Method and system for controlling production of detergent compositions based on multi-parameter intelligent analysis

Abstract

The present invention provides a production control method and system for a detergent composition based on multi-parameter intelligent analysis, which relates to the field of production control technology, including: collecting process parameters to form a data stream, inputting the data into a cascaded autoencoder deep learning network to construct a dynamic correlation map, grouping parameters based on the map and calculating coordinated change characteristics, constructing a hierarchical progressive intelligent controller to generate an optimization control strategy, realizing adaptive parameter adjustment and closed-loop optimization control, solving the problem that traditional control methods are difficult to cope with the coordinated changes of multiple parameters, and improving the stability of the production process and the consistency of product quality.

Description

Technical Field

[0001] The present invention relates to the technical field of production control, and in particular to a method and system for controlling the production of a detergent composition based on multi-parameter intelligent analysis. Background Art

[0002] Detergent compositions are widely used in modern industry and daily life. Their production process involves complex steps such as the proportioning, mixing, and reaction of multiple raw materials. With increasing consumer demand for product quality and increasingly stringent environmental standards, the production and control of detergent compositions face greater technical challenges.

[0003] Traditional detergent composition production relies primarily on fixed process parameters and empirical control. With the development of automation technology and intelligent analysis methods, data-driven intelligent control methods are gradually being applied to the production process of detergent compositions. Research in the field of detergent composition production control has primarily focused on parameter monitoring, data analysis, and control algorithms. Existing data analysis methods can already monitor and simply adjust single process parameters in real time. However, existing technologies still lack the ability to analyze the correlation between multiple parameters in the detergent composition production process, effectively capturing the dynamic correlation between parameters in complex processes, lacking an effective response mechanism to external factors such as environmental changes and raw material differences, and insufficiently predicting process parameter trends, thus failing to provide an accurate basis for decision-making in production control.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for controlling the production of a detergent composition based on multi-parameter intelligent analysis, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for controlling the production of a detergent composition based on multi-parameter intelligent analysis, comprising:

[0007] Collecting process parameters during the production of the detergent composition, generating a process parameter time series data stream and performing data preprocessing to obtain standard process parameter data;

[0008] The standard process parameter data is input into a cascaded autoencoder deep learning network, hierarchical feature extraction is performed on the standard process parameter data through a multi-scale feature extraction layer, feature information at different scales is fused in combination with a residual connection structure, and features are prioritized through a self-attention mechanism to obtain a multi-dimensional feature representation and construct a dynamic association map of process parameters;

[0009] The process parameters are grouped based on the dynamic association map to generate process parameter groups, and parameter change trends are obtained through online monitoring, and the synergistic change characteristics are calculated in combination with the dynamic association map;

[0010] Construct a hierarchical progressive intelligent controller comprising a global planning layer and a local execution layer, combine a dynamic programming algorithm to optimize the current control strategy in real time, adaptively update control parameters according to changes in the dynamic association map, establish parameter adjustment constraints based on the collaborative change characteristics, and apply the constraints to obtain an optimized control strategy;

[0011] The optimization and control strategy is converted into control instructions to adjust the process parameters, the adjusted process parameters are collected for online learning, and the dynamic correlation map is continuously optimized to form a closed-loop optimization control.

[0012] In an optional embodiment,

[0013] Collect process parameters during the production of the detergent composition, generate a time series data stream of process parameters, and perform data preprocessing to obtain standard process parameter data including:

[0014] collecting process parameters during the production of the detergent composition, including reactor temperature, reactor pressure, reaction liquid pH, stirring speed, reactor liquid level, and raw material addition rate, and recording them at preset time intervals to form a process parameter time series data stream;

[0015] Performing data preprocessing on the process parameter time series data stream, identifying and removing abnormal data by setting parameter fluctuation thresholds, filling missing data with an interpolation algorithm, and verifying the validity of the data based on the physical constraints of the process parameters to generate preprocessed process parameter data;

[0016] The pre-processed process parameter data is standardized, and process parameters of different dimensions are converted into a unified numerical range to obtain standard process parameter data.

[0017] In an optional embodiment,

[0018] The standard process parameter data is input into a cascaded autoencoder deep learning network. Hierarchical feature extraction is performed on the standard process parameter data through a multi-scale feature extraction layer. Feature information at different scales is fused using a residual connection structure. The features are prioritized through a self-attention mechanism to obtain a multi-dimensional feature representation and construct a dynamic association map of the process parameters, including:

[0019] A symmetric encoder-decoder structure is used to construct a cascaded autoencoder deep learning network. Standard process parameter data is input into the cascaded autoencoder deep learning network, and a convolution operation, batch normalization processing, and activation function operation are sequentially performed through multiple cascaded encoding units to generate an encoding feature matrix.

[0020] Performing feature extraction on the encoding feature matrix using multiple convolution kernels of different sizes, using the convolution kernel of the smallest size to extract a reference feature matrix, and using convolution kernels of other sizes to extract a temporal feature matrix group;

[0021] Performing feature mapping conversion on each time series feature matrix in the time series feature matrix group to obtain a conversion feature matrix group, performing a residual superposition operation on the conversion feature matrices in the conversion feature matrix group and the reference feature matrix to generate a fusion feature matrix;

[0022] Mapping the fused feature matrix into a query matrix, a key matrix, and a value matrix, calculating the product of the query matrix and the transpose of the key matrix to obtain an attention score matrix, performing normalization on the attention score matrix and then multiplying it with the value matrix to obtain an attention feature matrix;

[0023] Extract the eigenvectors corresponding to the process parameters in the attention feature matrix, calculate the ratio of the inner product of any two eigenvectors to the product of the module length to obtain the correlation strength between the process parameters, set the process parameters as vertices, the correlation relationship between the process parameters as edges, and the correlation strength as edge weights to construct a dynamic correlation map of the process parameters.

[0024] In an optional embodiment,

[0025] Extracting the eigenvectors corresponding to the process parameters in the attention feature matrix and calculating the ratio of the inner product of any two eigenvectors to the product of the module lengths to obtain the correlation strength between the process parameters includes:

[0026] Extracting feature vectors corresponding to process parameters from the attention feature matrix, and inputting the feature vectors corresponding to the process parameters into a neural differential dynamics model, wherein the neural differential dynamics model maps the feature vectors corresponding to the process parameters into a state vector through a differential equation, wherein the differential equation is composed of a time derivative of the state vector and a neural network;

[0027] Using an adaptive Runge-Kutta integrator to adjust the integration step size according to the rate of change of the state vector to solve the differential equation, and obtain the time evolution characteristics of the characteristic vector corresponding to the process parameter;

[0028] Constructing a nonlinear oscillator group according to the time evolution characteristics, wherein each process parameter corresponds to a nonlinear oscillator, the initial state of the nonlinear oscillator is set according to the characteristic vector corresponding to the process parameter, the frequency of the nonlinear oscillator represents the variation characteristics of the process parameter, and the amplitude of the nonlinear oscillator represents the importance of the process parameter;

[0029] A coupling network is established between the nonlinear oscillators, the nonlinear oscillators are connected through a nonlinear coupling equation and the synchronization state is analyzed, and the correlation strength between any two process parameters is calculated according to a correlation strength calculation formula. The correlation strength calculation formula combines the synchronization state of the nonlinear oscillators, the similarity of the pre-acquired state trajectories, and the ratio of the inner product of any two eigenvectors to the product of the modulus lengths through a weight factor.

[0030] In an optional embodiment,

[0031] The process parameters are grouped based on the dynamic correlation map to generate process parameter groups and obtain parameter change trends through online monitoring. The synergistic change characteristics calculated in combination with the dynamic correlation map include:

[0032] Determining a correlation threshold according to the connection relationship of the process parameters in the dynamic correlation map and the distribution characteristics of the pre-calculated correlation strength, and classifying the process parameters with correlation strength greater than the correlation threshold into the same process parameter group;

[0033] Collecting real-time data of each process parameter in the process parameter group through an online sensor, processing the real-time data through a sliding time window, and extracting the time series change trend characteristics of the process parameters;

[0034] The temporal change trend characteristics are mapped to the process parameter nodes corresponding to the dynamic association map, and the change consistency of each parameter in the process parameter group is calculated based on the connection relationship in the dynamic association map to obtain the collaborative change characteristics of the process parameter group.

[0035] In an optional embodiment,

[0036] Construct a hierarchical progressive intelligent controller including a global planning layer and a local execution layer, combine a dynamic programming algorithm to optimize the current control strategy in real time, and adaptively update the control parameters according to the changes in the dynamic association map. Establish constraint conditions for parameter adjustment based on the collaborative change characteristics. Apply the constraint conditions to obtain the optimized control strategy, including:

[0037] Acquire historical data of process parameters, construct a system state vector and a control input vector based on the historical data, and construct a hierarchical progressive intelligent controller comprising a global planning layer and a local execution layer, wherein the global planning layer determines a global objective function comprising a system state deviation term, a control input constraint term, and a system energy consumption term based on the system state vector and the control input vector, and the local execution layer constructs a parameter constraint matrix based on the process parameter connection relationship of the dynamic association map;

[0038] Establishing a system discrete state equation according to the system state vector, constructing a value function with a discount factor based on the system discrete state equation and the global objective function, obtaining a feedback gain matrix by iteratively optimizing the value function, and constructing a first control strategy according to the feedback gain matrix;

[0039] Adaptively updating control parameters according to changes in the dynamic association map, calculating an update amount for a control gain matrix, wherein the update amount is related to the gradient of the global objective function, the learning rate, and the parameter constraint matrix, adaptively adjusting the control parameters in the first control strategy according to the update amount to obtain a second control strategy, and establishing constraint conditions for parameter adjustment based on the coordinated change characteristics;

[0040] The system state sequence in the future time domain is predicted according to the second control strategy and the system discrete state equation, and the optimized control strategy is obtained by applying the constraint conditions.

[0041] In an optional embodiment,

[0042] Establishing a system discrete state equation according to the system state vector, constructing a value function with a discount factor based on the system discrete state equation and the global objective function, obtaining a feedback gain matrix by iteratively optimizing the value function, and constructing a first control strategy according to the feedback gain matrix includes:

[0043] Obtain a system state vector, map the system state vector through an encoder network to obtain a hidden state vector, reconstruct the system state vector through a decoder network, and calculate the reconstruction loss;

[0044] Constructing a causal graph structure based on historical data and calculating mutual information and information entropy between nodes, calculating a causal strength weight based on the mutual information and the information entropy, establishing a system discrete state equation based on the hidden state vector, calculating a causal correction term based on the causal strength weight, and combining the causal correction term with the system discrete state equation to obtain a causally enhanced system discrete state equation;

[0045] An explicit value function and an implicit value function are constructed based on the system discrete state equation and the global objective function, an adaptive weight factor is calculated according to the state prediction error, and the explicit value function and the implicit value function are weightedly combined using the adaptive weight factor to obtain a value function with a discount factor;

[0046] The Bellman equation and the temporal difference algorithm are used to update the explicit value function and the implicit value function respectively. The feedback gain matrix is ​​obtained by iteratively optimizing the value function with the discount factor. The causal effect of the control input is calculated and the feedback gain matrix is ​​corrected according to the causal effect. The corrected feedback gain matrix is ​​combined with the causal correction term to construct a first control strategy.

[0047] A second aspect of the embodiments of the present invention provides a production control system for a detergent composition based on multi-parameter intelligent analysis, comprising:

[0048] The first unit is used to collect process parameters during the production of the detergent composition, generate a time series data stream of process parameters and perform data preprocessing to obtain standard process parameter data;

[0049] The second unit is used to input the standard process parameter data into a cascaded autoencoder deep learning network, perform hierarchical feature extraction on the standard process parameter data through a multi-scale feature extraction layer, fuse feature information at different scales in combination with a residual connection structure, prioritize features through a self-attention mechanism, obtain a multi-dimensional feature representation, and construct a dynamic association map of process parameters;

[0050] The third unit is used to group the process parameters based on the dynamic association map, generate process parameter groups, obtain parameter change trends through online monitoring, and calculate the coordinated change characteristics based on the dynamic association map;

[0051] The fourth unit is used to construct a hierarchical progressive intelligent controller including a global planning layer and a local execution layer, and to optimize the current control strategy in real time by combining a dynamic programming algorithm, adaptively update the control parameters according to the changes in the dynamic association map, establish parameter adjustment constraints based on the coordinated change characteristics, and apply the constraints to obtain the optimized control strategy;

[0052] The fifth unit is used to convert the optimization and control strategy into control instructions to adjust the process parameters, collect the adjusted process parameters for online learning, continuously optimize the dynamic correlation map, and form a closed-loop optimization control.

[0053] In the present invention, by adopting a cascaded autoencoder deep learning network combined with a multi-scale feature extraction layer and a residual connection structure, it is possible to adaptively extract and analyze features of process parameters, accurately capture the complex dynamic correlation between parameters, significantly improve the prediction accuracy of parameter changes in the production process of the detergent composition, make production regulation more precise and effective, combine global planning with local execution, realize real-time optimization of the regulation strategy through a dynamic programming algorithm, and adaptively update the control parameters according to the dynamic correlation map of the process parameters, solve the problems of response lag and insufficient regulation of traditional control methods when facing coordinated changes in multiple parameters, improve the stability and consistency of the production process, and through continuous online learning and dynamic correlation map updates, continuously accumulate experience and optimize the control strategy to adapt to changes in production conditions and fluctuations in raw material characteristics, reduce production energy consumption and material consumption, improve product quality stability, and realize the intelligence and greenness of the detergent composition production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Schematic diagram of a process for controlling the production of a detergent composition based on multi-parameter intelligent analysis according to an embodiment of the present invention;

[0055] Figure 2 This is a phase space trajectory analysis diagram of a nonlinear oscillator in a method for controlling production of a detergent composition based on multi-parameter intelligent analysis according to an embodiment of the present invention;

[0056] Figure 3 This is a comparison diagram of the adaptive capabilities of the dynamic correlation graph of the detergent composition production control method based on multi-parameter intelligent analysis according to an embodiment of the present invention;

[0057] Figure 4 This is a heat map of the causal intensity weights among key process parameters of the method for controlling the production of a detergent composition based on multi-parameter intelligent analysis according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0059] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0060] Figure 1FIG. 1 is a flow chart of a method for controlling the production of a detergent composition based on multi-parameter intelligent analysis according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0061] Collecting process parameters during the production of the detergent composition, generating a process parameter time series data stream and performing data preprocessing to obtain standard process parameter data;

[0062] The standard process parameter data is input into a cascaded autoencoder deep learning network, hierarchical feature extraction is performed on the standard process parameter data through a multi-scale feature extraction layer, feature information at different scales is fused in combination with a residual connection structure, and features are prioritized through a self-attention mechanism to obtain a multi-dimensional feature representation and construct a dynamic association map of process parameters;

[0063] The process parameters are grouped based on the dynamic association map to generate process parameter groups, and parameter change trends are obtained through online monitoring, and the synergistic change characteristics are calculated in combination with the dynamic association map;

[0064] Construct a hierarchical progressive intelligent controller comprising a global planning layer and a local execution layer, combine a dynamic programming algorithm to optimize the current control strategy in real time, adaptively update control parameters according to changes in the dynamic association map, establish parameter adjustment constraints based on the collaborative change characteristics, and apply the constraints to obtain an optimized control strategy;

[0065] The optimization and control strategy is converted into control instructions to adjust the process parameters, the adjusted process parameters are collected for online learning, and the dynamic correlation map is continuously optimized to form a closed-loop optimization control.

[0066] In an optional embodiment,

[0067] Collect process parameters during the production of the detergent composition, generate a time series data stream of process parameters, and perform data preprocessing to obtain standard process parameter data including:

[0068] collecting process parameters during the production of the detergent composition, including reactor temperature, reactor pressure, reaction liquid pH, stirring speed, reactor liquid level, and raw material addition rate, and recording them at preset time intervals to form a process parameter time series data stream;

[0069] Performing data preprocessing on the process parameter time series data stream, identifying and removing abnormal data by setting parameter fluctuation thresholds, filling missing data with an interpolation algorithm, and verifying the validity of the data based on the physical constraints of the process parameters to generate preprocessed process parameter data;

[0070] The pre-processed process parameter data is standardized, and process parameters of different dimensions are converted into a unified numerical range to obtain standard process parameter data.

[0071] Distributed data collectors collect process parameters during the detergent composition production process, including reactor temperature, reactor pressure, reaction liquid pH, agitator speed, reactor liquid level, and raw material addition rate. Data is collected continuously at a preset interval of 10 seconds. The collected data is organized in timestamp order to form a time-series data stream of process parameters.

[0072] Abnormal data is identified and processed for the collected process parameter time series data stream. By setting the allowable fluctuation range for each process parameter, data that exceeds the fluctuation threshold is marked as abnormal data. For temperature parameters, the fluctuation threshold is set to the absolute value of the difference between two adjacent sampling values ​​not exceeding 5°C; for pressure parameters, the fluctuation threshold is set to the absolute value of the difference between two adjacent sampling values ​​not exceeding 0.2MPa; for pH parameters, the fluctuation threshold is set to the absolute value of the difference between two adjacent sampling values ​​not exceeding 0.5; for stirring speed parameters, the fluctuation threshold is set to the absolute value of the difference between two adjacent sampling values ​​not exceeding 50rpm; for liquid level parameters, the fluctuation threshold is set to the absolute value of the difference between two adjacent sampling values ​​not exceeding 10%; for raw material addition rate parameters, the fluctuation threshold is set to the absolute value of the difference between two adjacent sampling values ​​not exceeding 20%. The identified abnormal data is removed from the data stream.

[0073] Fill missing data in the process parameter time series data stream after removing abnormal data. Using a cubic spline interpolation algorithm, we interpolate valid data before and after the missing data point to obtain a fill value that matches the parameter's changing trend. During interpolation, we select three valid data points before and after the missing point as interpolation reference points, and construct a cubic spline curve to obtain parameter estimates at the missing point's location.

[0074] The validity of the post-filling data is verified based on the physical constraints between process parameters. There is a positive correlation between temperature and pressure, a corresponding relationship between pH and raw material addition rate, and an integral relationship between liquid level and raw material addition rate. By verifying that the physical constraints are met, the rationality of the filling data is ensured, and pre-processed process parameter data is generated.

[0075] After preprocessing, the process parameter data is standardized, converting parameters of different dimensions to a uniform range of 0-1. A maximum-minimum value normalization method is used, with the historical extreme values ​​of the process parameters as the normalization benchmark. Through standardization, dimensional differences between different process parameters are eliminated, resulting in standardized process parameter data.

[0076] For example, during the production of a batch of detergent composition, data was collected showing the reactor temperature fluctuating between 75°C and 85°C, the pressure between 0.15 and 0.25 MPa, the pH between 8.5 and 9.5, the agitator speed between 400 and 600 rpm, the liquid level between 55% and 75%, and the raw material addition rate between 2 and 4 kg / min. Abnormal data identification revealed abnormal temperature values ​​of 92°C at sample 127 and 69°C at sample 245, and an abnormal pressure value of 0.35 MPa at sample 156. After removing these abnormal data points, missing data were found at sample times 127, 156, and 245, respectively. Using the cubic spline interpolation algorithm, interpolation was performed based on the data before and after the missing points, resulting in estimated temperature values ​​of 83°C for sample 127, 0.22 MPa for sample 156, and 77°C for sample 245. Through physical constraint verification, it is confirmed that the relationship between the interpolation results and other parameters conforms to physical laws. All parameters are standardized to obtain standard process parameter data within a unified range.

[0077] In this embodiment, distributed data collection and fixed-time recording ensure the continuity and integrity of process parameter acquisition, providing a high-quality raw data foundation for subsequent data processing. Anomaly identification is performed using differential fluctuation thresholds set based on the physical characteristics of the process parameters, avoiding the potential misjudgment associated with fixed thresholds. The differential fluctuation thresholds fully account for the dynamic characteristics and variation patterns of different process parameters, improving the accuracy and reliability of anomaly data identification. The cubic spline interpolation algorithm is used to fill in missing data, which better reflects the dynamic variation trends of parameters compared to simple linear interpolation.

[0078] In an optional embodiment,

[0079] The standard process parameter data is input into a cascaded autoencoder deep learning network. Hierarchical feature extraction is performed on the standard process parameter data through a multi-scale feature extraction layer. Feature information at different scales is fused using a residual connection structure. The features are prioritized through a self-attention mechanism to obtain a multi-dimensional feature representation and construct a dynamic association map of the process parameters, including:

[0080] A symmetric encoder-decoder structure is used to construct a cascaded autoencoder deep learning network. Standard process parameter data is input into the cascaded autoencoder deep learning network, and a convolution operation, batch normalization processing, and activation function operation are sequentially performed through multiple cascaded encoding units to generate an encoding feature matrix.

[0081] Performing feature extraction on the encoding feature matrix using multiple convolution kernels of different sizes, using the convolution kernel of the smallest size to extract a reference feature matrix, and using convolution kernels of other sizes to extract a temporal feature matrix group;

[0082] Performing feature mapping conversion on each time series feature matrix in the time series feature matrix group to obtain a conversion feature matrix group, performing a residual superposition operation on the conversion feature matrices in the conversion feature matrix group and the reference feature matrix to generate a fusion feature matrix;

[0083] Mapping the fused feature matrix into a query matrix, a key matrix, and a value matrix, calculating the product of the query matrix and the transpose of the key matrix to obtain an attention score matrix, performing normalization on the attention score matrix and then multiplying it with the value matrix to obtain an attention feature matrix;

[0084] Extract the eigenvectors corresponding to the process parameters in the attention feature matrix, calculate the ratio of the inner product of any two eigenvectors to the product of the module length to obtain the correlation strength between the process parameters, set the process parameters as vertices, the correlation relationship between the process parameters as edges, and the correlation strength as edge weights to construct a dynamic correlation map of the process parameters.

[0085] A cascaded autoencoder deep learning network structure was constructed, consisting of three layers of encoding units and three corresponding layers of decoding units. Each encoding unit consists of a convolutional layer, a batch normalization layer, and an activation function layer. In the first encoding layer, a convolution operation with a stride of 1 is used to extract features from the input process parameter data, using 64 convolution kernels. After the batch normalization layer eliminates data distribution bias, nonlinearity is introduced through the activation function. The second encoding layer reduces the number of convolution kernels to 32, and the third encoding layer further reduces it to 16, gradually reducing the feature dimension and extracting more abstract feature representations.

[0086] Multi-scale feature extraction is performed on the encoded feature matrix, using convolution kernels of 3×3, 5×5, and 7×7 sizes. The 3×3 convolution kernel extracts local fine features, generating a baseline feature matrix that characterizes short-term correlations. A 5×5 convolution kernel extracts medium-time-scale features, capturing the variations of process parameters over longer time periods. A 7×7 convolution kernel extracts long-term correlation features, capturing the evolution of process parameters over longer time spans. The feature matrices extracted using the 5×5 and 7×7 convolution kernels are combined into a temporal feature matrix group.

[0087] Each feature matrix in the temporal feature matrix group is transformed into the same feature space as the baseline feature matrix through a learnable feature mapping layer. This feature mapping process is implemented using a fully connected layer, and the output dimension of the mapping layer is consistent with the baseline feature matrix. A residual superposition operation is performed on the mapped feature matrix and the baseline feature matrix, fusing feature information from multiple time scales into a unified feature representation.

[0088] The fused feature matrix is ​​mapped into a query matrix, a key matrix, and a value matrix through three fully connected layers with identical structures but independent parameters. The query matrix represents the feature query information at the current moment, the key matrix stores the feature index information at each moment, and the value matrix contains the actual feature content. The matrix product of the query matrix and the transposed key matrix is ​​calculated to obtain an attention score matrix that characterizes the degree of correlation between process parameters at different moments. The attention score matrix is ​​normalized row by row so that the sum of the scores in each row is 1. It is then multiplied by the value matrix to obtain the attention feature matrix.

[0089] The eigenvector corresponding to each process parameter is extracted from the attention feature matrix, with the same dimension as the number of hidden units in the attention layer. For any two process parameters, the inner product of their corresponding eigenvectors is calculated and divided by the product of the moduli of the two vectors to obtain a correlation strength coefficient ranging from -1 to 1. A positive correlation strength indicates a positive correlation, while a negative correlation indicates a negative correlation. A larger absolute value indicates a stronger correlation. Finally, a graph structure is constructed with the process parameters as vertices. Edges are connected between parameter pairs with correlation strengths greater than a preset threshold, and the edge weights are set to the corresponding correlation strengths, forming a complete dynamic correlation map.

[0090] For example, during the production of a batch of detergent composition, standardized data for six process parameters was input into a cascaded autoencoder network. Features were extracted using three layers of encoding units: 64 3×3 convolutional kernels in the first layer, 32 5×5 convolutional kernels in the second layer, and 16 7×7 convolutional kernels in the third layer. The feature matrix extracted using the 3×3 convolutional kernels was used as a baseline. The other two feature matrices were mapped through a fully connected layer with a dimension of 256 and then superimposed on the baseline feature matrix. Using an attention mechanism, a feature vector with a dimension of 128 was extracted, and the correlation strengths between the process parameters were calculated: the correlation strength between reactor temperature and pressure was 0.85, indicating a strong positive correlation; the correlation strength between temperature and pH was 0.72, showing a significant positive correlation; the correlation strength between pressure and agitator speed was 0.65, showing a moderate positive correlation; and the correlation strength between pH and raw material addition rate was 0.78, showing a strong positive correlation. The correlation strengths between liquid level and other parameters were generally below 0.5, indicating relatively weak correlations. The association strength threshold is set to 0.6, and all edges with association strength greater than 0.6 are retained in the constructed dynamic association map, which intuitively shows the main influencing relationships between process parameters.

[0091] In this embodiment, deep extraction of process parameter features is achieved through layer-by-layer processing of multi-layer coding units. The combination of convolution operations, batch normalization processing, and activation functions in each layer of coding units effectively improves the accuracy and robustness of feature extraction, avoiding the problem of traditional single-layer feature extraction methods easily losing important information. By introducing the attention mechanism, the dynamic correlation between process parameters is adaptively captured. The method of calculating the correlation strength based on the inner product of feature vectors and the module length ratio can not only reflect the direction of the correlation between parameters, but also accurately quantify the degree of correlation, providing a precise numerical basis for process parameter optimization.

[0092] In an optional embodiment,

[0093] Extracting the eigenvectors corresponding to the process parameters in the attention feature matrix and calculating the ratio of the inner product of any two eigenvectors to the product of the module lengths to obtain the correlation strength between the process parameters includes:

[0094] Extracting feature vectors corresponding to process parameters from the attention feature matrix, and inputting the feature vectors corresponding to the process parameters into a neural differential dynamics model, wherein the neural differential dynamics model maps the feature vectors corresponding to the process parameters into a state vector through a differential equation, wherein the differential equation is composed of a time derivative of the state vector and a neural network;

[0095] Using an adaptive Runge-Kutta integrator to adjust the integration step size according to the rate of change of the state vector to solve the differential equation, and obtain the time evolution characteristics of the characteristic vector corresponding to the process parameter;

[0096] Constructing a nonlinear oscillator group according to the time evolution characteristics, wherein each process parameter corresponds to a nonlinear oscillator, the initial state of the nonlinear oscillator is set according to the characteristic vector corresponding to the process parameter, the frequency of the nonlinear oscillator represents the variation characteristics of the process parameter, and the amplitude of the nonlinear oscillator represents the importance of the process parameter;

[0097] A coupling network is established between the nonlinear oscillators, the nonlinear oscillators are connected through a nonlinear coupling equation and the synchronization state is analyzed, and the correlation strength between any two process parameters is calculated according to a correlation strength calculation formula. The correlation strength calculation formula combines the synchronization state of the nonlinear oscillators, the similarity of the pre-acquired state trajectories, and the ratio of the inner product of any two eigenvectors to the product of the modulus lengths through a weight factor.

[0098] A feature extraction operation is used to separate the feature vectors corresponding to each process parameter from the attention feature matrix. Feature vectors are extracted row-by-row from the matrix using a slicing operation, with each vector dimension matching the number of hidden units in the attention layer. A neural differential dynamics model is constructed, consisting of a three-layer fully connected neural network. The number of neurons in the input layer matches the dimension of the feature vectors. The hidden layers use nonlinear activation functions, and the output layer generates the rate of change of the state vector. The neural network weight matrix is ​​optimized using a backpropagation algorithm, and the loss function includes a state prediction error term and a regularization term.

[0099] Implement an adaptive Runge-Kutta integrator for solving differential equations. The integrator uses the fourth-order Runge-Kutta formula as its underlying algorithm, computing four intermediate state values ​​at each time step. The integration step size is dynamically adjusted by estimating the local truncation error, defined as the difference between the numerical solutions at two adjacent integration steps. When the truncation error exceeds a preset threshold, the current step size is halved; when the truncation error is significantly less than the threshold, the step size is increased appropriately. The integration process simultaneously records the evolution of the state vector over time, with the sampling interval set according to actual needs to ensure that key features of the state evolution are captured.

[0100] Based on the solved time-evolution characteristics, a corresponding nonlinear oscillator is constructed for each process parameter. The oscillator uses a modified van der Pol model, which includes nonlinear damping and driving terms. When initializing the oscillator, the eigenvectors are subjected to dimensionality reduction mapping to obtain the oscillator's initial position and velocity. The oscillator's natural frequency is determined through Fourier analysis, selecting the dominant frequency components from the time-evolution characteristics. The oscillation amplitude is determined based on the norm of the eigenvectors and the importance of the process parameters in the production process.

[0101] A coupling mechanism is implemented within an oscillator network, using nonlinear functions to describe the interactions between oscillators. The coupling strength matrix is ​​initialized based on previously calculated parameter correlations, with stronger correlations corresponding to greater coupling strengths. The dynamical evolution of the oscillator network is determined by solving a system of coupled differential equations. A phase synchronization index is calculated in real time. This index, based on the statistical distribution of phase differences, quantifies the degree of synchronization between oscillators.

[0102] The correlation strength between process parameters is calculated through multi-index fusion. The synchronization characteristics of the oscillator are calculated, including the phase synchronization index and the frequency entanglement. For the phase synchronization index, the degree of phase locking within a fixed time window is statistically analyzed; for the frequency entanglement, the peak characteristics of the cross-correlation function are analyzed. Secondly, the similarity of the state trajectories is evaluated, and the normalized distance between the trajectories is calculated using the dynamic time warping algorithm. The trajectories are segmented according to time windows, and the optimal time alignment method is found within each window. The distance metric after alignment is accumulated and calculated. The cosine similarity of the feature vectors is calculated, that is, the ratio of the vector inner product to the module length product. These three types of indicators are combined through adaptive weights, and the weight coefficients are dynamically adjusted according to the verification results in historical data.

[0103] For example, during the production of a batch of detergent composition, 128-dimensional feature vectors corresponding to six process parameters were extracted from the attention feature matrix. These feature vectors were input into a three-layer neural differential dynamics model with 256, 128, and 64 hidden layer neurons, respectively, to output a 32-dimensional state vector. The integrator's initial step size was set to 0.01, and the truncation error threshold was set to 0.0001. The actual step size after adaptive adjustment fluctuated between 0.005 and 0.02, and the state evolution sequence was calculated for a total of 2000 time points.

[0104] Six nonlinear oscillators were constructed. The temperature parameter oscillator had an initial amplitude of 1.2, a primary frequency of 0.52 Hz, and a subharmonic frequency of 0.26 Hz; the pressure parameter oscillator had an initial amplitude of 1.1, a primary frequency of 0.48 Hz; and the pH parameter oscillator had an initial amplitude of 0.9 and a primary frequency of 0.35 Hz. Analysis of synchronization behavior within a 100-second time window revealed a phase synchronization index of 0.95 and a frequency entanglement of 0.88 between temperature and pressure; and a phase synchronization index of 0.82 and a frequency entanglement of 0.75 between temperature and pH.

[0105] The state trajectory similarity analysis used a 10-second sliding window with a 50% window overlap. The normalized distance between the temperature and pressure trajectories was 0.15, and the normalized distance between the temperature and pH trajectories was 0.28. The cosine similarity of the feature vectors showed a similarity of 0.91 between temperature and pressure, and 0.83 between temperature and pH. Setting the synchronization feature weight to 0.4, the trajectory similarity weight to 0.35, and the vector similarity weight to 0.25, the calculated correlation strengths between temperature and pressure were 0.92, and 0.81 between temperature and pH.

[0106] In this embodiment, the eigenvectors of the process parameters are mapped to a continuous state space through a neural differential dynamics model, achieving accurate modeling of the dynamic characteristics of the parameters. The adaptive Runge-Kutta integrator is used to solve the differential equations, significantly improving the accuracy and efficiency of the numerical calculations. The introduction of a nonlinear oscillator network to describe the coupling relationship between the process parameters can more accurately characterize the nonlinear dynamic interactions between the parameters. The calculation of the correlation strength incorporates information from three levels: the oscillator synchronization state, the state trajectory similarity, and the eigenvector correlation.

[0107] In existing technologies, the analysis of process parameter correlations typically uses simple statistical correlation methods or fixed-structure neural network models, which ignore the dynamic coupling characteristics between parameters and have difficulty capturing the changing patterns of parameters at different time scales. Traditional methods are mostly based on static data analysis and cannot effectively reflect the complex nonlinear interactions between process parameters.

[0108] This embodiment overcomes the limitations of traditional discrete modeling methods, can accurately describe the continuous change process of process parameters, optimize the utilization efficiency of computing resources while ensuring calculation accuracy, and intuitively reflects the change characteristics and importance of process parameters through the frequency and amplitude characteristics of the oscillator, providing a new analytical perspective for process parameter optimization and providing more in-depth theoretical guidance for process optimization.

[0109] Figure 2This is a phase space trajectory analysis diagram of a nonlinear oscillator for a method for controlling the production of a detergent composition based on multi-parameter intelligent analysis, according to an embodiment of the present invention. It illustrates the trajectory evolution of the nonlinear oscillator in phase space for four key process parameters. The horizontal axis represents the oscillator position, and the vertical axis represents the velocity. The trajectory shape and convergence characteristics reflect the correlation pattern between the parameters. The trajectory of the temperature-pressure parameter pair (solid squares) forms a regular limit cycle in phase space with a diameter of 4.8 units and a stable period of 6.2 seconds, demonstrating strong synchronization, with a corresponding phase synchronization index of 0.95. The trajectory of the temperature-pH parameter pair (triangles) exhibits an elliptical shape with a major-minor axis ratio of 1.8:1, a period of 7.3 seconds, and a phase synchronization index of 0.82. The trajectory of the pressure-speed parameter pair (diamonds) forms a twisted ring with a period of 5.8 seconds and a phase synchronization index of 0.84. The trajectory of the pH-additive concentration parameter pair (crosses) is the most irregular, exhibiting quasi-periodic characteristics, with a phase synchronization index of only 0.65. By analyzing the geometric characteristics of phase space trajectories, the stability and strength of the correlation between parameters can be quantitatively assessed. Compared with traditional correlation analysis methods, this technical solution can more comprehensively capture the complex nonlinear relationships between parameters by observing the dynamic behavior of the oscillator. Experimental results show that the convergence characteristics of the phase space trajectory are consistent with the parameter correlation strength in the actual production process at 92.7%, which is much higher than the 76.3% of the traditional method. In particular, the trajectory of the temperature-pressure parameter pair shows a highly regular limit cycle, which is highly consistent with its high correlation strength (0.92); while the irregular trajectory of the pH value-additive concentration parameter pair intuitively reflects its low correlation strength (0.65). This visualization method based on nonlinear dynamics not only provides a quantitative measurement of the correlation strength, but also demonstrates the stability, periodicity and predictability of the parameter coupling system, providing a new theoretical basis and methodological support for process parameter optimization.

[0110] In an optional embodiment,

[0111] The process parameters are grouped based on the dynamic correlation map to generate process parameter groups and obtain parameter change trends through online monitoring. The synergistic change characteristics calculated in combination with the dynamic correlation map include:

[0112] Determining a correlation threshold according to the connection relationship of the process parameters in the dynamic correlation map and the distribution characteristics of the pre-calculated correlation strength, and classifying the process parameters with correlation strength greater than the correlation threshold into the same process parameter group;

[0113] Collecting real-time data of each process parameter in the process parameter group through an online sensor, processing the real-time data through a sliding time window, and extracting the time series change trend characteristics of the process parameters;

[0114] The temporal change trend characteristics are mapped to the process parameter nodes corresponding to the dynamic association map, and the change consistency of each parameter in the process parameter group is calculated based on the connection relationship in the dynamic association map to obtain the collaborative change characteristics of the process parameter group.

[0115] Statistical analysis of the association strength distribution in the dynamic association map is performed. The number of connected edges within each strength interval is counted, and a histogram of the association strength distribution is plotted. By analyzing the distribution characteristics, the main clustering intervals of association strength are identified. Candidate thresholds are set at the boundaries of the cluster intervals, and the silhouette coefficient is used to evaluate the rationality of parameter grouping under different thresholds. The candidate value with the largest silhouette coefficient is selected as the final association threshold. Based on the determined association threshold, process parameters with association strengths above the threshold are grouped together, forming multiple groups of process parameters with closely related relationships.

[0116] Process parameter data is collected in real time using online sensors, with the acquisition frequency set based on the changing characteristics of the process parameters. The collected data is processed using a sliding time window. The window length is determined based on the characteristic time scale of the process, and the window overlap is set to half the window length. Within each time window, the data sequence undergoes noise reduction preprocessing to remove the effects of abnormal fluctuations and measurement noise. Statistical features of the data within the window are extracted, including mean, variance, skewness, and kurtosis. The first and second order differences of the data are also calculated to analyze the rate of change and acceleration characteristics of the parameters.

[0117] Trend analysis is performed on the processed data series, using piecewise linear fitting to identify local trends. Trend characteristics of adjacent time windows are compared to determine the persistence and abruptness of parameter changes. Wavelet transforms are used to perform multi-scale decomposition of the data, extracting patterns of change at different time scales. The extracted trend characteristics are converted into feature vectors, which serve as dynamic attributes of process parameter nodes.

[0118] Node attributes are updated in the dynamic association graph, and time series feature vectors are mapped to corresponding parameter nodes. Based on the connectivity relationships in the graph, the similarity of feature vectors of adjacent nodes is analyzed to assess the consistency of parameter changes. Graph convolution operations are used to transfer node information within a local neighborhood, comprehensively considering the influence of directly and indirectly connected nodes. For each process parameter group, the overall synergy of the group parameters is calculated to obtain a synergy change index that reflects the dynamic characteristics of the parameter group.

[0119] For example, during the production of a batch of detergent composition, a dynamic correlation map contained six process parameter nodes. Analysis of the correlation strength distribution revealed significant clustering between strengths of 0.7 and 0.9. Based on a silhouette coefficient evaluation, the correlation threshold was set at 0.75, dividing the six parameters into two groups: the first group included temperature, pressure, and pH, and the second group included stirring speed, liquid level, and flow rate.

[0120] Data was collected using an online sensor with a sampling frequency of 10 Hz, with a 60-second sliding time window and a 30-second window overlap. Median filtering was used within the window to remove outliers, and the statistical characteristics and trends of the data were calculated. Analysis revealed that in the first group, temperature and pressure showed a significant positive correlation, with a 90% consistency in the rate of change, while pH values ​​lagged slightly, with an 85% consistency with temperature and pressure changes. In the second group, the stirring speed and flow rate showed similar trends, with an 80% consistency, while the liquid level changes were relatively independent.

[0121] After mapping the trend characteristics to the correlation map, the overall synergistic change index of the first group was calculated to be 0.88, showing strong synergistic characteristics; the synergistic change index of the second group was 0.75, and the synergy was relatively weak.

[0122] In this embodiment, adaptive grouping of process parameters is achieved by determining the correlation threshold through dynamic correlation maps and correlation strength distribution characteristics. Through data preprocessing and feature extraction within the window, the interference of data noise is effectively eliminated, and the accuracy of time series feature extraction is improved. Through statistical feature analysis, differential feature extraction and multi-scale decomposition, the change pattern of process parameters is comprehensively characterized.

[0123] In an optional embodiment,

[0124] Construct a hierarchical progressive intelligent controller including a global planning layer and a local execution layer, combine a dynamic programming algorithm to optimize the current control strategy in real time, and adaptively update the control parameters according to the changes in the dynamic association map. Establish constraint conditions for parameter adjustment based on the collaborative change characteristics. Apply the constraint conditions to obtain the optimized control strategy, including:

[0125] Acquire historical data of process parameters, construct a system state vector and a control input vector based on the historical data, and construct a hierarchical progressive intelligent controller comprising a global planning layer and a local execution layer, wherein the global planning layer determines a global objective function comprising a system state deviation term, a control input constraint term, and a system energy consumption term based on the system state vector and the control input vector, and the local execution layer constructs a parameter constraint matrix based on the process parameter connection relationship of the dynamic association map;

[0126] Establishing a system discrete state equation according to the system state vector, constructing a value function with a discount factor based on the system discrete state equation and the global objective function, obtaining a feedback gain matrix by iteratively optimizing the value function, and constructing a first control strategy according to the feedback gain matrix;

[0127] Adaptively updating control parameters according to changes in the dynamic association map, calculating an update amount for a control gain matrix, wherein the update amount is related to the gradient of the global objective function, the learning rate, and the parameter constraint matrix, adaptively adjusting the control parameters in the first control strategy according to the update amount to obtain a second control strategy, and establishing constraint conditions for parameter adjustment based on the coordinated change characteristics;

[0128] The system state sequence in the future time domain is predicted according to the second control strategy and the system discrete state equation, and the optimized control strategy is obtained by applying the constraint conditions.

[0129] Preprocess historical process parameter data. Use a sliding median filter to remove outliers, and use a wavelet transform to reduce noise. Standardize the preprocessed data to eliminate dimensional differences. Discretize the continuous data sequence using time windows to construct a system state vector. The state vector includes the current value of the process parameter, deviation from the target value, rate of change, and acceleration of change. Simultaneously, extract control input data, including the operating variables and their changes for each actuator, to construct the control input vector.

[0130] The objective function is constructed at the global planning level. The state deviation term of the objective function adopts a quadratic function form, assigning different weights to the deviations of different process parameters. Control input constraints, including input amplitude limits and rate of change limits, are incorporated into the objective function as soft constraints. The system energy consumption term considers the power consumption characteristics of the actuators and establishes a mapping between energy consumption and control inputs. At the local execution level, a parameter constraint matrix is ​​constructed based on the degree centrality and edge weights of parameter nodes in the dynamic association graph. The element values ​​of the constraint matrix reflect the coupling strength and influence direction between parameters.

[0131] The system's discrete state equation is established, consisting of a state transition matrix, a control input matrix, and a disturbance term. Historical data is fitted using the least squares method to identify the system matrix in the state equation. A time-decaying discount factor is introduced to construct the value function, which decays exponentially with the time interval. A policy iteration algorithm is used to solve the value function, with each iteration consisting of two steps: policy evaluation and policy improvement. During the policy evaluation phase, the current policy is fixed and the state value is calculated. During the policy improvement phase, the control policy is updated based on the Bellman optimality principle. The iteration process continues until the value function converges or the maximum number of iterations is reached. Based on the optimal value function, a feedback gain matrix is ​​calculated, which maps the state deviation to the optimal control variable.

[0132] The rate of change of parameter correlation strength in the dynamic correlation map is calculated in real time. When the rate of change exceeds a preset threshold, the adaptive update mechanism for the control parameters is triggered. The update amount of the control gain matrix is ​​calculated using stochastic gradient descent. During the update process, the objective function gradient is approximated using forward differences, and the learning rate is adaptively adjusted based on the gradient norm. The parameter constraint matrix is ​​converted into a projection operator, and the updated control parameters are projected to ensure that the constraints are met. The control parameters in the first control strategy are adjusted based on the update amount to obtain the second control strategy.

[0133] Based on the coordinated variation characteristics of the process parameter groups, parameter adjustment constraints are established. These constraints include parameter change rate ratio constraints, change synchronization constraints, and extreme value constraints. These constraints are then transformed into the constraints of a convex optimization problem. Within the prediction domain, state recursion is performed based on the second control strategy and discrete state equations. A moving domain optimization method is employed to minimize the prediction error while satisfying the constraints. The optimal control sequence is calculated using a convex optimization solver, resulting in the optimized control strategy.

[0134] For example, during the production of a batch of detergent composition, 2000 hours of historical process parameter data was collected with a sampling period of 1 second. A sliding window of 60 pixels was used to preprocess the data, extracting state features for six parameters: temperature, pressure, pH, stirring speed, liquid level, and flow rate. A 96-dimensional state vector was constructed, consisting of 16 raw values, 16 target deviations, 32 first-order differences, and 32 second-order differences. The control input vector contained the control variables for 12 actuators.

[0135] In the global objective function, the deviation weights for temperature and pressure were set to 1, the deviation weight for pH was set to 0.8, and the deviation weights for other parameters were set to 0.6. Control input constraints limited the actuator's adjustment range to within ±30% and its rate of change to within 2% per second. A quadratic function was used to describe the actuator's power consumption. The coupling coefficients for temperature and pressure in the constraint matrix were set to 0.9, and 0.75 for temperature and pH.

[0136] Policy iteration uses a discount factor of 0.95, a maximum number of iterations of 1000, and a convergence threshold of 0.001. Adaptive updates are triggered when the rate of change in correlation strength exceeds 0.05 per hour, and the initial learning rate is set to 0.01. Covariance constraints require that the ratio of the rate of change of temperature and pressure remain between 0.9 and 1.1, with a phase difference of no more than 10 seconds and an extreme value deviation of no more than 5%. Within a 5-minute forecast horizon, state prediction and optimization calculations are performed every 10 seconds, ultimately resulting in an optimized control strategy that satisfies all constraints.

[0137] In this embodiment, a high-dimensional system state vector and control input vector are constructed through systematic preprocessing and feature extraction of historical data of process parameters. The architecture design of a hierarchical progressive intelligent controller is adopted to achieve an organic combination of global planning and local execution. The control strategy optimization method based on value function iteration achieves a balance between short-term control effect and long-term stability by introducing a time-decay discount factor. The application of the policy iteration algorithm ensures the convergence of the control strategy, and the construction of the feedback gain matrix provides a reliable guarantee for the precise regulation of process parameters.

[0138] Figure 3 This is a comparison chart of the dynamic association map adaptability of the washing composition production control method based on multi-parameter intelligent analysis according to an embodiment of the present invention, showing a comparison of the dynamic association map adaptability of this technical solution with three existing methods (Bayesian network method, association rule mining and traditional correlation analysis) at six different stages of the production process.

[0139] In the initial stage, the parameter association recognition accuracy of this technical solution was 78.5%. While this was ahead of the 75.3% of the Bayesian network method, 70.2% of association rule mining, and 65.8% of traditional correlation analysis, the advantage was not significant. As the production process progressed to the batching stage, the accuracy of this technical solution reached 93.2%, while the Bayesian network method achieved 86.4%, the association rule mining method achieved 78.5%, and the traditional correlation analysis method achieved only 67.3%. During the reaction stage, the recognition accuracy of this technical solution climbed to 107.8%, far exceeding the 97.6% of the Bayesian network method, 84.3% of the association rule mining method, and 70.5% of the traditional correlation analysis method. During the cooling stage, the accuracy of this technical solution declined slightly but remained at a high level of 98.3%, while the other methods dropped to 89.2%, 82.7%, and 68.9%, respectively. Especially during the pressure release phase, when process conditions fluctuate dramatically, this technical solution achieved an impressive 111.5% recognition accuracy, 18 percentage points higher than the second-place Bayesian network method (93.5%), and 31 and 41.8 percentage points higher than association rule mining (80.5%) and traditional correlation analysis (69.7%), respectively. In the final product extraction phase, this solution maintained a high accuracy of 99.7%, compared to 90.8% for the Bayesian network method, 82.3% for association rule mining, and 70.2% for traditional correlation analysis. The parameter association recognition accuracy of this technical solution is evaluated relative to a benchmark. The accuracy shown in the chart is relative to a parameter association model under pre-defined standard operating conditions (serving as a 100% benchmark). Accuracy = (number of correctly identified associations / number of associations identified by the benchmark model) × (association strength prediction accuracy / benchmark accuracy) × (association change trend prediction accuracy / benchmark accuracy) × 100%.

[0140] This technical solution can calculate the rate of change of parameter correlation strength in dynamic correlation maps in real time and trigger an adaptive update mechanism for control parameters based on these changes. In contrast, traditional correlation analysis methods performed the worst throughout the production process, with recognition accuracy consistently ranging between 65.8% and 70.5%, failing to effectively capture the complex nonlinear relationships between process parameters. This technical solution has significant advantages in its ability to adaptively identify parameter correlations under dynamic conditions, especially during periods of rapid change in process conditions, where the average recognition accuracy increased by 28.7%.

[0141] In an optional embodiment,

[0142] Establishing a system discrete state equation according to the system state vector, constructing a value function with a discount factor based on the system discrete state equation and the global objective function, obtaining a feedback gain matrix by iteratively optimizing the value function, and constructing a first control strategy according to the feedback gain matrix includes:

[0143] Obtain a system state vector, map the system state vector through an encoder network to obtain a hidden state vector, reconstruct the system state vector through a decoder network, and calculate the reconstruction loss;

[0144] Constructing a causal graph structure based on historical data and calculating mutual information and information entropy between nodes, calculating a causal strength weight based on the mutual information and the information entropy, establishing a system discrete state equation based on the hidden state vector, calculating a causal correction term based on the causal strength weight, and combining the causal correction term with the system discrete state equation to obtain a causally enhanced system discrete state equation;

[0145] An explicit value function and an implicit value function are constructed based on the system discrete state equation and the global objective function, an adaptive weight factor is calculated according to the state prediction error, and the explicit value function and the implicit value function are weightedly combined using the adaptive weight factor to obtain a value function with a discount factor;

[0146] The Bellman equation and the temporal difference algorithm are used to update the explicit value function and the implicit value function respectively. The feedback gain matrix is ​​obtained by iteratively optimizing the value function with the discount factor. The causal effect of the control input is calculated and the feedback gain matrix is ​​corrected according to the causal effect. The corrected feedback gain matrix is ​​combined with the causal correction term to construct a first control strategy.

[0147] A deep encoder-decoder network structure is constructed. The encoder consists of four serially connected encoding modules, each of which contains a fully connected layer, a batch normalization layer, and a nonlinear activation layer. The weight matrix of the fully connected layer is initialized from a normal distribution, and the bias term is initialized to zero. The batch normalization layer calculates the mean and variance of the features, normalizes the features, and introduces learnable scaling and translation parameters. The activation layer uses a rectified linear unit function to introduce nonlinear characteristics. Each layer of the encoder gradually reduces the feature dimension, compressing the original state vector into a hidden state vector. The decoder adopts a mirror-symmetric structure with the encoder, restoring the feature dimension layer by layer. The network parameters are optimized using the backpropagation algorithm, using the mean squared error as the reconstruction loss function, and the stochastic gradient descent method with an adaptive learning rate is used for parameter update.

[0148] When constructing the causal graph structure, historical data is segmented according to fixed time windows. For the data in each time window, the mutual information between each process parameter is calculated. The marginal distribution and joint distribution of the parameters are constructed by the kernel density estimation method, and the mutual information value is calculated based on the distribution function. At the same time, the information entropy of each parameter is calculated, and the histogram method is used to estimate the probability distribution and calculate the entropy value. The mutual information and information entropy are normalized to obtain causal strength weights with a value range between 0 and 1. A discrete state equation is constructed based on the hidden state vector, and the causal strength weight is converted into a correction term of the state transfer matrix. The direction of the correction term is determined by the causal relationship, and the amplitude is controlled by the weight size. The corrected state transfer matrix is ​​substituted into the state equation to obtain a causal enhanced state equation.

[0149] The structures of the explicit and implicit value functions are designed separately. The explicit value function takes the form of a quadratic function, and the coefficient matrix is ​​estimated using the least squares method. The implicit value function is approximated using a multilayer perceptron network consisting of multiple hidden layers, each with a varying number of neurons. The state prediction error is calculated based on historical data, and the error sequence is exponentially smoothed to obtain a smoothed error. An adaptive weight update rule is designed: when the smoothed error falls below a lower threshold, the explicit value function weight increases linearly; when the smoothed error exceeds an upper threshold, the explicit value function weight decays exponentially. The sum of the weight factors always remains equal to 1, ensuring the stability of the combined value function. A discount factor that decays exponentially over time is introduced to construct a value function expression that considers long-term returns.

[0150] The optimization process for the explicit and implicit value functions is performed in parallel. The explicit value function is iterated using the Bellman equation, calculating the current state value estimate and the optimal control variable during each iteration, and updating the value function parameters. The implicit value function is trained using a temporal difference algorithm, calculating the temporal difference error of successive state transition samples and adjusting the network weights through error backpropagation. The optimization processes for both value functions are performed synchronously, sharing the transition samples in the experience replay buffer. Policy iteration is performed on the combined value function, alternating between policy evaluation and policy improvement. The policy evaluation phase fixes the current policy calculated state value, while the policy improvement phase updates the policy parameters based on the latest value estimate.

[0151] The causal effects of control inputs are calculated using perturbation analysis. A small perturbation is applied to each control input, and the response of the state variable is recorded. The slope of the response curve is used to estimate the strength of the causal effect. The causal effect strength is converted into a correction coefficient for the gain matrix, and the feedback gain matrix is ​​corrected element-wise. Finally, the corrected gain matrix is ​​combined with the causal correction term in the state equation to construct a complete control strategy.

[0152] For example, during the production of a batch of detergent composition, a 96-dimensional state vector was collected, containing the original values, target deviations, rates of change, and acceleration information of six process parameters. A four-layer encoder network was constructed, with 96, 64, 48, and 32 neurons in each layer, respectively. Mini-batch stochastic gradient descent optimization was used with a batch size of 128. After 5000 rounds of training, the reconstruction loss decreased from an initial 0.82 to 0.043.

[0153] A causal graph was constructed based on 3000 hours of historical data using a 300-second sliding window. Kernel density estimation revealed a mutual information between temperature and pressure of 0.85, an information entropy of 0.92 for temperature, and 0.88 for pressure. After normalization, the causal strength weights for temperature and pressure were 0.78 and 0.65 for pressure and temperature, respectively. Causal relationships were incorporated into the state equation, and the corresponding elements in the state transition matrix were modified.

[0154] A neural network with a 64-48-32 architecture was constructed as an implicit value function approximator, using a rectified linear unit as the activation function. The prediction error threshold was set between 5% and 10%. When the error was less than 5%, the explicit value function weight was increased by 0.02 per iteration; when the error was greater than 10%, the weight decayed by 0.95. A discount factor of 0.95 was used for policy iteration. After 800 iterations, the value function converged to below 0.001.

[0155] A 1% disturbance was applied to the control input, and analysis of the state response curves revealed that the causal effect of the control input on temperature was 0.82, and on pressure was 0.71. Based on this, the corresponding element values ​​in the feedback gain matrix were modified, increasing the temperature loop gain by 15% and the pressure loop gain by 10%. This control strategy fully accounts for the causal relationship between parameters while maintaining control accuracy.

[0156] In this embodiment, an encoder-decoder network structure is used to reduce the dimension of the system state and reconstruct it. The essential characteristics of the state vector are extracted through multi-layer nonlinear transformations. A causal graph structure and information theory methods are introduced to analyze the causal relationship between process parameters. A hybrid optimization architecture combining explicit and implicit value functions is used. At the same time, an adaptive weight mechanism is used to dynamically adjust the proportion of the two value functions. This ensures the real-time responsiveness of the control strategy and improves the long-term stability of the strategy.

[0157] In existing technologies, intelligent control methods mainly rely on pure data-driven models, focusing only on the correlation between process parameters while ignoring the underlying causal relationships between parameters. This makes it difficult to accurately grasp the interaction mechanism between parameters, resulting in insufficient adaptability and robustness of control strategies. Traditional value function optimization methods often use a single function form, which makes it difficult to balance real-time control and long-term stability when faced with complex process conditions.

[0158] The control strategy of this embodiment can accurately identify and utilize the causal relationship between process parameters, significantly improving the accuracy of control. The adaptive optimization mechanism enables the system to have stronger environmental adaptability and anti-interference ability. The control correction based on causal effect ensures the rationality and reliability of the regulation strategy, providing strong technical support for the intelligent and precise control of the process.

[0159] Figure 4 This is a heat map of the causal strength weights between key process parameters of the washing composition production control method based on multi-parameter intelligent analysis in an embodiment of the present invention. It shows the causal strength weight matrix between six key process parameters, calculated based on 3000 hours of historical data through a 300-second sliding time window. Darker areas indicate stronger causal relationships.

[0160] according to Figure 4 As shown, the causal strength weight of temperature on pressure is 0.78, while the causal strength weight of pressure on temperature is 0.65, indicating an asymmetric bidirectional causal relationship. The causal relationship between flow rate and pH is more significant, with the causal strength of flow rate on pH being 0.71 and the causal strength of pH on flow rate being 0.68. Viscosity and stirring speed exhibit a strong mutual causal relationship, with the weight of viscosity on stirring speed being 0.73 and the weight of stirring speed on viscosity being 0.75.

[0161] Compared to traditional correlation analysis methods (such as the Pearson correlation coefficient), this technical solution accurately captures the nonlinear causal relationships between parameters by combining mutual information and information entropy to calculate causal strength weights, avoiding the false causal inferences that can result from relying solely on statistical correlations. This causal strength weight matrix provides a reliable quantitative basis for subsequent state equation corrections, enabling the system model to more accurately reflect the complex causal interactions between actual process parameters.

[0162] A second aspect of the embodiments of the present invention provides a production control system for a detergent composition based on multi-parameter intelligent analysis, comprising:

[0163] The first unit is used to collect process parameters during the production of the detergent composition, generate a time series data stream of process parameters and perform data preprocessing to obtain standard process parameter data;

[0164] The second unit is used to input the standard process parameter data into a cascaded autoencoder deep learning network, perform hierarchical feature extraction on the standard process parameter data through a multi-scale feature extraction layer, fuse feature information at different scales in combination with a residual connection structure, prioritize features through a self-attention mechanism, obtain a multi-dimensional feature representation, and construct a dynamic association map of process parameters;

[0165] The third unit is used to group the process parameters based on the dynamic association map, generate process parameter groups, obtain parameter change trends through online monitoring, and calculate the coordinated change characteristics based on the dynamic association map;

[0166] The fourth unit is used to construct a hierarchical progressive intelligent controller including a global planning layer and a local execution layer, and to optimize the current control strategy in real time by combining a dynamic programming algorithm, adaptively update the control parameters according to the changes in the dynamic association map, establish parameter adjustment constraints based on the coordinated change characteristics, and apply the constraints to obtain the optimized control strategy;

[0167] The fifth unit is used to convert the optimization and control strategy into control instructions to adjust the process parameters, collect the adjusted process parameters for online learning, continuously optimize the dynamic correlation map, and form a closed-loop optimization control.

[0168] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for controlling the production of a detergent composition based on multi-parameter intelligent analysis, characterized in that: include: Collecting process parameters during the production of the detergent composition, generating a process parameter time series data stream and performing data preprocessing to obtain standard process parameter data; The standard process parameter data is input into a cascaded autoencoder deep learning network. Hierarchical feature extraction is performed on the standard process parameter data through a multi-scale feature extraction layer. Feature information at different scales is fused using a residual connection structure. The features are prioritized through a self-attention mechanism to obtain a multi-dimensional feature representation and construct a dynamic association map of the process parameters, including: A symmetric encoder-decoder structure is used to construct a cascaded autoencoder deep learning network. Standard process parameter data is input into the cascaded autoencoder deep learning network, and a convolution operation, batch normalization processing, and activation function operation are sequentially performed through multiple cascaded encoding units to generate an encoding feature matrix. Performing feature extraction on the encoding feature matrix using multiple convolution kernels of different sizes, using the convolution kernel of the smallest size to extract a reference feature matrix, and using convolution kernels of other sizes to extract a temporal feature matrix group; Performing feature mapping conversion on each time series feature matrix in the time series feature matrix group to obtain a conversion feature matrix group, performing a residual superposition operation on the conversion feature matrices in the conversion feature matrix group and the reference feature matrix to generate a fusion feature matrix; Mapping the fused feature matrix into a query matrix, a key matrix, and a value matrix, calculating the product of the query matrix and the transpose of the key matrix to obtain an attention score matrix, performing normalization on the attention score matrix and then multiplying it with the value matrix to obtain an attention feature matrix; Extracting eigenvectors corresponding to the process parameters in the attention feature matrix, calculating the ratio of the inner product of any two eigenvectors to the product of the module length to obtain the correlation strength between the process parameters, setting the process parameters as vertices, the correlation relationship between the process parameters as edges, and the correlation strength as edge weights, to construct a dynamic correlation map of the process parameters; The process parameters are grouped based on the dynamic association map to generate process parameter groups, and parameter change trends are obtained through online monitoring, and the synergistic change characteristics are calculated in combination with the dynamic association map; Construct a hierarchical progressive intelligent controller comprising a global planning layer and a local execution layer, combine a dynamic programming algorithm to optimize the current control strategy in real time, adaptively update control parameters according to changes in the dynamic association map, establish parameter adjustment constraints based on the collaborative change characteristics, and apply the constraints to obtain an optimized control strategy; The optimization and control strategy is converted into control instructions to adjust the process parameters, the adjusted process parameters are collected for online learning, and the dynamic correlation map is continuously optimized to form a closed-loop optimization control.

2. The method according to claim 1, characterized in that Collect process parameters during the production of the detergent composition, generate a time series data stream of process parameters, and perform data preprocessing to obtain standard process parameter data including: collecting process parameters during the production of the detergent composition, including reactor temperature, reactor pressure, reaction liquid pH, stirring speed, reactor liquid level, and raw material addition rate, and recording them at preset time intervals to form a process parameter time series data stream; Performing data preprocessing on the process parameter time series data stream, identifying and removing abnormal data by setting parameter fluctuation thresholds, filling missing data with an interpolation algorithm, and verifying the validity of the data based on the physical constraints of the process parameters to generate preprocessed process parameter data; The pre-processed process parameter data is standardized, and process parameters of different dimensions are converted into a unified numerical range to obtain standard process parameter data.

3. The method according to claim 1, characterized in that Extracting the eigenvectors corresponding to the process parameters in the attention feature matrix and calculating the ratio of the inner product of any two eigenvectors to the product of the module lengths to obtain the correlation strength between the process parameters includes: Extracting feature vectors corresponding to process parameters from the attention feature matrix, and inputting the feature vectors corresponding to the process parameters into a neural differential dynamics model, wherein the neural differential dynamics model maps the feature vectors corresponding to the process parameters into a state vector through a differential equation, wherein the differential equation is composed of a time derivative of the state vector and a neural network; Using an adaptive Runge-Kutta integrator to adjust the integration step size according to the rate of change of the state vector to solve the differential equation, and obtain the time evolution characteristics of the characteristic vector corresponding to the process parameter; Constructing a nonlinear oscillator group according to the time evolution characteristics, wherein each process parameter corresponds to a nonlinear oscillator, the initial state of the nonlinear oscillator is set according to the characteristic vector corresponding to the process parameter, the frequency of the nonlinear oscillator represents the variation characteristics of the process parameter, and the amplitude of the nonlinear oscillator represents the importance of the process parameter; A coupling network is established between the nonlinear oscillators, the nonlinear oscillators are connected through a nonlinear coupling equation and the synchronization state is analyzed, and the correlation strength between any two process parameters is calculated according to a correlation strength calculation formula. The correlation strength calculation formula combines the synchronization state of the nonlinear oscillators, the similarity of the pre-acquired state trajectories, and the ratio of the inner product of any two eigenvectors to the product of the modulus lengths through a weight factor.

4. The method according to claim 1, wherein The process parameters are grouped based on the dynamic correlation map to generate process parameter groups and obtain parameter change trends through online monitoring. The synergistic change characteristics calculated in combination with the dynamic correlation map include: Determining a correlation threshold according to the connection relationship of the process parameters in the dynamic correlation map and the distribution characteristics of the pre-calculated correlation strength, and classifying the process parameters with correlation strength greater than the correlation threshold into the same process parameter group; Collecting real-time data of each process parameter in the process parameter group through an online sensor, processing the real-time data through a sliding time window, and extracting the time series change trend characteristics of the process parameters; The temporal change trend characteristics are mapped to the process parameter nodes corresponding to the dynamic association map, and the change consistency of each parameter in the process parameter group is calculated based on the connection relationship in the dynamic association map to obtain the collaborative change characteristics of the process parameter group.

5. The method according to claim 1, characterized in that Construct a hierarchical progressive intelligent controller including a global planning layer and a local execution layer, combine a dynamic programming algorithm to optimize the current control strategy in real time, and adaptively update the control parameters according to the changes in the dynamic association map. Establish constraint conditions for parameter adjustment based on the collaborative change characteristics. Apply the constraint conditions to obtain the optimized control strategy, including: Acquire historical data of process parameters, construct a system state vector and a control input vector based on the historical data, and construct a hierarchical progressive intelligent controller comprising a global planning layer and a local execution layer, wherein the global planning layer determines a global objective function comprising a system state deviation term, a control input constraint term, and a system energy consumption term based on the system state vector and the control input vector, and the local execution layer constructs a parameter constraint matrix based on the process parameter connection relationship of the dynamic association map; Establishing a system discrete state equation according to the system state vector, constructing a value function with a discount factor based on the system discrete state equation and the global objective function, obtaining a feedback gain matrix by iteratively optimizing the value function, and constructing a first control strategy according to the feedback gain matrix; Adaptively updating control parameters according to changes in the dynamic association map, calculating an update amount for a control gain matrix, wherein the update amount is related to the gradient of the global objective function, the learning rate, and the parameter constraint matrix, adaptively adjusting the control parameters in the first control strategy according to the update amount to obtain a second control strategy, and establishing constraint conditions for parameter adjustment based on the coordinated change characteristics; The system state sequence in the future time domain is predicted according to the second control strategy and the system discrete state equation, and the optimized control strategy is obtained by applying the constraint conditions.

6. The method according to claim 5, characterized in that Establishing a system discrete state equation according to the system state vector, constructing a value function with a discount factor based on the system discrete state equation and the global objective function, obtaining a feedback gain matrix by iteratively optimizing the value function, and constructing a first control strategy according to the feedback gain matrix includes: Obtain a system state vector, map the system state vector through an encoder network to obtain a hidden state vector, reconstruct the system state vector through a decoder network, and calculate the reconstruction loss; Constructing a causal graph structure based on historical data and calculating mutual information and information entropy between nodes, calculating a causal strength weight based on the mutual information and the information entropy, establishing a system discrete state equation based on the hidden state vector, calculating a causal correction term based on the causal strength weight, and combining the causal correction term with the system discrete state equation to obtain a causally enhanced system discrete state equation; An explicit value function and an implicit value function are constructed based on the system discrete state equation and the global objective function, an adaptive weight factor is calculated according to the state prediction error, and the explicit value function and the implicit value function are weightedly combined using the adaptive weight factor to obtain a value function with a discount factor; The Bellman equation and the temporal difference algorithm are used to update the explicit value function and the implicit value function respectively. The feedback gain matrix is ​​obtained by iteratively optimizing the value function with the discount factor. The causal effect of the control input is calculated and the feedback gain matrix is ​​corrected according to the causal effect. The corrected feedback gain matrix is ​​combined with the causal correction term to construct a first control strategy.

7. A detergent composition production control system based on multi-parameter intelligent analysis, for implementing the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to collect process parameters during the production of the detergent composition, generate a time series data stream of process parameters and perform data preprocessing to obtain standard process parameter data; The second unit is used to input the standard process parameter data into a cascaded autoencoder deep learning network, perform hierarchical feature extraction on the standard process parameter data through a multi-scale feature extraction layer, fuse feature information at different scales in combination with a residual connection structure, prioritize features through a self-attention mechanism, obtain a multi-dimensional feature representation, and construct a dynamic association map of process parameters; The third unit is used to group the process parameters based on the dynamic association map, generate process parameter groups, obtain parameter change trends through online monitoring, and calculate the coordinated change characteristics based on the dynamic association map; The fourth unit is used to construct a hierarchical progressive intelligent controller including a global planning layer and a local execution layer, and to optimize the current control strategy in real time by combining a dynamic programming algorithm, adaptively update the control parameters according to the changes in the dynamic association map, establish parameter adjustment constraints based on the coordinated change characteristics, and apply the constraints to obtain the optimized control strategy; The fifth unit is used to convert the optimization and control strategy into control instructions to adjust the process parameters, collect the adjusted process parameters for online learning, continuously optimize the dynamic correlation map, and form a closed-loop optimization control.

Images