature engineering method to convert each time series dataset into a one-dimensional feature vector, specifically extracting the general statistical features, distribution shape features, task-related features, and exogenous knowledge embedding features of each time series dataset, and using the one-dimensional feature vector as the dataset feature vector corresponding to the time series dataset ,in represents the feature vector of the i-th data set, K represents the number of data set features; the data set feature vectors of N time series data sets are concatenated together to obtain the data set feature matrix .

[0099] Furthermore, common statistical features include the total number of samples in the dataset , total number of data set variables m, mean value of data set , median , Standard Deviation , maximum value , minimum , four sample quantiles , interquartile range (Interquartile Range, IQR), normalized mean , normalized median , sample range, sample Gini coefficient, absolute median deviation , absolute mean deviation , Quantile coefficient dispersion , Variance Coefficient , in and The number of abnormal points outside and The proportion of abnormal points outside and The number of abnormal points outside and The number of abnormal points outside the range of 1% to 99% and the number of abnormal points outside the range of 1% to 99% and the number of abnormal points outside the range of 1% to 99% and The number of abnormal points other than The proportion of abnormal points other than , the mean of the sparsity of each variable, the maximum value of the sparsity of each variable, the minimum value of the sparsity of each variable, the standard deviation of the sparsity of each variable, the skewness of the sparsity of each variable, the kurtosis of the sparsity of each variable, the mean of the entropy of each variable, the maximum value of the entropy of each variable, the minimum value of the entropy of each variable, the standard deviation of the entropy of each variable, the skewness of the entropy of each variable, the kurtosis of the entropy of each variable, the mean of the covariance between variables, the maximum value of the covariance between variables, the minimum value of the covariance between variables, the standard deviation of the covariance between variables, the variables The skewness of the covariance between variables, the kurtosis of the covariance between variables, the mean of the correlation between variables, the maximum value of the correlation between variables, the minimum value of the correlation between variables, the standard deviation of the correlation between variables, the skewness of the correlation between variables, the kurtosis of the correlation between variables, the mean of the P-value (p-value) of the ANOVA of each variable, the maximum value of the P-value of the ANOVA of each variable, the minimum value of the P-value of the ANOVA of each variable, the standard deviation of the P-value of the ANOVA of each variable, the skewness of the P-value of the ANOVA of each variable, and the kurtosis of the P-value of the ANOVA of each variable.

[0100] It should be understood that the calculation methods of the above general statistical features are well known. For example, the sample Gini coefficient can be calculated as , where n is the number of sample values, is the ith value, is the average value, data usage All values, not all values ​​of a certain variable. The values ​​are sorted from low to high. All calculation methods are not described in detail here.

[0101] Furthermore, the distribution shape characteristics include normality test, the mean of the skewness of each variable, the maximum value of the skewness of each variable, the minimum value of the skewness of each variable, the standard deviation of the skewness of each variable, the skewness of the skewness of each variable, the kurtosis of each variable, the mean of the kurtosis of each variable, the maximum value of the kurtosis of each variable, the minimum value of the kurtosis of each variable, the standard deviation of the kurtosis of each variable, the skewness of the kurtosis of each variable, the kurtosis of each variable, the mean of the fifth-order moment of each variable, the maximum value of the fifth-order moment of each variable, the minimum value of the fifth-order moment of each variable, the standard deviation of the fifth-order moment of each variable, the skewness of the fifth-order moment of each variable, the kurtosis of the fifth-order moment of each variable, the mean of the sixth-order moment of each variable, the maximum value of the sixth-order moment of each variable, the minimum value of the sixth-order moment of each variable, the standard deviation of the sixth-order moment of each variable, the skewness of the sixth-order moment of each variable, the skewness of the sixth-order moment of each variable Kurtosis, mean of the seventh-order moment of each variable, maximum value of the seventh-order moment of each variable, minimum value of the seventh-order moment of each variable, standard deviation of the seventh-order moment of each variable, skewness of the seventh-order moment of each variable, kurtosis of the seventh-order moment of each variable, mean of the eighth-order moment of each variable, maximum value of the eighth-order moment of each variable, minimum value of the eighth-order moment of each variable, standard deviation of the eighth-order moment of each variable, skewness of the eighth-order moment of each variable, kurtosis of the eighth-order moment of each variable, mean of the ninth-order moment of each variable, maximum value of the ninth-order moment of each variable, minimum value of the ninth-order moment of each variable, standard deviation of the ninth-order moment of each variable, skewness of the ninth-order moment of each variable, kurtosis of the ninth-order moment of each variable, mean of the tenth-order moment of each variable, maximum value of the tenth-order moment of each variable, minimum value of the tenth-order moment of each variable, standard deviation of the tenth-order moment of each variable, skewness of the tenth-order moment of each variable, kurtosis of the tenth-order moment of each variable.

[0102] It should be understood that the calculation methods of the above distribution shape characteristics are well known. For example, for a single variable, its skewness can be calculated as , where n is the number of values ​​of the variable, is the ith value, is the average value; the mean of the skewness of each variable is calculated by averaging the skewness of all variables. All calculation methods are not specifically introduced here.

[0103] Furthermore, the task-related characteristics include the standard deviation of the first principal component, the skewness of the first principal component, the kurtosis of the first principal component, the proportion of explained variance of the first three principal components, the proportion of singular values ​​of the first three canonical correlation components, the ADF test of the first three stationary components, the proportion of singular values ​​of the first three stationary components, the imputation error of linear regression, the imputation error of nonlinear regression, and the linearity based on the proportion of imputation error.

[0104] It should be understood that the calculation methods of the above task-related features are well known. For example, the explained variance percentage of the first three principal components can be calculated as ,in It represents the eigenvalue of the jth principal component. We will not explain all the calculation methods in detail here.

[0105] It should be noted that the method of the present invention can be extended to model recommendation for other industrial tasks, including but not limited to fault diagnosis, soft measurement, data filling, etc., by adjusting the design of task-related features.

[0106] Furthermore, the exogenous knowledge embedding features are specifically obtained through the following method: the industrial processes and equipment to which the time series data set belongs are input into the large language model, and the answers of the large language model to five questions are encoded 0-1 as exogenous knowledge embedding features; the five questions are: whether it is a multi-stage industrial process, whether there are fluctuations in raw materials, whether there are changes in set points, whether it is a closed-loop system, and whether complex chemical reactions occur.

[0107] It should be understood that the large language model here can use an existing pre-trained large language model, such as ChatGPT. If the exogenous knowledge embedding feature output by the large language model is 01110, it means that the industrial process corresponding to the time series data set is not a multi-stage industrial process, its raw materials fluctuate, there are changes in the set points, it is a closed-loop system, and no complex chemical reactions occur.

[0108] (2) For each time series data set, use the training set data in the time series data set to train the preset M different models, and test them on the test set data to obtain the data set score vector composed of the test results of all models corresponding to the time series data set , and concatenate all the dataset score vectors to get the dataset score matrix ,in Represents the score vector of the i-th dataset.

[0109] It should be noted that each time series data set is randomly divided into training sets and test sets, which are preset in the system after the division is completed; at the same time, M different models are preset in the system's model library, and the scoring rules corresponding to each model are preset in the system's scoring rule library, which can be directly called when needed.

[0110] (3) Using the dataset feature matrix and the dataset score matrix, a width learning network is trained to establish the mapping relationship between the feature matrix and the score matrix to obtain a trained width learning network; and a triple incremental strategy is used to incrementally update and optimize the trained width learning network to obtain an updated width learning network.

[0111] It should be noted that the width learning network is to establish a mapping relationship between the feature matrix and the score matrix. Therefore, the feature matrix of the data set can be used as the input of the width learning network, and the corresponding score matrix of the data set can be used as the output of the width learning network. The width learning network is a single-layer mapping. and , and then and Put them together and map them to the corresponding dataset score matrix through a matrix.

[0112] Furthermore, the dataset feature matrix and the dataset score matrix are used to train a width learning network for establishing a mapping relationship between the feature matrix and the score matrix to obtain a trained width learning network, specifically including:

[0113] (3.1) First, the feature matrix of the dataset is mapped to feature nodes through the matrix to be mapped, which is expressed as:

[0114]

[0115] in, represents the rth feature node, represents the activation function, represents the feature matrix of the dataset, represents a vector whose elements are all 1. Represents a randomly initialized orthogonal matrix corresponding to the feature node.

[0116] (3.2) Then, the R groups of feature nodes obtained in step (3.1) are spliced ​​to obtain the spliced ​​feature nodes , which is used as input for computing the boost node.

[0117] (3.3) Reuse the spliced ​​feature nodes Compute the enhanced node, expressed as:

[0118]

[0119] in, represents an enhanced node, represents the activation function, Represents a randomly initialized orthogonal matrix corresponding to the enhanced node, E is The number of columns indicates the output dimension of the enhanced node.

[0120] (3.4) Concatenate the feature nodes and the enhanced nodes, denoted as , and use it as the input data of the width learning network. Then the objective function of the width learning network is:

[0121]

[0122] in, is the linear mapping to be solved, represents the dataset score matrix, represents the score vector of the i-th data set, Represents the penalty coefficient.

[0123] (3.5) Linear mapping to be solved in the objective function of the width learning network , first use the ridge regression method to solve the matrix The pseudo-inverse of The pseudo-inverse of the dataset score matrix is ​​solved and expressed as:

[0124]

[0125]

[0126] in, Representation Matrix The pseudo-reversal, Representation Matrix The transpose of Represents the identity matrix.

[0127] In this embodiment, a triple incremental strategy is adopted to incrementally update and optimize the trained width learning network to obtain an updated width learning network, specifically including: incrementally updating and optimizing the width learning network for three incremental situations that may exist in reality, namely, incremental monitoring cases, incremental candidate models, and incremental recommended nodes. The triple incremental strategy includes update strategies for incremental monitoring cases, incremental candidate models, and incremental recommended nodes.

[0128] Furthermore, the update strategy for monitoring case increments specifically includes: when new cases (i.e., new time series data sets) are collected, such as Figure 2 As shown, first go through the above steps (1) and (2) to obtain the new data set feature matrix And the dataset score matrix , then the feature nodes and enhanced nodes corresponding to the newly added time series dataset are expressed as:

[0129]

[0130]

[0131] in, Represents the concatenated feature nodes corresponding to the newly added time series dataset. Indicates the number of newly added time series datasets. Indicates the enhanced node corresponding to the newly added time series data set. , then the updated complete input data (i.e. the augmentation of the historical data set and the newly added data set) is:

[0132]

[0133] in, represents the updated complete input data, then its pseudo-inverse is expressed as:

[0134]

[0135]

[0136]

[0137] in, for The pseudo-reversal, and is an intermediate calculation variable and has no practical significance. Therefore, the updated linear mapping to be solved can be calculated and expressed as:

[0138]

[0139] in, Represents the updated linear mapping to be solved.

[0140] Furthermore, the update strategy of the candidate model increment specifically includes: when the number of candidate models increases, such as Figure 3 As shown, the dataset score matrix is ​​transformed from Expanded to ,in Represents the number of newly added models. At this time, the linear mapping to be solved is updated as follows:

[0141]

[0142] in, Indicates newly added The dataset score matrix corresponding to the candidate models. The candidate models are After training each case with the training data, the test data of each case is used to train the The models are tested and the obtained F1 scores form the score matrix of the new model. Since the cases and feature nodes are not affected by the increment of candidate models, the increment of candidate models With no increment Be consistent.

[0143] It should be noted that the formula for the candidate model increment is obtained The dimension is ,therefore The output dimension is , so the model's score predictions cover model, thus realizing the candidate model from the original To The model increments.

[0144] Furthermore, the update strategy of the recommended node increment specifically includes: the recommended node increment is as follows: Figure 4 As shown in the figure, before the feature nodes and enhancement nodes are incremented, their respective numbers are and , assuming that the number of newly added feature nodes is , the number of newly added enhanced nodes is , then the newly added feature nodes and enhanced nodes are expressed as:

[0145]

[0146]

[0147] in, Represents the newly added feature node after splicing. Indicates a newly added enhanced node. Indicates the number of nodes in a newly added set of feature nodes, Indicates The newly added feature nodes are linearly mapped. Indicates a newly added enhanced node, then the updated input data It can be expressed as:

[0148]

[0149] in, Represents the newly added input data. Based on the above formula, solve the matrix The pseudo-inverse of is calculated as:

[0150]

[0151]

[0152]

[0153] in, Representation Matrix The pseudo-reversal, and It is an intermediate calculation variable and has no practical significance. Finally, according to the matrix The pseudo-inverse of and the dataset score matrix obtain the solution of the updated linear mapping in the case of incremental recommendation nodes, expressed as:

[0154]

[0155] in, Represents the updated linear mapping.

[0156] (4) When applied online, the target dataset is processed using a dataset feature engineering method to obtain a dataset feature vector corresponding to the target dataset, which is then input into a trained or updated width learning network to obtain a dataset score vector corresponding to all models. The model with the highest score is selected as the recommended model.

[0157] Specifically, when applied online, the industrial dataset that requires model recommendation is used as the target dataset. The target dataset is first processed using the dataset feature engineering method to obtain the dataset feature vector corresponding to the target dataset. The dataset feature vector is then input into the trained width learning network or the updated width learning network to obtain the dataset score vector corresponding to all models. In the dataset score vector, each element represents the score of a model. Therefore, the scores of all models can be obtained according to the dataset score vector, and the model with the highest score is selected as the recommended model. Alternatively, the G models with the highest scores can be selected as the preliminary recommendation results, and the data in the target dataset are used to further test the G models, and the model with the best performance is selected as the final recommended model.

[0158] It is worth mentioning that the embodiment of the present invention also provides a recommendation system for implementing the industrial process monitoring model intelligent recommendation method of the above embodiment. Figure 5 As shown, the system includes a case library, a candidate model library, a scoring rule library, a data set feature matrix acquisition module, a data set score matrix acquisition module, a width learning network training and updating module and an online application module.

[0159] In this embodiment, the case library consists of various fault cases of typical industrial equipment involved in typical industrial processes, and is used to provide fault cases for the recommendation system.

[0160] Furthermore, typical industrial processes include coal-fired power generation processes, chemical processes, biomass fuel pelletizing processes, etc. Faults involve various industrial systems in industrial processes. Taking the coal-fired power generation process as an example, it includes steam turbine systems, power generation systems, etc. The faults of the steam turbine system include low vibration of the turbine bearing and high temperature of the pre-pump bearing. The faults of the power generation system include abnormal pressure difference of the coal mill and high temperature of the fan bearing of the blower. The fault cases cover typical fault types, including step faults, intermittent faults, initial faults, slow-changing faults, drift faults, and jamming faults. Taking the coal-fired power generation process as an example, the high temperature of the pre-pump bearing, abnormal flue gas pressure of the air preheater, and high temperature of the blower bearing are all slow-changing faults, and the high temperature of the lower bearing of the condensate pump motor and abnormal steam flow of the waste heat boiler are step faults. Each fault case contains training data, test data, and fault occurrence time. Among them, the training data is used for model training of the model library and the construction of the data set feature vector in the data set feature engineering method; the test data is used to evaluate the performance of each model in the model library according to the fault occurrence time and the scoring rule library.

[0161] Specifically, in this embodiment, there are 69 fault cases in the case library, including coal-fired power generation process, gas-fired power generation process, Tennessee-Eastman process, three-phase flow process, and biomass fuel pelletizing process; covering typical fault types, including step, intermittent fault, initial fault, slow-changing fault, drift, jamming and other fault types. The corresponding relationship between each industrial process and the case is shown in Table 1, where some industrial processes are described as follows: the coal-fired power generation process is a real industrial process of a million-kilowatt supercritical thermal power unit. The equipment involved in the fault includes coal mill, blower, primary fan, induced draft fan, etc. The number of variables in different fault cases is different, the sampling interval is 1 minute or 10 minutes, and a total of 26 faults are included; the three-phase flow process is a process in which the flow of water, oil and air in a pressurized system changes dynamically during the control process. Considering its 16 faults, the type of fault can be early, intermittent or sudden. The data set has 23 variables and the sampling interval is 1 second. The present invention has strong versatility. Although it has only been verified in multiple groups of cases of the above various industrial processes, it can actually be applied to various industrial fields including but not limited to energy, chemical industry, injection molding, etc. Some industrial processes involve a variety of industrial systems. Taking the coal-fired power generation process as an example, the various industrial systems involved in the coal-fired power generation process are shown in Table 2. The faults include low vibration of the turbine bearing, low vibration of the turbine mechanical bearing, high stator temperature of the pre-pump motor, high temperature of the pre-pump bearing, etc. of the steam turbine system, abnormal pressure difference of the coal mill in the power generation system, low temperature of the secondary air of the air preheater, and high temperature of the fan bearing of the blower; the fault cases cover typical fault types, including step, intermittent fault, initial fault, slow-changing fault, drift, jamming and other fault types. Taking the coal-fired power generation process as an example, as shown in Table 3, the high temperature of the pre-pump bearing, abnormal flue gas pressure of the air preheater, and high temperature of the blower bearing are all slow-changing faults, and the high temperature of the lower bearing of the condensate pump motor and the abnormal steam flow of the waste heat boiler are step faults.

[0162] Table 1: Number of cases for each industrial process

[0163]

[0164] Table 2: Industrial systems of coal-fired power generation process

[0165]

[0166] Table 3: Fault types in coal-fired power generation

[0167]

[0168] In this embodiment, the candidate model library is composed of industrial process monitoring models of different types and different hyperparameter settings, which are used to provide different preset models for the recommendation system. Among them, different types of models can be classified into models based on typical statistical analysis methods and models based on deep learning methods. Typical statistical analysis methods include but are not limited to principal component analysis methods, slow feature analysis methods, stationary subspace analysis methods, etc., and deep learning methods include but are not limited to autoencoders, recurrent neural networks, Transformers, etc.; different hyperparameters include but are not limited to cumulative variance contribution, confidence, training rounds, number of neurons, model learning rate, etc. For example, the principal component analysis method can set different numbers of principal components, and the autoencoder can set different numbers of neurons and training rounds. The model trained under each hyperparameter setting is used as an independent model in the model library. These monitoring models of different types and different hyperparameters are trained using the training data of each fault case in the case library, and the model performance is evaluated based on the fault occurrence time and scoring rules on the test data of each fault case in the case library.

[0169] Furthermore, the candidate model library includes but is not limited to nine basic models: principal component analysis (PCA) model, slow feature analysis (SFA) model, stationary subspace analysis (SSA) model, canonical correlation analysis (CVA) model, autoencoder (AE) model, variational autoencoder (VAE) model, convolutional autoencoder (convAE) model, recurrent neural network (GRU) model, and Transformer model. In the PCA setting, a total of 5 models were defined, the confidence level was fixed at 0.99, and the cumulative variance contribution rates were set to 0.75, 0.80, 0.85, 0.90 and 0.95, respectively, to select different numbers of principal components; the hyperparameter settings of SFA and SSA include a confidence level fixed at 0.99, latent feature dimensions of 2, 3, 4, 5 and 6, respectively, with a total of 5 models each, of which SSA also fixed the time window size to 20; CVA defined 5 models, with a confidence level of 0.99, a cumulative contribution rate fixed at 0.9, and the number of past and future time series expansion steps set to 2, 3, 4, 5 and 6, respectively. MLPAE defines 6 models with 2, 3 and 4 network layers, fixed neuron growth factor of 0.8 for each layer, learning rate of 2e-4, and tested under 50 and 100 training rounds; VAE also defines 6 models, similar to MLPAE, and tests the performance of 2, 3 and 4 layers of networks under the same neuron growth factor, learning rate and training rounds; convAE sets 6 models with convolutional layer depth of 1, 2 and 3, and experiments are conducted under 50 and 100 rounds of training respectively; GRU has 6 models with hidden feature dimensions set to 8, 16 and 32, fixed network layers to 3, and tested under 50 and 100 rounds respectively; Transformer model defines 6 models with feature dimensions of 8, 16 and 32, fixed multi-head attention heads to 8, encoder layers to 3, forward propagation layer dimension to 128, and tested under 50 and 100 rounds of training respectively. A total of 50 model settings with different hyperparameters are obtained.

[0170] In this embodiment, the scoring rule library includes a variety of scoring rules, and the scoring rules are used to monitor the performance evaluation of models of different types and different hyperparameters in the model library, and obtain the performance evaluation results as the data set score matrix.

[0171] Specifically, since the industrial process monitoring task requires the model to accurately alarm on fault data and reduce false alarms on normal data, the model's alarm rate and false alarm rate are combined to design the model's scoring rules. The F1 score of the model is used as the evaluation result. The calculation formula of the F1 score is as follows:

[0172]

[0173] Among them, Precision is the model accuracy, which indicates the proportion of real fault samples among samples predicted to be faults. It can reflect the alarm rate of process monitoring tasks. The higher the precision, the higher the alarm rate. Recall is the model recall rate, which indicates the proportion of samples predicted to be faults among real fault samples. It can reflect the false alarm rate of process monitoring tasks. The higher the recall rate, the lower the false alarm rate.

[0174] In the process of building the model recommendation system, each model in the model library is trained with the training data of each fault case in the case library, and then the F1 score is calculated based on the scoring rules on the test data of the fault case. Each model will get a score for each fault case, thus forming a score matrix, where the number of rows in the matrix is ​​equal to the number of cases and the number of columns is equal to the number of models. This score matrix can be used to predict the target in the training process of the intelligent recommendation method.

[0175] In the application process of the model recommendation system, the case that needs model recommendation is input into the recommendation method of the recommendation system, the F1 score of each model predicted by the recommendation method for the case is obtained, and the model with the highest F1 score is selected as the recommended model.

[0176] In this embodiment, the data set feature matrix acquisition module is used to call the fault case library to obtain the time series data sets of N industrial objects, and use the data set feature engineering method to process all the time series data sets to obtain the data set feature vector corresponding to each time series data set, and all the data set feature vectors are spliced ​​together to obtain the data set feature matrix.

[0177] In this embodiment, the data set score matrix acquisition module is used to use the training set data in the fault case library to train different models in the candidate model library, and use the test set data in the fault case library to perform performance evaluation on different models according to the scoring rules in the scoring rule library, to obtain a data set score vector composed of the performance evaluation results of all models, and to concatenate all the data set score vectors to obtain a data set score matrix.

[0178] In this embodiment, the width learning network training and update module is used to use the data set feature matrix and the data set score matrix to train the width learning network used to establish the mapping relationship between the feature matrix and the score matrix to obtain a trained width learning network; and a triple incremental strategy is used to incrementally update and optimize the trained width learning network to obtain an updated width learning network.

[0179] In this embodiment, the online application module is used to process the target data set using a data set feature engineering method to obtain a data set feature vector corresponding to the target data set, and input it into a trained or updated width learning network to obtain data set score vectors corresponding to all models, and select the model with the highest score as the recommended model.

[0180] The present invention is applicable to model recommendation including but not limited to thermal power generation process, chemical process, and pelletizing process of biomass fuel. The following examples demonstrate the versatility and superiority of the model recommendation capability of the present invention in thermal power generation process, chemical process, and pelletizing process of biomass fuel. Then, the weight of the width network of the model recommendation method is visualized, which helps to understand the specific process of model recommendation; finally, an experiment is designed to demonstrate the effectiveness of the triple increment strategy, including case increment, candidate model increment, and recommended node increment. In this embodiment, the dimension of the mapping feature node in the width network of the present invention is set to 128, the number of nodes is set to 10, the dimension of the enhanced feature node is set to 1280, and the activation functions are all set to ReLU.

[0181] Example 1: Thermal power generation process

[0182] Thermal power generation is the process of using fuel (such as coal or natural gas) to generate heat through combustion, driving steam turbines and generators for energy conversion. It is one of the main sources of global electricity supply. Thermal power generation is an important pillar of energy security and economic development, meeting the energy needs of social production and life. However, the thermal power generation process involves complex dynamic processes such as combustion, heat exchange and generator operation, which are highly sensitive to equipment performance and environmental impact. Therefore, the selection of efficient and accurate monitoring models is of great significance to improving the operating efficiency of thermal power generation systems, ensuring equipment safety and reducing environmental pollution.

[0183] In order to verify the effectiveness of the model recommendation of the present invention in the thermal power generation process, 10 cases were randomly selected from the coal-fired power generation process and the gas-fired power generation process in the case library as the test set, and a total of 5 groups were randomly selected. For each group, the average performance of each method on the target case was first calculated, which will be used as a benchmark for comparison. Table 4 shows the results of the model recommendation of the present invention in the thermal power generation process. First, it can be seen that no model is always leading in the 5 groups of experiments. For example, GRU is better than convAE in the first group, but convAE has a higher F1 score and a more advanced average ranking than GRU in the second group. This reflects the necessity of model recommendation because no model can always lead. Secondly, the performance of the model recommended by the present invention is higher than that of all single methods. This reflects that the present invention can recommend suitable models according to the characteristics of different process data in the group, which enables different and more effective models to be set for different processes in the group, and the average F1 score is higher than the score of using only a single method, and the average ranking is better than the average ranking of all single methods.

[0184] Table 4: Performance test results of the model recommended by the present invention in the process of thermal power generation

[0185]

[0186] Example 2: Chemical Process

[0187] Chemical processes are industrial processes that transform raw materials into chemicals, energy or other useful products through physical and chemical changes. They are an important part of the modern industrial system. Their importance lies in meeting society's diverse needs for materials, energy and chemical products, and promoting economic growth and technological progress. Chemical processes usually involve complex reaction kinetics, heat and mass transfer, and multivariable coupling control, which have a significant impact on production efficiency, product quality and safety. Therefore, the selection of efficient and accurate monitoring models is crucial to improving the stability of chemical processes, reducing energy consumption and environmental load, and ensuring operational safety.

[0188] In order to verify the effectiveness of the model recommendation of the present invention in the chemical process, 10 cases were randomly selected from the Tennessee-Eastman process and the three-phase flow process in the case library as test sets, and a total of 5 groups were randomly selected. For each group, the average performance of each method on the target case was first calculated, which will be used as a benchmark for comparison. Table 5 shows the results of the model recommendation of the present invention on the chemical process. It can be seen that similar to the thermal power generation process, the model recommendation ability of the present invention is also applicable and effective on the objects of the chemical process. The average F1 score within the group is higher than all single methods, and the average ranking is better than all single methods.

[0189] Table 5: Test results of the performance of the model recommended by the present invention in the chemical process

[0190]

[0191] Example 3: Biomass fuel pelletizing process

[0192] The pelletizing process of biomass fuel is the process of processing biomass raw materials into biomass pellet fuel with high density and standard size, which is an important link in the utilization of biomass energy. Its process includes raw material pretreatment, crushing, drying, compression molding, cooling and packaging. In the pelletizing process, the lignin inside the raw material is softened by high-pressure extrusion and heating and acts as a natural binder to stably shape the pellets. Monitoring of the pelletizing process is essential to ensure product quality and production efficiency. Selecting a monitoring model based on key process parameters not only helps to improve the stability of equipment operation, but also effectively reduces energy consumption and promotes the sustainable development of the biomass fuel industry.

[0193] In order to verify the effectiveness of the model recommendation of the present invention in the pelletizing process of biomass fuels, two cases in the case library were used as test sets, each case was tested once as a group, and the training cases were the other remaining cases. For each group, the F1 score and the average ranking of the random selection method from the candidate models were first calculated, which will be used as a benchmark for comparison. Table 6 shows the results of the model recommendation of the present invention in the pelletizing process of biomass fuels. It can be seen that the F1 score of the recommended monitoring model is significantly higher than that of the random selection model, and the ranking is also better than that of the random selection model. This reflects the applicability and effectiveness of the present invention in this type of process.

[0194] Table 6: Test results of the performance of the model recommended by the present invention in the pelletizing process of biomass fuel

[0195]

[0196] In order to increase the intuitive understanding of the intermediate process of model recommendation, the The heat map of the weights is visualized, and the results are as follows Figure 6 As shown, the larger the value, the higher the influence of the intermediate node on the recommendation result of the current model. In this embodiment, the number of candidate models M is 50, the number of nodes R is 10, the dimension D of the mapping feature node is 128, and the dimension E of the enhanced feature node is 1280. Without loss of generality, only the first dimension of 10 groups of nodes is selected here, so the dimension of the visualized matrix is ​​10 rows and 50 columns. Figure 6 It can be seen that intermediate nodes 3 and 4 tend to exclude deep learning algorithms, while intermediate node 2 tends to recommend deep learning algorithms and exclude most statistical analysis methods. This reflects that different intermediate nodes have different preferences for monitoring models, and the activation of different intermediate nodes will affect the recommendation results.

[0197] Example 4: Effectiveness of the triple-increment strategy

[0198] In order to fully evaluate the effectiveness of the triple incremental strategy, a 70%-30% training / test split was used, i.e., 59 cases were set as training cases and 10 cases were set as target cases. These cases were divided into 10 groups, each containing 59 cases for training the model and 10 cases for testing. For each group, the average performance of each method on the target cases was first calculated, which will serve as the baseline for comparison.

[0199] The impact of case increment on the performance of the recommended algorithm under different groups is listed. Three training sets are designed for each group, and the number of cases in the training set increases in order of 10. The proposed method trained with the data with the least number of cases is used as the basic model to demonstrate the impact of case increment on the recommendation ability, and the results are shown in Table 7. From the results in Table 7, it can be seen that as the number of cases increases, the performance of the model recommended by the proposed method in different groups is constantly improving, both in terms of F1 score and ranking. This reflects the positive impact of monitoring case increment on the recommendation ability, which is due to the fact that the wide network learns more mapping relationships between features and scores.

[0200] Table 7: Impact of case increment on the performance of the recommended method under different groups

[0201]

[0202] Table 8 lists the performance impact of the increment of candidate models on the proposed method under different groups. The number of candidate models was increased for each group in turn, resulting in four groups of results. The models that were increased in turn were convAE, GRU, and Transformer. The proposed method with the least number of candidate models was used as the base model to demonstrate the impact of the increase in candidate models on the recommendation ability. Due to the different number of candidate models, only the F1 score was considered without considering the rank. From the results, it can be seen that as the number of candidate models increases, the recommendation effect of the proposed method becomes better and better. This is because as the number of candidate models increases, more models suitable for the target process may appear.

[0203] Table 8: Impact of candidate model increments on the performance of the recommended algorithms under different groups

[0204]

[0205] Figure 7The curve of the impact of the incremental recommendation node on the recommendation performance in the industrial model recommendation method is shown. It can be seen that for most groups, when the number of nodes increases from 1 to 12, the recommendation performance gradually improves, which shows that the increase in nodes helps to learn new cross-features. However, after the number of nodes increases further, the recommendation performance becomes unstable. This trend of first rising and then falling is reasonable, because the increase of a large number of nodes will lead to an increase in model complexity and the number of parameters, thereby increasing the risk of overfitting. In general, when the number of nodes is small, the increase in nodes can help the wide network learn more cross-features, but when the number of nodes is too large, it will also increase the risk of overfitting.

[0206] In summary, the method and system described in the present invention are universal and are not limited to the industrial processes and fault types mentioned in the embodiments. In addition, the concept of the invention is universal and can be extended to model recommendations for other industrial tasks by adjusting the design of task-related features, including but not limited to fault diagnosis, soft measurement, data filling, etc.

[0207] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent recommendation method for industrial process monitoring models, characterized in that: The following steps are involved: (1) Obtain time series data sets of N industrial objects, process all time series data sets using data set feature engineering methods to obtain data set feature vectors corresponding to each time series data set, and concatenate all data set feature vectors together to obtain a data set feature matrix; wherein the data set feature vector corresponding to each time series data set includes the general statistical features, distribution shape features, task-related features, and exogenous knowledge embedding features of the corresponding time series data set; (2) For each time series data set, use the training set data in the time series data set to train M different preset models, and test them on the test set data to obtain a data set score vector composed of the test results of all models corresponding to the time series data set, and concatenate all the data set score vectors to obtain a data set score matrix; (3) using the dataset feature matrix and the dataset score matrix, training a width learning network for establishing a mapping relationship between the feature matrix and the score matrix to obtain a trained width learning network; and using a triple incremental strategy to incrementally update and optimize the trained width learning network to obtain an updated width learning network; wherein the triple incremental strategy includes an update strategy of monitoring case increment, candidate model increment, and recommendation node increment; The update strategy of the monitoring case increment specifically includes: When a new time series data set is collected, first obtain the feature matrix of the newly added data set And the dataset score matrix The feature nodes and enhanced nodes corresponding to the newly added time series dataset are expressed as: Among them, Z R,a represents the concatenated feature nodes corresponding to the newly added time series dataset, N′ represents the number of newly added time series datasets, and H a Indicates the enhanced node corresponding to the newly added time series dataset; Note the newly added input data A a =[Z R,a |H a ], the updated complete input data is: Among them, A x+a represents the updated complete input data, then its pseudo-inverse is expressed as: Q=A a A + Among them, (A x+a ) + A x+a The pseudo-inverse of , B and Q are intermediate calculation variables and have no practical significance; Then the updated linear mapping to be solved is: Where W′ represents the updated linear mapping to be solved; The update strategy of the candidate model increment specifically includes: As the number of candidate models increases, the dataset score matrix changes from Expanded to Where M′ represents the number of newly added models, the linear mapping to be solved is updated as: Among them, S b Represents the dataset score matrix corresponding to the newly added M′ candidate models; (4) When applied online, the target dataset is processed using a dataset feature engineering method to obtain a dataset feature vector corresponding to the target dataset, which is then input into a trained or updated width learning network to obtain a dataset score vector corresponding to all models, and the model with the highest score is selected as the recommended model.

2. The intelligent recommendation method for industrial process monitoring model according to claim 1 is characterized in that: The step (1) specifically comprises: Get a time series dataset of N industrial objects and record each time series dataset as Where m represents the total number of variables in the time series dataset, and Nr represents the total number of samples in the time series dataset. All time series datasets are processed using the dataset feature engineering method to convert each time series dataset into a one-dimensional feature vector. Specifically, the common statistical features, distribution shape features, task-related features, and exogenous knowledge embedding features of each time series dataset are extracted, and the one-dimensional feature vector is used as the dataset feature vector corresponding to the time series dataset. where d i represents the feature vector of the i-th data set, K represents the number of data set features; the data set feature vectors of N time series data sets are concatenated together to obtain the data set feature matrix 3. The intelligent recommendation method for industrial process monitoring model according to claim 2 is characterized in that: The general statistical features include the total number of data set samples Nr, the total number of data set variables m, the data set mean μ, the median median(X), the standard deviation σ, the maximum value max(X), the minimum value min(X), the four sample quantiles q1, q25, q50, q75, the interquartile range q75-q25, the normalized mean μ / max(X), the normalized median median(X) / max(X), the sample range, the sample Gini coefficient, the absolute median deviation median(X-median(X)), the absolute mean deviation m ean(X-median(X)), quantile coefficient dispersion (q75-q25) / (q75+q25), variance coefficient σ / μ, number of outliers outside q25-1.5IQR and q75+1.5IQR, percentage of outliers outside q25-1.5IQR and q75+1.5IQR, number of outliers outside q25-3IQR and q75+3IQR, percentage of outliers outside q25-3IQR and q75+3IQR, percentage of outliers outside 1% to 99% Number, percentage of outliers outside the range of 1% to 99%, number of outliers outside 3σ, percentage of outliers outside 3σ, mean of sparsity of each variable, maximum of sparsity of each variable, minimum of sparsity of each variable, standard deviation of sparsity of each variable, skewness of sparsity of each variable, kurtosis of sparsity of each variable, mean of entropy of each variable, maximum of entropy of each variable, minimum of entropy of each variable, standard deviation of entropy of each variable, skewness of entropy of each variable, kurtosis of entropy of each variable, mean of covariance between variables, maximum of covariance between variables, covariance between variables The minimum value of the inter-variable covariance, the standard deviation of the inter-variable covariance, the skewness of the inter-variable covariance, the kurtosis of the inter-variable covariance, the mean of the inter-variable correlation, the maximum value of the inter-variable correlation, the minimum value of the inter-variable correlation, the standard deviation of the inter-variable correlation, the skewness of the inter-variable correlation, the kurtosis of the inter-variable correlation, the mean of the P value of the variance analysis of each variable, the maximum value of the P value of the variance analysis of each variable, the minimum value of the P value of the variance analysis of each variable, the standard deviation of the P value of the variance analysis of each variable, the skewness of the P value of the variance analysis of each variable, the kurtosis of the P value of the variance analysis of each variable; where IQR represents the interquartile range; The distribution shape characteristics include normality test, mean of skewness of each variable, maximum value of skewness of each variable, minimum value of skewness of each variable, standard deviation of skewness of each variable, skewness of skewness of each variable, kurtosis of skewness of each variable, mean of kurtosis of each variable, maximum value of kurtosis of each variable, minimum value of kurtosis of each variable, standard deviation of kurtosis of each variable, skewness of kurtosis of each variable, kurtosis of kurtosis of each variable, mean of fifth-order moment of each variable, maximum value of fifth-order moment of each variable, minimum value of fifth-order moment of each variable, standard deviation of fifth-order moment of each variable, skewness of fifth-order moment of each variable, kurtosis of fifth-order moment of each variable, mean of sixth-order moment of each variable, maximum value of sixth-order moment of each variable, minimum value of sixth-order moment of each variable, standard deviation of sixth-order moment of each variable, skewness of sixth-order moment of each variable, kurtosis of sixth-order moment of each variable degree, the mean of the seventh-order moment of each variable, the maximum value of the seventh-order moment of each variable, the minimum value of the seventh-order moment of each variable, the standard deviation of the seventh-order moment of each variable, the skewness of the seventh-order moment of each variable, the kurtosis of the seventh-order moment of each variable, the mean of the eighth-order moment of each variable, the maximum value of the eighth-order moment of each variable, the minimum value of the eighth-order moment of each variable, the standard deviation of the eighth-order moment of each variable, the skewness of the eighth-order moment of each variable, the kurtosis of the eighth-order moment of each variable, the mean of the ninth-order moment of each variable, the maximum value of the ninth-order moment of each variable, the minimum value of the ninth-order moment of each variable, the standard deviation of the ninth-order moment of each variable, the skewness of the ninth-order moment of each variable, the kurtosis of the ninth-order moment of each variable, the mean of the tenth-order moment of each variable, the maximum value of the tenth-order moment of each variable, the minimum value of the tenth-order moment of each variable, the standard deviation of the tenth-order moment of each variable, the skewness of the tenth-order moment of each variable, the kurtosis of the tenth-order moment of each variable; The task-related characteristics include the standard deviation of the first principal component, the skewness of the first principal component, the kurtosis of the first principal component, the proportion of explained variance of the first three principal components, the proportion of singular values ​​of the first three typical correlation components, the ADF test of the first three stationary components, the proportion of singular values ​​of the first three stationary components, the filling error of linear regression, the filling error of nonlinear regression, and the linearity based on the proportion of filling error; The exogenous knowledge embedding feature is specifically obtained by the following method: the industrial process and equipment to which the time series data set belongs are input into the large language model, and the answers of the large language model to the five questions are encoded by 0-1 as the exogenous knowledge embedding feature; The five questions are: whether it is a multi-stage industrial process, whether there are fluctuations in raw materials, whether there are changes in set points, whether it is a closed-loop system, and whether complex chemical reactions occur.

4. The method for intelligent recommendation of industrial process monitoring models according to claim 1, characterized in that: The method of using the dataset feature matrix and the dataset score matrix to train a width learning network for establishing a mapping relationship between the feature matrix and the score matrix to obtain a trained width learning network specifically includes: (3.1) The feature matrix of the data set is mapped to feature nodes through the matrix to be mapped, which is expressed as: Z r =φ([D,1]W r ) Among them, Z r represents the rth feature node, φ() represents the activation function, D represents the feature matrix of the data set, represents a vector whose elements are all 1. Represents a randomly initialized orthogonal matrix corresponding to the feature node; (3.2) Concatenate the R groups of feature nodes obtained in step (3.1) to obtain the concatenated feature nodes (3.3) Using the concatenated feature node Z R Compute the enhanced node, expressed as: H=ψ([Z R ,1]In h ) Among them, H represents the enhancement node, ψ() represents the activation function, represents a randomly initialized orthogonal matrix corresponding to the enhanced node, and E represents the output dimension of the enhanced node; (3.4) The feature nodes and enhanced nodes are concatenated and recorded as Using it as the input data of the width learning network, the objective function of the width learning network is: in, is the linear mapping to be solved, Represents the dataset score matrix, s i represents the score vector of the i-th data set, and λ represents the penalty coefficient; (3.5) For the linear mapping W to be solved in the objective function of the width learning network, the ridge regression method is first used to solve the pseudo-inverse of the matrix A, and then the pseudo-inverse of the matrix A and the data set score matrix are used to solve it, which is expressed as: W=A + S Among them, A + represents the pseudo-inverse of matrix A, represents the transpose of matrix A, and I represents the identity matrix.

5. The method for intelligent recommendation of industrial process monitoring models according to claim 1, characterized in that: The update strategy of the recommended node increment specifically includes: Before the feature nodes and enhanced nodes are incremented, their respective numbers are R and E. Assuming that the number of newly added feature nodes is R′ and the number of newly added enhanced nodes is E′, the newly added feature nodes and enhanced nodes are expressed as: Among them, Z R′ represents the newly added feature node after splicing, H′ represents the newly added enhanced node, and D′ represents the number of nodes in the newly added set of feature nodes. represents the linear mapping of the r′th newly added feature node, represents a newly added enhanced node, then the updated input data A - It is expressed as: Where A′=[Z R′ ,H′] represents the newly added input data; based on the above formula, solve the matrix A - The pseudo-inverse of is calculated as: Q=(A + )A′ Among them, (A * ) + Represents the matrix A * The pseudo-inverse of * The pseudo-inverse of and the dataset score matrix obtain the solution of the updated linear mapping in the case of incremental recommendation nodes, expressed as: Where W′ represents the updated linear mapping.

6. A recommendation system for implementing the intelligent recommendation method for industrial process monitoring model according to any one of claims 1 to 5, characterized in that: include: A case library, consisting of various failure cases of industrial equipment involved in the industrial process, for providing failure cases for the recommendation system; A candidate model library, consisting of industrial process monitoring models of different types and with different hyperparameter settings, for providing different preset models for the recommendation system; A scoring rule library, including a plurality of scoring rules, wherein the scoring rules are used to evaluate the monitoring performance of models of different types and different hyperparameters in the model library; The data set feature matrix acquisition module is used to call the fault case library to obtain the time series data sets of N industrial objects, and use the data set feature engineering method to process all the time series data sets to obtain the data set feature vectors corresponding to each time series data set, and splice all the data set feature vectors together to obtain the data set feature matrix; The data set score matrix acquisition module is used to train different models in the candidate model library using the training set data in the fault case library, and use the test set data in the fault case library to perform performance evaluation on different models according to the scoring rules in the scoring rule library, and obtain the data set score vector composed of the performance evaluation results of all models, and splice all the data set score vectors to obtain the data set score matrix; A width learning network training and updating module is used to train a width learning network for establishing a mapping relationship between a feature matrix and a score matrix using a data set feature matrix and a data set score matrix, so as to obtain a trained width learning network; A triple incremental strategy is used to incrementally update and optimize the trained width learning network to obtain an updated width learning network. and The online application module is used to process the target dataset using the dataset feature engineering method to obtain the dataset feature vector corresponding to the target dataset, and input it into the trained or updated width learning network to obtain the dataset score vectors corresponding to all models, and select the model with the highest score as the recommended model.

7. The recommendation system according to claim 6, characterized in that The fault cases cover a variety of fault types, including step faults, intermittent faults, initial faults, slow-changing faults, drift faults, and stuck faults; each fault case contains corresponding training data, test data, and fault occurrence time, wherein the training data is used for model training of the model library and the construction of data set feature vectors, and the test data is used to evaluate the performance of each model in the model library according to the fault occurrence time and the scoring rule library; The candidate model library includes principal component analysis model, slow feature analysis model, stationary subspace analysis model, canonical correlation analysis model, autoencoder model, variational autoencoder model, convolutional autoencoder model, recurrent neural network model, and Transformer model; The specific scoring rule is: use the F1 score of the model as the evaluation result, and its calculation formula is: Among them, Precision is the model accuracy, which indicates the proportion of real fault samples among the samples predicted to be faults; Recall is the model recall rate, which indicates the proportion of samples predicted to be faults among the real fault samples.

Images