Knowledge guide-based soft measurement method and system for evaporation quality index of alumina

Abstract

The application provides an alumina evaporation quality index soft measurement method and system based on knowledge guidance, the method first collects multivariate time series data of the evaporation process, and divides the multivariate time series data into three variable subsets of material, energy and disturbance; then a multiscale feature extraction module is constructed, different scale sliding windows are used to capture local dynamic characteristics and perform fusion; then a parallel processing path corresponding to the three types of variables is built, each path takes LSTM as the core, and the foregoing fused features are introduced as inputs; then specific gate guide factors are designed in combination with the process mechanism of each subsystem to finely regulate and control the memory updating mechanism of LSTM in the corresponding path, so that the hidden state with stronger representation is obtained; finally, the outputs of each path are fused for consistency, and the precise prediction of the key quality index is realized; the application effectively fuses data driving and process mechanism knowledge, and can ensure that the prediction model has high precision and good physical interpretability under complex and variable working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial process control and soft sensor technology, and particularly to a soft sensor method and system for alumina evaporation quality index based on knowledge guidance. Background Art

[0002] In the alumina production process, the evaporation process is the core link affecting product quality, production efficiency and energy consumption. Among them, the real-time and accurate monitoring of the discharge concentration is crucial for realizing process optimization control. However, due to the harsh field environment, it is difficult for physical sensors to measure such key quality indicators stably online for a long time, and the maintenance cost is high. Therefore, the soft sensor technology that constructs a mathematical model based on easily measurable process parameters (such as temperature, pressure, flow, etc.) to estimate quality variables has been widely used in the industrial field.

[0003] However, the alumina evaporation process is a typical complex industrial process, showing significant non-linear coupling characteristics and multi-time scale dynamic characteristics. The material flow, energy transfer and external disturbances are intertwined, bringing great challenges to traditional soft sensor modeling. Although existing deep learning methods (such as RNN, LSTM) have made progress in processing time series data, they are still insufficient when applied to this specific process. Existing models usually process data on a single time scale, and it is difficult to capture the dynamic dependence of process variables under different short, medium and long-term horizons at the same time, and cannot adapt to the inherent heterogeneous time evolution characteristics in the industrial process. In addition, these models are mostly "black box" structures and lack explicit combination with industrial domain knowledge. Although some models can decompose the time scale to judge "when" the process changes, they cannot explain the internal physical mechanism of "why" it changes.

[0004] Specifically manifested as: in terms of material dynamics, traditional LSTM units are not sensitive enough to variable changes, ignoring the indication effect of concentration difference changes on system state switching, and it is difficult to handle the common buffer lag phenomenon in industrial systems; in terms of energy, the absolute value of energy variables fluctuates greatly and has strong hysteresis, and directly using the absolute value for modeling is difficult to truly reflect the driving efficiency of heat input; in terms of disturbance response, in the face of sudden random disturbances such as equipment startup and shutdown, the traditional inference mechanism based on historical sequences responds slowly and cannot make timely adjustments at the moment of disturbance.

[0005] To sum up, the existing technology is difficult to balance high prediction accuracy and good physical consistency under changing complex working conditions, and there is an urgent need for a new soft sensor method that can integrate multi-scale feature extraction and domain knowledge guidance. Summary of the Invention

[0006] To address the aforementioned shortcomings of existing technologies, the present invention aims to provide a knowledge-guided soft measurement method and system for alumina evaporation quality indicators. This method and system are intended to solve the technical problem that existing soft measurement models struggle to integrate multi-scale features and domain knowledge, resulting in low prediction accuracy and poor physical consistency under complex operating conditions.

[0007] To achieve the above objectives, in a first aspect, the present invention provides a knowledge-guided soft measurement method for alumina evaporation quality indicators, the steps of which include:

[0008] S1. Obtain multivariate time series data of the alumina evaporation process, and divide the multivariate time series data into a material variable subset, an energy variable subset, and a disturbance variable subset based on the physical mechanism;

[0009] S2. Construct a multi-scale feature extraction module, which constructs sliding windows of different time scales for the multivariate time series data and extracts local evolution features within each scale. Then, the features of each scale are spliced ​​and fused to obtain a unified multi-scale time feature representation.

[0010] S3. Construct parallel material paths, energy paths, and disturbance paths corresponding to each subset of variables. Each path uses a Long Short-Term Memory (LSTM) network as its core unit and uses the unified multi-scale time feature representation as the common input of each path.

[0011] S4. Calculate specific gating guidance factors based on the mechanistic characteristics of the material, energy, and disturbance subsystems respectively, and use the gating guidance factors to adjust the memory update process of the LSTM unit under the corresponding path to obtain the hidden state of each path.

[0012] S5. Consistently integrate the hidden states of material, energy, and disturbance path outputs to generate predicted values ​​of key quality indicators for the alumina evaporation process.

[0013] As a further improvement to the above scheme, in step S1, the multivariate time series data is represented as follows: ,in Indicates the time step of a time series. Indicates the dimension of the variable;

[0014] Variables directly involved in reaction and transport are divided into a subset of material variables;

[0015] The variables that provide the heat source to drive evaporation are divided into a subset of energy variables;

[0016] Variables related to equipment start-up and shutdown and environmental fluctuations are divided into a subset of disturbance variables.

[0017] As a further improvement to the above scheme, the specific implementation steps of step S2 include:

[0018] S21. Set a sliding window with three time scales. For each process variable Extract the corresponding sliding time window slices to form the sliding time window slice tensor. ,in ;in, The sliding window lengths are for three different time scales, corresponding to sequence analysis at short, medium, and long time scales, respectively. Let be the observed value of the i-th process variable at time t, such as temperature, pressure, flow rate, and other industrial process variables. Let be the sliding slice tensor of the i-th process variable at the s-th scale at time t, which is composed of historical observations within the window;

[0019] S22. Input the sliding window sequences at each scale into a one-dimensional convolutional layer, extract local evolution features by sensing the sequence pattern within a specific window size through the convolutional kernel, and enhance the feature expression through a nonlinear activation function.

[0020] S23. Concatenate and fuse the extracted feature vectors at different scales into a multi-scale feature vector. This is to preserve the dynamic complementarity between different scales.

[0021] As a further improvement to the above scheme, in step S22, the feature representation is enhanced by a nonlinear activation function, the specific expression formula of which is shown below:

[0022] ;

[0023] in, For the i-th process variable at the s-th scale, the single-scale feature extracted at time t; To modify the linear unit activation function, nonlinearity is introduced to enhance the model's feature representation ability and avoid gradient vanishing; For the one-dimensional convolution operation corresponding to the s-th scale, the convolution kernel slides on the sliding window to perceive the sequence pattern and local dependencies within a specific time window. For the i-th process variable from The time series segment up to t is used as the input to the one-dimensional convolutional layer.

[0024] The multi-scale feature vector is specifically shown in the following formula:

[0025] ;

[0026] in, Let be the final feature vector of the i-th process variable at time t after multi-scale feature splicing and fusion. To concatenate single-scale features extracted from three different time scales along the channel dimension, the independence of each scale feature is preserved, while providing complementary dynamic input to the downstream network.

[0027] As a further improvement to the above scheme, in step S4, the material path memory update process is guided by a concentration differential gating mechanism, specifically including:

[0028] Calculate the concentration difference term between the current time and historical time. ;

[0029] Concentration gating factor is constructed based on the concentration difference component. When the concentration changes drastically, the adjustment model increases the proportion of absorption of the current input; when the concentration changes steadily, it increases the proportion of retention of historical states.

[0030] The state update equation of the material path LSTM is modified using the concentration gating factor.

[0031] As a further improvement to the above scheme, the concentration gating factor The calculation formula is as follows:

[0032] ;

[0033] in, The sigmoid activation function maps the output to the [0,1] interval, enabling the "on / off" adjustment of the gating. This is the material path sensitivity coefficient.

[0034] As a further improvement to the above scheme, the state update equation of the material path LSTM is specifically shown in the following equation:

[0035] ;

[0036] in, The corrected material path LSTM cell state at time t represents the long-term memory unit after incorporating concentration difference gating information; Forget gate, controls the cell state at the previous moment. The proportion of information that needs to be "forgotten" in the middle; This represents the cell state at the previous time step t-1, i.e., the long-term memory retained by the model in previous time steps; This is the input gate, which controls the current candidate cell state. The proportion of information that needs to be updated to the current cell state; The candidate cell state is calculated from the current input and the hidden state at the previous time step, representing the candidate memory information to be updated. The Hadamard product represents the operation of multiplying elements at corresponding positions.

[0037] As a further improvement to the above scheme, in step S4, the memory update process of the energy path is guided by a local heat input ratio gating mechanism, specifically including:

[0038] Calculate the ratio of the current heat input to the previous time window moving average heat input. ;in The moving average heat input is the preceding time window. This represents the material heat input at the current time t.

[0039] Based on the ratio Constructing hot input gating factors This is used to distinguish between the energy ramp-up period and the stable period, and to adjust the memory update intensity of the LSTM accordingly;

[0040] Using the heat input gating factor The state update equation of the energy path LSTM is modified.

[0041] As a further improvement to the above scheme, the formula for calculating the heat input gating factor is as follows:

[0042] ;

[0043] in, The sigmoid activation function maps the output to the [0,1] interval, enabling the "on / off" adjustment of the gating. This is a sensitivity coefficient for energy paths, used to adjust the heat input ratio. Gating factor The intensity of the impact.

[0044] As a further improvement to the above solution, Thermoinfluence input gating factors are used to dynamically control the proportion of new information incorporated into the cellular state, and the specific regulatory mechanism is as follows:

[0045] when At that time, that is, when the heat input was significantly higher than the historical average, Gating allows new inputs to update the memory state at a relatively high rate;

[0046] when At that time, that is, when the heat input is close to the historical average, The gating system integrates new and old information in a moderate proportion;

[0047] when At that time, that is, when the heat input was significantly lower than the historical average, Gating inhibits the incorporation of new information.

[0048] As a further improvement to the above scheme, the state update equation of the energy path LSTM is specifically shown in the following equation:

[0049] ;

[0050] The modified energy path LSTM cell state at time t represents the long-term memory unit after incorporating thermal input gating information, used to characterize the dynamic evolution of materials under the influence of thermal input. Forget gate, controls the cell state at the previous moment. The proportion of information that needs to be "forgotten" in the middle; This represents the cell state at the previous time step t-1, i.e., the long-term memory retained by the model in previous time steps; This is the input gate, which controls the current candidate cell state. The proportion of information that needs to be updated to the current cell state; The candidate cell state is calculated from the current input and the hidden state at the previous time step, representing the candidate memory information to be updated. The Hadamard product represents the operation of multiplying elements at corresponding positions.

[0051] As a further improvement to the above scheme, in step S4, the memory update process of the perturbation path is guided by a perturbation mutation detection gating mechanism, and the specific implementation steps include:

[0052] Calculate the degree of perturbation change The numerator is the first-order difference magnitude of the perturbation variable vector, used to capture the magnitude of the perturbation change; the denominator is the normalization factor of the perturbation magnitude, used to suppress the influence of abnormal peaks.

[0053] Constructing perturbation gating factors This is used to dynamically adjust the model response based on the intensity of the perturbation mutation.

[0054] The state update equation of the perturbation path LSTM is corrected using the perturbation gating factor:

[0055] ;

[0056] in, The vector of perturbation variables consists of the observed values ​​of perturbation variables such as feed temperature, vacuum degree, and condensate level at time t, representing the perturbation state at the current time. The observed value of the perturbation variable vector at time t-1 represents the perturbation state at the previous time. Extremely small positive numbers, such as 10 -6 This prevents the denominator from being zero and avoids calculation errors. This represents the normalized abrupt change in perturbation, reflecting the severity of the current perturbation relative to its historical state.

[0057] The Sigmoid activation function maps the output to the [0,1] interval, enabling continuous adjustment of the gating. This is the perturbation path sensitivity coefficient, used to adjust the degree of perturbation abruptness. Gating factor The intensity of the impact; As the perturbation gating factor, when the perturbation abruptness is high, The value tends towards 1, enabling the model to respond quickly to disturbances; when the disturbance is stable, The value tends towards 0, suppressing interference from invalid perturbation information;

[0058] The cell state of the corrected perturbation path LSTM at time t represents the long-term memory unit after incorporating perturbation gating information, used to characterize the impact of perturbation on the evaporation process.

[0059] The forget gate of the perturbation path LSTM controls the cell state at the previous time step. The proportion of information that needs to be "forgotten" in the middle;

[0060] This represents the cell state at the previous time step t-1, i.e., the perturbation-related long-term memory retained by the model in previous time steps;

[0061] This serves as the input gate for the perturbation path LSTM, controlling the current candidate cell state. The proportion of information that needs to be updated to the current cell state;

[0062] The candidate cell states of the perturbation path LSTM are calculated from the current input and the hidden state at the previous time step, representing the perturbation-related candidate memory information to be updated.

[0063] The Hadamard product is an element-wise multiplication used to achieve precise adjustment of the corresponding feature information by the gating signal.

[0064] As a further improvement to the above scheme, the specific implementation steps of step S5 include:

[0065] The hidden state vectors output from the three paths are concatenated in terms of dimensions.

[0066] The concatenated vector is input into a multilayer perceptron with at least two hidden layers for adaptive fusion;

[0067] The final predicted value of the mother liquor concentration in the alumina output is output through a linear regression layer.

[0068] Secondly, the present invention also provides a knowledge-guided soft measurement system for alumina evaporation quality indicators, comprising:

[0069] The data acquisition and segmentation module is used to acquire multivariate data of the alumina evaporation process and segment it into material, energy and disturbance variable subsets based on the physical mechanism.

[0070] The multi-scale feature extraction module is used to extract multi-scale dynamic features of process data through sliding windows of different lengths and one-dimensional convolutional layers.

[0071] The knowledge-guided multi-path prediction module includes parallel material, energy, and perturbation LSTM paths, with each path's state update adjusted by a built-in physical mechanism gating factor.

[0072] The consistency fusion output module is used to adaptively weight and fuse the features of each path, and output the real-time monitoring results of the quality indicators.

[0073] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the knowledge-guided soft measurement method for alumina evaporation quality indicators as described in the first aspect.

[0074] Fourthly, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the program to implement the steps of the knowledge-guided soft measurement method for alumina evaporation quality indicators described in the first aspect.

[0075] Because the present invention adopts the above technical solutions, the beneficial effects of the present invention are as follows:

[0076] This invention provides a knowledge-guided soft measurement method for alumina evaporation quality indicators. Addressing the technical problems of existing soft measurement models for alumina evaporation quality indicators, such as difficulty in integrating multi-scale process characteristics with evaporation mechanism knowledge, low prediction accuracy under complex unsteady-state conditions, poor model physical consistency, and insufficient interpretability, this invention achieves beneficial effects in the following aspects through a complete technical design that incorporates pre-guided mechanistic knowledge, multi-scale feature extraction, and path-specific mechanism constraint modeling:

[0077] (1) Effectively improves the model's ability to express the dynamic features of the process at multiple scales. Specifically, this invention constructs a multi-scale sliding window and a multi-scale one-dimensional convolution feature extraction structure. Targeting the heterogeneous time evolution characteristics of different variables in the alumina evaporation process, it can simultaneously extract the short-term fluctuation features of parameters such as feed flow rate and steam pressure in the process data, as well as the long-term trend features of parameters such as liquid concentration and boiling point rise. This effectively solves the problem that a single time window cannot simultaneously cover fast-changing disturbances and slow-changing trends, and is difficult to adapt to the dynamic characteristics of industrial processes at multiple time scales. Through multi-scale feature splicing and fusion, it provides the downstream prediction network with dynamic feature inputs with richer contextual information, laying a solid feature foundation for the model to accurately fit the dynamic laws of the process.

[0078] (2) It enhances the physical interpretability and operational transparency of the model. Specifically, this invention does not adopt the common pure data-driven "black box" LSTM structure. Instead, based on the physical mechanism of the alumina evaporation process, it first divides the process variables into subsets according to material, energy, and disturbance attributes, and then explicitly transforms the mechanism laws of each subsystem into corresponding gating guidance strategies, embedding them into the core state update link of the corresponding path LSTM network. This design gives the model's memory update mechanism a clear physical meaning. The model's state retention and update behavior directly correspond to actual physical conditions such as changes in material concentration, fluctuations in heat input, and external disturbances. This breaks the black box defect of traditional soft measurement model gating logic without physical constraints, effectively improves the traceability and physical consistency of the model prediction process, and avoids abnormal results that violate the basic physical laws of the evaporation process.

[0079] (3) This method significantly improves the prediction accuracy and robustness of the model under unsteady conditions. Specifically, this invention addresses the strong nonlinear coupling of multiple variables in the alumina evaporation process by constructing independent parallel modeling paths for the three core subsystems of materials, energy, and disturbances. Based on the mechanistic characteristics of each subsystem, targeted gating guidance is implemented, effectively decoupling the nonlinear coupling interference between different physical property variables. This allows each path to accurately fit the independent dynamic evolution law of its corresponding subsystem. Furthermore, through the consistent fusion of the hidden states of multiple paths, the coupling influence of the three subsystems on the evaporation quality indicators is comprehensively considered. Compared with traditional soft measurement methods without mechanistic guidance, this method demonstrates higher prediction accuracy and stronger adaptability to various operating conditions when dealing with complex unsteady conditions commonly encountered in alumina evaporation industrial settings, such as fluctuations in feed composition, sudden changes in steam pressure, and large-scale load adjustments. It can reliably meet the soft measurement requirements of full-condition operation in industrial settings.

[0080] This invention enables the deep integration of mechanistic knowledge in the field of alumina evaporation process with data-driven soft measurement models. While improving the model's prediction accuracy and robustness, it also ensures the model's physical interpretability and consistency, providing reliable soft measurement support for real-time monitoring and process control of alumina evaporation process quality indicators. Attached Figure Description

[0081] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0082] Figure 1 This is a schematic diagram of an alumina six-effect four-flash structure evaporation process equipment disclosed in this invention;

[0083] Figure 2 This is a flowchart illustrating a knowledge-guided soft measurement method for alumina evaporation quality indicators disclosed in this invention.

[0084] Figure 3 This is a schematic diagram of the multi-scale, multi-path LSTM model structure for a knowledge-guided soft measurement method for alumina evaporation quality indicators disclosed in this invention.

[0085] Figure 4 This is a schematic diagram illustrating the predicted results of the outlet mother liquor quality index in the evaporation process of a knowledge-guided soft measurement method for alumina evaporation quality index disclosed in this invention.

[0086] The realization of the objective, functional characteristics and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0087] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0088] It should be noted that the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0089] Example 1

[0090] See Figures 1-4 To address the technical problems in existing soft measurement techniques for alumina evaporation process quality indicators, which struggle to integrate multi-scale dynamic characteristics of the process with evaporation mechanism knowledge, resulting in low prediction accuracy, poor physical consistency, and insufficient interpretability under complex unsteady conditions, this invention provides a knowledge-guided soft measurement method for alumina evaporation quality indicators. The specific implementation steps are as follows:

[0091] S1. Process data acquisition and mechanism-oriented variable classification:

[0092] First, multivariate time series data of the evaporation process are acquired through the distributed control system (DCS) of the alumina evaporation workshop. The sampling period is set to 1 minute / time, and the data covers the operating parameters of the entire alumina evaporation process. The collected raw data is preprocessed to eliminate the interference of data noise on modeling.

[0093] Based on the physical mechanism of the alumina evaporation process, these variables are divided into three independent subsets: a subset of material variables reflecting material inventory and transformation, a subset of energy variables reflecting heat supply and exchange, and a subset of disturbance variables reflecting sudden impacts such as equipment start-up and shutdown.

[0094] Specifically, the subset of material variables includes parameters directly related to material transport and concentration evolution, such as feed mother liquor flow rate, circulating mother liquor density, and raw liquid temperature;

[0095] The subset of energy variables includes parameters directly related to steam input and heat transfer, such as new steam pressure, new steam flow rate, and thermodynamic parameters such as the temperature and temperature difference at each stage from the first effect to the last effect.

[0096] The subset of disturbance variables includes uncontrollable external fluctuation parameters, such as vacuum fluctuations, flash evaporator liquid level, and circulating pump current frequency, which reflect the environmental and equipment operating status.

[0097] This step, through mechanism-oriented variable partitioning, aligns with the core operational laws of the alumina evaporation process at the data source level, avoiding the feature interference problem between different physical property variables caused by indiscriminate mixed inputs. At the same time, it lays the classification foundation for subsequent mechanism-guided modeling of different paths, ensuring the physical consistency of the model from the source.

[0098] S2. Construction of Multi-Scale Feature Extraction Module and Feature Fusion:

[0099] A multi-scale feature extraction module is constructed to address the characteristics of industrial process data, which contains both high-frequency noise and long-term trends. This step involves creating multiple sliding windows of varying lengths from the original time series data. For example, short-time windows capture instantaneous fluctuations, while long-time windows extract slowly evolving trend features. Local evolutionary features are extracted within each window, and these feature vectors from different scales are then concatenated and fused to form a unified multi-scale temporal feature representation. This design overcomes the limitations of single-time-scale analysis, enabling the model to simultaneously perceive short-term fluctuations and long-term trends, providing richer contextual information for downstream forecasting.

[0100] S3. Parallel LSTM path construction based on mechanism matching:

[0101] Parallel paths corresponding one-to-one with the three variable subsets in S1 are constructed, namely the material path, the energy path, and the disturbance path. All three paths use the Long Short-Term Memory (LSTM) network as the core computing unit, and the unified multi-scale time feature representation output by S2 is used as the common input of the three paths.

[0102] By employing a parallel LSTM path that matches the mechanism of the variable subset, we can perform dedicated modeling for the independent dynamic characteristics of the three subsystems of material, energy, and disturbance. This effectively decouples the interference caused by the multivariate nonlinear coupling in the alumina evaporation process, avoids the mutual masking of features of different subsystems in the hybrid modeling, and provides an independent modeling carrier for the accurate embedding of subsequent mechanism gating.

[0103] S4. Mechanism-guided gating factor calculation and LSTM state update adjustment:

[0104] Based on the mechanistic characteristics of the material, energy, and perturbation subsystems, specific gating factors are calculated for each path. These gating factors are then used to regulate the memory update process of the LSTM units within that path, ultimately outputting the hidden states of each path. For example, in the material path, factors can be designed based on differential information about concentration changes to sense system state transitions; in the energy path, factors can be designed based on the cumulative effect of heat input to measure the actual driving force of energy. These gating factors are applied to the LSTM units within the corresponding paths, dynamically regulating the update and forgetting processes of their internal cell states. In this way, the physical mechanisms are explicitly transformed into control signals for the network, giving the model's memory update process a clear physical meaning and effectively solving the problem of poor physical interpretability in traditional "black box" models.

[0105] S5. Multi-path hidden state consistency fusion and quality index prediction:

[0106] The hidden states output from three parallel material, energy, and disturbance paths are concatenated along the feature dimension to achieve consistent fusion of multi-path information. The fused feature vector is then input into a fully connected layer and processed through linear mapping and activation functions to generate predicted values ​​for key quality indicators in the alumina evaporation process. These key quality indicators can be the caustic ratio of the evaporated mother liquor. Because each path has effectively handled specific types of disturbances and dynamics through physical guidance, this fusion result exhibits higher prediction accuracy and stronger robustness than traditional methods under complex and variable industrial conditions.

[0107] In a preferred embodiment, the implementation steps of the multi-scale feature extraction module construction and feature fusion in step S2 are as follows:

[0108] S21. Set three time scales for sliding windows, and extract corresponding sliding time window slices for each process variable to form a sliding time window slice tensor. In this embodiment, the process data of the alumina evaporation process is collected through the workshop DCS system, with a sampling period of 1 minute / time. Based on this, three different time scales of sliding window lengths are set: short window... This corresponds to a 5-minute duration; medium window This corresponds to a 20-minute duration; long window Corresponding to a duration of 60 minutes, the three windows represent short, medium, and long timescales for series analysis, respectively, and are matched to the parameter characteristics of the alumina evaporation process: short windows are suitable for rapidly changing disturbance parameters such as live steam pressure and feed flow rate; medium windows are suitable for parameters with medium rates of change such as evaporator liquid level and feed temperature; and long windows are suitable for slowly changing trend parameters such as feed concentration and feed solids content. For each process variable... Slices of the corresponding sliding time windows are extracted according to the three time scales mentioned above, forming sliding time window slice tensors: For the s-th scale, s∈{1,2,3}, corresponding to short, medium, and long time scales respectively, the sliding slice tensor of the i-th process variable at time t at the s-th scale is: ,in Let be the observed value of the i-th process variable at time t, including industrial process variables such as temperature, pressure, flow rate, liquid level, and concentration during the alumina evaporation process. The sliding slice tensor is composed of continuous historical observations within the corresponding window arranged in chronological order. By setting a sliding window that matches the characteristics of the process parameters at different scales, the entire dynamic range of transient fluctuations and long-term trends during the alumina evaporation process can be covered simultaneously. This solves the problem that a single time window cannot simultaneously capture the characteristics of rapid disturbances and slow evolutions, laying a data foundation for the subsequent accurate extraction of multi-scale features.

[0109] S22. The sliding window sequences at each scale are input into a one-dimensional convolutional layer. The convolutional kernel perceives the sequence patterns within a specific window size to extract local evolutionary features, and a nonlinear activation function is used to enhance feature representation. In this embodiment, independent one-dimensional convolutional layers are configured for the three time scales. The kernel size of each one-dimensional convolutional layer is adapted to the corresponding window length. The kernel sizes for short, medium, and long windows are set to 3, 5, and 7, respectively, and the convolution stride is set to 1. Boundary padding uses equal-length padding to ensure that the length of the output sequence is consistent with the length of the input sequence. The sliding time window slice tensors obtained in S21 are input into the corresponding one-dimensional convolutional layers. Through the sliding operation of the convolutional kernel on the time series, the local correlation patterns and evolutionary rules of the sequence within the corresponding window size are perceived, and the initial local evolutionary features of each scale are extracted. Then, each initial local evolutionary feature is input into the ReLU nonlinear activation function. Through nonlinear transformation, invalid information is filtered out and the expressive power of effective features is enhanced, finally obtaining the single-scale feature vectors corresponding to each scale. By configuring dedicated one-dimensional convolutional layers for different scales, specific sequence features within the corresponding time scale can be extracted, avoiding the problem of insufficient adaptability of a single convolutional kernel to extract features of sequences of different lengths. At the same time, the non-linear activation function further enhances the model's ability to capture the dynamic characteristics of the process.

[0110] S23. The extracted feature vectors at different scales are concatenated and fused into a multi-scale feature vector to preserve the dynamic complementarity between different scales. In this embodiment, the single-scale feature vectors at short, medium, and long scales extracted in S22 are concatenated end-to-end along the feature dimension to form a unified multi-scale feature vector, which serves as the common input to the subsequent parallel path LSTM network. This feature concatenation fusion method fully preserves the independence and dynamic complementarity of features at different time scales, avoiding mutual masking of information at different scales during feature fusion and providing the downstream prediction network with more contextually rich dynamic feature inputs, effectively improving the model's adaptability to fluctuations under different operating conditions.

[0111] In a preferred embodiment, in step S22, single-scale feature extraction is performed by combining one-dimensional convolution operations with a nonlinear activation function. The specific feature extraction expression is as follows:

[0112] ;

[0113] In the formula, For the i-th process variable from The time series segment up to t is used as the input to a one-dimensional convolutional layer; where The sliding window length for the s-th scale is set in step S21. s∈{1,2,3} corresponds to the short, medium and long time scales respectively. By dividing the input sequence segments according to the scale, the input range of the convolution operation is guaranteed to be completely matched with the dynamic characteristics of the corresponding time scale. This avoids the local sequence pattern extraction deviation caused by cross-scale input and lays the input foundation for accurate extraction of features at the corresponding scale.

[0114] For the one-dimensional convolution operation corresponding to the s-th scale, in this embodiment, independent one-dimensional convolutional layers are configured for each of the three time scales. The parameters of each convolutional layer are adapted to the corresponding window length: the kernel size for the short scale is 3, the kernel size for the medium scale is 5, and the kernel size for the long scale is 7. The stride of all convolutional layers is set to 1, and equal-length padding is used to ensure that the length of the output sequence is consistent with the length of the input sequence. By sliding the convolutional kernel on the corresponding sliding window, the local correlation pattern and temporal dependency of the sequence within a specific time window can be accurately perceived, and the specific sequence features of the corresponding scale can be extracted in a targeted manner: the short-scale convolution captures the transient fluctuation features of rapidly changing parameters such as live steam pressure and feed flow rate, while the long-scale convolution captures the long-term evolution trend of slowly changing parameters such as feed liquid concentration and feed solid content. This effectively solves the problem that a single convolutional kernel is not adaptable to sequences of different lengths and cannot fully cover the multi-scale dynamic characteristics of the process.

[0115] To correct the linear unit activation function, in this embodiment, the initial features output from the one-dimensional convolution operation are input into the activation function for nonlinear transformation. The transformation rule is to set negative features to zero and retain positive features. This activation function introduces nonlinearity into the feature extraction process, effectively fitting the strong nonlinear coupling relationship between multiple variables in the alumina evaporation process and enhancing the model's feature representation ability. Furthermore, it effectively alleviates the gradient vanishing problem during deep neural network training, improving the model's training efficiency and convergence stability.

[0116] The single-scale feature extracted at time t for the i-th process variable at the s-th scale serves as the basic unit for subsequent multi-scale feature fusion.

[0117] Furthermore, after extracting single-scale features at each scale, multi-scale feature fusion is achieved through feature concatenation to obtain the final multi-scale feature vector, as shown in the following expression:

[0118] ;

[0119] In this embodiment, in the formula Let be the final feature vector of the i-th process variable after multi-scale feature splicing and fusion at time t; To integrate single-scale features extracted from three different time scales, the features are concatenated along the feature channel dimension. This concatenation and fusion method fully preserves the independence of features at each scale, avoiding the problem of weak-scale features being masked by strong-scale features, which can occur with weighted fusion or global pooling fusion. Simultaneously, it integrates complementary dynamic features from short, medium, and long scales into a unified feature representation, providing more comprehensive dynamic input with contextual information for the downstream parallel path LSTM network. This effectively improves the model's adaptability to fluctuations across the entire alumina evaporation process, laying a solid feature foundation for accurate prediction of quality indicators.

[0120] In a preferred embodiment, in step S4, the material path memory update process is guided by a concentration differential gating mechanism, and the specific implementation steps include:

[0121] First, calculate the concentration difference term between the current time and historical time. In this embodiment, This represents the measured concentration of the feed solution at time t during the alumina evaporation process. Let ΔC(t) be the measured concentration of the feed liquid at the previous sampling time t-1, and ΔC(t) be the change in feed liquid concentration between adjacent sampling times. ΔC(t) directly characterizes the operational fluctuation state of the material subsystem in the alumina evaporation process: changes in feed liquid concentration directly correspond to changes in feed flow rate, discharge flow rate, and evaporation efficiency, and are core state parameters of the material conservation mechanism in the alumina evaporation process. Using the concentration difference term as the driving input of the gating mechanism fundamentally guides the state update process of the LSTM model based on knowledge of material conservation, avoiding the black-box defect of purely data-driven gating logic lacking physical meaning.

[0122] Secondly, a concentration gating factor is constructed based on the concentration difference component. The gating factor enables the mechanistic adjustment of model state updates: when the concentration changes drastically, the adjustment model increases the absorption ratio of the current input to quickly adapt to the fluctuation of the operating conditions; when the concentration changes steadily, the retention ratio of historical states is increased to match the inertial characteristics of the continuous evaporation process of the material.

[0123] The concentration gating factor The calculation formula is as follows:

[0124] ;

[0125] in, The sigmoid activation function maps the output to the [0,1] interval, enabling the "on / off" adjustment of the gating. The material path sensitivity coefficient can be calibrated according to the actual operating characteristics of the alumina evaporator. In this embodiment, the value range is 0.8~1.5, which is used to adjust the influence of concentration change on the gate factor and adapt to the dynamic characteristics of different specifications of evaporators. This is the absolute value of the concentration difference term, used to quantify the severity of fluctuations in the feed solution concentration. Through the above formula, the output of the concentration gating factor is directly linked to the feed solution concentration fluctuation state: when the feed solution concentration changes drastically, As the value increases, The concentration tends towards 0; when the concentration of the feed solution changes steadily, The value approaches 0. The gating logic tends towards 1, giving the model's gating logic a clear physical meaning and breaking the limitation of traditional LSTM gating without physical constraints.

[0126] Finally, the state update equation of the material path LSTM is corrected using the concentration gating factor. The specific state update equation of the material path LSTM is as follows:

[0127] ;

[0128] The physical meaning and implementation method of each parameter in the formula are as follows:

[0129] The cell state of the corrected material path LSTM at time t represents the long-term memory unit after incorporating concentration difference gating information, used to characterize the long-term dynamic characteristics of the material subsystem.

[0130] The output of the forget gate of the LSTM for the material path is used to control the cell state at the previous time step. The proportion of information that needs to be retained or forgotten;

[0131] This represents the cell state at the previous time step t-1, which is the long-term memory of the historical operating state of the corresponding material subsystem retained by the model at previous time steps.

[0132] The input gate output of the material path LSTM is used to control the current candidate cell state. The proportion of information that needs to be updated to the current cell state;

[0133] The candidate cell states of the material path LSTM are calculated from the multi-scale feature input at the current time and the hidden state at the previous time. They represent the candidate memory information corresponding to the current material working condition that needs to be updated to the cell state.

[0134] The Hadamard product is an operation that multiplies corresponding elements of a matrix. It is used to achieve precise element-by-element adjustment of corresponding feature information by gating signals.

[0135] In this embodiment, concentration-gated factors are used. By weighting and correcting the new information update branch of the LSTM, state update control that perfectly conforms to the material conservation mechanism is achieved: when the concentration of the feed solution changes drastically and the operating conditions fluctuate greatly, The model tends towards 1, weakening the memory interference of historical stable states through gating adjustment, increasing the absorption ratio of the current input, and quickly adapting to changes in unsteady operating conditions, effectively improving the prediction accuracy when the operating conditions fluctuate; when the feed concentration changes steadily and the operating conditions are stable, The model approaches 0.5, which smoothly integrates current input information while retaining historical valid states. It conforms to the material transport inertia characteristics of the continuous evaporation process of alumina, ensuring the continuity and physical consistency of the model's prediction results and avoiding abnormal results that do not conform to the material balance law.

[0136] Through the above concentration difference gating mechanism, this embodiment explicitly embeds the material conservation mechanism of the alumina evaporation process into the core state update stage of the LSTM network, so that the model's memory update process has a clear physical meaning. While improving the prediction accuracy of the model under unsteady conditions, it can also enhance the physical interpretability and operational reliability of the model.

[0137] In a preferred embodiment, in step S4, the memory update process of the energy path is guided by a local heat input ratio gating mechanism, and the specific implementation steps include:

[0138] First, calculate the ratio of the current heat input to the previous time window moving average heat input. ,in The moving average heat input is the preceding time window. The material heat input at the current time t is calculated from the current live steam flow rate and live steam enthalpy, directly representing the real-time energy input level of the alumina evaporation process;

[0139] In this embodiment, the length of the preceding sliding window is set to 12 sampling steps, obtained by averaging the historical thermal input data within the window, and used to characterize the recent steady-state operating level of the energy subsystem. As a local heat input ratio, it can directly quantify the degree of deviation of the current heat input from the historical steady-state level, accurately reflect the operating status of the energy subsystem, and use this as the driving input of the gating mechanism. It fundamentally realizes the guidance of the knowledge of energy conservation in the alumina evaporation process on the state update process of the LSTM model, avoiding the black box defect of pure data-driven gating logic without physical meaning.

[0140] Secondly, based on the ratio Constructing hot input gating factors This is used to distinguish between the ramp-up, steady-state, and load-reduction phases of the energy subsystem, and to adjust the memory update intensity of the LSTM accordingly. The heat input gating factor... The calculation formula is as follows:

[0141] ;

[0142] In the formula, σ is the sigmoid activation function, which can smoothly map the calculation result to the [0,1] interval, realize the continuous "on / off" adjustment of the gating, adapt to the characteristics of continuous operation of industrial processes, and avoid the model output jump caused by hard-on gating; The sensitivity coefficient for the energy path can be calibrated based on the actual thermal inertia characteristics of the alumina evaporator. In this embodiment, the value ranges from 1.0 to 2.0, and it is used to adjust the heat input ratio. Gating factor The influence intensity is adapted to the dynamic characteristics of evaporation devices of different specifications.

[0143] The heat input gating factor is used to dynamically control the proportion of new information incorporated into the cell state, and the specific regulatory mechanism is as follows: when... At this time, that is, when the heat input is significantly higher than the historical average, corresponding to the condition of increased load on the evaporator and a significant increase in live steam input, Gating allows new inputs to update the memory state at a relatively large rate, enabling the model to quickly capture the impact of sudden changes in thermal input on the evaporation process; when At this time, that is, when the heat input is close to the historical average, corresponding to the steady-state operation of the evaporator, at this time... The gating system integrates new and old information in a moderate proportion, taking into account both historical steady-state memory and minor fluctuations in current operating conditions; when At this time, that is, when the heat input is significantly lower than the historical average, corresponding to the evaporator load reduction and a significant decrease in live steam input, Gating suppresses the incorporation of new information, preserves historical steady-state memory, and aligns with the high inertia of heat transfer during alumina evaporation. Through this adjustment mechanism, the model's memory update intensity is perfectly matched to the actual physical conditions of the energy subsystem, overcoming the shortcomings of traditional LSTMs, such as lag in response to sudden changes in thermal input and large output fluctuations during steady-state operation.

[0144] Finally, using the aforementioned heat input gating factor The state update equation of the energy path LSTM is modified, and the specific state update equation of the energy path LSTM is shown in the following equation:

[0145] ;

[0146] The physical meaning and implementation method of each parameter in the formula are as follows:

[0147] The modified energy path LSTM cell state at time t represents the long-term memory unit after incorporating heat input gating information, used to characterize the dynamic evolution process of materials under the influence of heat input, including the core states of energy subsystems such as changes in liquid temperature and fluctuations in evaporation efficiency.

[0148] The forget gate output of the energy path LSTM controls the cell state at the previous time step. The proportion of information that needs to be "forgotten" in the middle;

[0149] The cell state at the previous time step t-1 is the long-term memory of the corresponding energy subsystem's historical operating state, which is retained by the model at previous time steps.

[0150] The input gate output of the energy path LSTM controls the current candidate cell state. The proportion of information that needs to be updated to the current cell state;

[0151] The candidate cell states of the energy path LSTM are calculated from the multi-scale feature input at the current time and the hidden state at the previous time. They represent the candidate memory information corresponding to the current heat input condition that needs to be updated to the cell state.

[0152] The Hadamard product is an operation that multiplies corresponding elements of a matrix. It is used to achieve precise element-by-element adjustment of corresponding feature information by gating signals.

[0153] Through the modified state update equations described above, this embodiment explicitly embeds the energy conservation mechanism of the alumina evaporation process into the core state update stage of the LSTM network. A heat input gating factor is used to weight and modify the new information update branch of the LSTM: Under unsteady conditions with abrupt changes in heat input, the model can quickly respond to changes in conditions, significantly improving the fitting accuracy of the dynamic characteristics of the energy subsystem, thereby improving the prediction accuracy of quality indicators; under steady-state conditions with stable heat input, the model can retain effective historical memory, conforming to the inertial characteristics of heat transfer, ensuring the continuity and physical consistency of prediction results, and avoiding abnormal results that do not conform to the law of energy conservation. Simultaneously, this design gives the model's gating adjustment process a clear physical meaning, enhancing the model's physical interpretability and operational reliability.

[0154] In a preferred embodiment, in step S4, the memory update process of the perturbation path is guided by a perturbation mutation detection gating mechanism, and the specific implementation process is as follows:

[0155] First, calculate the abrupt change degree of the perturbation. The calculation formula is as follows:

[0156] ;

[0157] In this embodiment, The vector of perturbation variables consists of the observed values ​​of perturbation variables such as feed temperature, final vacuum, condensate level, and circulating pump frequency at time t, representing the perturbation state at the current time. The observed value of the perturbation variable vector at time t-1 represents the perturbation state at the previous time. For extremely small positive numbers, such as 10 -6 This is used to prevent the denominator from being zero and to avoid calculation errors.

[0158] molecular The first-order difference magnitude of the disturbance variable vector is used to accurately capture the overall magnitude of the disturbance by comprehensively quantifying the changes of multiple disturbance variables; the denominator is... It is a normalization factor for the magnitude of the perturbation, used to eliminate the difference in magnitude between different perturbation variables and suppress the influence of abnormal peaks on the calculation of abrupt change. The normalized disturbance abruptness directly reflects the severity of the current disturbance relative to the historical state. Using this as the driving input for the gating mechanism, the disturbance condition characteristics guide the LSTM model state update process from the root, avoiding the black box defect of pure data-driven gating logic without physical meaning.

[0159] Secondly, based on the degree of perturbation change Constructing perturbation gating factors The calculation formula is as follows:

[0160] ;

[0161] In this embodiment, The Sigmoid activation function can smoothly map the calculation result to the [0,1] interval, realize continuous adjustment of the gating, adapt to the characteristics of continuous operation of industrial processes, and avoid the model output jump caused by hard-switching gating; The disturbance path sensitivity coefficient can be calibrated based on the actual disturbance tolerance characteristics of the alumina evaporation device. In this embodiment, the value ranges from 0.5 to 1.2, which is used to adjust the degree of disturbance abruptness. Gating factor The impact intensity is determined to adapt to the disturbance response requirements of evaporation devices of different specifications. The disturbance gating factor... Used to dynamically adjust the model response based on the intensity of disturbance abrupt changes: when the disturbance abruptness is high, such as a sudden change in feed temperature or a large fluctuation in vacuum, The value tends towards 1, enabling the model to respond quickly to disturbances and promptly capture their impact on the evaporation process; when the disturbance is stable, such as with stable vacuum and no significant change in condensate level, the value tends towards 1. The value tends to 0, suppressing interference from invalid disturbance information and preventing the model from deviating from its core operating principles due to excessive focus on stationary disturbances.

[0162] Finally, the state update equation of the perturbation path LSTM is corrected using the perturbation gating factor, as shown below:

[0163] ;

[0164] The physical meaning and implementation method of each parameter in the formula are as follows:

[0165] The Sigmoid activation function maps the output to the [0,1] interval, enabling continuous adjustment of the gating. This is the perturbation path sensitivity coefficient, used to adjust the degree of perturbation abruptness. Gating factor The intensity of the impact; As the perturbation gating factor, when the perturbation abruptness is high, The value tends towards 1, enabling the model to respond quickly to disturbances; when the disturbance is stable, The value tends towards 0, suppressing interference from invalid perturbation information;

[0166] The cell state of the corrected perturbation path LSTM at time t represents the long-term memory unit after incorporating perturbation gating information, used to characterize the impact of perturbation on the evaporation process.

[0167] The forget gate of the perturbation path LSTM controls the cell state at the previous time step. The proportion of information that needs to be "forgotten" in the middle;

[0168] This represents the cell state at the previous time step t-1, i.e., the perturbation-related long-term memory retained by the model in previous time steps;

[0169] This serves as the input gate for the perturbation path LSTM, controlling the current candidate cell state. The proportion of information that needs to be updated to the current cell state;

[0170] The candidate cell states of the perturbation path LSTM are calculated from the current input and the hidden state at the previous time step, representing the perturbation-related candidate memory information to be updated.

[0171] The Hadamard product is an element-wise multiplication used to achieve precise adjustment of the corresponding feature information by the gating signal.

[0172] This invention explicitly embeds the disturbance mutation detection mechanism into the core state update stage of the LSTM network. A disturbance gating factor is used to weight and correct the new information update branch of the LSTM. Under sudden disturbance conditions, the model can quickly respond to disturbance changes, significantly improving the accuracy of capturing the impact of disturbances and thus enhancing the prediction accuracy of quality indicators. Under stable disturbance conditions, the model can suppress invalid information interference, ensuring the stability and physical consistency of prediction results and avoiding abnormal results that do not conform to the actual operating laws of the evaporation process. Simultaneously, this design gives the model's gating adjustment process a clear physical meaning, enhancing the model's physical interpretability and operational reliability.

[0173] In a preferred embodiment, step S5 specifically includes the following steps:

[0174] S51. Concatenate the dimensions of the hidden state vectors output by the three paths.

[0175] In this embodiment, the three paths are, respectively, the material path constrained by the concentration difference gating mechanism, the energy path constrained by the heat input ratio gating mechanism, and the disturbance path constrained by the disturbance fluctuation gating mechanism. The latent state vectors output by the three paths are the exclusive feature representations of their respective subsystems after mechanistic guidance, fully carrying the dynamic operation rules and physical constraint information of each subsystem. The latent state vectors output by the three paths are concatenated along the feature channel dimension to obtain the joint feature vector before fusion. Through the dimensional concatenation preprocessing method, the independence and physical meaning of the features of each path are fully preserved, avoiding the mutual obscuring or loss of mechanistic constraint information of different subsystems during the fusion process, laying a feature foundation with both physical consistency and information integrity for subsequent adaptive fusion.

[0176] S52. Input the concatenated vector into a multilayer perceptron with at least two hidden layers for adaptive fusion.

[0177] In this embodiment, the multilayer perceptron is configured with two fully connected hidden layers. The number of neurons in the first hidden layer matches the dimension of the spliced ​​joint feature vector, and the number of neurons in the second hidden layer is set to half that of the first hidden layer. Both hidden layers use the Sigmoid nonlinear activation function. The spliced ​​joint feature vector is input into the multilayer perceptron. Through the nonlinear transformation of the two fully connected layers, the nonlinear coupling effect of the three subsystems of material, energy, and perturbation on the concentration of the mother liquor from alumina evaporation is adaptively fitted, and the deep adaptive fusion of multipath features is completed to obtain the fused high-dimensional feature vector.

[0178] Since the fused input features have been pre-constrained by the mechanisms of each subsystem, the nonlinear fitting process of the multilayer perceptron is always constrained by the underlying physical mechanism, avoiding the problems of loss of physical consistency and poor interpretability caused by traditional black-box fusion. While improving the effectiveness of feature fusion, it ensures the physical rationality of the fusion process.

[0179] S53. Output the final predicted value of the mother liquor concentration of alumina output through the linear regression layer.

[0180] In this embodiment, the linear regression layer is a fully connected linear layer, and its output dimension matches the dimension of the quality index to be predicted. In this embodiment, the alumina discharge mother liquor concentration index to be predicted includes the caustic ratio of the discharge mother liquor, and the output dimension of the corresponding linear regression layer is set to 1. The fused high-dimensional feature vector output by the multilayer perceptron is input into the linear regression layer, and the high-dimensional feature is converted into a prediction result that matches the actual industrial index value range through linear mapping. Finally, the soft measurement prediction value of the alumina discharge mother liquor concentration is output.

[0181] By using a linear regression layer for the final prediction output, the numerical continuity and industrial adaptability of the prediction results can be guaranteed, which aligns with the continuous operation characteristics of the alumina evaporation process. At the same time, based on the feature input constrained by the front-end mechanism and the adaptive fusion results, the final output prediction value has both high prediction accuracy and conforms to the basic laws of material and energy conservation in the alumina evaporation process. It can stably adapt to the soft measurement requirements of the entire alumina evaporation process, providing reliable quality index data support for the real-time control and optimized operation of the process.

[0182] To further illustrate the inventive concept of this invention, a six-effect, four-flash evaporation process for alumina is used as an application scenario, and the schematic diagram of the process equipment is shown below. Figure 1 As shown, the core equipment of the evaporation process includes seven countercurrent falling film evaporators (including one spare evaporator), one forced evaporator, four flash evaporators, and a matching condensate tank. The core operating process is as follows: the mother liquor inside the evaporator is indirectly heated by external steam through the heating tubes. After the mother liquor absorbs heat and boils, the concentrated liquid is discharged from the bottom of the evaporation chamber to the next effect for further concentration. The separated secondary steam is discharged from the steam port and used as a heat source for the next effect evaporator.

[0183] In the alumina evaporation process, the concentration of caustic alkali in the outlet mother liquor is a core quality indicator and a key control target for on-site manual operations. Currently, this indicator value can only be obtained through laboratory testing every 4 hours, which is severely lagging and cannot provide timely guidance for production operations. (See also...) Figure 2 This embodiment is based on a knowledge-guided soft measurement method for alumina evaporation quality indicators provided by the present invention, which predicts the quality indicators in real time. The specific implementation process is as follows:

[0184] Step 1: Construction of soft measurement dataset for alumina evaporation process:

[0185] Collect historical multivariate time series data from the production site to construct a soft measurement dataset.

[0186] The specific process is as follows:

[0187] 1.1 Exporting historical time series data from the alumina plant's distributed control system (DCS):

[0188] Based on the physical mechanism of the evaporation process, process variables are divided into three categories:

[0189] Material variable subset These include material flow parameters that directly participate in the material concentration process, such as feed mother liquor flow rate, circulating mother liquor density, and raw liquor temperature.

[0190] energy variable subset This includes thermodynamic parameters such as new steam pressure, new steam flow rate, and the temperature and temperature difference at each stage from the first effect to the last effect;

[0191] Perturbation variable subset These include parameters that reflect the environmental and equipment operating status, such as vacuum fluctuations, flash evaporator liquid level, and circulating pump current frequency.

[0192] 1.2 Data Preprocessing: The Laida criterion is used to remove obvious outliers in the data, and linear interpolation is used to fill in missing values ​​in the data to eliminate the interference of data noise and missing values ​​on modeling;

[0193] 1.3 Data normalization: The Min-Max normalization method is used to map the values ​​of all variables to the interval [0,1]. The maximum and minimum values ​​used for normalization are taken from the statistical values ​​of the corresponding variables in the training set.

[0194] Step 2: Construction of the multi-scale feature extraction module:

[0195] Three time window steps—short, medium, and long—are set to construct a multi-scale feature extraction module. The specific process is as follows:

[0196] 2.1. Set the sliding window size for the three time scales. For the i-th variable at time t, three sliding time window slices are extracted respectively. This forms a sliding slice tensor of the corresponding scale;

[0197] 2.2. Input the sliding slice tensor of each scale into the corresponding one-dimensional convolutional layer. In this embodiment, the kernel size is set to 3, the stride is 1, and the padding method is "max" padding. The convolution operation formula is:

[0198] ;

[0199] A batch normalization layer is added before the ReLU activation function to prevent gradient vanishing and accelerate model training, thereby completing the extraction of single-scale local evolution features.

[0200] 2.3 Extracting feature vectors from the three scales The features are concatenated along the feature dimension to form a unified multi-scale feature vector with expanded dimensions. This serves as the input for the subsequent LSTM network.

[0201] Step 3: Building a domain knowledge-guided multi-path LSTM network:

[0202] Three parallel LSTM subnetworks are constructed, namely the material path, energy path and disturbance path. The three paths correspond one-to-one with the three subsets of variables divided in step 1. All three paths use LSTM as the core computing unit, and the unified multi-scale feature vector output in step 2 is used as the common input of the three paths.

[0203] Step 4: Mechanism-guided gating factor calculation and LSTM state update:

[0204] Each path's corresponding physical subsystem has a dedicated gating guidance factor calculated based on its mechanistic characteristics. This gating guidance factor is then used to adjust the memory update process of the corresponding path's LSTM unit, yielding the hidden state of each path. The specific process is as follows:

[0205] 4.1 Material path gating adjustment:

[0206] 4.1.1 Calculate the concentration difference ;

[0207] 4.1.2 Calculate the concentration difference gating factor and sensitivity coefficient. Initialized to 1.0 and automatically optimized during training using gradient descent; this gating logic ensures that the concentration changes are controlled within a certain range. When the concentration is relatively high and the operating conditions are not steady-state, the gating factor tends to be 1, and the model tends to absorb new information through the input gate; when the concentration changes steadily, the gating factor tends to be 0.5, and the model retains long-term memory.

[0208] 4.1.3. Correct the state update equation of the material path LSTM using the concentration gating factor. Complete the status update of the material path.

[0209] 4.2 Energy path gating regulation:

[0210] 4.2.1 Calculate the moving average value of heat input Set the sliding window size to 12, and calculate the sliding average value of the heat input within the preceding window. ;

[0211] 4.2.2 Calculate the local heat input ratio In actual calculations, you can add to the denominator. To prevent division by zero of small constants and avoid calculation errors;

[0212] 4.2.3 Calculate the heat input gating factor By guiding the LSTM update through the relative rate of change of heat input, the model can effectively distinguish between the ramp-up period and the steady period of the energy subsystem, and adapt to the dynamic characteristics of different working conditions.

[0213] 4.2.4. Correcting the state update equation of the energy path LSTM using the aforementioned heat input gating factor. Complete the state update of the energy path.

[0214] 4.3 Disturbance path gating adjustment:

[0215] 4.3.1 Calculate the abrupt change degree of the perturbation The input is a perturbation vector consisting of r=5 perturbation variables. The vector difference magnitude is calculated using the Euclidean norm and normalized by dividing by the historical magnitude to eliminate the influence of the magnitude difference between different variables.

[0216] 4.3.2 Calculate the disturbance fluctuation gating factor When the abrupt change in perturbation When the threshold value is exceeded, the gate value increases significantly, forcing the LSTM unit to pass through. The weighted terms quickly update the cell state, enabling immediate response to unexpected operating conditions;

[0217] 4.3.3. Correct the state update equation of the perturbation path LSTM using the perturbation fluctuation gating factor. Complete the state update of the disturbance path.

[0218] After completing the state updates for the three paths, output the hidden state of each path at time t.

[0219] Step 5: Consistency Integration and Quality Indicator Prediction:

[0220] A consistency fusion layer is constructed to complete multi-path feature fusion and final prediction. The specific process is as follows:

[0221] 5.1. The hidden state vectors output by the three paths at time t are concatenated in dimension and input into a multilayer perceptron containing two hidden layers for adaptive fusion. In this embodiment, the number of hidden layer nodes is set to [32, 16] and the activation function is Sigmoid.

[0222] 5.2 The last layer of a multilayer perceptron is a linear regression layer, outputting a univariate scalar. That is, the predicted concentration value of the mother liquor from the alumina evaporation.

[0223] Step 6: Supervised training of the model:

[0224] Using the multivariate process data from the training and validation sets divided in step 1 as input, and the corresponding alumina evaporation process discharge concentration test values ​​as labels, the following method is employed: Figure 3 The multi-scale, multi-path LSTM neural network structure shown is subjected to supervised training. For details, see [link to documentation]. Figure 3 The model first extracts and concatenates multi-scale features from the input data through windowing and one-dimensional convolutional modules to capture dynamic features from different temporal perspectives. Then, the features are fed into three parallel mechanism-guided paths: the material path, the energy path, and the perturbation path. These three paths utilize physical gating mechanisms such as concentration difference, local heat input ratio, and perturbation mutation detection to perform differentiated state updates on the LSTM. Finally, the output features from the three paths converge to a fully connected layer for consistent fusion, thereby generating the final predicted values ​​for the mass variables. In this embodiment, the training hyperparameters are configured as follows: the Adam optimizer is used for parameter updates, the initial learning rate is set to 0.001, the batch size is set to 32, and an L2 regularization term is set to prevent overfitting. The minimum mean squared error loss function is used during training, ultimately resulting in a trained soft measurement prediction model.

[0225] Step 7, Model Application and Performance Evaluation:

[0226] The process data of the test set samples are input into the trained soft measurement model. After multi-scale feature extraction, mechanism-guided multi-path memory update and consistency fusion, the final quality index prediction value is output. The prediction value can be compared with the actual on-site test value to complete the model performance evaluation.

[0227] Figure 4 This paper presents comparative results on the predictive performance of the model provided by this invention on key mass variables in the alumina evaporation process. From... Figure 4 As can be seen, the predicted curve closely follows the trend of the true value. Even under conditions where the sample values ​​fluctuate significantly or change drastically, the model still maintains extremely high tracking accuracy and fitting ability, thus intuitively verifying the effectiveness and robustness of this method in actual industrial soft measurement tasks.

[0228] Example 2

[0229] This invention also provides a knowledge-guided soft measurement system for alumina evaporation quality indicators. Its structural design and functional modules work collaboratively to achieve high-precision, highly interpretable soft measurement of quality indicators in complex industrial processes. The system includes:

[0230] The data acquisition and segmentation module is used to collect multivariate time-series data during the operation of the alumina evaporation unit, such as easily measurable parameters like temperature, pressure, and flow rate. Based on the physical mechanism of the alumina evaporation process, the original variables are divided into three independent subsets:

[0231] A subset of material variables used to reflect the material inventory, transformation, and flow status;

[0232] A subset of energy variables used to reflect the characteristics of heat energy supply, exchange, and dissipation;

[0233] A subset of disturbance variables is used to reflect non-steady-state factors such as equipment start-up and shutdown, and external disturbances;

[0234] This classification method is based on the process mechanism, which lays the foundation for subsequent differentiated modeling of subsystems with different physical properties, effectively reduces the interference caused by multivariate nonlinear coupling, and improves the adaptability of the model under unsteady conditions.

[0235] The multi-scale feature extraction module slices the process data using sliding windows at different time scales, such as short, medium, and long scales, and inputs each slice into a one-dimensional convolutional layer to extract local evolutionary features at each scale. The convolutional kernel size is matched to the window length to perceive sequence patterns within a specific time span, and a non-linear activation function further enhances the feature representation. Finally, the features from each scale are concatenated and fused to form a unified multi-scale temporal feature representation. This design enables the model to simultaneously capture short-term fluctuations and long-term trends, solving the problem that a single time window cannot adapt to the heterogeneous temporal evolution of industrial processes, and providing context-rich dynamic input for downstream prediction.

[0236] The knowledge-guided multi-path prediction module comprises three parallel processing paths: a material path, an energy path, and a perturbation path. Each path uses a Long Short-Term Memory (LSTM) network as its core unit and shares input from the multi-scale feature extraction module. Each path incorporates physical mechanism gating factors, which dynamically adjust the cell state update process of the LSTM unit based on the process characteristics of the corresponding subsystem. For example, in the material path, the gating factor can trigger state retention or update based on concentration difference values, providing the model with a clear physical basis for judgment during process state switching. This design explicitly embeds the physical mechanism into the network structure, giving the model's memory update process a clear physical meaning, improving the model's physical interpretability and transparency, and avoiding the uninterpretability problem of traditional "black box" LSTM models.

[0237] The consistent fusion output module receives the hidden state vectors from three parallel paths. It first integrates the multi-source information through dimensional concatenation, then inputs it into a multilayer perceptron (MLP) containing at least two hidden layers for adaptive fusion to capture the complex nonlinear relationships between paths. Finally, a linear regression layer outputs predicted values ​​for key quality indicators of the alumina evaporation process (such as the concentration of the mother liquor in the effluent). This fusion structure ensures full information integration while maintaining a concise and interpretable output through the linear layer, effectively improving prediction accuracy and robustness under complex operating conditions.

[0238] This system, through modular design, organically combines multi-scale feature extraction, physical mechanism-guided path modeling, and consistency fusion, achieving end-to-end optimization from raw data to quality indicator prediction. Each module has a clearly defined function and works collaboratively and efficiently, solving the problem of traditional soft measurement models struggling to integrate multi-scale features and domain knowledge. Furthermore, it exhibits higher prediction accuracy and stronger physical consistency under non-steady-state conditions, meeting the dual requirements of industrial sites for model reliability and interpretability.

[0239] Example 3

[0240] The present invention also provides a readable storage medium having a computer program stored thereon, wherein, when the program is executed, it controls the device where the storage medium is located to perform some or all of the steps in Embodiment 1.

[0241] The computer-readable storage medium may include high-speed RAM memory and may also include nonvolatile memory (NVM), such as at least one disk storage device. It is understood that the storage medium can be any machine-readable medium capable of storing program code, such as random access memory (RAM), magnetic disk, hard disk, solid state disk (SSD), or nonvolatile memory.

[0242] Those skilled in the art will understand that embodiments of the present invention can be provided as methods or storage media. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0243] Example 4

[0244] The present invention also provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the computer program instructions are executed by the processor, the device is triggered to perform some or all of the steps in Embodiment 1;

[0245] A processor may include one or more processing units, such as an application processor (AP), a modem processor, a graphics processing unit (GPU), an image signal processor (ISP), a controller, a video codec, a digital signal processor (DSP), a baseband processor, and / or a neural network processing unit (NPU). Different processing units may be independent devices or integrated into one or more processors.

[0246] The controller can serve as the nerve center and command center of an electronic device. Based on the instruction opcode and timing signals, the controller generates operation control signals to control the fetching and execution of instructions.

[0247] The processor may also include memory for storing instructions and data. In some embodiments, the memory in the processor is a cache memory. This memory can store instructions or data that the processor has just used or that are used repeatedly. If the processor needs to use the instruction or data again, it can retrieve it directly from the memory. This avoids repeated accesses, reduces processor waiting time, and thus improves system efficiency.

[0248] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct or indirect applications in other related technical fields, are within the patent protection scope of the present invention.

Claims

1. A knowledge-guided soft measurement method for alumina evaporation quality indicators, characterized in that, The steps include: S1. Obtain multivariate time series data of the alumina evaporation process, and divide the multivariate time series data into a material variable subset, an energy variable subset, and a disturbance variable subset based on the physical mechanism; S2. Construct a multi-scale feature extraction module, which constructs sliding windows of different time scales for the multivariate time series data and extracts local evolution features within each scale. Then, the features of each scale are spliced ​​and fused to obtain a unified multi-scale time feature representation. S3. Construct parallel material paths, energy paths, and disturbance paths corresponding to each subset of variables. Each path uses a Long Short-Term Memory (LSTM) network as its core unit and uses the unified multi-scale time feature representation as the common input of each path. S4. Calculate specific gating guidance factors based on the mechanistic characteristics of the material, energy, and disturbance subsystems respectively, and use the gating guidance factors to adjust the memory update process of the LSTM unit under the corresponding path to obtain the hidden state of each path. The material path memory update process is guided by a concentration differential gating mechanism, specifically including: Calculate the concentration difference term between the current time and historical time. ; Concentration gating factor is constructed based on the concentration difference component. When the concentration changes drastically, the adjustment model increases the proportion of absorption of the current input; when the concentration changes steadily, it increases the proportion of retention of historical states. The state update equation of the material path LSTM is corrected using the concentration gating factor. The memory update process of the energy path is guided by a local heat input ratio gating mechanism, which specifically includes: Calculate the ratio of the current heat input to the previous time window moving average heat input. ;in The moving average heat input is the preceding time window. This represents the material heat input at the current time t. Based on the ratio Constructing hot input gating factors This is used to distinguish between the energy ramp-up period and the stable period, and to adjust the memory update intensity of the LSTM accordingly; Using the heat input gating factor The state update equation of the energy path LSTM is modified; The memory update process of the perturbation path is guided by a perturbation mutation detection gating mechanism, and the specific implementation steps include: Calculate the degree of perturbation change The numerator is the first-order difference magnitude of the perturbation variable vector, used to capture the magnitude of the perturbation change; the denominator is the normalization factor of the perturbation magnitude, used to suppress the influence of abnormal peaks. Constructing perturbation gating factors This is used to dynamically adjust the model response based on the intensity of the perturbation mutation. The state update equation of the perturbation path LSTM is corrected using the perturbation gating factor: ； in, For the perturbation variable vector; Let be the observed value of the perturbation variable vector at time t-1; 10 -6 ; This represents the normalized perturbation abrupt change degree. Use the Sigmoid activation function; This is the disturbance path sensitivity coefficient; The disturbance gating factor; The cell state of the corrected perturbation path LSTM at time t; Forget gate for perturbation path LSTM; This represents the cell state at the previous time step t-1; This is the input gate for the perturbation path LSTM; Candidate cell states for the perturbation path LSTM; For Hadamah accumulation; S5. Consistently integrate the hidden states of material, energy, and disturbance path outputs to generate predicted values ​​of key quality indicators for the alumina evaporation process.

2. The knowledge-guided soft measurement method for alumina evaporation quality indicators according to claim 1, characterized in that, In step S1, the multivariate time series data is represented as follows: ,in Indicates the time step of a time series. Indicates the dimension of the variable; Variables directly involved in reaction and transport are divided into a subset of material variables; The variables that provide the heat source to drive evaporation are divided into a subset of energy variables; Variables related to equipment start-up and shutdown and environmental fluctuations are divided into a subset of disturbance variables.

3. The knowledge-guided soft measurement method for alumina evaporation quality indicators according to claim 1, characterized in that, The specific steps for implementing step S2 include: S21. Set a sliding window with three time scales. For each process variable Extract the corresponding sliding time window slices to form a sliding time window slice tensor; where Sequence analysis corresponding to short, medium, and long time scales respectively; S22. Input the sliding window sequences at each scale into a one-dimensional convolutional layer, extract local evolution features by sensing the sequence pattern within a specific window size through the convolutional kernel, and enhance the feature expression through a nonlinear activation function. S23. Concatenate and fuse the extracted feature vectors at different scales into a multi-scale feature vector. This is to preserve the dynamic complementarity between different scales.

4. The knowledge-guided soft measurement method for alumina evaporation quality indicators according to claim 1, characterized in that, The concentration gating factor The calculation formula is as follows: ； in, It is the sigmoid activation function. This is the material path sensitivity coefficient; The specific state update equation for the material path LSTM is shown in the following formula: ； in, The cell state of the corrected material path LSTM at time t; Forgotten Gate; This represents the cell state at the previous time step t-1; For input gates; Candidate cell state; For Hadama accumulation.

5. The knowledge-guided soft measurement method for alumina evaporation quality indicators according to claim 1, characterized in that, The formula for calculating the heat input gating factor is as follows: ； in, It is the sigmoid activation function. This is a sensitivity coefficient for energy paths, used to adjust the heat input ratio. Gating factor The intensity of the impact; The state update equation for the energy path LSTM is shown in the following equation: ； The modified energy path LSTM cell state at time t; Forgotten Gate; This represents the cell state at the previous time step t-1; For input gates; Candidate cell state, For Hadama accumulation.

6. A knowledge-guided soft measurement method for alumina evaporation quality indicators according to any one of claims 1-5, characterized in that, The specific steps for implementing step S5 include: The hidden state vectors output from the three paths are concatenated in terms of dimensions. The concatenated vector is input into a multilayer perceptron with at least two hidden layers for adaptive fusion; The final predicted value of the mother liquor concentration in the alumina output is output through a linear regression layer.

7. A knowledge-guided soft measurement system for alumina evaporation quality indicators, employing the measurement method described in any one of claims 1-6, characterized in that, include: The data acquisition and segmentation module is used to acquire multivariate data of the alumina evaporation process and segment it into material, energy and disturbance variable subsets based on the physical mechanism. The multi-scale feature extraction module is used to extract multi-scale dynamic features of process data through sliding windows of different lengths and one-dimensional convolutional layers. The knowledge-guided multi-path prediction module includes parallel material, energy, and perturbation LSTM paths, with each path's state update adjusted by a built-in physical mechanism gating factor. The consistency fusion output module is used to adaptively weight and fuse the features of each path, and output the real-time monitoring results of the quality indicators.

Images