A Soft Measurement Method for Process Indicators Based on Structured Full-Perception Deep Learning

Abstract

This invention relates to the field of industrial process indicator monitoring technology, specifically disclosing a soft measurement method for process indicators based on structured full-perception deep learning. Based on structured full-perception modeling, it is no longer limited to single indicator prediction, but instead, by organizing process data, mechanistic knowledge, and multi-channel dynamic features in a structured manner, it synchronously outputs the probability distribution of operating conditions, the set of mechanistic parameters adapted to the operating conditions, and the inference results of global process indicators. This achieves deep and broad collaborative perception of complex industrial systems, fully explores and utilizes the distribution and mechanistic features contained in the process data, perceives different operating states of the reactor and their corresponding mechanistic parameters, and describes the interaction between variables through a directed topological network, realizing the comprehensive output of core information such as operating conditions, mechanistic parameters, and process indicators.

Description

Technical Field

[0001] This application relates to the field of industrial process indicator monitoring technology, and specifically discloses a soft measurement method for process indicators based on structured full-perception deep learning. Background Technology

[0002] Intelligent manufacturing is a key path for modern industry to achieve high-level production and optimize the overall quality, output, material consumption, environmental protection, and safety. The foundation for intelligent manufacturing lies in establishing accurate mechanistic models. An effective mechanistic model needs to be able to analyze the correlation between variables, clarify the physical constraints between reaction mechanisms, and provide a theoretical basis for intelligent optimization and autonomous control.

[0003] For continuous industrial processes characterized by high integration, strong coupling, and complex spatiotemporal distribution, their operating state is jointly characterized by the dynamics of a large number of process variables, and each typical operating condition corresponds to a specific intrinsic reaction mechanism. Although the mechanistic models constructed by traditional methods have physical interpretability, the mechanistic parameters they contain are usually set as static or piecewise constants, which makes it difficult to reflect the evolution of reaction mechanisms caused by raw material fluctuations, equipment degradation, or control interventions in the actual operation of continuous industrial processes.

[0004] For example, the paper "Hybrid modeling for vehicle lateral dynamics via AGRU with a dual-attention mechanism under limited data" by Jianwei Chen et al., published in *Control Engineering Practice*, October 2024 (journal number: print 0967-0661, electronic version 1873-6939, Vol. 151, article number 106015), constructs a linear lateral dynamics model based on prior knowledge of vehicle dynamics, treating it as an interpretable prior mechanism model. It introduces a gated recurrent unit network with a dual attention mechanism and adaptive initial hidden states to perform sequential modeling and prediction of the residuals between the output of the prior mechanism model and the actual measured values, thereby achieving dynamic recalibration of the vehicle dynamics model. Although its hybrid modeling approach, which combines mechanistic modeling with data-driven residual compensation, significantly improves the prediction accuracy and adaptability of vehicle lateral dynamics under limited data conditions, it is essentially still a series connection of mechanistic model and data model. It only compensates for the output error of the mechanistic model and does not make in-depth corrections to the mechanistic model at the structural or parameter level. Problems such as time-varying mechanistic parameters and mechanistic deviations under complex working conditions remain unresolved.

[0005] Another example is the paper "Fault Diagnosis Model of Digital Twin Technology for CNC Machine-Tool Feed System Based on CA-GRU-PINN" by Xu Zhang et al., published in *The International Journal of Advanced Manufacturing Technology*, October 26, 2025 (print version 0268-3768, electronic version 1433-3015, pages 1-19). This paper proposes a digital twin fault diagnosis method that integrates Convolutional Attention Gated Recurrent Unit (CA-GRU) and Cyber-Physical Network (PINN). It embeds a multibody dynamics model as a physical constraint into PINN and uses the CA-GRU module to extract transient impact features from vibration signals using convolution and attention mechanisms for deep temporal modeling. Although this method still achieves high-precision identification of fault characteristic periods, characteristic frequencies and fault types when fault samples are scarce, it mainly establishes a mapping relationship between "input vibration signal to fault characteristics / fault category" and does not systematically characterize the interaction mode between multiple variables. It cannot clearly describe the mechanism by which the change of one variable affects other variables and thus changes key diagnostic indicators, making it difficult to support higher-level process understanding and global optimization control.

[0006] In summary, the two aforementioned methods for constructing mechanistic models in intelligent manufacturing processes struggle to fully leverage the distributional and mechanistic characteristics inherent in process data. This is particularly true when applied to core production units with multiple cascading reaction processes. Due to the complexity of the internal reaction mechanisms, including the interweaving of main and side reactions, forward and reverse reactions, time-varying operating conditions and mechanistic parameters, and deep coupling between high-frequency and low-frequency variables, highly chaotic dynamic systems are formed. Existing mechanistic model construction methods fail to perceive different reactor operating states and corresponding mechanistic parameters, cannot characterize the interaction between input and output variables, and cannot comprehensively output core information such as operating conditions, mechanistic parameters, and process indicators.

[0007] This invention provides a soft measurement method for process parameters based on structured full-sensory deep learning to solve the above-mentioned problems. Summary of the Invention

[0008] The purpose of this invention is to provide a soft measurement method for process indicators based on structured full-perception deep learning. By organizing process data, mechanistic knowledge, and multi-channel dynamic features in a structured manner, it synchronously outputs the probability distribution of the operating conditions, the set of mechanistic parameters adapted to the operating conditions, and the inference results of the process indicators throughout the entire process, thereby achieving deep and broad collaborative perception of complex industrial systems.

[0009] To achieve the above objectives, the basic solution of this invention provides a soft measurement method for process indicators based on structured full-sensory deep learning, comprising the following steps:

[0010] Step A1: Based on the knowledge of the field to which the process belongs, construct local mechanism models of each reaction process in the reactor, and obtain the relationship between the input variables and mechanism parameters and the output variables under the influence of fitting error;

[0011] Step A2: Based on the dependency relationship between the output and input variables of the cascaded reaction process, organize all local mechanism models into a directed topological network that includes the variable set and variable dependency relationship in the cascade order, and calculate the values ​​of all variables;

[0012] Step A3: Obtain the historical state variables of the reactor, divide them into multiple channels to obtain multi-channel state variables, and extract multi-channel dynamic features by performing dynamic spatiotemporal feature extraction on each channel.

[0013] Step A4: After enhancing the multi-channel dynamic features through cross-attention, project them onto a unified probability space and perform adaptive weighted fusion to obtain the working condition probability vector representing the global operating state;

[0014] Step A5: The working condition probability vector is weighted and estimated through a pre-constructed mechanism parameter dictionary matrix to obtain a dynamic mechanism parameter set;

[0015] Step A6: Iterate the above steps until the end. Then, at each time step, execute steps A3 to A5 in sequence to obtain the probability distribution of operating conditions and the set of mechanism parameters at continuous time points. Substitute the set of mechanism parameters into the directed topology network according to the time points to obtain the global process indicators at all time points.

[0016] Furthermore, in step A1, the reactor is one of a cascaded reactor in the same continuous process.

[0017] Furthermore, in step A2, the set of variables includes a set of input variables and a set of output variables.

[0018] Furthermore, in step A2, when the output variable of the local mechanism model constructed for the i-th reaction process... It is the input variable of the local mechanism model constructed for the j-th reaction process. Then, in the directed topology network, there exists a variable from the output variable. To input variables The directed edge.

[0019] Furthermore, in step A3, the obtained historical state variables are divided to obtain the following multi-channel state variables: entry state variables, reaction state variables, and exit state variables. Dynamic spatiotemporal features are extracted through a trend attention graph convolutional recurrent network to obtain multi-channel dynamic features, which include entry features, reaction features, and exit features.

[0020] Furthermore, in step A3, the trend attention graph convolutional recurrent network includes multiple cascaded parameter-shared TAGCRN blocks. Each parameter-shared TAGCRN block is responsible for processing features at one time step, and each parameter-shared TAGCRN block is equipped with a trend attention graph convolutional module for assisting in capturing time-varying spatial correlations of state variables.

[0021] Furthermore, in step A4, the multi-channel dynamic features are stacked column-wise to obtain a fused feature matrix.

[0022] Furthermore, in step A5, each row in the mechanism parameter dictionary matrix corresponds to a working condition and its matching mechanism parameter group.

[0023] The principle and effect of this basic scheme are as follows:

[0024] 1. Compared with existing technologies, this invention establishes local mechanism models for multiple reaction processes in complex industrial reactors and organizes them into a directed topological network according to variable dependencies. Then, it uses a trend attention graph convolutional recurrent network to extract reactor operation features from multivariate time series data to form a structured framework. This structured framework is no longer limited to single indicator prediction, but through the structured organization of process data, mechanism knowledge, and multi-channel dynamic features, it synchronously outputs the probability distribution of operating conditions, the set of mechanism parameters adapted to the operating conditions, and the inference results of the entire process indicators, thereby achieving a deep and broad collaborative perception of complex industrial systems.

[0025] 2. Compared with the prior art, the directed topology network established by this invention not only retains the local model's ability to accurately model specific units, but also reveals the multi-level interaction relationship between variables through global structural constraints. Ultimately, it achieves a structured description and detailed characterization of the complex mechanism inside the reactor, solving the problem that a single local mechanism model can only describe the mechanism behavior of a limited area inside the reactor, and it is difficult to fully characterize the complex action paths and reaction details between variables.

[0026] 3. Compared with existing technologies, this invention organizes multiple local mechanism models with clear physical meanings into a directed topological network according to the reaction mechanism. It achieves accurate perception of spatiotemporal dynamics through trend attention graph convolutional recurrent network, achieves adaptive binding of mechanism parameters through multi-channel working condition identification, and achieves reliable output of global state through mechanism constraint reasoning. It constructs a closed-loop comprehensive perception path in continuous process industrial modeling, which can provide an interpretable and generalizable modeling foundation for intelligent industrial agents, autonomous control and cross-unit collaboration. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 A flowchart of the process index soft measurement method based on structured full-sensory deep learning proposed in an embodiment of this application is shown;

[0029] Figure 2 A schematic diagram of the directed topology network proposed in an embodiment of this application is shown;

[0030] Figure 3 A schematic diagram of the multi-channel dynamic feature extraction process proposed in an embodiment of this application is shown;

[0031] Figure 4 A schematic diagram of the working condition perception process based on multi-channel dynamic feature fusion proposed in an embodiment of this application is shown;

[0032] Figure 5 A schematic diagram of the cobalt removal purification process in the wet zinc smelting method proposed in this application embodiment is shown;

[0033] Figure 6 The following diagrams illustrate the multi-index prediction performance of the full-sensing model proposed in this application, where Figure (a) shows the predicted value of ORP and Figure (b) shows the predicted value of the outlet cobalt ion concentration.

[0034] Figure 7 The probability distribution diagram of different samples belonging to different working conditions proposed in the embodiments of this application is shown;

[0035] Figure 8 The diagram shows the variation curves of mechanistic parameters for different samples proposed in the embodiments of this application;

[0036] Figure 9The illustration shows a schematic diagram of different outputs for the same input under different operating conditions according to the embodiments of this application. Figure (a) is a schematic diagram of different outputs for the first sample pair under different operating conditions, and Figure (b) is a schematic diagram of different outputs for the second sample pair under different operating conditions. Detailed Implementation

[0037] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0038] A soft measurement method for process parameters based on structured full-sensory deep learning is implemented, for example... Figure 1 As shown, it includes the following steps:

[0039] Step A1: Based on knowledge of the relevant technological domain, construct local mechanistic models of each reaction process within the reactor to obtain the relationships between input variables and mechanistic parameters and output variables under the influence of fitting error. Here, the reactor is one of a cascaded reactor in the same continuous process.

[0040] Specifically, such as Figure 2 As shown, in a complex reaction system, n reaction processes occur inside the reactor. This discussion focuses on the i-th reaction process within the reactor. The constructed local mechanism model is as follows:

[0041] ;

[0042] In the formula, This represents the mechanistic term constructed from explicit physical laws. This represents the input variable for the i-th reaction process. Represents the mechanistic parameters characterizing the fundamental features of the i-th reaction process. This represents the output variable of the i-th reaction process. , Represents the set of input variables. , Represents the set of mechanism parameters. This represents the prediction error of the local mechanism model, serving as an unknown term in the data-driven fitting process. , where n represents the total number of reaction processes.

[0043] In this embodiment, the mechanistic term ensures the physical rationality of the local mechanistic model. Simultaneously, the prediction error of the mechanistic model supplements the unknown factors at the data-driven level, enabling the local mechanistic model to capture complex nonlinear behaviors while maintaining a clear structure. By combining the physical constraints of the mechanistic term with the flexibility of the data-driven method through the constructed local mechanistic model, the interpretability of the mechanistic model and its ability to adapt to the dynamic changes of complex industrial systems are improved.

[0044] Step A2: Based on the dependencies between output and input variables in the cascaded reaction process, organize all local mechanism models into a directed topological network that includes variable sets and variable dependencies in a cascaded order, and calculate all variable values. The variable sets include the input variable set and the output variable set.

[0045] Specifically, such as Figure 2 As shown, the dependency between reactor output variables and input variables is based on cascaded reaction processes, which seeks the dependency between the output variables of the previous reaction process and the input variables of the next reaction process in two cascaded reaction processes.

[0046] For example, in a continuous cobalt removal purification process, the various reaction processes in the reactor are cascaded. The output variables of a local mechanism model built based on the previous reaction process can be directly used as the input variables of a local mechanism model built for the next reaction process. Based on the above dependency, such as Figure 2 As shown, all local mechanism models are organized into the following directed topology network according to the arrangement of cascade reactions:

[0047] ;

[0048] In the formula, This represents a set of variables, which includes a set of input variables and a set of output variables. This indicates the dependency between output and input variables in a cascaded reaction process.

[0049] When the output variable of the local mechanism model constructed for the i-th reaction process It is the input variable of the local mechanism model constructed for the j-th reaction process. Then, in the directed topology network, there exists a variable from the output variable. To input variables The directed edges, such as Figure 2 In the middle, the reaction process Output variables of the constructed local mechanism model These are the reaction processes. and reaction process Input variables of the constructed local mechanism model and input variables Then an output variable is constructed. The same as the input variables and input variables The directed edge.

[0050] Furthermore, after obtaining the directed topology network, the local mechanism model constructed based on the reaction process with known input variables in the cascade reaction process is started. The above local mechanism model is calculated in parallel to obtain its corresponding output variables. When these output variables have directed edges in the directed topology network, these output variables are added to the input variables of the next reactor connected by the directed edge. The input of the remaining local mechanism model is iteratively supplemented, and the above process is repeated until all variable values ​​are obtained.

[0051] Step A3: Obtain the historical state variables of the reactor, divide them into multi-channel state variables, and extract multi-channel dynamic features by performing dynamic spatiotemporal feature extraction. Specifically, a trend attention graph convolutional recurrent network is used to extract dynamic spatiotemporal features to obtain multi-channel dynamic features.

[0052] In this embodiment, the obtained historical state variables are divided into the following multi-channel state variables: entry state variables, reaction state variables, and exit state variables. The entry state variables are then processed using a trend attention graph convolutional recurrent network. Reaction state variables Export state variables Dynamic spatiotemporal feature extraction was performed on the three types of variables to obtain entry features. Reaction characteristics and export characteristics ,in, Indicates the time step. , where m represents time m.

[0053] like Figure 3 As shown, the Trend Attention Graph Convolutional Recurrent Network (TAGCRN network) consists of multiple cascaded parameter-shared TAGCRN blocks. Each parameter-shared TAGCRN block serves as the computational unit of the Trend Attention Graph Convolutional Recurrent Network, and each parameter-shared TAGCRN block is responsible for processing the features at one time step. Each parameter-shared TAGCRN block is equipped with a trend attention graph convolution module to help capture the time-varying spatial correlation of the input state variables.

[0054] Each trend attention graph convolutional module is from the previous time step: Hidden state of time Taking the current node state and the current node embedding matrix as input, and based on the Data Adaptive Graph Generation (DAGG) module, the system autonomously learns the optimal variable association pattern using the trainable node embedding matrix. The expression is as follows:

[0055] ;

[0056] In the formula, This represents the node embedding matrix at the current time. This is the matrix transpose operator.

[0057] Here, the node at the current moment refers to one of the entry state variable, reaction state variable, or exit state variable at the current moment, and the current moment refers to the current time t.

[0058] Furthermore, for the node state at the current moment, its first-order difference is calculated using the following formula. The first-order difference of the node state at the current moment is used to characterize the magnitude and direction of the change in the state variable, and the first-order difference of the node state at the current moment is used as the trend feature to dynamically simulate the spatiotemporal dependency:

[0059] ;

[0060] In the formula, Indicates the current The node state at any given moment. This represents all nodes in the graph. express The node state at any given moment.

[0061] The node state and trend feature are concatenated to obtain the state-trend concatenation matrix at the current moment. Based on the state-trend concatenation matrix, the dynamic attention weight matrix is ​​calculated using the following formula. :

[0062] ;

[0063] In the formula, This represents the state-trend cascade matrix at the current moment. , express The feature dimension. Introducing a trend term can perceive the magnitude of changes in state variables and more accurately capture local correlation patterns under transient operating conditions.

[0064] Furthermore, by adaptively fusing the variable association patterns with the attention weight matrix, a time-varying graph convolution matrix is ​​obtained. This matrix preserves steady-state topological constraints while adaptively adjusting connection weights under transient conditions, thereby enhancing the robustness of modeling non-stationary processes. Specifically, the time-varying graph convolution matrix... The expression is as follows:

[0065] ;

[0066] In the formula, Represents the learnable fusion coefficient. .

[0067] Furthermore, in the parameter-sharing TAGCRN block, the trend attention graph convolution operation is combined with a gated recursive unit (GRU) to perform the following calculation process in sequence, in order to effectively capture the temporal dependencies between state variables:

[0068] ;

[0069] ;

[0070] ;

[0071] ;

[0072] In the formula, To update the door, To reset the door, Let represent the candidate hidden state at time t. express The hidden state at all times These represent the weights of each transformation. These represent the deviations of each transformation. It represents the Hadamardi (or Hadama) stack.

[0073] In the Trend Attention Graph Convolutional Recurrent Network, the first parameter shares the hidden state input to the TAGCRN block as an all-zero vector.

[0074] The hidden state of the TAGCRN block output is shared by the last parameter in the trend attention graph convolutional recurrent network. To obtain the dynamic spatiotemporal characteristics of the corresponding channels and state variables, multi-channel dynamic features are obtained.

[0075] Step A4: After enhancing the multi-channel dynamic features through cross-attention, project them onto a unified probability space and perform adaptive weighted fusion to obtain a working condition probability vector representing the global operating state.

[0076] like Figure 4 As shown, the multi-channel dynamic features obtained in step A3—inlet features, reaction features, and outlet features—are stacked column-wise to obtain a fused feature matrix, expressed as follows:

[0077] .

[0078] Subsequently, the attention weights between channels are calculated using the established cross-attention module to enhance feature interactions between channels. In this embodiment, the cross-attention module achieves information flow between features through similarity weighting, allowing each channel to focus on important information from other dimensions and uncover potential correlations between features. The expression is as follows:

[0079] ;

[0080] In the formula, This represents the enhanced fusion feature matrix. , This indicates the enhanced entry feature. This indicates the enhanced reaction characteristics. This indicates the enhanced export characteristics. This represents the scaling factor.

[0081] The enhanced fusion feature matrix is ​​input into a multilayer perceptron (MLP) with shared weights, and mapped to the probability space through a softmax activation function to obtain the working condition probability distribution of each channel, as shown in the following expression:

[0082] ;

[0083] In the formula, Represents the probability distribution of operating conditions. .

[0084] Furthermore, to improve the robustness and decision confidence of the model, this embodiment constructs a probabilistic consistency weighting module based on Jensen Shannon divergence (JSD) to calculate the weights of each channel, as shown in the following expression:

[0085] ;

[0086] In the formula, Represents the average probability distribution. , Indicates the channel index. The temperature parameter represents the sharpness of the weight distribution. , which represents the JSD distribution coefficient.

[0087] Then, the probability distributions of each channel are summed according to their weights to obtain the final operating condition probability vector. This enables condition perception based on multi-channel dynamic feature fusion, and generates condition probability vectors. The expression is as follows:

[0088] .

[0089] Step A5: The working condition probability vector is weighted and estimated through a pre-constructed mechanism parameter dictionary matrix to obtain a dynamic mechanism parameter set.

[0090] Using the operating condition probability vector as an intermediate carrier connecting data patterns and physical states, each operating condition corresponds to a set of mechanism parameters. Through a weighted estimation method, the operating condition probability vector is mapped to a trainable mechanism parameter dictionary matrix to obtain dynamically updated mechanism parameters.

[0091] Based on the operating condition probability vector, the mechanism parameter identification process is described as follows:

[0092] .

[0093] In the formula, This represents a vector of mechanistic parameters identified at time t. It is a trainable mechanism parameter dictionary matrix, where the mechanism parameter dictionary matrix is ​​a learnable matrix, and each row corresponds to a working condition and its matching mechanism parameter set. After probability weighting by the working condition probability vector, the unique mechanism parameter set under the corresponding working condition probability distribution is obtained.

[0094] Step A6: Iterate the above steps until the end. Then, at each time step, execute steps A3 to A5 in sequence to obtain the probability distribution of operating conditions and the set of mechanism parameters at continuous time points. Substitute the set of mechanism parameters into the directed topology network according to the time points to obtain the global process indicators at all time points.

[0095] Furthermore, in this embodiment, a multi-objective loss function is constructed to perform gradient backpropagation during iteration steps A1 to A5, and the iteration ends when the maximum number of iterations is reached. The expression for the multi-objective loss function is as follows:

[0096] ;

[0097] In the formula, This indicates that in order to balance the dynamic weights of local fitting, The dynamic weights represent the losses in the balancing inference process, where, , , Indicates the current training round. This indicates the maximum number of training rounds.

[0098] In the formula, The local fitting loss function is used to ensure that each local mechanistic model accurately fits its corresponding observation data. The expression is as follows:

[0099] ;

[0100] in, This represents the actual output variable. Indicates the predicted output variable. This indicates the total number of reaction processes.

[0101] In the formula, The loss function based on physical constraints is expressed as follows:

[0102] ;

[0103] Among them, the projection operator parameters Project back to a physically feasible domain.

[0104] In the formula, The inference loss function based on the output variables of the directed topological network is used to limit the potential accumulation of errors during inference. The expression is as follows:

[0105] ;

[0106] in, Represents the actual observed value. This represents the predicted value obtained from network inference.

[0107] In the formula, The regularization loss function is used to limit the excessive contribution of data-driven components to the overall model. The expression is as follows:

[0108] .

[0109] Meanwhile, during the iteration process, dynamic mechanism parameters are substituted into the aforementioned directed topology network. The output variables of each local mechanism model are integrated through a differentiable computation graph. Collaborative reasoning is performed under physical constraints such as conservation equations and transfer equations. The probability distribution of operating conditions, the set of mechanism parameters adapted to the operating conditions, and the global process indicators are output synchronously to complete the soft measurement of process indicators.

[0110] In this embodiment, the cobalt removal process in the continuous cobalt removal purification step of the hydrometallurgical zinc smelting process is taken as an example. This cobalt removal process involves multiple cascaded cobalt removal reactors, which complete the cobalt removal purification process. The core task of the cobalt removal purification process is to remove impurity ions from the zinc sulfate solution through a redox reaction. Zinc powder is used as the key reducing agent in the purification process, and its concentration in the zinc sulfate solution directly determines the driving force of the redox reaction, making it a key operational variable for regulating the reaction rate. Therefore, controlling the concentration of zinc powder is crucial during the purification process. The concentration of zinc powder in the zinc sulfate solution determines the redox environment within the reactor, which is typically characterized by the redox potential (ORP) in industrial production.

[0111] In summary, the impurity removal process in zinc purification can be described as follows: the addition of zinc powder adjusts its cumulative concentration in the solution, thereby affecting the redox potential. This potential determines the impurity removal reaction rate, thus affecting the concentration of impurity ions in the outlet solution. Simultaneously, the consumption of zinc powder during impurity removal further affects its cumulative concentration in the solution, forming a dynamic closed-loop process. In this process, input variables include the zinc powder addition rate, the zinc powder concentration in the reactor inlet solution, the impurity particle concentration at the reactor inlet, and the system flow rate. Output variables include the redox potential, the zinc powder concentration in the reactor outlet solution, and the impurity particle concentration at the reactor outlet. Based on steps A1 to A6, the final global process parameters obtained include the outlet cobalt ion concentration and the redox potential.

[0112] The following is an example of the application of the process index soft measurement method based on structured full-perception deep learning provided by this invention in the purification process of a zinc smelter. Figure 5 As shown, the purification process consists of five cascaded cobalt removal reactors, including reactors 1 through 5. Taking reactor 1 as an example, seven models are used to predict the following key process parameters: outlet cobalt ion concentration and redox potential (ORP).

[0113] The seven models are: Gated Recurrent Unit (GRU), Adaptive Graph Convolutional Recurrent Network (AGCRN), Diffusion Convolutional Recurrent Neural Network (DCRNN), Transformer, Informer, Spatiotemporal Transformer Network (STTN), and Structured Full Perception Model (SCPM) constructed based on the method provided in this invention.

[0114] Mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination ( The accuracy of the model is measured by the calculation of ), as shown below:

[0115] ;

[0116] ;

[0117] ;

[0118] ;

[0119] In the formula, This represents the total number of samples in the test dataset. Indicates the first The true value of each sample Indicates the first The predicted value for each sample, The mean of the true values ​​of the sample. This represents the mean of the predicted values ​​for the sample, where the test dataset includes actual production data collected over seven consecutive months.

[0120] The performance of the seven models across four metrics is shown in Table 1, and the prediction results of the structured full-perception model are shown in the figure below. Figure 6 As shown, the predicted value of ORP is as follows: Figure 6 As shown in Figure (a), the predicted values ​​of the outlet cobalt ion (Co) concentration are as follows: Figure 6 As shown in Figure (b).

[0121] Table 1. Calculation Results of Indicators

[0122]

[0123] According to Table 1 and Figure 6 It can be seen that the structured full-sensing model proposed in this invention has achieved optimal performance in soft measurement tasks of all key process indicators.

[0124] However, focusing solely on the predictive power of process indicators is insufficient to fully reflect the core innovation of the structured full-perception model in its modeling mechanism. The key breakthrough of the structured full-perception model established in this invention lies in its ability to simultaneously perceive operating conditions and mechanistic parameters that are difficult to measure directly, and to endow the prediction results with physical interpretability, such as… Figure 7 and Figure 8 The probability distributions and perceived mechanism parameter curves for different samples belonging to three different operating conditions: operating condition 1, operating condition 2, and operating condition 3 are displayed respectively. A total of 500 samples are included. Figure 7 In this model, 0.0 to 1.0 represent probabilities. These indicators are difficult to measure directly. However, the structured full-sensing model organically couples the identification of mechanistic parameters with the prediction of process indicators. While completing high-precision soft measurement, it can also reflect the changes in the internal state of the system in real time.

[0125] at the same time, Figure 9 The data presents the average values ​​of key variables for two representative sample pairs over two hours, illustrating the phenomenon that the same input can produce different outputs under different operating conditions. For example... Figure 9 As shown in Figure (a), taking the first sample pair as the object, the first sample pair includes samples 74 and 294. The first sample pair is highly similar in terms of feed flow rate and zinc powder addition amount, but their ORPs are significantly different. Figure 9As shown in Figure (b), taking the second sample pair as the object, which includes samples 109 and 189, the second sample pair showed significantly different outlet impurity concentrations despite similar feed flow rates, ORP, and inlet impurity concentrations. This phenomenon typically reflects the "multi-condition coexistence" characteristic in industrial processes, that is, the same or similar operational inputs may lead to drastically different output results due to differences in internal states.

[0126] Furthermore, Table 2 summarizes the average mechanistic parameters obtained by the structured full-sensory model from the perception of four samples over two hours.

[0127] For the first sample pair (sample 74 and sample 294), the zinc ion concentrations inferred by the structured full-sensory model established in this invention are 0.1769 g / L and 0.1704 g / L, respectively. The ratios were 945.2 and 959.5, and the inference results suggest that this ratio was not the primary cause of the ORP difference. On the contrary, and The significant differences reveal a fundamental difference in their electrochemical environments: Sample 74 exhibits stronger reducing properties, with zinc powder contributing more significantly. For the second sample pair (Samples 109 and 189), the reaction rates were calculated to be 0.86 and 0.74, respectively. Figure 9 Analysis of the results shown indicates that the pre-exponential factor and cathode transfer coefficient This is a key factor leading to the rate difference, which is beneficial to the impurity removal efficiency of sample 109.

[0128] Table 2. Average values ​​of mechanism parameters

[0129]

[0130] In summary, the structured full-perception model possesses a deep understanding of the intrinsic mechanisms of complex industrial processes. Compared to traditional purely data-driven models, the proposed solution not only achieves high-precision predictions but also reveals the physical causes behind typical operating conditions such as "same input, different outputs," providing a new paradigm for industrial process monitoring, diagnosis, and optimization that combines accuracy and interpretability.

[0131] Based on the entire process from steps A1 to A6, this invention first establishes a local mechanism model for multiple cascaded reactors in a continuous process and organizes them into a directed topological network according to variable dependencies. Then, using a trend attention graph convolutional recurrent network, it extracts reactor operating features from multivariate time series data to form a structured framework to determine the probability of belonging to different operating conditions. Based on this, it drives the adaptive update of mechanism parameters, perceives the current operating state, and dynamically adjusts its own behavior. At the same time, it gives the output of the data model a clear physical semantics.

[0132] Meanwhile, this invention introduces a data adaptation mechanism on the basis of the trend attention graph convolutional recurrent network, dynamically adjusting the connection weights between nodes according to real-time running data, so as to accurately capture the effective interaction between variables even under transient perturbations. In addition, TAGCRN embeds a trend attention mechanism, focusing on the consistency of variable evolution direction, effectively enhancing the sensitivity to key dynamic patterns, capturing complex spatiotemporal joint dynamics with high fidelity, and providing a robust feature foundation for working condition identification.

[0133] Furthermore, this invention designs a multi-channel dynamic feature fusion-based condition perception module, which extracts process features from multiple dynamic perspectives in parallel. It achieves information alignment between channels through a cross-attention mechanism and introduces a distribution consistency weighting strategy based on Jensen-Shannon divergence to dynamically fuse the outputs of each channel, generating a probability distribution vector of the current operating conditions. This distribution vector is transformed into mechanistic parameters through weighted mapping, realizing an interpretable transformation from "data pattern" to "mechanism state." This directly solves the model mismatch problem caused by static parameters in traditional mechanistic models, while simultaneously endowing the data-driven model with physical semantics.

[0134] Finally, to ensure the physical consistency of global inference, this invention constructs a cooperative inference network subject to directed topological constraints. This network formalizes the mechanistic equations into a differentiable computational graph and performs joint operations with the original variables and parameters to directly output process indices.

[0135] During the training process, this invention employs a hierarchical loss function to optimize the accuracy of index prediction, constrain the mechanism residuals, and introduces a working condition consistency regularization term to prevent parameter updates from deviating from a reasonable physical range.

[0136] In summary, this invention organizes multiple local mechanistic models with clear physical meanings into a directed topological network according to reaction mechanisms. It achieves precise spatiotemporal dynamic perception through trend attention graph convolutional recurrent networks, adaptive binding of mechanistic parameters through multi-channel operating condition identification, and reliable output of the global state through mechanistic constraint inference. This constructs a closed-loop integrated perception pathway in continuous process industrial modeling, providing an interpretable and generalizable modeling foundation for intelligent industrial agents, autonomous control, and cross-unit collaboration. Simultaneously, the established directed topological network retains the accurate modeling capability of local models for specific units while revealing multi-level interaction relationships between variables through global structural constraints. Ultimately, it achieves a structured description and detailed characterization of the complex mechanisms within the reactor, solving the problem that a single local mechanistic model can only describe the mechanistic behavior of a limited region within the reactor, making it difficult to comprehensively characterize the complex interaction paths and reaction details between variables.

[0137] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any indirect modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A soft measurement method for process parameters based on structured full-sensory deep learning, characterized in that, Includes the following steps: Step A1: Based on the knowledge of the field to which the process belongs, construct local mechanism models of each reaction process in the reactor, and obtain the relationship between the input variables and mechanism parameters and the output variables under the influence of fitting error; In step A1, the reactor is one of a cascaded reactor in the same continuous process; Step A2: Based on the dependency relationship between the output and input variables of the cascaded reaction process, organize all local mechanism models into a directed topological network that includes the variable set and variable dependency relationship in the cascade order, and calculate the values ​​of all variables; In step A2, the variable set includes an input variable set and an output variable set; When the output variable of the local mechanism model constructed for the i-th reaction process It is the input variable of the local mechanism model constructed for the j-th reaction process. Then, in the directed topology network, there exists a variable from the output variable. To input variables The directed edges; Step A3: Obtain the historical state variables of the reactor, divide them into multiple channels to obtain multi-channel state variables, and extract multi-channel dynamic features by performing dynamic spatiotemporal feature extraction on each channel. The obtained historical state variables are divided into the following multi-channel state variables: inlet state variables, reaction state variables, and outlet state variables; Step A4: After enhancing the multi-channel dynamic features through cross-attention, project them onto a unified probability space and perform adaptive weighted fusion to obtain the working condition probability vector representing the global operating state; Step A5: The working condition probability vector is weighted and estimated through a pre-constructed mechanism parameter dictionary matrix to obtain a dynamic mechanism parameter set; In step A5, each row of the mechanism parameter dictionary matrix corresponds to a working condition and its matching mechanism parameter set. After probability weighting by the working condition probability vector, the unique mechanism parameter set under the corresponding working condition probability distribution is obtained. Step A6: Iterate the above steps until the end. Then, at each time step, execute steps A3 to A5 in sequence to obtain the probability distribution of operating conditions and the set of mechanism parameters at continuous time points. Substitute the set of mechanism parameters into the directed topology network according to the time points to obtain the global process indicators at all time points.

2. The method for soft measurement of process indicators based on structured full-sensory deep learning according to claim 1, characterized in that, In step A3, dynamic spatiotemporal features are extracted using a trend attention graph convolutional recurrent network to obtain multi-channel dynamic features, which include ingress features, reaction features, and exit features. In step A3, the trend attention graph convolutional recurrent network includes multiple cascaded parameter-shared TAGCRN blocks. Each parameter-shared TAGCRN block is responsible for processing features at one time step. Each parameter-shared TAGCRN block is equipped with a trend attention graph convolutional module for assisting in capturing the time-varying spatial correlation of state variables. Each trend attention graph convolutional module is from the previous time step: Hidden state of time Taking the current node state and the current node embedding matrix as input, the system, based on the data adaptive graph generation module, autonomously learns the optimal variable association pattern using the trainable node embedding matrix. The expression is as follows: ； In the formula, The node embedding matrix represents the node embedding at the current time. This is the matrix transpose operator; Here, the node at the current moment refers to one of the entry state variable, reaction state variable, and exit state variable at the current moment, and the current moment refers to the current time t. Furthermore, for the node state at the current moment, its first-order difference is calculated using the following formula. The first-order difference of the node state at the current moment is used as the trend feature: ； In the formula, Indicates the current The node state at any given moment. This represents all nodes in the graph. express The node state at any given moment; The node state and trend feature are concatenated to obtain the state-trend concatenation matrix at the current moment. Based on the state-trend concatenation matrix, the dynamic attention weight matrix is ​​calculated using the following formula. : ； In the formula, This represents the state-trend cascade matrix at the current moment. express Feature dimensions; Furthermore, by adaptively fusing the variable association patterns with the attention weight matrix, a time-varying graph convolution matrix is ​​obtained. The expression is as follows: ； In the formula, Represents the learnable fusion coefficient. ; Furthermore, the following calculation process is performed sequentially within the parameter-sharing TAGCRN block to capture the temporal dependencies between state variables: ； ； ； ； In the formula, To update the door, To reset the door, Let represent the candidate hidden state at time t. express The hidden state at all times These represent the weights of each transformation. These represent the deviations of each transformation. It represents the Hadamardi (or Hadama) stack.

3. The method for soft measurement of process indicators based on structured full-sensory deep learning according to claim 1, characterized in that, In step A4, the multi-channel dynamic features are stacked column-wise to obtain a fused feature matrix.

Images