Intelligent control method for product material outlet index

Abstract

The embodiment of the application discloses a kind of in-process material outlet index intelligent control methods, comprising: collecting variable and obtaining characteristic vector by normalizing processing;Characteristic vector is input into recurrent neural network to obtain control, environmental characteristics, after the fusion of two kinds of characteristics, input into fusion fully connected layer, output prediction curve;Determine that head is in transition section or steady running section;Head is in steady running section, freeze double-branch main stem, fine-tune prediction head, set the solvable variable, construct objective function to solve solvable variable, obtain optimal control amount sequence;The first step control amount in optimal control amount sequence is carried out reverse normalization processing, and is issued to actuating mechanism, and actuating mechanism adjusts processing process, changes next cycle variable;The present method solves the problem that control response characteristics and environmental disturbance characteristics are coupled to each other, resulting in insufficient prediction accuracy and further restricting advanced control, and realizes advanced closed-loop control on outlet index by combining double-branch decoupling prediction and rolling optimization solution.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for industrial processes, specifically relating to an intelligent control method and system for the moisture content and temperature at the outlet of tobacco processing materials based on a combination of deep time-series prediction and model predictive control (MPC).

[0002] The technical solution of this invention spans two sub-fields: "deep learning temporal prediction" and "process optimization control": on the prediction side, a dual-branch long short-term memory network is employed. This method performs multi-step advance prediction of the hot and wet processing process. By decoupling the physical causality between the control branch and the environmental branch, the dynamic responses of operable controllable variables and indirect environmental disturbances are modeled separately. On the control side, the frozen prediction model is embedded into a rolling optimization framework in a differentiable manner, and gradient descent replaces traditional grid search or black-box optimization to achieve advance intervention of control variables. The above prediction-update-control closed loop is not dependent on specific machine models and can be applied to typical hot and wet processing links such as the outlet moisture / hot air temperature control of the feeding process in tobacco processing lines and the outlet moisture / temperature control of the re-drying and moistening process. Background Technology

[0003] In the tobacco processing of raw materials, the export index is a key process parameter for measuring product quality. Traditional export index control methods suffer from insufficient predictive model accuracy, lag in control response, lack of proactive control capabilities, and difficulty in handling large lag processes. Existing technologies often rely on fixed initial setpoints and human experience for export index control, resulting in low robustness and poor adaptability. While some solutions introduce data-driven predictive models to forecast export indices, these models typically mix control and environmental variables in the same network branch for feature extraction. These variables differ in their temporal characteristics, frequency of change, and mechanisms of influence on export indices. Mixing these two types of variables leads to the coupling of control response characteristics and environmental disturbance characteristics within the network, making effective separation difficult and ultimately limiting the accuracy of export index prediction. Traditional feedback control methods only adjust after the export index deviates from the setpoint, representing a "post-compensation" type of control, which is insufficient to effectively address the lag phenomenon under pure feedback control. Although predictive control methods based on predictive models have emerged in the existing technology, the reliability of proactive control is difficult to guarantee due to the insufficient accuracy of the prediction itself, and it is impossible to achieve true "ex-ante control" of export indicators. Summary of the Invention

[0004] The main objective of this invention is to propose an intelligent control method for the outlet indicators of materials under production, which enables high-precision prediction of outlet moisture / temperature in multiple steps, overcoming the large inertia and lag of hot and wet processing; automatically identifying and isolating non-stationary transition segments, ensuring that online updates and control are only executed in the steady-state segment; and adapting to material batch differences and environmental drift without stopping the machine, maintaining long-term prediction accuracy.

[0005] To achieve the above objectives, embodiments of the present invention propose an intelligent control method for the outlet indicators of in-process materials, comprising:

[0006] Data collection steps: Collect process variables during the material processing process at a fixed sampling period, construct a sliding history window, and normalize the process variables to obtain a normalized input feature vector. The input feature vector includes control variables and environmental variables. The control variables are directly adjustable setpoints, and the environmental variables are process disturbances that cannot be directly manipulated.

[0007] Prediction steps: Input the control quantity and environmental quantity in the input feature vector into the control branch recurrent neural network respectively to obtain the corresponding control features and environmental features. Concatenate and fuse the control features and environmental features, input them into the fusion fully connected layer to map to the prediction time domain dimension, and output the prediction curve of the exit index for multiple future sampling periods.

[0008] Segmented judgment steps: Based on the changes in the state of the material flow rate and the control quantity in the process variables, determine whether the current stage is the material head transition stage or the steady-state operation stage;

[0009] Online update steps: When it is determined that the system is in the steady-state operation segment, the parameters of the control branch recurrent neural network and the environment branch recurrent neural network are frozen. The parameters of the fused fully connected layer are updated by gradient according to the prediction error. The offset of the updated exit index prediction curve relative to the previous curve is smoothed and corrected to obtain the corrected exit index prediction curve. The control solution step constructs the objective function based on the corrected exit index prediction curve.

[0010] Control solution steps: When it is determined that the steady-state operation segment is in progress, the control quantity sequence in the future control time domain is set as an optimizable variable. An objective function is constructed based on the deviation between the predicted exit index curve and the set value of the exit index. Differentiable optimization is performed on the optimizable variable to obtain the optimal control quantity sequence.

[0011] Issuance Step: The first step control quantity in the optimal control quantity sequence is denormalized and issued to the actuator corresponding to the control quantity. The actuator adjusts the in-process material processing to change the process variable in the next sampling period.

[0012] Furthermore, the prediction step specifically includes:

[0013] Control input sub-step: Extract the control quantity sequence within the historical lookback steps from the sliding history window and input it into the control branch recurrent neural network;

[0014] Environmental quantity input sub-step: Extract the environmental quantity sequence within the historical lookback step from the sliding history window and input it into the environmental branch recurrent neural network;

[0015] The dual-branch feature extraction sub-step involves extracting the control features of the control quantity sequence based on the control branch recurrent neural network, and extracting the environmental features of the environmental quantity sequence based on the environmental branch recurrent neural network.

[0016] The splicing and fusion prediction sub-step involves splicing the control features and the environmental features in the channel dimension, inputting them into the fusion fully connected layer, mapping them to the prediction step dimension, and obtaining the prediction curve of the exit index for the next multiple sampling periods.

[0017] Furthermore, it achieves physical causal decoupling and independent feature extraction of control and environmental variables, enabling the fusion of control and environmental features in the channel dimension after extraction, thus avoiding feature coupling problems in a single network.

[0018] Furthermore, in the dual-branch feature extraction sub-step, both the control branch recurrent neural network and the environment branch recurrent neural network are two-layer stacked structures with the same hidden state dimension; the control branch recurrent neural network outputs the control features using the historical lookback step number as the time expansion step number; the environment branch recurrent neural network outputs the environment features using the historical lookback step number as the time expansion step number.

[0019] Furthermore, the specific structural parameters of the dual-branch recurrent neural network were determined to ensure that the feature extraction capabilities of the two paths are equal and to avoid the decline in fusion effect due to the imbalance of feature expression capabilities.

[0020] Furthermore, in the splicing and fusion prediction sub-step, the input dimension of the fusion fully connected layer is the sum of the dimension of the control feature and the dimension of the environmental feature, and the output dimension is the prediction step number; the output of the fusion fully connected layer is also superimposed with the residual term corresponding to the inlet material residual in the environmental quantity to obtain the predicted curve of the outlet index after superimposing the residual.

[0021] Furthermore, by directly correcting the steady-state offset by superimposing the inlet material residual term, the predictive ability to adapt to inlet operating condition offsets is enhanced, thus solving the technical problem of insufficient response of the outlet index prediction curve to the steady-state offset of inlet material residuals.

[0022] Furthermore, in the online update step, smoothing the offset of the updated export indicator prediction curve relative to the previous curve includes:

[0023] Window offset estimation sub-step: Maintain a sliding window, record the base model predicted value and the corresponding measured value of the export index prediction curve at multiple sampling times within the sliding window, and calculate the original value of the system offset at the current sampling time based on the residual sequence between the base model predicted value and the measured value of the export index within the sliding window.

[0024] Exponential smoothing sub-step: Perform first-order exponential smoothing on the original system offset value to obtain the system offset value at the current sampling time;

[0025] Offset superposition sub-step: The system offset value is superimposed on the base model prediction value of the export index prediction curve to obtain the corrected export index prediction curve.

[0026] Furthermore, by combining sliding window statistics with first-order exponential smoothing, high-frequency noise is filtered out while following batch-level slow drift, thus solving the technical problem of system offset estimation jitter caused by single-step residual noise.

[0027] Furthermore, in the control solution step, the objective function includes a control tracking loss term and a control constraint penalty term. The control tracking loss term is the mean square error between the predicted curve of the export index and the set value of the export index, and the control constraint penalty term is the penalty for the optimizable variable exceeding the preset upper and lower limits.

[0028] Furthermore, by introducing an out-of-bounds penalty term into the objective function, the constraint solution results fall within the preset safe range, thus solving the technical problem that the control quantity may exceed the safe range in the rolling optimization solution.

[0029] Furthermore, in the control solution step, the differentiable optimization solution for the optimizable variables includes:

[0030] The optimizable variable is initialized to zero and set as a learnable parameter, and the parameters of the recurrent neural network and the fused fully connected layer in the prediction step are frozen.

[0031] The optimizable variable is updated iteratively multiple times using stochastic gradient descent with momentum. During each iteration, the gradient of the optimizable variable is pruned according to its norm, and the learning rate is adjusted based on the loss improvement.

[0032] Record the values ​​of the objective function during multiple iterations, and select the optimizable variable corresponding to the minimum value of the objective function as the optimal control variable sequence.

[0033] Furthermore, by employing gradient pruning and learning rate adjustment strategies, stable convergence of the differentiable optimization solution is ensured, thus solving the technical problem of unstable solution caused by uncontrollable gradient update magnitude during the solution of control variables.

[0034] Furthermore, the control quantity includes at least one of a water addition setpoint or a steam setpoint, and the outlet index includes at least one of an outlet moisture content or an outlet temperature; the method, by adjusting the dimension of the control quantity and the outlet index setpoint, can be adapted to the outlet index control scenarios of different processing steps.

[0035] Furthermore, by jointly solving multiple control variables and designing a transferable architecture, we can achieve coordinated control of moisture and temperature and cross-process deployment. This solves the technical problems of the inability to coordinate optimization when outlet moisture and outlet temperature are controlled independently, as well as the technical problems of the inability to transfer the method across processes.

[0036] Accordingly, embodiments of the present invention also provide an intelligent control method for the outlet index of in-process materials, including:

[0037] Data collection steps: Collect process variables during the material processing process at a fixed sampling period, and normalize the process variables to obtain a normalized input feature vector. The input feature vector includes control variables and environmental variables. The control variables are directly adjustable setpoints, and the environmental variables are process disturbances that cannot be directly manipulated.

[0038] Prediction steps: Input the control quantity and environmental quantity in the input feature vector into the control branch recurrent neural network respectively to obtain the corresponding control features and environmental features. Concatenate and fuse the control features and environmental features, input them into the fusion fully connected layer to map to the prediction time domain dimension, and output the exit index prediction curve for multiple future sampling periods.

[0039] Flow rate determination step: Determine whether the material flow rate is lower than a preset flow rate threshold. If it is lower than the preset flow rate threshold, skip the execution of subsequent steps.

[0040] Starting point recording steps: When the material flow rate is not lower than the preset flow rate threshold, determine whether the control quantity changes from zero to non-zero. When it changes from zero to non-zero, record the current sampling time as the material head starting point.

[0041] Stage determination steps: Based on whether the time difference between the current sampling time and the starting point of the feed head reaches the preset feed head transition period, determine whether the current stage is in the feed head transition period or the steady-state operation period;

[0042] Triggering steps: When it is determined that the system is in the steady-state operation segment, the update of the prediction model and the solution of the control quantity sequence in the future control time domain are triggered; when it is determined that the system is in the material head transition segment, the update and the solution are not triggered.

[0043] Issuance Step: The first control quantity in the obtained control quantity sequence is issued to the actuator corresponding to the control quantity, and the actuator adjusts the processing of the work-in-process material.

[0044] Furthermore, by automatically identifying the transition segment and isolating the triggering conditions through flow rate and setpoint status, the closed loop is ensured to operate only in the steady state segment, thus solving the technical problem of non-stable transition segment data polluting online updates and model predictive control during the start-up, shutdown, and batch change of the processing process.

[0045] Furthermore, the step between the stage determination step and the triggering step also includes:

[0046] Duration conversion steps: Convert the preset material head transition segment duration into the corresponding number of sampling steps, and determine the determination result of the stage determination step based on whether the number of sampling steps experienced relative to the starting point of the material head at the current sampling time has reached the number of sampling steps.

[0047] Furthermore, by converting the number of steps, a segmented determination matching the fixed sampling period is achieved, thus solving the technical problem that the determination of the transition segment duration using continuous time is inaccurate in discrete sampling scenarios.

[0048] Furthermore, the step between the data acquisition step and the prediction step also includes:

[0049] Window maintenance steps: Add the collected process variables to the sliding history window. After the length of the sliding history window reaches the preset window length, discard the earliest collected process variable in the sliding history window each time a new collected process variable is added, so as to keep the length of the sliding history window constant.

[0050] Furthermore, by using a fixed-length sliding window mechanism to maintain a constant window length, the technical problem of increased storage and computing overhead caused by the continuous growth of historical data windows over runtime is solved.

[0051] This invention proposes an intelligent control method for the output index of in-process materials. First, process variables during the in-process material processing are collected periodically to construct a sliding history window. The process variables are then normalized to obtain an input feature vector. Next, the control and environmental variables from the input feature vector are input into a control branch recurrent neural network to obtain corresponding features. These features are then concatenated and fused, and input to a fusion fully connected layer to map to the prediction time domain, outputting an output index prediction curve for multiple future sampling periods. Then, it is determined whether the current stage is the material transition phase or the steady-state operation phase. If the steady-state operation phase is determined, the sequence of control variables in the future control time domain is set as an optimizable variable. An objective function is constructed based on the deviation between the output index prediction curve and the output index setpoint. The optimizable variable is then solved using differentiable optimization to obtain the optimal control variable sequence. Finally, the first step control variable in the optimal control variable sequence is denormalized and sent to the corresponding actuator. The actuator adjusts the in-process material processing to change the process variables in the next sampling period. Overcoming the technical problem that insufficient prediction accuracy is caused by the coupling between control response characteristics and environmental disturbance characteristics, which in turn restricts advanced control, the advanced closed-loop control of the exit index is achieved by combining bi-branch decoupled prediction with rolling optimization solution. Attached Figure Description

[0052] Figure 1 This is a system overall flowchart of an embodiment of the present invention.

[0053] Figure 2 This is a flowchart of the data acquisition process according to an embodiment of the present invention.

[0054] Explanation of proper nouns:

[0055] Long Short-Term Memory (LSTM) network is a type of recurrent neural network.

[0056] Model Predictive Control;

[0057] SGD: Stochastic Gradient Descent;

[0058] Norm clipping: When the norm of the gradient exceeds a preset clipping threshold, the gradient is scaled proportionally to ensure that its norm does not exceed the clipping threshold, thus preventing gradient explosion.

[0059] Control branch: In a two-branch predictive model, this branch is specifically used to handle operable control variables. Side road;

[0060] Environmental Branch: In the two-branch prediction model, environmental disturbances that cannot be directly manipulated are specifically handled. Side road;

[0061] Prediction Head: A fully connected fusion layer used to generate prediction curves after the outputs of the two branches are spliced ​​together;

[0062] Material head transition section: A non-smooth transition period after the control setpoint is established when processing starts, stops, or batches are changed;

[0063] Steady-state phase: The relatively stable operating phase that begins after the material head transition phase ends;

[0064] Online update: Lightweight adaptive updates to the prediction model without downtime during operation;

[0065] Freeze the main trunk—fine-tune the header: fix during online updates The strategy is to update only the main parameters and the prediction head parameters.

[0066] Differentiable optimization: An optimization method that sets the control variable as a learnable variable and solves it using gradient descent on a differentiable predictive model;

[0067] Control time domain: The number of future control variables to be optimized in model predictive control;

[0068] Prediction time domain: The number of future prediction steps output by the prediction model;

[0069] Materials under processing: Materials (leaves, tobacco shreds, etc.) that are in the process of tobacco processing, such as tobacco shredding and re-drying. Detailed Implementation

[0070] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0071] In this invention, the use of terms such as "first" and "second" is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features.

[0072] Please refer to Figures 1-2The method described in this embodiment of the invention operates on an industrial control computer or edge computing terminal (hereinafter collectively referred to as the "control terminal"). The control terminal connects to process sensors (such as infrared moisture meters, resistance temperature sensors, vortex flow meters, and pressure transmitters) via analog input modules or fieldbus interfaces (such as 4-20mA analog acquisition modules, Modbus RTU / TCP communication interfaces, and PROFIBUS bus interfaces) to collect process variables during material processing. It also connects to actuators (such as pneumatic diaphragm valves, electric regulating valves—i.e., water proportional valves, and steam regulating valves) via analog output modules or fieldbus interfaces (such as 4-20mA analog output modules and Modbus communication interfaces) to send the calculated control quantities to the actuators. The actuators then adjust process parameters such as water flow rate and steam flow rate, thereby changing the outlet indicators. The process sensors then collect the adjusted outlet indicators again and feed them back to the control terminal, forming a complete hardware closed loop of "sensor acquisition → control terminal calculation → actuator action → sensor feedback".

[0073] The control terminal described in this specific embodiment corresponds to the computing unit, and its hardware form includes, but is not limited to, an industrial control computer, an edge computing terminal, or a programmable logic controller equipped with a corresponding computing module. The control terminal communicates with actuators such as water proportional valves and steam regulating valves by issuing control commands via analog output modules in the form of 4-20mA current signals or 0-10V voltage signals, or via fieldbus (such as Modbus or Profibus) in the form of digital signals. The control terminal also communicates with sensors such as flow meters, infrared thermometers, and moisture detectors by receiving collected data via analog input modules or fieldbus interfaces. The above hardware carrier and interaction methods remain consistent throughout all embodiments and will not be described repeatedly.

[0074] Example 1: Basic Implementation

[0075] Basic Scheme: This embodiment takes the control of moisture content at the outlet of the feeding process in a tobacco processing line as an example to illustrate how the data acquisition step, prediction step, segmented judgment step, online update step, control solution step, and distribution step are specifically implemented on the control terminal.

[0076] 1. Data collection steps

[0077] The control terminal collects process variables during the material processing at a fixed sampling period (5 seconds in this embodiment, but can be selected within the range of 1 to 10 seconds depending on the dynamic characteristics of the process). The process variables include control variables and environmental variables.

[0078] Control quantity: Water addition setpoint (unit: L / h), which is an adjustable setpoint that the control terminal can directly send to the water addition proportional valve;

[0079] Environmental parameters (7 dimensions in total): inlet material flow rate (kg / h), feed pump flow rate (kg / h), steam flow rate (kg / h), infrared temperature measurement value of the front cylinder wall (°C), infrared temperature measurement value of the middle cylinder wall (°C), infrared temperature measurement value of the rear cylinder wall (°C), and inlet moisture content (%) are collected by the corresponding flow meters, infrared temperature measuring devices, and moisture detection devices and transmitted to the control terminal via the analog input module.

[0080] The control terminal constructs a sliding history window using process variables from multiple recently collected sampling periods, with the window length being the total number of steps N in the history window. h =21 steps (corresponding to 105s of historical data, which can be selected within the range of 20 to 40 steps depending on process inertia). Each time a new set of data is collected, the earliest data at the beginning of the window is discarded to keep the window length constant.

[0081] The control terminal performs Z-score normalization on each process variable within the window:

[0082] ;

[0083] Where x is the original process variable, Let be the mean of this dimension of the variable. Let ε be the standard deviation of this dimension variable, and ε be a small constant excluding zero (which can be taken as 1×10⁻⁶). -6 The normalization parameters are stored in the parameter storage unit of the control terminal after offline training is completed. The same set of parameters is reused in both the online running phase and the online update phase to ensure data scale consistency.

[0084] 2. Prediction Steps

[0085] The control terminal inputs the control quantity sequence from the normalized input feature vector to the control branch Long Short-Term Memory network (a two-layer stacked structure, with 128 hidden state dimensions per layer), and outputs control features reflecting the dynamic adjustment process of the water addition setpoint. ctl ∈ℝ¹² 8 The environmental quantity sequence is input into the environmental branch of the Long Short-Term Memory network (also a two-layer stacked structure, with 128 dimensions of hidden states in each layer), and the output reflects the environmental features reflecting the process perturbation. env ∈ℝ¹² 8 The two branch network structures are independent of each other and their parameters are not shared, thus ensuring that the dynamic response characteristics of the control quantity and the process disturbance characteristics of the environmental quantity do not interfere with each other before entering the fusion stage, providing uncoupled and uncontaminated feature inputs for subsequent fusion.

[0086] The control terminal splices the control features and environmental features along the channel dimension to obtain the fused features [o ctl ;o env ]∈ℝ²56 The input is fed into the fusion fully connected layer to obtain the basic predicted value. Then, the inlet water content residual m is superimposed. in,t Output the future N p Predicted outlet moisture content curves for each sampling period:

[0087] ;

[0088] in Indicates channel-dimensional splicing. To integrate the fully connected layer, residual terms The introduction of this feature enables the prediction to directly correct for deviations in inlet moisture content, where t represents the current sampling time, t+1 represents the next sampling time, and t+N represents the next sampling time. p This represents the time from t+1 to t+N. p Continuous total N p N future sampling times p This indicates the number of prediction steps.

[0089] In this embodiment, the prediction steps are N. p Take 11 steps (corresponding to a 55s prediction time domain) because the response of the moisture content at the outlet of the feeding process to the adjustment of the water setting value has a pure lag of 30s to 60s. The prediction time domain needs to cover this lag interval, so as to provide effective leading information for subsequent control solution steps.

[0090] 3. Segmentation Judgment Steps

[0091] The control terminal determines the current stage based on the changes in material flow rate and water addition setpoint within the sliding history window:

[0092] When the material flow rate is lower than the preset flow rate threshold (which can be set according to the normal processing flow rate range of the production line), it is determined that no material has been fed and the subsequent steps are skipped.

[0093] When the water addition setpoint changes from zero to non-zero, record the current sampling time as the material head start point t. start ;

[0094] When the current time t is equal to the starting point t of the feed head start The time difference between them Δt ≥ t start +N head Time (where N) head The preset material head transition period duration is set to 40 steps in this embodiment, corresponding to 200 seconds. It can be selected within the range of 30 to 60 steps according to the process. Otherwise, it is determined that the material head transition period has been entered.

[0095] 4. Online update steps

[0096] Backbone Freezing and Prediction Head Fine-tuning: The parameters of the control branch Long Short-Term Memory Network and the environment branch Long Short-Term Memory Network are frozen at the control terminal (keeping the dynamic response feature extraction capability learned in the offline stage unchanged). Only the parameters of the fused fully connected layer (i.e., the prediction head) are adjusted based on the prediction error at the current sampling time, using a preset learning rate (1×10⁻⁻⁶ in this embodiment). 4 It can be used in 1×10⁻ 5 Gradient updates are performed within the range of 1×10⁻³ (optimizer is Adam).

[0097] Offset estimation and smoothing correction: The control terminal maintains a sliding window of length N (N=20 steps in this embodiment), and records the predicted values ​​of the basis model within the window. base Calculate the original system offset value using the measured value y:

[0098] ;

[0099] in At the current sampling time, This serves as an index for historical samples within the sliding window. The length of the sliding window. To prevent the division by zero small constant, y i For the first in the sliding window Normalized measured outlet moisture content at each sampling time This is the normalized prediction value of the base model at the i-th sampling time within the sliding window;

[0100] Then, perform first-order exponential smoothing on the original system offset value to obtain the system offset value:

[0101] ;

[0102] in The exponential smoothing coefficient is set to 0.1 in this embodiment, but can be selected within the range of 0.1 to 1. Finally, the system offset is superimposed on the base model predictions to obtain the corrected prediction curve:

[0103] ;

[0104] 5. Controlling the solution steps

[0105] When the system is determined to be in a steady-state operating phase, the control terminal sets the sequence of control variables in the future control time domain as an optimizable variable. Freeze prediction model ( Neither the backbone nor the prediction head is updated; only optimization is performed. .

[0106] The control terminal constructs the objective function:

[0107] L=λ·Lconstraint +L track ;

[0108] Among them, the tracking loss term (2) is the mean square error between the predicted curve and the set value:

[0109] L track =MSE( y sp );

[0110] Constraint penalty item (1) is the penalty for exceeding the upper and lower limits of the control quantity:

[0111] ;

[0112] Where λ is the penalty weight coefficient (0.1 in this embodiment, but can be selected in the range of 0.01 to 1.0). and These are the lower and upper limits of the control quantity after normalization, respectively. It is a linear rectified function; when Below or higher A positive penalty is applied to the corresponding term when it is within the specified interval, and the term is zero when it is within the specified interval.

[0113] The conventional gradient-based iterative optimization method in this field is adopted: stochastic gradient descent ( Learning rate =0.1, momentum 0.9), combined with a learning rate decay strategy (the learning rate is halved when the loss does not improve within 10 steps); iterate 200 times, recording the loss throughout, and take the control sequence corresponding to the moment of minimum loss as the optimal solution; perform norm clipping on the gradient of the optimization variables (maximum norm 1.0) to ensure solution stability; take the first step of the optimal control sequence, and perform inverse normalization (the formula and parameters are consistent with the normalization process in the acquisition step) and change limiting. The parameters are then rolled out to the actuators. In this embodiment, parameter selection mainly relies on grid / random / Bayesian search on the validation set, and the optimal parameter is selected by comprehensively considering accuracy, stability, and cost. The optimal parameter can be selected based on the actual situation.

[0114] 6. Issuance Steps

[0115] The control terminal selects the first control quantity from the optimal control quantity sequence, performs inverse normalization and change limiting processing, and sends it to the water addition proportional valve. The water addition proportional valve adjusts the opening and changes the water addition amount, thereby adjusting the outlet moisture of the next sampling cycle.

[0116] Technical effects: By using dual-branch decoupling prediction, control response features and perturbation features are extracted separately, avoiding prediction distortion caused by feature coupling in a single network. The prediction RMSE is reduced by more than 15% compared to a single LSTM. By using differentiable optimization MPC, the prediction model is embedded into a differentiable computation graph, enabling proactive intervention and achieving higher accuracy in solving control variables than black-box search.

[0117] Example (II): Preferred Example

[0118] Solution Description: This embodiment takes the outlet temperature control of the re-drying and moistening leaf process as an example. The control quantity is replaced by the steam setpoint, and the environmental quantity includes all temperature-related sensor data (inlet material flow rate, inlet moisture content, ambient temperature, ambient humidity, and infrared thermometry values ​​of the front, middle, and rear cylinder walls). The prediction time domain is adjusted from 11 steps in Embodiment (I) to 15 steps (corresponding to 75s) to adapt to the relatively longer thermal inertia response time of the re-drying and moistening leaf process. The structure of the acquisition step, prediction step, segmented judgment step, control solution step, and distribution step executed by the control terminal is the same as that in Embodiment (I), the only difference being the specific physical meaning of the control quantity and environmental quantity and the value of the prediction time domain step. The optimal steam setpoint obtained by the control solution step is denormalized and then distributed to the steam regulating valve, which adjusts the steam flow rate, thereby changing the outlet temperature of the re-drying and moistening leaf process in the next sampling cycle.

[0119] The difference from Example (I): Example (I) predicts 11 steps in the time domain, adapting to the 30 to 60 seconds pure lag in the feeding process; this example predicts 15 steps in the time domain, adapting to the longer heat conduction lag range of the re-drying and moistening process. Both examples demonstrate that the statement "multiple sampling periods in the future" is not limited to a single step value, but rather is a higher-level range that can be flexibly configured according to the actual lag characteristics of different processes. The rationality of this range is supported by the specific applications of the two different values ​​in this example and Example (I).

[0120] Example 3: Control Branch and Environment Branch Network Structure

[0121] Basic Scheme: Based on Example (I), this embodiment further explains the specific structural limitations of the control branch long short-term memory network and the environment branch long short-term memory network. Both adopt a two-layer stacked structure, that is, each branch contains two sequentially connected long short-term memory network layers, and the hidden state output of the previous layer serves as the input of the next layer; the hidden state dimension of each layer is 128, that is, the two are completely equivalent in terms of network depth and hidden state dimension.

[0122] Technical effect: By configuring network parameters with structural equivalence for the control branch and the environment branch, the two branches do not produce an expression ability bias due to structural differences in feature extraction capabilities. This ensures that the control features and environment features have similar information densities before splicing and fusion, providing a consistent basis for the accurate fusion of the fully connected layer.

[0123] Example (IV): Construction of the Objective Function

[0124] Basic Scheme: This embodiment, based on embodiment (I), further explains the specific composition of the objective function in the control solution step. The objective function constructed by the control terminal consists of two parts:

[0125] Control tracking loss term: the mean square error between the predicted curve and the setpoint.

[0126] L track =MSE( y sp );

[0127] Control constraint penalty term: This term optimizes the penalty for variables exceeding preset upper and lower limits.

[0128] ;

[0129] Overall objective function: L = λ·L constraint +L track ;

[0130] in, The future control input sequence to be optimized. and These are the lower and upper limits of the control quantity after normalization, respectively. It is a linear rectified function; when Below or higher A positive penalty is applied to the corresponding item when it is within the interval, and the item is zero when it is within the interval; λ is a preset weight coefficient (which can be selected in the range of 0.01 to 1.0).

[0131] Technical effect: By introducing a control constraint penalty term into the objective function, the control solution steps minimize the prediction deviation while ensuring that the control quantity sequence obtained by the constraint solution does not exceed the upper and lower limits of safe operation of the equipment, thus avoiding the solution result from causing action commands that exceed the safe range for actuators such as water supply proportional valves.

[0132] As one implementation method: λ is set to 0.5, and the remaining configuration is the same as the basic embodiment. This configuration is suitable for application scenarios with stricter control constraints (such as a narrower response range for steam regulating valves than for water proportional valves), and together with the basic embodiment, it demonstrates that the value of λ can be flexibly adjusted within a reasonable range.

[0133] Example (5): Differentiable Optimization Solution Details

[0134] Basic Scheme: This embodiment, based on embodiment (IV), further explains the specific method for solving the differentiable optimization problem. The control terminal uses stochastic gradient descent with momentum to iteratively update the optimizable variables multiple times (200 times in this embodiment, which can be adjusted within the range of 100 to 500 times according to real-time requirements). The preset learning rate is initially set to 0.1, and the preset momentum is set to 0.9 during the iteration process. During the iteration process:

[0135] 1. In each iteration, the current control input sequence is input into the frozen prediction model to calculate the objective function value;

[0136] 2. Calculate the gradient of the objective function with respect to the control sequence through backpropagation;

[0137] 3. Perform norm clipping on the gradient (the clipping threshold is set to 1.0, and can be selected within the range of 0.5 to 5.0) to prevent gradient explosion;

[0138] 4. Update the control variable sequence;

[0139] 5. If the objective function value does not improve within a preset number of consecutive steps, halve the learning rate;

[0140] 6. After 200 iterations, the control sequence with the minimum objective function value is taken as the optimal control sequence.

[0141] Technical effects: By combining variable SGD with a dynamic learning rate adjustment strategy, the differentiable optimization solution process converges faster and more stably on the objective function surface; gradient norm pruning avoids divergence in the solution process caused by abrupt changes in gradient magnitude; since the iterative process directly calculates the gradient of the optimizable variables in the same computational graph, the gradient can be directly obtained from the prediction error through backpropagation, which can obtain more accurate control values ​​in the same solution time compared to the black-box search method of traversing candidate control variables one by one.

[0142] As one implementation method: the number of iterations is adjusted to 300, the initial learning rate is set to 0.05, the momentum is set to 0.85, and the pruning threshold is set to 0.5. This configuration is suitable for scenarios with higher requirements for solution accuracy and a relatively large tolerance for the time consumption of a single solution. Together with the basic implementation, it demonstrates the rationality of functional expressions such as "multiple iterations" and "preset norm" covering different specific values.

[0143] Example (6): Material Head-Steady-State Segmented Trigger Control Method

[0144] Basic Scheme: This embodiment illustrates the specific implementation of the material head-steady-state segmented trigger control method, which refines the segmented judgment into multiple independent sub-steps to form an independently implementable technical solution.

[0145] Explanation of the relationship with the segmented judgment steps in Example (I): The segmented judgment steps in Example (I) are a higher-level generalization of the technical feature of "judging the current stage based on the changes in material flow rate and control quantity status". The flow rate judgment sub-step, the starting point recording sub-step, and the stage determination sub-step in this example are further detailed expansions of this higher-level generalization. The two are technical solutions at different levels of generalization under the same inventive concept, rather than independent and unrelated technical features.

[0146] 1. Data collection steps: The control terminal collects process variables of the material processing process at a fixed sampling period (5 seconds in this embodiment). The process variables include control quantities and environmental quantities.

[0147] 2. Window maintenance sub-step: The control terminal adds the newly collected process variables in each sampling cycle to the sliding history window; after the length of the sliding history window reaches the preset window length (21 steps in this embodiment), the control terminal discards the earliest collected process variable in the sliding history window every time a new process variable is added, so as to keep the length of the sliding history window always at the preset window length and not continue to grow as the running time goes by.

[0148] Technical effect: By adopting a first-in-first-out fixed-length sliding window maintenance method, the historical data storage space and corresponding computational overhead used for prediction steps and online update steps are kept constant during long-term continuous operation of the control terminal. This avoids the problem of storage resource occupation and single-step calculation time continuously increasing due to the continuous accumulation of historical data windows over running time.

[0149] 3. Flow rate judgment sub-step: The control terminal judges whether the material flow rate at the current sampling time is lower than the preset flow rate threshold (in this embodiment, 50 is taken under the normalized space, corresponding to the original physical dimension kg / h); if it is lower than the preset flow rate threshold, it is determined that no material is being fed at present, and the subsequent online update of the prediction curve and solution of the control quantity are skipped. The control terminal re-executes the flow rate judgment sub-step in the next sampling cycle.

[0150] 4. Starting point recording sub-step: When the material flow rate is not lower than the preset flow rate threshold, the control terminal determines whether the water addition setting value at the current sampling time changes from zero to non-zero; if it changes from zero to non-zero, the control terminal records the current sampling time as the material head starting point and stores it in the status storage unit of the control terminal.

[0151] 5. Stage Determination Sub-step: The control terminal determines whether it is currently in the material head transition segment or the steady-state operation segment based on whether the time difference between the current sampling time and the aforementioned material head starting point reaches the preset material head transition segment duration (200 seconds in this embodiment). If the time difference does not reach the preset material head transition segment duration, it is determined that it is currently in the material head transition segment; if it has reached or exceeded the preset material head transition segment duration, it is determined that it is currently in the steady-state operation segment.

[0152] 6. Duration Conversion Sub-step: The control terminal pre-converts the preset material head transition segment duration into the corresponding preset sampling steps. The conversion method is to divide the preset material head transition segment duration by a fixed sampling period. In this embodiment, the preset material head transition segment duration is 200 seconds, and the fixed sampling period is 5 seconds (which can be replaced according to actual conditions), resulting in a preset sampling step of 40 steps. In the stage determination sub-step, the control terminal counts the number of sampling steps that have been performed relative to the starting point of the material head at the current sampling time, and determines whether the number of sampling steps that have been performed reaches the preset sampling step obtained above. This determines the determination result of the stage determination sub-step, that is, the comparison of sampling steps replaces the direct comparison of continuous duration.

[0153] Technical effect: Since the control terminal collects data discretely according to a fixed sampling period, if the continuous time difference is directly compared with the preset material head transition period, there may be a conversion error between the sampling time and the judgment time. By pre-converting the preset material head transition period into a fixed number of sampling steps and comparing it with the number of sampling steps, the judgment result of the stage judgment sub-step is completely aligned with the actual data acquisition rhythm under the fixed sampling period, thus avoiding the judgment error introduced by the continuous time comparison method.

[0154] 7. Triggering steps: When the stage determination sub-step determines that the current state is in a steady-state operation phase, the online update of the prediction model and the solution of the water addition setpoint sequence in the future control time domain are triggered; when the stage determination sub-step determines that the current state is in the material head transition phase, the above online update and solution are not triggered. At this time, the control terminal only executes the prediction steps without changing the issued water addition setpoint.

[0155] 8. Issuance Step: The first water addition setting value in the water addition setting value sequence obtained in the triggering step is issued to the water addition proportion valve, which adjusts the water addition amount, thereby changing the process variable corresponding to the moisture content at the outlet of the feeding process in the next sampling cycle.

[0156] Technical benefits: By identifying the no-feed state through the flow rate judgment sub-step, invalid online updates to the model are avoided when no material is passing through; by automatically identifying the start and end times of the material head transition section through the start point recording sub-step and the stage judgment sub-step, the control terminal can distinguish between the highly non-stationary material head transition section data and the relatively stable steady-state operation section data; by triggering the step to perform online updates and control solutions only in the steady-state operation section, the non-stationary data of the material head transition section is prevented from polluting the online update and control quantity solution process of the model, ensuring that closed-loop control only takes effect during the stable data operation phase.

[0157] Example (7): Preferred Example

[0158] Solution Description: This embodiment takes the re-drying and moistening leaf process as an example. The preset flow rate threshold is adjusted to 40 in the normalized space, and the preset material head transition time is adjusted to 300 seconds. The execution methods of the remaining flow rate judgment sub-steps, starting point recording sub-steps, stage judgment sub-steps, triggering steps, and issuing steps are the same as in embodiment (VI). This configuration is suitable for scenarios where the material flow rate benchmark and material head transition time in the re-drying and moistening leaf process differ from those in the feeding process.

[0159] The difference from Example (VI): Example (VI) is for the feeding process, with a preset flow rate threshold of 50 and a preset material head transition time of 200 seconds; this example is for the re-drying and moistening leaf process, with a preset flow rate threshold of 40 and a preset material head transition time of 300 seconds. Both examples demonstrate that the "preset flow rate threshold" and "preset material head transition time" are not limited to a single value, but are functional parameters that can be flexibly configured according to the material flow rate benchmark and transition characteristics of different processes.

[0160] Example 8: Joint Optimization of Multiple Control Variables (Shared Environment Branch)

[0161] Basic scheme: Based on Examples (I) and (II), this embodiment further explains how the prediction step and the control solution step jointly process the two control variables when the control variables are expanded to two items: water addition set value and steam set value.

[0162] 1. Shared Environmental Branch Feature Extraction: For the outlet moisture content control target corresponding to the water setpoint, the control terminal constructs a first dual-branch time-series prediction model, whose control branch takes the water setpoint sequence as input; for the outlet temperature control target corresponding to the steam setpoint, a second dual-branch time-series prediction model is constructed, whose control branch takes the steam setpoint sequence as input. The environmental branches of the first and second dual-branch time-series prediction models share the same set of environmental feature extraction results. That is, environmental quantities (inlet material flow rate, inlet moisture content, ambient temperature, ambient humidity, cylinder wall infrared thermometry value, etc.) only need to be extracted once through one environmental branch long short-term memory network, and then provided to two fused fully connected layers to generate the outlet moisture content prediction curve and the outlet temperature prediction curve, thereby avoiding repeated extraction of environmental features.

[0163] 2. Joint Prediction and Solution: In the prediction step, the control terminal obtains the outlet moisture prediction curve and the outlet temperature prediction curve respectively; in the control solution step, the control terminal sets the water supply setpoint sequence and the steam setpoint sequence as the optimizable variables to be optimized, constructs a joint objective function including the outlet moisture tracking loss term, the outlet temperature tracking loss term, and two control constraint penalty terms, and iteratively solves the joint objective function to obtain two optimal control quantity sequences corresponding to the water supply setpoint and the steam setpoint respectively.

[0164] 3. Separate issuance: In the issuance step, the control terminal issues the first step control quantity from the two optimal control quantity sequences separately—the water setting value is issued to the water proportional valve, and the steam setting value is issued to the steam regulating valve. The two actuators adjust the water supply and steam flow respectively, thereby simultaneously changing the process variables corresponding to the outlet moisture and outlet temperature of the re-drying and moistening process in the next sampling cycle.

[0165] Technical benefits: By sharing the results of environmental branch feature extraction, computational redundancy caused by repeatedly extracting environmental features for moisture control and temperature control is avoided; by incorporating the solution of the two control variables into the same joint objective function for joint iterative solution, the adjustment of the water supply setpoint and steam setpoint can simultaneously take into account the tracking deviation of the two indicators of outlet moisture and outlet temperature, avoiding the mutual interference that may occur when the two control variables are solved independently, and realizing coordinated control of outlet moisture and outlet temperature.

[0166] Technical and effect differences from Example (II): Example (II) only applies to the scenario of a single control target and a single control quantity (steam setpoint) for outlet temperature; this example further extends to the scenario of a joint solution of dual control targets and dual control quantities (water addition setpoint and steam setpoint) for outlet moisture and outlet temperature. Both of them jointly verify that the prediction-update-control closed-loop architecture of the present invention can be implemented in both single control quantity and multi-control quantity scenarios, and has cross-index and cross-control quantity dimensions.

[0167] Furthermore, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0168] Furthermore, a portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. The program instructions invoking the methods of the invention may be stored in a fixed or removable recording medium, and / or transmitted via a data stream in a broadcast or other signal-carrying medium, and / or stored in the working memory of a computer device operating according to the program instructions. Here, an embodiment of the invention includes an apparatus comprising a memory for storing computer program instructions and a processor for executing the program instructions, wherein, when the computer program instructions are executed by the processor, the apparatus is triggered to operate the methods and / or technical solutions based on the foregoing embodiments of the invention.

[0169] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent control of material outlet indicators, characterized in that, include: Data collection steps: Collect process variables during the material processing process at a fixed sampling period, construct a sliding history window, and normalize the process variables to obtain a normalized input feature vector. The input feature vector includes control variables and environmental variables. The control variables are directly adjustable setpoints, and the environmental variables are process disturbances that cannot be directly manipulated. Prediction steps: Input the control quantity and the environmental quantity in the input feature vector into the control branch recurrent neural network respectively to obtain the corresponding control features and environmental features. Concatenate and fuse the control features and the environmental features, and input them into the fusion fully connected layer to map to the prediction time domain dimension, and output the exit index prediction curve for multiple future sampling periods. Segmented judgment steps: Based on the changes in the state of the material flow rate and the control quantity in the process variables, determine whether the current stage is the material head transition stage or the steady-state operation stage; Online update steps: When it is determined that the system is in the steady-state operation segment, the parameters of the control branch recurrent neural network and the environment branch recurrent neural network are frozen. The parameters of the fused fully connected layer are updated by gradient according to the prediction error. The offset of the updated exit index prediction curve relative to the previous one is smoothed and corrected to obtain the corrected exit index prediction curve. Control solution steps: When it is determined that the steady-state operation segment is in progress, the control quantity sequence in the future control time domain is set as an optimizable variable. An objective function is constructed based on the deviation between the predicted exit index curve and the set value of the exit index. Differentiable optimization is performed on the optimizable variable to obtain the optimal control quantity sequence. Issuance Step: The first step control quantity in the optimal control quantity sequence is denormalized and issued to the actuator corresponding to the control quantity. The actuator adjusts the in-process material processing to change the process variable in the next sampling period.

2. The method as described in claim 1, characterized in that, The prediction steps include: Control input sub-step: Extract the control quantity sequence within the historical lookback steps from the sliding history window and input it into the control branch recurrent neural network; Environmental quantity input sub-step: Extract the environmental quantity sequence within the historical lookback step from the sliding history window and input it into the environmental branch recurrent neural network; The dual-branch feature extraction sub-step involves extracting the control features of the control quantity sequence based on the control branch recurrent neural network, and extracting the environmental features of the environmental quantity sequence based on the environmental branch recurrent neural network. The splicing and fusion prediction sub-step involves splicing the control features and the environmental features in the channel dimension, inputting them into the fusion fully connected layer, mapping them to the prediction step dimension, and obtaining the prediction curve of the exit index for the next multiple sampling periods.

3. The method as described in claim 2, characterized in that, In the dual-branch feature extraction sub-step: Both the control branch recurrent neural network and the environment branch recurrent neural network are two-layer stacked structures with the same hidden state dimension. The control branch recurrent neural network expands the control features by taking the historical lookback steps as the time expansion steps; The environmental branch recurrent neural network expands the time step by the historical lookback step and outputs the environmental features.

4. The method as described in claim 2, characterized in that, In the splicing and fusion prediction sub-step: The input dimension of the fused fully connected layer is the sum of the dimensions of the control features and the dimensions of the environment features, and the output dimension is the number of prediction steps; The output of the fused fully connected layer is also superimposed with the residual term corresponding to the inlet material residual in the environmental quantity, to obtain the predicted curve of the outlet index after superimposing the residual.

5. The method as described in claim 1, characterized in that, The online update step includes smoothing and correcting the offset of the updated export indicator prediction curve relative to the previous curve, which includes: Window offset estimation sub-step: Maintain a sliding window, record the base model predicted value and the corresponding measured value of the export index prediction curve at multiple sampling times within the sliding window, and estimate the original value of the system offset at the current sampling time based on the residual sequence between the base model predicted value and the measured value of the export index within the sliding window. Exponential smoothing sub-step: Perform first-order exponential smoothing on the original system offset value to obtain the system offset value at the current sampling time; Offset superposition sub-step: The system offset value is superimposed on the base model prediction value of the export index prediction curve to obtain the corrected export index prediction curve.

6. The method as described in claim 1, characterized in that, In the control solution step, the objective function includes a control tracking loss term and a control constraint penalty term. The control tracking loss term is the mean square error between the predicted curve of the export index and the set value of the export index. The control constraint penalty term is the penalty for the optimizable variable exceeding the preset upper and lower limits.

7. The method as described in claim 1, characterized in that, The control solution step includes differentiable optimization of the optimizable variables, which includes: The optimizable variable is initialized to zero and set as a learnable parameter, and the parameters of the recurrent neural network and the fused fully connected layer in the prediction step are frozen. The optimizable variable is updated iteratively multiple times using stochastic gradient descent with momentum. During each iteration, the gradient of the optimizable variable is pruned according to its norm, and the learning rate is adjusted based on the loss improvement. Record the values ​​of the objective function during multiple iterations, and select the optimizable variable corresponding to the minimum value of the objective function as the optimal control variable sequence.

8. A method for intelligent control of material outlet indicators, characterized in that, include: Data collection steps: Collect process variables during the material processing process at a fixed sampling period, normalize the process variables to obtain a normalized input feature vector, the input feature vector includes control variables and environmental variables, the control variables are directly adjustable set values, and the environmental variables are process disturbances that cannot be directly operated; Prediction steps: Input the control quantity and environmental quantity in the input feature vector into the control branch recurrent neural network respectively to obtain the corresponding control features and environmental features. Concatenate and fuse the control features and environmental features, input them into the fusion fully connected layer to map to the prediction time domain dimension, and output the exit index prediction curve for multiple future sampling periods. Flow rate determination step: Determine whether the material flow rate is lower than the preset flow rate threshold. If it is lower than the preset flow rate threshold, it is determined that no material is being fed and the process waits to proceed to the next step. Starting point recording steps: When the material flow rate is not lower than the preset flow rate threshold, determine whether the control quantity changes from zero to non-zero. When it changes from zero to non-zero, record the current sampling time as the material head starting point. Stage determination steps: Based on whether the time difference between the current sampling time and the starting point of the feed head reaches the preset feed head transition period, determine whether the current stage is in the feed head transition period or the steady-state operation period; Triggering steps: When it is determined that the system is in the steady-state operating segment, the system will trigger the update of the prediction model and the solution of the control quantity sequence in the future control time domain. Issuance Step: The first control quantity in the obtained control quantity sequence is issued to the actuator corresponding to the control quantity, and the actuator adjusts the processing of the work-in-process material.

9. The method as described in claim 8, characterized in that, The step determination step and the triggering step also include: Duration conversion steps: Convert the preset material head transition segment duration into the corresponding number of sampling steps, and determine the determination result of the stage determination step based on whether the number of sampling steps experienced relative to the starting point of the material head at the current sampling time has reached the number of sampling steps.

10. The method as described in claim 8, characterized in that, The step between the data acquisition step and the prediction step also includes: Window maintenance steps: Add the collected process variables to the sliding history window. After the length of the sliding history window reaches the preset window length, discard the earliest collected process variable in the sliding history window each time a new collected process variable is added, so as to keep the length of the sliding history window constant.

Images