A chemical soft measurement modeling method based on fuzzy heterogeneous random configuration network

Abstract

The application discloses a chemical soft measurement modeling method based on a fuzzy heterogeneous random configuration network, and belongs to the field of intelligent analysis of industrial data and modeling of complex chemical processes. The application is used for online soft measurement and intelligent prediction of key quality variables of a complex chemical process. The method constructs a heterogeneous base function space to describe multimodal characteristics in chemical process data; morphological variation of a network residual sequence is extracted and is fuzzed into rough and smooth membership degrees; zero-order Takagi-Sugeno fuzzy reasoning is used to dynamically calculate generation probability quotas of each base function; node competition and optimization are performed in combination with morphological preference factors and supervised constraint scores to realize adaptive configuration of hidden layer nodes; finally, Moore-Penrose pseudo-inverse method is used to solve output weights to obtain a final soft measurement model for online prediction. The application realizes differentiated configuration of nodes through residual morphological perception and fuzzy routing, and improves prediction accuracy and structural compactness of the model.

Description

Technical Field

[0001] This invention belongs to the field of intelligent analysis of industrial data and modeling of complex chemical processes, and specifically relates to a chemical soft measurement modeling method based on fuzzy heterogeneous random configuration network. Background Technology

[0002] As modern chemical systems become increasingly complex and large-scale, process data streams exhibit highly complex inherent multimodal characteristics. In soft-sensor modeling of chemical processes, the relationship between mass variables and process variables is no longer limited to simple nonlinear mapping, but has evolved into a deep coupling of inherent multimodal characteristic information. Specifically, the inherent information in the data includes both smooth, slowly varying modes dominated by long-term thermodynamic evolution and rough oscillatory modes induced by fluid disturbances or step loads; the former reflects low-frequency information, while the latter describes high-frequency information. This cross-spectral, heterogeneous dynamic evolutionary behavior poses a severe challenge to the feature decoupling and trajectory approximation capabilities of data-driven soft-sensor models.

[0003] In recent years, Stochastic Configuration Networks (SCNs) have demonstrated significant advantages in industrial soft sensor modeling due to their general approximation capabilities in function space and rapid incremental modeling performance, becoming one of the mainstream solutions. However, when dealing with the complex chemical process data mentioned above, existing SCN frameworks have revealed two limitations.

[0004] First, there is the frequency domain mismatch caused by the isomorphic space assumption. Classical SCN typically uses only a single sigmoid activation function to construct the feature space. This function has the ability to continuously and smoothly map nonlinear patterns, and can accurately fit monotonically and slowly varying steady-state time series trends. However, fluid disturbances and step loads in chemical processes can excite high-frequency transient fluctuations, and a single sigmoid function cannot simultaneously capture low-frequency smooth modes and high-frequency coarse modes.

[0005] Secondly, the differences in modal information are not fully considered in the configuration of hidden layer nodes. Traditional SCN networks, when incrementally configuring hidden layer nodes, ignore the different types of modal information within the current model residuals, relying entirely on a large candidate pool for random selection with uniform probability. This hidden layer node generation paradigm, lacking guidance from modal information, not only increases the complexity of the model but also makes it difficult to effectively capture data trend information, easily leading to redundant node allocation and overfitting risks.

[0006] Therefore, how to fully consider the multi-modal information inherent in the data and use the modal information to guide the differentiated configuration of hidden layer nodes is a very worthwhile research problem in the field of intelligent modeling of soft measurement of chemical processes based on random configuration networks. Summary of the Invention

[0007] To address the aforementioned technical problems in existing technologies, this invention proposes a chemical engineering soft measurement modeling method based on fuzzy heterogeneous random configuration networks. This method solves the technical problems of traditional random configuration networks being unable to simultaneously fit low-frequency smooth modes and high-frequency coarse modes in chemical process data due to frequency domain mismatch of isomorphic activation functions, as well as the blind configuration of nodes, redundant model structure, and overfitting caused by the lack of residual morphology perception mechanisms.

[0008] To achieve the above objectives, the present invention adopts the following technical solution;

[0009] A chemical engineering soft measurement modeling method based on fuzzy heterogeneous random configuration networks includes the following steps;

[0010] Step S1: Acquisition of normal operating data and network initialization of chemical process;

[0011] Collect and preprocess normal operating data of chemical process, construct input feature matrix and output quality variable matrix, and initialize fuzzy heterogeneous random configuration network model parameters.

[0012] Step S2: Construct a heterogeneous basis function space for multimodal feature description;

[0013] A heterogeneous basis function space composed of Sigmoid, Legendre and Bessel functions is constructed to describe the inherent multimodal characteristics of chemical process data.

[0014] Step S3: Configuration of fuzzy nodes driven by residual morphology;

[0015] Adaptive configuration of hidden layer nodes in the network is achieved through an iterative approach;

[0016] Step S4: Online application of the soft measurement model;

[0017] Collect and preprocess online operational variable data, input it into a trained fuzzy heterogeneous randomized network model, and output the predicted value of the target quality variable.

[0018] Preferably, the heterogeneous basis function space in step S2 is composed of three complementary activation functions: the Sigmoid function, the Legendre function, and the Bessel function; wherein the Sigmoid function provides basic nonlinear mapping capability, the Legendre function characterizes long-period steady-state drift characteristics, and the Bessel function describes high-frequency transient fluctuations in chemical processes.

[0019] Preferably, step S3 specifically includes the following steps;

[0020] S3.1 Residual morphology observation and fuzzification;

[0021] Calculate the morphological variational operator of the current residual sequence, and map the morphological variational operator to coarse morphological membership and smooth morphological membership through a linear membership function;

[0022] S3.2 Modal routing calculation based on zero-order TS fuzzy inference;

[0023] Using the coarse morphological membership degree and smooth morphological membership degree as input, the generation probability quota of various heterogeneous basis functions is calculated by the zero-order Takagi-Sugeno fuzzy inference method;

[0024] S3.3, Morphology-driven candidate node competition and selection mechanism;

[0025] Guided by the generation probability quota, a set of candidate nodes is generated in the heterogeneous basis function space, the supervision constraint score of each candidate node is calculated, and the comprehensive competition score is calculated in combination with the morphological preference factor. The optimal node is selected as the new hidden layer node.

[0026] S3.4 Network state update and model convergence judgment;

[0027] The Moore-Penrose pseudo-inverse method is used to solve for the globally optimal output weights, update the network output and residual sequence, and determine whether the iteration termination condition is met; if not, return to step S3.1 to continue the iteration.

[0028] Preferably, the morphological variational operator in step S3.1 is obtained by calculating the cumulative sum of the absolute differences between adjacent samples of the residual sequence and normalizing it by the amplitude standard deviation, and is used to quantitatively describe the roughness or smoothness of the residual sequence.

[0029] Preferably, the fuzzy rule base of the zero-order Takagi-Sugeno fuzzy inference method described in step S3.2 includes;

[0030] Rule 1: If the residual shape is coarse, then the Bessel basis function generation probability is the maximum probability value, and the Legendre basis function generation probability is the minimum probability value.

[0031] Rule 2: If the residual shape is smooth, then the Bessel basis function generation probability is the minimum probability value, and the Legendre basis function generation probability is the maximum probability value.

[0032] Using the membership degrees of the coarse and smooth morphologies as weights, the generation probabilities of the Bessel and Legendre basis functions are calculated using a weighted average method. The generation probability of the Sigmoid basis function is determined by subtracting the sum of the generation probabilities of the Bessel and Legendre basis functions from 1.

[0033] Preferably, the supervision constraint score in step S3.3 is used to quantify the correlation between the feature mapping of the candidate node and the current residual, and to screen the effective nodes that satisfy the inequality constraints; the morphological preference factor is used to perform differentiated weighting based on the basis function type to which the candidate node belongs and the morphological membership degree of the current residual, and is used to strengthen and calibrate the basis function selection process; the comprehensive competition score is obtained by multiplying the morphological preference factor and the supervision constraint score, and the candidate node corresponding to the maximum comprehensive competition score is used as the newly added hidden layer node.

[0034] Preferably, the iteration termination condition in step S3.4 is: the current residual norm reaches a preset tolerance, or the number of hidden layer nodes reaches a preset maximum value; the global optimal output weight is obtained by solving the least squares optimization objective using the Moore-Penrose pseudo-inverse method.

[0035] Preferably, the online application of the soft measurement model in step S4 includes: collecting online operational variable data after normalization preprocessing, inputting the trained fuzzy heterogeneous random configuration network model, and outputting the predicted value of the target quality variable after forward computation, which is used to guide the operation and control of the chemical production process.

[0036] Furthermore, this invention also mentions a chemical engineering soft measurement modeling system based on a fuzzy heterogeneous random configuration network, comprising:

[0037] The data acquisition and preprocessing module is used to collect normal operating data of chemical processes and perform normalization processing to construct the input feature matrix and the output mass variable matrix.

[0038] The heterogeneous basis function space construction module is used to construct a heterogeneous basis function space composed of the Sigmoid function, Legendre function, and Bessel function;

[0039] The residual morphology analysis module is used to calculate the morphological variational operator of the current residual sequence and map it to coarse morphological membership degree and smooth morphological membership degree through a linear membership function.

[0040] The fuzzy inference routing module is used to calculate the generation probability quotas of various heterogeneous basis functions based on fuzzy membership degrees using the zero-order Takagi-Sugeno fuzzy inference method.

[0041] The node competition and selection module is used to generate a candidate node pool in the heterogeneous basis function space under the guidance of the generation probability quota, calculate the supervision constraint score, and combine the morphological preference factor to calculate the comprehensive competition score, and select the optimal node as the new hidden layer node.

[0042] The network update and convergence judgment module is used to solve the globally optimal output weights using the Moore-Penrose pseudo-inverse method, update the network output and residual sequence, and determine whether the iteration termination condition is met.

[0043] The online prediction module is used to collect and preprocess online operational variable data, input the trained fuzzy heterogeneous randomized network model, and output the predicted value of the target quality variable.

[0044] The beneficial technical effects brought about by this invention;

[0045] 1. Frequency domain complementarity of heterogeneous basis function space: This invention constructs a heterogeneous basis function space composed of Sigmoid (global trend), Legendre (low-frequency steady state), and Bessel (high-frequency transient), breaking the frequency domain limitation of the traditional SCN single activation function, and realizing the accurate description and decoupling of smooth and slowly varying modes and coarse oscillating modes in chemical process data.

[0046] 2. Fuzzy driving mechanism for residual morphology perception: This invention introduces fuzzy logic into the node configuration process of a randomly configured network for the first time. It uses a morphological variational operator to perceive the geometric morphological features of the residual in real time and fuzzifies them into coarse membership degree and smooth membership degree, thus establishing an intelligent mapping relationship between residual morphology and basis function selection.

[0047] 3. Closed-loop competitive selection paradigm for node configuration; This invention establishes a progressive closed-loop growth logic of "residual observation → fuzzy routing → node competition". The zero-order TS fuzzy inference system dynamically adjusts the generation probability quota of each basis function according to the residual morphology. The morphology preference factor further enhances the pertinence of basis function selection, ensuring that the newly added nodes highly match the true morphology of the current signal.

[0048] 4. Significantly improved model compactness and robustness: This invention introduces a modal information guidance mechanism in node configuration, avoiding redundant allocation and overfitting risks caused by blind node search while significantly compressing the candidate pool size. Experimental results show that the RMSE standard deviation of the method in this invention is reduced by nearly an order of magnitude compared to the traditional SCN, exhibiting higher structural stability and robustness for industrial deployment.

[0049] 5. Prediction accuracy is significantly improved; validation results on an industrial butane removal tower dataset show that the determination coefficient of the method of this invention is significantly improved. The mean square error (RMSE) reached 0.8014, and the root mean square error (RMSE) decreased to 0.0766, which is significantly better than the classic SCN, MASCN and FSCN comparison methods. Attached Figure Description

[0050] Figure 1This is a flowchart of the soft measurement method based on Fuzzy Heterogeneous Random Configuration Network (FHSCN) of the present invention;

[0051] Figure 2 This is a graph showing the prediction fit of the classic SCN (Stochastic Configuration Network) model on a test set of a butane removal tower. The horizontal axis represents the test set sample number, and the vertical axis represents the normalized predicted bottom butane concentration. The graph contains two curves: the solid line represents the actual bottom butane concentration, and the dashed line represents the predicted bottom butane concentration by the classic SCN model.

[0052] Figure 3 This is a graph showing the prediction fit of the MASCN (Multi-Activation Stochastic Configuration Network) model on a test set of a butanizer tower. The horizontal axis represents the test set sample number, and the vertical axis represents the normalized predicted bottom butane concentration. The graph contains two curves: the solid line represents the actual bottom butane concentration, and the dashed line represents the MASCN model's prediction of the bottom butane concentration.

[0053] Figure 4 This is a graph showing the prediction fit of the FSCN (Fuzzy Stochastic Configuration Network) model on a test set of a butane removal tower. The horizontal axis represents the test set sample number, and the vertical axis represents the normalized predicted bottom butane concentration. The graph contains two curves: the solid line represents the actual bottom butane concentration, and the dashed line represents the FSCN model's prediction of the bottom butane concentration.

[0054] Figure 5 This is a graph showing the prediction fitting curve of the FHSCN (Fuzzy Heterogeneous Stochastic Configuration Network) model on the test set of a butane removal tower, which represents the method proposed in this invention. The horizontal axis of the graph represents the sample number of the test set, and the vertical axis represents the normalized predicted value of the bottom butane concentration. The graph contains two curves: the solid line represents the actual value curve of the bottom butane concentration, and the dashed line represents the prediction value curve of the bottom butane concentration by the FHSCN model. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0056] Example 1: The complete process of the FHSCN soft measurement modeling method.

[0057] This embodiment provides a chemical engineering soft measurement modeling method based on fuzzy heterogeneous random configuration networks, the process of which is as follows: Figure 1 As shown, the specific steps are as follows.

[0058] Step S1: Data acquisition and network initialization for normal operating conditions of chemical processes.

[0059] A dataset containing N samples and D variables from a chemical process is collected and normalized to obtain the input feature matrix. and output mass variable matrix .in, This represents the total number of training samples. This represents the input feature dimension (i.e., the number of auxiliary variables) for each sample. express OK The space of real matrices in columns, express A real vector space with 1 row and 1 column.

[0060] Initialize the number of hidden layer nodes in the FHSCN model Set the maximum number of hidden layer nodes. Candidate pool size Target error tolerance Maximum probability value Minimum probability value Initial residual sequence .

[0061] Step S2: Construct a heterogeneous basis function space for multimodal feature description.

[0062] To overcome the limitation of a single activation function in traditional randomly allocated networks, a heterogeneous basis function space consisting of three complementary activation functions is constructed. ;

[0063] ;

[0064] in, This represents the direct sum operation, which is the direct combination of the three types of basis function spaces; , , Let represent the Sigmoid function space, the Legendre polynomial function space, and the Bessel function space, respectively.

[0065] The Sigmoid function provides basic nonlinear mapping capabilities and serves as a fundamental function for describing the nonlinear characteristics of industrial processes.

[0066] ;

[0067] in, This indicates that the Sigmoid function operates on the input... The output value at that location has a range of 1. ; Represents the input feature vector The result of the affine transformation; is the base of the natural logarithm.

[0068] ;

[0069] in, The weight vector is randomly configured, and its dimensions are the same as the input features. same; A randomly configured bias scalar; express and The inner product (dot product); Represents the input feature matrix A single line of data in the sample 3D feature vectors.

[0070] Using the definition in The Legendre polynomial function on the π represents the long-period steady-state drift characteristics and is generated by the Bonnet recursive formula;

[0071] ;

[0072] in, The order index of the Legendre polynomial is usually taken as 1. That is, using zero-order, first-order, and second-order Legendre functions as... basis functions; express Legendre polynomial of order in The function value at that location; , This serves as the initial condition for the recursion; The result of the affine transformation of the input variables needs to be normalized to... The interval is used to ensure the orthogonality of Legendre polynomials.

[0073] The Bessel function, which has damped oscillation characteristics, is used to describe high-frequency transient fluctuations in chemical processes.

[0074] ;

[0075] in, Let be the order of the Bessel function, usually taken as . That is, using zero-order and first-order Bessel functions as basis functions; To sum the index variable, the summation is carried over from 0 to infinity (in actual calculations, it is truncated to sufficient precision). express Bessel function of order in The function value at that location; This represents the factorial operation; Providing alternating positive and negative signs makes the function exhibit oscillatory characteristics; The term is a power term, which determines the decay rate of the function.

[0076] Step S3: Configuration of fuzzy nodes driven by residual morphology;

[0077] This step achieves adaptive configuration of hidden layer nodes in the network through an iterative approach, specifically including the following sub-steps;

[0078] S3.1 Residual morphology observation and fuzzification.

[0079] make Extracting residual sequences Geometric features, computational morphological variational operators ;

[0080] ;

[0081] in, This is a morphological variational operator used to quantitatively describe the volatility of the residual sequence. A larger value indicates a coarser residual morphology (high-frequency oscillation), while a smaller value indicates a smoother morphology (low-frequency gradual variation). For inclusion The residual sequence at each hidden layer node, i.e., the difference vector between the actual output and the model's predicted output, has a length of... ; For residual sequence The global standard deviation is used to normalize the residual fluctuation amplitude and eliminate the influence of the difference in residual magnitude at different training stages; The total number of training samples; To prevent extremely small positive numbers with a denominator of zero; The representation model contains The residual sequence of the nth node The residual values ​​of each sample; This represents the absolute value of the difference between adjacent sample residuals from arrive The summation reflects the overall fluctuation range of the residual sequence.

[0082] The morphological variational operator is mapped to a rough morphological membership degree using a linear membership function with anti-saturation constraints. Membership degree of smooth shape ;

[0083] ; ;

[0084] in, The membership degree represents the rough shape of the residual, and its value range is [value range missing]. The closer the value is to 1, the more the residual tends to have a coarser (high-frequency) shape; The membership degree represents the smooth shape of the residuals, and its value ranges from [value range missing]. The closer the value is to 1, the more the residual tends to have a smooth (low-frequency) shape; This is the lower limit of the shape reference (0 for perfect smoothness). This is the upper limit of saturation (set as the noise limit value; exceeding this value is considered completely coarse). As an anti-saturation cutoff function, it ensures that the membership values ​​are constrained to... Within the interval. This step maps the original residual to a range within the interval. The fuzzy membership degree, and satisfying .

[0085] S3.2 Modal routing calculation based on zero-order TS fuzzy inference.

[0086] Fuzzy membership , The driving force is the zero-order Takagi-Sugeno fuzzy inference method, which is used to calculate the generation probability of different basis functions. , The upper and lower bounds of the probability are preset by humans, and the values ​​are fixed in the implementation examples. , The fuzzy rule base is as follows;

[0087] Rule 1: If the residual shape is coarse (dominantly transient fluctuations), then the Bessel basis function generation probability... Legendre basis function generation probability .

[0088] Rule 2: If the residual shape is smooth (dominant steady-state drift), then the Bessel basis function generation probability... Legendre basis function generation probability .

[0089] Among them, superscript This indicates that the current number is the [number]. Secondary node configuration iteration, index , These represent the probability values ​​corresponding to Rule 1 and Rule 2, respectively.

[0090] The membership degree is used as the weight, and the generation probability of each basis function is calculated by weighted average method;

[0091] ; ; ;

[0092] in, , , The first In the next iteration, the generation probability quotas of the Bessel, Legendre, and Sigmoid basis functions sum to 1. and These are the preset upper and lower bounds for the probability.

[0093] S3.3, Morphology-driven candidate node competition and selection mechanism.

[0094] In the probability matrix Under guidance, randomly generate functions containing in the heterogeneous basis function space. A set of candidate nodes ;in, Indicates the first The candidate node set generated in the next iteration has a size of . ; superscript This represents the transpose of a vector.

[0095] Calculate the supervised constraint score ;

[0096] ;

[0097] in, For the first The supervision constraint score of each candidate node is used to quantify the correlation between the candidate node and the current residual. The higher the score, the greater the potential contribution of the node to reducing the residual. heterogeneous substrate space The first one is randomly generated in the middle. 1 candidate basis function (i.e., candidate node); Represents the residual vector With candidate basis functions The inner product of the two measures their correlation. Candidate basis functions The square of the L2 norm is used for normalization; These are convergence control parameters used to control the correlation requirements for new nodes; To adaptively adjust parameters, depending on the number of nodes Increase and gradually decrease, satisfying ; Let L2 norm be the square of the residual vector, representing the total energy of the current modeling error.

[0098] Further design of morphology preference factors , used for enhanced calibration in the basis function selection process;

[0099] ;

[0100] in, For the first The morphological preference factor for each candidate node is used to adjust the basis functions according to the current residual morphology.

[0101] The intensity of preference; The hyperparameter used to control the strength of preference is the one with the largest value, which indicates a stronger preference effect. The smooth membership degree of the current residual is used to enhance the preference weights for the Legendre basis functions; The coarse membership degree of the current residual is used to enhance the preference weights for the Bessel basis function; the preference factor of the Sigmoid basis function is fixed at 1 as the benchmark.

[0102] Calculate the overall competitive score Select the node with the highest score from the candidate pool. The corresponding candidate nodes are added as hidden layer nodes. For the first The overall competitive score of each candidate node, which combines the supervision constraint score and the morphological preference factor, is used to determine the winner.

[0103] S3.4 Network state update and model convergence judgment.

[0104] After the optimal construct node is generated, the Moore-Penrose pseudoinverse method is used to solve for the globally optimal output weights, and the network output and the current residual sequence are updated. If the current residual norm The expected tolerance has been met. or number of nodes Reaching the upper limit If the offline model training ends, the offline model training is complete; otherwise, return to step S3.1 to continue iterative computation.

[0105] Specifically, the node with the highest score in the candidate pool is selected as follows: As the first of the randomly configured networks There are n hidden layer nodes, and their corresponding basis functions are denoted as . At this time, possessing The network model with randomly configured hidden nodes is represented as follows:

[0106] ;

[0107] in, For inclusion A network of hidden nodes receives input. Predicted output at; The globally optimal output weight vector The corresponding number in the middle The output weights of each node are solved through global optimization. For the first The basis functions of each hidden layer node. For the first The result of the input affine transformation of each node; and The first The weight vector and bias scalar are randomly configured for each node.

[0108] Based on the known training data and Establish an optimization objective function;

[0109] ;

[0110] in, This is the globally optimal output weight vector; For model prediction output Compared with actual output The squared L2 norm of the error between them; This indicates the search for the function that minimizes the objective function. value.

[0111] The above optimization problem is solved using the Moore-Penrose pseudoinverse method to obtain the updated globally optimal output weight vector. .

[0112] Based on the optimization results, the current residual sequence is further calculated;

[0113] ;

[0114] in, For inclusion The residual vector at each hidden layer node has a length of... , where represents the prediction error for each training sample.

[0115] If the current residual norm Target tolerance reached or number of nodes Reaching the upper limit If the offline model training ends, the offline model training is complete; otherwise, return to step S3.1 to continue iterative computation.

[0116] After the iteration meets the termination condition, the final offline model is obtained. ;

[0117] ;

[0118] in, The final number of hidden layer nodes determined for the model (when the termination condition is met). value); For the final trained FHSCN soft sensor model, input The predicted output function at that point.

[0119] Step S4: Online application of the soft measurement model.

[0120] When the online operational variable data has undergone normalization preprocessing Upon arrival, the predicted values ​​of the target quality variables are obtained by inputting them into the FHSCN model. ;

[0121] ;

[0122] in, The input feature vector is collected online and preprocessed using normalization. The predicted values ​​of the target quality variables output by the model; This refers to the final FHSCN model function trained in step S3.4.

[0123] Example 2: Application Case Verification.

[0124] To further verify the effectiveness and superiority of the proposed soft measurement method based on residual morphological fuzzy-driven heterogeneous random configuration network (FHSCN) in actual complex chemical processes, this embodiment uses a typical industrial debutanizer column process (DCP) as the test object for application verification.

[0125] Butane debutanizers are key separation equipment in refineries and petrochemical plants. Their internal physicochemical reactions are complex, and the underlying data streams naturally have complex characteristics such as high nonlinearity, time-varying delay, and multimodal transient offset.

[0126] 1. Data processing and experimental setup;

[0127] First, the collected real industrial operation data of DCP were preprocessed. Considering the dynamic time-series dependence of industrial processes, a 20th-order lag (Lag) extension technique was used to extract time-series features from the data. The input variables were historical window data of various process variables (such as temperature, pressure, flow rate, etc.), and the output was key quality variables (such as bottom butane concentration).

[0128] To simulate the online prediction needs of real industrial sites, the dataset was strictly divided according to time order: the first 1800 data samples were used as the training set for offline construction of the FHSCN soft measurement model; the remaining 400 data samples were used as the test set for online evaluation of the model's generalization prediction performance.

[0129] 2. Compare the experiments and results;

[0130] To comprehensively evaluate the performance of the method of this invention, it was compared and analyzed with three existing mainstream SCN soft sensor models: classic SCN, MASCN (Multi-Activity Randomized Network), and FSCN (Fuzzy Randomized Network). To ensure the statistical reliability of the experimental results, multiple independent repeated tests were conducted on network models containing random weights, and the mean and standard deviation were recorded.

[0131] The statistical results of the prediction error indices (root mean square error RMSE, mean absolute error MAE, and coefficient of determination R²) of each model on the DCP test set are shown in Table 1.

[0132] Table 1: Prediction Performance of Four Models;

[0133] .

[0134] 3. Results analysis and proof of technical advantages.

[0135] As shown in Table 1, the FHSCN model proposed in this invention achieves optimal performance across all quantitative indicators. Its coefficient of determination (R²) reaches a maximum of 0.8014, while the root mean square error (RMSE) is reduced to a minimum of 0.0766. Notably, compared to the traditional open-loop classical SCN model (whose RMSE standard deviation is 0.0081), the standard deviation of the method in this invention is rigorously compressed to 0.0012, a reduction of nearly an order of magnitude. This rigorously demonstrates from a statistical perspective that the "residual morphological fuzzy-driven" evolution mechanism introduced in this invention effectively eliminates the structural uncertainty caused by the blind search of traditional random networks, endowing the model with higher structural stability and robustness for industrial deployment.

[0136] Further combining the prediction fitting curves of each model Figures 2-5It is known that the DCP test set sequence exhibits drastic high-frequency transient changes in specific sample intervals. Compared with MASCN, FSCN, and the classic SCN model, all three reveal varying degrees of phase lag, amplitude attenuation, and high-frequency jitter when facing these peaks and troughs, making it difficult to accurately match complex frequency domain features. However, the proposed method (FHSCN) benefits from the complementary advantages of heterogeneous basis functions (Sigmoid, Legendre, Bessel) in the frequency domain and the real-time adaptive modal routing of the fuzzy inference engine, demonstrating extremely high tracking fidelity in the most volatile transient intervals, with the predicted curve almost perfectly coinciding with the true value curve.

[0137] The above experimental results fully demonstrate that the method of the present invention is not only significantly superior to the comparative model in terms of overall statistical accuracy and robustness, but also can accurately decouple and track the multimodal dynamic changes of complex chemical processes. It effectively solves the technical problems of frequency domain mismatch and easy overfitting in traditional industrial soft measurement models, and provides a highly reliable solution for online monitoring of key variables in actual industrial processes.

[0138] In summary, this invention establishes a progressive closed-loop growth logic of "residual observation - fuzzy routing - node competition". Its innovation lies not only in constructing a heterogeneous basis function library (Sigmoid, Legendre, and Bessel) to accurately describe the intrinsic multimodal information of the data, but also in using a zero-order TS fuzzy system to guide the competitive generation of basis functions, ensuring that the newly added nodes closely match the true form of the current signal. Under the premise of significantly compressing the size of the computing pool, it avoids the risks of redundant allocation and overfitting caused by blind node search, and achieves extremely high model compactness and strong industrial robustness.

[0139] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A chemical engineering soft measurement modeling method based on fuzzy heterogeneous random configuration networks, characterized in that, Includes the following steps; Step S1: Acquisition of normal operating data and network initialization of chemical process; Collect and preprocess normal operating data of chemical process, construct input feature matrix and output quality variable matrix, and initialize fuzzy heterogeneous random configuration network model parameters. Step S2: Construct a heterogeneous basis function space for multimodal feature description; A heterogeneous basis function space composed of Sigmoid, Legendre and Bessel functions is constructed to describe the inherent multimodal characteristics of chemical process data. Step S3: Configuration of fuzzy nodes driven by residual morphology; Adaptive configuration of hidden layer nodes in the network is achieved through an iterative approach; Step S4: Online application of the soft measurement model; Collect and preprocess online operational variable data, input it into a trained fuzzy heterogeneous randomized network model, and output the predicted value of the target quality variable.

2. The chemical engineering soft measurement modeling method based on fuzzy heterogeneous random configuration network according to claim 1, characterized in that, The heterogeneous basis function space mentioned in step S2 consists of three complementary activation functions: the Sigmoid function, the Legendre function, and the Bessel function. The Sigmoid function provides basic nonlinear mapping capability, the Legendre function characterizes long-period steady-state drift characteristics, and the Bessel function describes high-frequency transient fluctuations in chemical processes.

3. The chemical engineering soft measurement modeling method based on fuzzy heterogeneous random configuration network according to claim 1, characterized in that, Step S3 specifically includes the following steps: S3.1 Residual morphology observation and fuzzification; Calculate the morphological variational operator of the current residual sequence, and map the morphological variational operator to coarse morphological membership and smooth morphological membership through a linear membership function; S3.2 Modal routing calculation based on zero-order TS fuzzy inference; Using the coarse morphological membership degree and smooth morphological membership degree as input, the generation probability quota of various heterogeneous basis functions is calculated by the zero-order Takagi-Sugeno fuzzy inference method; S3.3, Morphology-driven candidate node competition and selection mechanism; Guided by the generation probability quota, a set of candidate nodes is generated in the heterogeneous basis function space, the supervision constraint score of each candidate node is calculated, and the comprehensive competition score is calculated in combination with the morphological preference factor. The optimal node is selected as the new hidden layer node. S3.4 Network state update and model convergence judgment; The Moore-Penrose pseudo-inverse method is used to solve for the globally optimal output weights, update the network output and residual sequence, and determine whether the iteration termination condition is met; if not, return to step S3.1 to continue the iteration.

4. The chemical engineering soft measurement modeling method based on fuzzy heterogeneous random configuration network according to claim 3, characterized in that, The morphological variational operator described in step S3.1 is obtained by calculating the cumulative sum of the absolute differences between adjacent samples of the residual sequence and normalizing it by the amplitude standard deviation. It is used to quantitatively describe the roughness or smoothness of the residual sequence.

5. The chemical engineering soft measurement modeling method based on fuzzy heterogeneous random configuration network according to claim 3, characterized in that, The fuzzy rule base of the zero-order Takagi-Sugeno fuzzy inference method described in step S3.2 includes: Rule 1: If the residual shape is coarse, then the Bessel basis function generation probability is the maximum probability value, and the Legendre basis function generation probability is the minimum probability value. Rule 2: If the residual shape is smooth, then the Bessel basis function generation probability is the minimum probability value, and the Legendre basis function generation probability is the maximum probability value. Using the membership degrees of the coarse and smooth morphologies as weights, the generation probabilities of the Bessel and Legendre basis functions are calculated using a weighted average method. The generation probability of the Sigmoid basis function is determined by subtracting the sum of the generation probabilities of the Bessel and Legendre basis functions from 1.

6. The chemical engineering soft measurement modeling method based on fuzzy heterogeneous random configuration network according to claim 3, characterized in that, In step S3.3, the supervised constraint score is used to quantify the correlation between the feature mapping of the candidate node and the current residual, and to screen the effective nodes that satisfy the inequality constraints; the morphological preference factor is used to perform differentiated weighting based on the basis function type to which the candidate node belongs and the morphological membership degree of the current residual, and is used to strengthen and calibrate the basis function selection process; the comprehensive competition score is obtained by multiplying the morphological preference factor and the supervised constraint score, and the candidate node corresponding to the maximum comprehensive competition score is used as the new hidden layer node.

7. The chemical engineering soft measurement modeling method based on fuzzy heterogeneous random configuration network according to claim 3, characterized in that, The iteration termination condition in step S3.4 is: the current residual norm reaches the preset tolerance, or the number of hidden layer nodes reaches the preset maximum value; the global optimal output weight is obtained by solving the least squares optimization objective using the Moore-Penrose pseudo-inverse method.

8. The chemical engineering soft measurement modeling method based on fuzzy heterogeneous random configuration network according to claim 1, characterized in that, The online application of the soft measurement model in step S4 includes: collecting online operational variable data after normalization preprocessing, inputting it into a trained fuzzy heterogeneous randomized network model, and outputting the predicted value of the target quality variable after forward computation, which is used to guide the operation and control of chemical production processes.

9. A chemical engineering soft measurement modeling system based on fuzzy heterogeneous random configuration networks, characterized in that, include: The data acquisition and preprocessing module is used to collect normal operating data of chemical processes and perform normalization processing to construct the input feature matrix and the output mass variable matrix. The heterogeneous basis function space construction module is used to construct a heterogeneous basis function space composed of the Sigmoid function, Legendre function, and Bessel function; The residual morphology analysis module is used to calculate the morphological variational operator of the current residual sequence and map it to coarse morphological membership degree and smooth morphological membership degree through a linear membership function. The fuzzy inference routing module is used to calculate the generation probability quotas of various heterogeneous basis functions based on fuzzy membership degrees using the zero-order Takagi-Sugeno fuzzy inference method. The node competition and selection module is used to generate a candidate node pool in the heterogeneous basis function space under the guidance of the generation probability quota, calculate the supervision constraint score, and combine the morphological preference factor to calculate the comprehensive competition score, and select the optimal node as the new hidden layer node. The network update and convergence judgment module is used to solve the globally optimal output weights using the Moore-Penrose pseudo-inverse method, update the network output and residual sequence, and determine whether the iteration termination condition is met. The online prediction module is used to collect and preprocess online operational variable data, input the trained fuzzy heterogeneous randomized network model, and output the predicted value of the target quality variable.

Images